LUNG SOUND ANALYSIS SYSTEM

TECHNICAL FIELD

The present invention relates to a lung sound analysis system, a lung sound analysis method, and a storage medium, for supporting diagnosis of heart failure.

BACKGROUND ART

Heart failure is a clinical syndrome in which as a result that cardiac dysfunction, that is, an organic and/or functional dysfunction, occurred in the heart and compensation mechanism of a heart pump function failed, dyspnea, malaise, or an edema appears, which is accompanied by a drop of exercise tolerability. A patient who suffered from heart failure always has a risk of exacerbation even though the patient has been treated and reached remission. When acute exacerbation occurs in the patient due to excessive water or salt intake, forgetting to take medicines, too much exercise, and the like, the patient must be hospitalized again. Therefore, it is important to prevent acute exacerbation by finding heart failure exacerbation of a patient discharged from hospital in an early stage and giving treatment intervention.

One method of diagnosing heart failure is a lung sound examination by auscultation. Such an examination is a method usable for diagnosing health condition of lungs and also heart failure, in a safe and easy manner. However, it is difficult for those other than skilled medical specialists to obtain a detailed and accurate diagnosis result. Therefore, in the rounds by general nurses or caring staff and in the visiting care sites, it is impossible to obtain a detailed diagnosis.

In order to cope with such a problem, a system has been proposed (for example, see Patent Literatures 1 to 6). The system automatically determines presence or absence of abnormal sounds called adventitious sounds in the lung sounds collected by an electronic stethoscope. Further, In Patent Literature 10, abnormality learning data created from a plurality of units of respiratory sound data including various types of adventitious sounds and normality learning data created from a plurality of units of respiratory sound data not including adventitious sounds are used to detect adventitious sounds from analysis object respiratory sound data. Further, Patent Literature 11 describes giving a correct answer such as “it is normal respiratory sound” or “it is abnormal respiratory sound” as a teacher signal to each respiratory sound sample, and generating a parameter of a model having the highest performance in identifying normality and abnormality of respiration by using an optimization method such as a steepest descent method or Newton's method.

Patent Literature 1: JP 2014-4018 A
Patent Literature 2: JP 2002-538921 A
Patent Literature 3: JP 2017-536905 A
Patent Literature 4: WO 2010/044452 A
Patent Literature 5: JP 2008-113936 A
Patent Literature 6: JP 4849424 B
Patent Literature 7: WO 2019/220609 A
Patent Literature 8: WO 2019/220620 A
Patent Literature 9: JP 2007-190081 A
Patent Literature 10: JP 2005-66044 A
Patent Literature 11: JP 5585428 B

SUMMARY

In order to improve accuracy in detecting abnormal lung sounds, it is necessary to collect a large quantity of abnormal lung sound data to be used as learning data. However, it is difficult to accurately determine whether or not lung sound data is abnormal, for persons other than well-trained medical specialists. Therefore, it is difficult to efficiently collect abnormal lung sound data in a clinic without a medical specialist or at home of a patient.

An object of the present invention is to provide a lung sound analysis system that solves the above-described problem.

A lung sound analysis system, according to one aspect of the present invention, is configured to include

- an acquisition means for acquiring time-series acoustic signals including lung sounds of a heart failure patient,
- a detection means for detecting abnormal lung sounds from the acquired time-series acoustic signals;
- an observation acquisition means for transmitting, to a terminal device of a medical specialist, analysis object lung sound information in which the acquired time-series acoustic signals and a result of the detection are associated with each other, and receiving, from the terminal device, analysis object lung sound information to which an observation by the medial specialist on the time-series acoustic signals is added; and
- a generation means for generating learning data for detecting abnormal lung sounds on the basis of the analysis object lung sound information to which the observation by the medial specialist is added.

A lung sound analysis method, according to another aspect of the present invention, is configured to include

- acquiring time-series acoustic signals including lung sounds of a heart failure patient,
- detecting abnormal lung sounds from the acquired time-series acoustic signals,
- transmitting, to a terminal device of a medical specialist, analysis object lung sound information in which the acquired time-series acoustic signals and a result of the detection are associated with each other, and receiving, from the terminal device, analysis object lung sound information to which an observation by the medial specialist on the time-series acoustic signals is added, and
- generating learning data for detecting abnormal lung sounds on the basis of the analysis object lung sound information to which the observation by the medial specialist is added.

Further, a computer-readable medium according to another aspect of the present invention, is configured to store thereon a program for causing a computer to execute processing to

- acquire time-series acoustic signals including lung sounds of a heart failure patient,
- detect abnormal lung sounds from the acquired time-series acoustic signals,
- transmit, to a terminal device of a medical specialist, analysis object lung sound information in which the acquired time-series acoustic signals and a result of the detection are associated with each other, and receive, from the terminal device, analysis object lung sound information to which an observation by the medial specialist on the time-series acoustic signals is added, and
- generate learning data for detecting abnormal lung sounds on the basis of the analysis object lung sound information to which the observation by the medial specialist is added.

Since the present invention has the configurations as described above, it is possible to efficiently collect learning data for detecting abnormal lung sounds in a clinic without a medical specialist or at home of a patient.

BRIEF DESCRIPTION OF DRAWINGS

FIG. 1 is a block diagram of a lung sound analysis device according to a first exemplary embodiment of the present invention.

FIG. 2 is a diagram illustrating an exemplary configuration of a lung sound record stored in the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 3 illustrates auscultation positions (1) to (12) of examination by an electronic stethoscope in the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 4 is a diagram illustrating an exemplary configuration of analysis object lung sound information stored in the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 5 illustrates an exemplary configuration of target personal information previously set in the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 6 illustrates an exemplary configuration of analysis object lung sound information in which auscultation observations by a medical specialist are recorded, stored in the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 7 is a flowchart illustrating an example of a previous operation of the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 8 illustrates learning using learning data in the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 9 is a flowchart illustrating an example of an analysis operation of the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 10 is a flowchart illustrating details of the analysis operation of the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 11 illustrates examples of abnormality frequency at each auscultation position of a patient and the auscultation sequence determined based on the abnormality frequency, calculated by the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 12 is a schematic diagram illustrating a waveform of time-series acoustic signals including lung sounds output from an electronic stethoscope in the first exemplary embodiment of the present invention.

FIG. 13 illustrates an example of a determination table for determining the severity of heart failure from an analysis result of lung sound data for each auscultation position by the lung sound analysis device according to the first exemplary embodiment of the present invention.

FIG. 14 is a block diagram of a lung sound analysis system according to a second exemplary embodiment of the present invention.

FIG. 15 is a block diagram of a lung sound analysis system according to a third exemplary embodiment of the present invention.

EXEMPLARY EMBODIMENTS

Next, exemplary embodiments of the present invention will be described with reference to the drawings.

First Exemplary Embodiment

FIG. 1 is a block diagram of a lung sound analysis device 10 according to a first exemplary embodiment of the present invention. The lung sound analysis device 10 is an information processing device that acquires lung sounds from a heart failure patient, and analyzes them. The lung sound analysis device 10 may be a smartphone, a tablet terminal, a personal digital assistant (PDA), a laptop personal computer, or the like, but is not limited thereto. Hereinafter, a patient whose lung sounds are to be analyzed using the lung sound analysis device 10 is referred to as a patient A.

The lung sound analysis device 10 includes an electronic stethoscope 11, a communication OF unit 12, an operation input unit 13, a screen display unit 14, a storage unit 15, and an arithmetic processing unit 16.

The electronic stethoscope 11 is configured to convert the lung sounds of the patient A, obtained when the chest piece of the stethoscope is attached to the posterior side of the chest or the anterior side of the chest of the patient A, into digital signals, and transfer them to the arithmetic processing unit 16 in a wireless or wired manner.

The communication IN unit 12 is configured of, for example, a dedicated data communication circuit, and is configured to perform data communication with various devices such as a server device connected in a wired or wireless manner.

The operation input unit 13 includes operation input devices such as a keyboard and a mouse, and is configured to detect an operation by an operator and output it to the arithmetic processing unit 16. An operator is a person who performs an operation of obtaining lung sounds of the patient A by using the lung sound analysis device 10. An operator may be, for example, a doctor of a clinic, a medical professional such as a nurse, caring staff such as a care worker, or family of the patient A.

The screen display unit 14 is configured of a screen display device such as a liquid crystal display (LCD) or a plasma display panel (PDP), and is configured to display, on a screen, various types of information such as an analysis result according to an instruction from the arithmetic processing unit 16.

The storage unit 15 includes storage devices such as a hard disk and a memory, and is configured to store processing information and a program 151 necessary for various types of processing to be performed in the arithmetic processing unit 16.

The program 151 is a program that is read and executed by the arithmetic processing unit 16 to thereby implement various processing units. The program 151 is read, in advance, from an external device (not illustrated) or a storage medium (not illustrated) via a data input and output function of the communication IN unit 12 or the like, and is stored in the storage unit 15.

The main processing information stored in the storage unit 15 includes a lung sound record 152, analysis object lung sound information 153, and a learning data database (DB) 154.

The lung sound record 152 is a record of lung sounds of the patient A. The lung sound record 152 is a record of medical practice including auscultation performed on the patient A for heart failure treatment during hospitalization, for example. FIG. 2 illustrates an exemplary configuration of the lung sound record 152. In this example, the lung sound record 152 is configured of a patient ID 1521, one or more pieces of auscultation information 1527, personal information 1525, and a contact email address 1526. In the field of patient ID 1521, an ID uniquely identifying the patient A is set.

The field of auscultation information 1527 is configured of auscultation date/time 1522, a doctor in charge 1523, and lung sound information 1524. In the field of auscultation date/time 1522, the date/time on which diagnosis including auscultation is performed is recorded. The fields of one or more pieces of auscultation information 1527 are aligned in the descending order of the auscultation date/time 1522. The auscultation information 1527 at the bottom (auscultation information immediately before the personal information 1525) is the latest one. In the field of the doctor in charge 1523, the name of the doctor who made a diagnosis is recorded.

The field of lung sound information 1524 is provided for each auscultation position. The auscultation position is a location on the patient body on which a chest piece of a stethoscope for auscultating the lung sounds is put. That is, the auscultation position is a position for acquiring the lung sounds. In the example of FIG. 2, twelve positions in total from an auscultation position (1) to an auscultation position (12) are set (in FIG. 2, auscultation positions (2) to (11) are omitted). FIG. 3 is a schematic diagram for explaining the auscultation positions (1) to (12).

Referring to FIG. 3, the auscultation positions (1) and (2) are set at left and right of the upper lung field in the posterior side of the chest. The auscultation positions (3) and (4) are set at left and right of the middle lung field in the posterior side of the chest. The auscultation positions (5) and (6) are set at left and right of the lower lung field in the posterior side of the chest. The auscultation positions (7) and (8) are set at left and right of the upper lung field in the anterior side of the chest. The auscultation positions (9) and (10) are set at left and right of the middle lung field in the anterior side of the chest. The auscultation positions (11) and (12) are set at left and right of the lower lung field in the anterior side of the chest. The auscultation positions are not limited to the number and the positions described above. For example, not only the posterior side of the chest and the anterior side of the chest, it is also acceptable to set auscultation positions in the upper lung field, the middle lung field, and the lower lung field of the left and right sides of the chest to have eighteen positions in total. Alternatively, some of the above-described auscultation positions may be excluded. For example, it is acceptable to exclude the auscultation positions (3) to (6), (9), and (10) to thereby limit the positions to six positions in total, that is, the auscultation positions (1), (2), (7), (8), (11), and (12).

Referring to FIG. 2 again, the field of lung sound information 1524 for each auscultation position includes at least one set of a lung sound data field and an auscultation observation field. In the lung sound data field, digital time-series acoustic signals including lung sounds obtained by an electronic stethoscope from an auscultation position of the patient A are recorded. The posture of the patient at the time of auscultation is roughly classified into a lying position and a sitting position. The auscultation of the posterior side of the chest and the anterior side of the chest is generally performed in a sitting position. The signal length of one piece of lung sound data (for example, lung sound data 1) may have any length. For example, one piece of lung sound data may include signals of consecutive N times of breathing of the patient A. Here, N represents a positive integer of 1 or larger. The lung sound data may be signals to which processing such as removal of time-series acoustic signals in the period of pause phase, noise removal, and application of breath timing is performed on the time-series acoustic signals obtained from an electronic stethoscope.

In the auscultation observation field, the auscultation observation by a medical specialist on the lung sound data is recorded. In the auscultation observation, presence or absence of abnormal lung sounds and the type of abnormal sounds if any (rales or the like) are recorded.

In the field of the personal information 1525, information such as sex, age, weight, somatotype (BMI), and anamnesis of the patient A is recorded.

In the field of the contact email address 1526, at least one email address of a person to whom an analysis result is to be sent is recorded. The contact email address may be an email address of the hospital where the patient is hospitalized, a medical specialist in heart failure, the family doctor of the patient A, or the like. Note that the method of sending an analysis result is not limited to email, and may be another communication method such as a messaging function of groupware, business chat, or the like.

Referring to FIG. 1 again, in the analysis object lung sound information 153, the lung sound information obtained from the patient A by using the electronic stethoscope 11 and the analysis result thereof are recorded. FIG. 4 illustrates an exemplary configuration of the analysis object lung sound information 153. In this example, the analysis object lung sound information 153 is configured of a patient ID 1531, analysis date/time 1532, a person in charge 1533, lung sound information 1534, an emergency level 1535, an informative matter at analysis 1536, and consent information 1537.

In the field of patient ID 1531, an ID uniquely identifying the patient A recorded in the field of patient ID 1521 of the lung sound record 152 is recorded. In the field of analysis date/time 1532, date/time on which the lung sounds of the patient A were acquired and analyzed is recorded. In the field of person in charge 1533, an ID uniquely identifying an operator who performed an operation of obtaining the lung sounds of the patient A is recorded.

The field of lung sound information 1534 is provided for each auscultation position. In the example of FIG. 4, twelve positions in total from the auscultation position (1) to the auscultation position (12), described with reference to FIG. 3, are set (in FIG. 4, the auscultation positions (2) to (11) are omitted). The field of lung sound information 1534 for each auscultation position includes at least one set of a lung sound data field and an analysis result field. In the lung sound data field, digital time-series acoustic signals including lung sounds obtained by the electronic stethoscope 11 from the auscultation position of the patient A are recorded. The signal length of one piece of lung sound data (for example, lung sound data 1) may have any length. For example, one piece of lung sound data may include signals of consecutive N times of breathing of the patient A. Here, N represents a positive integer of 1 or larger. The lung sound data may be signals to which processing such as removal of time-series acoustic signals in a period of pause phase, noise removal, and application of breath timing is performed on the time-series acoustic signals obtained from the electronic stethoscope 11.

The analysis result field contains a result of mechanically analyzing the lung sound data. In the analysis result, a numerical value indicating whether or not the lung sound data is abnormal lung sound data is recorded. For example, the analysis result field may contain a binary value, that is, a value 0 indicating normal lung sounds or a value 1 indicating abnormal lung sounds. Alternatively, the analysis result field may contain a numerical value representing the abnormal degree of the lung sound data. Regarding the abnormal degree, an abnormal degree that is equal to or less than a preset threshold represents that the lung sound data is normal lung sounds, and an abnormal degree exceeding the threshold represents that the lung sound data is abnormal lung sounds. Alternatively, the analysis result field may contain the details of detected abnormal sounds (for example, type or characteristics, acquired timing, or the like).

The field of emergency level 1535 contains an emergency level calculated by comprehensively determining the respective analysis results of the auscultation positions (1) to (12). The emergency level is an index indicating how seriously the patient condition is emergent. In other words, the emergency level is an index indicating a degree of time allowance that can prevent or reduce a crisis of readmission into hospital due to acute exacerbation by performing appropriate heart failure treatment within some time. By including such an emergency level in the analysis result, it is possible to take action according to the emergency level by a medical professional or the like who recognizes the analysis result.

The field of informative matter at analysis 1536 contains the conditions of the patient A on the analysis date. The conditions of the patient A include, for example, weight, blood pressure, pulse, subjective symptoms (short breath when goes out, edema, cough, anorexia, and the like), medication, and water intake amount.

The field of consent information 1537 contains information of whether or not consent to use of whole or part of lung sound data of the patient A, recorded in the analysis object lung sound information 153, is given by the patient A.

Referring to FIG. 1 again, the learning data DB 154 records learning data for detecting abnormal lung sounds from the lung sound data. The learning data is configured of lung sound data and a correct answer label, and is also referred to as teacher data. In the learning data DB 154, a plurality of pieces of learning data are recoded for each auscultation position. Further, a plurality of pieces of learning data of one auscultation position include, for example, a plurality of pieces of learning data including normal lung sound data and a plurality of pieces of learning data including abnormal lung sound data. A label of each piece of learning data indicates whether the corresponding lung sound data is abnormal lung sound or normal lung sound. A label indicating abnormal lung sounds may further include information about the abnormal lung sound (for example, type, characteristics, abnormal degree, and the like).

The arithmetic processing unit 16 includes a microprocessor such as a CPU and the peripheral circuits thereof, and is configured to read and execute the program 151 from the storage unit 15 to allow the hardware and the program 151 to cooperate with each other to thereby implement the various processing units. The main processing units implemented by the arithmetic processing unit 16 include a lung sound record acquisition means 161, an analysis object lung sound acquisition means 162, a lung sound abnormality detection means 163, an analysis result output means 164, a learning data generation means 165, and a learning means 166.

The lung sound record acquisition means 161 is configured to acquire the lung sound record 152 of the patient A from an external device (not illustrated) or a storage medium (not illustrated) via a data input/output function of the communication I/F unit 12 or the like, and record it on the storage unit 15.

The analysis object lung sound acquisition means 162 is configured to acquire digital time-series acoustic signals including the lung sounds of the patient A and other information. The analysis object lung sound acquisition means 162 acquires the digital time-series acoustic signals including the lung sounds of the patient A from the electronic stethoscope 11, in accordance with an instruction by the operator input from the operation input unit 13 or the like. As other information, the analysis object lung sound acquisition means 162 acquires the patient ID, the analysis date/time, the person in charge, the informative matter at analysis, and the consent information, from the operator via the operation input unit 13 or from the lung sound record 152 stored in the storage unit 15. The analysis object lung sound acquisition means 162 also generates the analysis object lung sound information 153 from the acquired digital time-series acoustic signals and the other information, and stores it in the storage unit 15. The analysis object lung sound information 153 to be stored in the storage unit 15 by the analysis object lung sound acquisition means 162 is configured to have a format as illustrated in FIG. 4 for example. At the time of storage by the analysis object lung sound acquisition means 162, the field of each analysis result of the lung sound information 1534 and the field of the emergency level 1535 have a NULL value.

The lung sound abnormality detection means 163 is configured to detect whether or not the lung sound data is abnormal lung sounds. There are various methods for detecting abnormality in the lung sounds. In the present embodiment, the lung sound abnormality detection means 163 uses an abnormality detection method by means of supervised learning. That is, the lung sound abnormality detection means 163 learns abnormal sounds such as coarse crackles and fine crackles that are discontinuous rales and wheezes and rhonochi that are continuous rales in advance, and detects abnormal sounds based on the learning result.

For example, as supervised learning, the lung sound abnormality detection means 163 may use deep learning with respect to, for example, learning data of collected abnormal sounds to create a model having learned the characteristics and determination criteria of the input sound data (input data), and perform detection by checking whether or not input data matches the model. The lung sound abnormality detection means 163 can use, for leaning and input data for example, a spectrum program in which sounds are applied with fast Fourier transform (FFT) or log-FFT for each certain section to be aligned in a time-series manner, and for deep learning, use recurrent neural network (RNN) or convolutive neural network (CNN).

Further, the lung sound abnormality detection means 163 may use a method in which a lung sounds waveform of learning and input data is transformed into a short-time feature amount such as zero-cross coefficient or mel-frequency cepstral coefficient (MFCC), and abnormal sounds are detected by machine learning. For example, the lung sound abnormality detection means 163 may perform modeling by mixed Gaussian distribution (GMM) at the time of learning using learning data, and check whether the input data matches the model at the time of detection. Further, the lung sound abnormality detection means 163 may learn the identification surface of an identifier such as a support vector machine (SVM) by using learning data and use the identifying surface to identify whether or not the input data corresponds to the abnormal sounds. The lung sound abnormality detection means 163 may generate the feature amount by using the data itself like non-negative matrix factorization (NMF) or principal component analysis (PCA), other than the method of directly calculating the feature amount as described above.

Further, the lung sound abnormality detection means 163 may detect abnormal sounds by the decision tree using statistical features of an input waveform such as long-time power distribution of input signals, distribution of component amount/component ratio of a specific frequency bin range, or the like. In that case, as items of the decision tree, the lung sound abnormality detection means 163 may use statistical features (for example, when a process frame larger than 36 is generated by Gaussian approximation), rather than a direct value (for example, when the power exceeds 20 mW for three consecutive frames). Further, the lung sound abnormality detection means 163 may detect abnormal sounds by not using the input signal itself but modeling it through auto-regression (AR) process or the like and detecting abnormal sounds when some of the model parameters exceed a threshold. These methods may not include a learning process, but include observations of abnormal sounds that are object signals in the configuration of the decision tree or determination of a threshold. Therefore, they are included in supervised learning for the sake of convenience.

The lung sound abnormality detection means 163 uses the abnormality detection method by means of supervised learning as described above to analyze the lung sound data of each auscultation position of the patient A recorded in the analysis object lung sound information 153, and records the analysis result in the field of analysis result of the lung sound record 152 of each auscultation position. The lung sound abnormality detection means 163 also calculates the emergency level on the basis of the analysis result of the lung sound data of each auscultation position, and records it in the field of the emergency level 1535.

The analysis result output means 164 is configured to output the analysis object lung sound information 153 in order to notify the concerned persons of the the heart failure condition of the patient A. For example, the analysis result output means 164 is configured to read the analysis object lung sound information 153 from the storage unit 15, and display the analysis object lung sound information 153 on the screen display unit 14. The analysis result output means 164 is also configured to send an email to which the analysis object lung sound information 153 read from the storage unit 15 is attached as a file, to the contact email address 1526 of the lung sound record 152 via the communication IN unit 12, in accordance with an instruction from the operation input unit 13 or automatically. At that time, the analysis result output means 164 may determine the destination of the analysis object lung sound information 153 on the basis of the emergency level 1535 of the analysis object lung sound information 153.

The analysis result output means 164 is also configured to transmit the analysis object lung sound information 153 to a terminal device of a medical specialist in heart failure via the communication OF unit 12 in order to acquire observations by the medical specialist in heart failure on the lung sound data acquired from the patient A to create learning data.

At that time, the analysis result output means 164 may determine whether or not to transmit the analysis object lung sound information 153 to a medical specialist in heart failure, on the basis of the analysis result of each auscultation position. For example, the analysis result output means 164 may not transmit the analysis object lung sound information 153 in which lung sound data is normal at every auscultation position, and transmit the analysis object lung sound information 153 in which lung sound data of one or more auscultation positions is abnormal. In general, learning data including normal lung sound data can easily acquired from a large number of healthy persons, but learning data including abnormal lung sound data can be acquired only from some persons. Therefore, by only transmitting the analysis object lung sound information 153 including abnormal lung sound data, it is possible to reduce the burden on the medical specialist in heart failure. Further, the analysis result output means 164 may transmit only the analysis object lung sound information 153 in which the type of abnormal lung sounds in the analysis result matches a predetermined type. The predetermined type of abnormal lung sounds may be a type of lung sounds whose learning data is insufficient. As a result, it is possible to create learning data of abnormal lung sounds of an insufficient type without unnecessarily increasing the burden on the medical specialist in heart failure.

Further, the analysis result output means 164 may determine whether or not to transmit the analysis object lung sound information 153 to a medical specialist in heart failure on the basis of the personal information of the patient A. For example, the analysis result output means 164 may transmit only analysis object lung sound information of a patient having personal information matching the predetermined target personal information (sex, age group, BMI, anamnesis, or the like) to a medical specialist in heart failure. As predetermined target personal information, one of sex, age group, BMI, anamnesis, and the like or a combination of two or more of them in which learning data thereof is insufficient, may be used. FIG. 5 illustrates an exemplary configuration of target personal information. In this example, a patient in which sex is female, the age group is 40 to 60 years old, BMI is 20 or lower, and anamnesis is hypertension, is taken as a target. By focusing the target in this way, it is possible to efficiently collect insufficient learning data.

The analysis object lung sound information 153 transmitted to a medical specialist in heart failure is analyzed by the medical specialist. For example, a medical specialist replays the lung sound data of each auscultation position recorded in the analysis object lung sound information 153 by a personal computer or the like, and diagnoses whether or not adventitious sounds such as rales are heard from the lung sounds of the patient A. Then, the medical specialist creates auscultation observations on the lung sound data of the respective auscultation positions, and records them in the analysis object lung sound information 153. The analysis object lung sound information 153 in which the auscultation observations of the medical specialist are recorded as described above is returned to the lung sound analysis device 10, that is, the transmission source, by means of a communication means such as an email. Hereinafter, analysis object lung sound information in which auscultation observations by a medical specialist are recorded is referred to as analysis object lung sound information with auscultation observations. FIG. 6 illustrates an exemplary configuration of the analysis object lung sound information with auscultation observations.

When the learning data creation means 165 receives the analysis object lung sound information with auscultation observations 153 from the medical specialist in heart failure via the communication I/F unit 12, the learning data creation means 165 generates learning data based on the received analysis object lung sound information with auscultation observations 153. Further, the learning data creation means 165 records the generated learning data in the learning data DB 154.

The learning means 166 is configured to use the learning data recorded in the learning data DB to learn a model for detecting abnormal lung sounds. The learning means 166 is also configured to re-learn the model for detecting abnormal lung sounds after new learning data is added to the learning data DB 154.

Next, operation of the lung sound analysis device 10 will be described. The operation of the lung sound analysis device 10 is roughly classified into a previous operation, an analysis operation to be performed thereafter, and a learning data creation and learning operation.

First, a previous operation will be described. FIG. 7 is a flowchart showing an example of a previous operation. The previous operation is started when the lung sound record acquisition means 161 is activated by operating the start button of the previous operation shown on the screen display unit 14, for example.

Referring to FIG. 7, when the lung sound record acquisition means 161 is activated, it acquires the lung sound record 152 of the patient A from an external device (not illustrated) or a storage medium (not illustrated) via a data input/output function of the communication I/F unit 12 or the like, and records it on the storage unit 15 (step S1). FIG. 2 illustrates an exemplary configuration of the lung sound record 152 acquired in this manner. The lung sound record 152 includes at least the past lung sound data and the history of the auscultation observations of the patient A.

Upon completion of the above-described operation by the lung sound record acquisition means 161, the learning means 166 is activated automatically or according to an instruction from the operation input unit 13. The learning means 166 reads the learning data DB 154 from the storage unit 15, learns a model for detecting abnormality in lung sounds on the basis of the learning data recorded in the learning data DB 154, and stores the learned model in the lung sound abnormality detection means 163 (step S2).

FIG. 8 illustrates model learning. Referring to FIG. 8, the learning means 166 reads learning data of the auscultation position (1) from the learning data DB 154, and uses the readout learning data to create a model 171-1 for detecting abnormality in the lung sound data of the auscultation position (1). Similarly, the learning means 166 creates models 171-2 to 171-12 for detecting abnormality in the lung sound data of the auscultation positions (2) to (12), on the basis of the learning data of the auscultation positions (2) to (12) recorded in the learning data DB 154. In the above example, the learning means 166 learned a model for detecting abnormality in the lung sound data for each auscultation position. However, the learning means 166 may learn one common model for detecting abnormality in the lung sound data of a plurality of auscultation positions. In the case of learning a model common to a plurality of auscultation positions, a threshold set for detecting abnormality by using the model (for example, threshold set for detecting abnormality using a decision tree), may be set commonly for a plurality of auscultation positions or may be set individually for each auscultation position.

Next, the analysis operation will be described. FIG. 9 is a flowchart showing an example of the analysis operation. The analysis operation is performed at a place other than a specialized hospital, such as a clinic or home of the patient A. However, the analysis operation may be used to assist diagnosis performed by a doctor at a specialized hospital or the like. The analysis operation is started when the analysis object lung sound acquisition means 162 is activated by operating the start button of the analysis operation shown on the screen display unit 14, for example.

Referring to FIG. 9, when activated, the analysis object lung sound acquisition means 162 creates the analysis object lung sound information 153 in which necessary matters are put in the respective fields of the patient ID 1531, the analysis date/time 1532, the person in charge 1533, the informative matter at analysis 1536, and the consent information 1537, and a NULL value is put in the other fields, and stores it in the storage unit 15 (step S11). For example, the analysis object lung sound acquisition means 162 acquires the patient ID 1531 from the patient ID 1521 of the lung sound record 152 stored in the storage unit 15. The analysis object lung sound acquisition means 162 also acquires the analysis date/time 1532, the person in charge 1533, the informative matter at analysis 1536, and the consent information 1537 from the operator via the operation input unit 13. The consent information 1537 may include electronic signature indicating that the patient A gives consent to use of the own lung sound data for learning.

Then, the analysis object lung sound acquisition means 162 acquires digital time-series acoustic signals including the lung sounds of each auscultation position of the patient A from the electronic stethoscope 11, and records it in the analysis object lung sound information 153 in association with the auscultation position (step S12). Any method may be used to acquire the lung sounds of each auscultation position of the patient by the electronic stethoscope and record it in association with the auscultation position. For example, as described in Patent Literature 1, 4, 6 or the like, a method in which a guidance screen for giving guidance on the auscultation position to an operator who uses the electronic stethoscope 11 is shown on the screen display unit 14, or the like may be used. Moreover, at step S12, the lung sound abnormality detection means 163 reads the analysis object lung sound information 153 from the storage unit 15, analyzes the lung sound data of each auscultation position of the patient A recorded in the lung sound information 1534 of the analysis object lung sound information 153 by using the model created in advance, and records the analysis result in the field of analysis result for each auscultation position of the lung sound information 1534. Furthermore, at step S12, the analysis result output means 164 appropriately displays the analysis result performed by the lung sound abnormality detection means 163 on the screen display unit 14.

Then, the lung sound abnormality detection means 163 calculates the emergency level 1535 on the basis of the analysis result of the lung sound data of each auscultation position, and records it in the field of the emergency level 1535 of the analysis object lung sound information 153 (step S13). Then, the analysis result output means 164 reads the analysis object lung sound information 153 from the storage unit 15, displays the analysis object lung sound information 153 on the screen display unit 14, and sends an email to which the analysis object lung sound information 153 is attached as a file, to the contact email address 1526 of the lung sound record 152 via the communication I/F unit 12 (step S14). In step S14, the analysis result output means 164 may output the analysis object lung sound information 153 for the purpose of notifying the concerned persons of the heart failure condition of the patient A, and transmit the analysis object lung sound information 153 to a terminal of a medical specialist in heart failure for the purpose of using the lung sound data acquired from the patient A as learning data.

Next, the details of step S12 for performing acquisition of analysis object lung sounds and detection of abnormality will be described with reference to the flowchart of FIG. 10. FIG. 10 is a flowchart showing an example of a detailed procedure of step S12 of FIG. 9.

Referring to FIG. 10, the analysis object lung sound acquisition means 162 calculates the frequency of appearance of an abnormal sound for each auscultation position, on the basis of presence or absence of an abnormal sound recorded in the auscultation observation of each auscultation position in one or more pieces of the auscultation information 1527 recorded in the lung sound record 152 of the patient A (step S21). Specifically, first, the analysis object lung sound acquisition means 162 initializes the frequency counter of each of the auscultation positions (1) to (12) to zero. Then, the analysis object lung sound acquisition means 162 focuses on the auscultation information 1527 in which the auscultation date/time recorded in the lung sound record 152 is the latest. Then, when there is at least one auscultation observation describing presence of an abnormal sound in the auscultation observations recorded for the auscultation position (1), the analysis object lung sound acquisition means 162 increments the frequency count corresponding to the auscultation position (1) by one. Then, when there is at least one auscultation observation describing presence of an abnormal sound in the auscultation observations recorded for the auscultation positions (2) to (12), the analysis object lung sound acquisition means 162 increments the frequency count corresponding to the auscultation positions (2) to (12) by one. Then, the analysis object lung sound acquisition means 162 focuses on the auscultation information 1527 whose auscultation date/time is immediately before the latest date/time, and performs the same operation as that performed using the auscultation information 1527 of the latest date/time on the frequency count of each of the auscultation positions (1) to (12). Thereafter, the analysis object lung sound acquisition means 162 repeats the above-described operation until the processing performed on the predetermined number of auscultation information 1527 ends or until the processing performed on the oldest auscultation information 1527 recorded in the lung sound record 152 ends, whichever earlier. Then, the analysis object lung sound acquisition means 162 determines the value of the frequency count of each of the auscultation positions (1) to (12) to be the abnormal sound appearance frequency of each of the auscultation positions (1) to (12).

Then, the analysis object lung sound acquisition means 162 determines the sequence (order) of the auscultation positions for auscultating the lung sounds from the patient A, on the basis of the abnormality frequency of each of the auscultation positions (1) to (12) of the patient A (step S22). When there is a difference in the occurrence frequency of abnormal sounds such as rales among the auscultation positions (1) to (12) of the patient A, it means that the patient A has an auscultation position where an abnormal sound is likely to occur relatively and an auscultation position where an abnormal sound is not likely to occur. Therefore, by performing auscultation according to the sequence of the auscultation positions on the basis of the past abnormality frequency of the auscultation positions (1) to (12) of the patient A, even if auscultation is interrupted for any reason such as circumstances of the patient A, and the heart failure condition of the patient A is to be determined based on the analysis result of the lung sound data of some auscultation positions in which auscultation has been performed, it is possible to reduce the probability of overlooking exacerbation of the heart failure.

The analysis object lung sound acquisition means 162 may determine the sequence of the auscultation positions only based on the abnormality frequency of each auscultation position of the patient A. In that case, the analysis object lung sound acquisition means 162 may determine the result of sorting the auscultation positions in the descending order (order from the highest to the lowest) of abnormality frequency to be the sequence of the auscultation positions. In the case where the abnormality frequency of each of the auscultation positions (1) to (12) of the patient A is as illustrated in FIG. 11, an example of sequence of the auscultation positions based on the result of sorting the auscultation positions in the descending order of the abnormality frequency is as illustrated in an auscultation sequence 1 of FIG. 11. In the auscultation sequence 1, the auscultation position (11) in which the abnormality frequency is 4, that is, the largest, is auscultated first. The auscultation positions whose abnormality frequency is the second largest are auscultation positions (6) and (12) in which the abnormality frequency is 3. Since there is no difference in the abnormality frequency, in the auscultation sequence 1, the auscultation position (12) in the anterior side of the chest that is the same as the first one, is set to be the second, and the auscultation position (6) in the posterior side of the chest is set to be the third. Thereafter, the sequence of the auscultation positions is set to (5), (9), (10), (7), (1), (2), (3), (4), and (8) in the similar manner.

By determining the sequence of the auscultation positions only based on the abnormality frequency of each auscultation position of the patient as described above, it is possible to acquire the lung sound data in order from an auscultation position having higher probability of abnormal lung sound. However, depending on the distribution of abnormality frequencies, auscultation on the anterior side of the chest and auscultation on the posterior side of the chest must be changed some times, which may cause a burden on the patient and the operator.

Therefore, it is possible to determine the sequence of the auscultation positions while considering not only the abnormality frequency of each auscultation position of the patient but also reduction of a burden on the patient and the operator. For example, the analysis object lung sound acquisition means 162 determines that a side where an auscultation position having the highest abnormality frequency is present, of the posterior side of the chest and the anterior side of the chest, to be a site that is auscultated first, and determine the side opposite to such a site to be a site that is auscultated next. Further, the analysis object lung sound acquisition means 162 determines, for each site, a result of sorting the abnormality frequencies of all auscultation positions of the site in the descending order to be the sequence of the auscultation positions of the site. An example of the auscultation sequence determined by this determination method will be shown as an auscultation sequence 2 in FIG. 11.

In the auscultation sequence 2, the anterior side of the chest where the auscultation position (11) whose abnormality frequency is 4, that is, the largest, is present is determined to be the site to be auscultated first, and the sequence of the auscultation positions (7) to (12) in the anterior side of the chest is determined to be a sequence of the auscultation positions (11), (12), (9), (10), (7), and (8), in accordance with the result of sorting the auscultation sequence in the descending order of the abnormality frequency. Further, in the auscultation sequence 2, after completion of auscultation of all auscultation positions on the anterior side of the chest, auscultation is switched to the posterior side of the chest, and the sequence of the auscultation positions (1) to (6) in the posterior side of the chest is determined to be a sequence of the auscultation positions (6), (5), (1), (2), (3), and (4), in accordance with the result of sorting the auscultation sequence in the descending order of abnormality frequency.

Referring to FIG. 10 again, the analysis object lung sound acquisition means 162 focuses on the first auscultation position in the sequence (step S23). Then, the analysis object lung sound acquisition means 162 acquires digital time-series acoustic signals including the lung sounds of the focused auscultation position from the electronic stethoscope 11 (step S24). At that time, the analysis object lung sound acquisition means 162 may display, on the screen display unit 14, a guidance screen for giving guidance on the focused auscultation position to the operator using the electronic stethoscope 11 to support acquisition of the lung sounds. Moreover, the analysis object lung sound acquisition means 162 may play a guidance sound, from a loudspeaker not illustrated, for giving guidance on the focused auscultation position to the operator using the electronic stethoscope 11 to support acquisition of the lung sounds. In this way, the analysis object lung sound acquisition means 162 gives guidance on the auscultation position to which the electronic stethoscope 11 should be applied, to the operator by using an image or a sound, and acquires digital time-series acoustic signals including lung sounds of the guided auscultation position from the electronic stethoscope 11.

Then, the analysis object lung sound acquisition means 162 measures the quality of the acquired lung sounds (step S25). In general, time-series acoustic signals output from the electronic stethoscope 11 include lung sounds of the patient A in the frequency band of 100 Hz to about 2 kHz, and the background noise (stationary noise) is also included in the same frequency band. For example, environment sounds, person's voice, metal noise, and the like entering from the outside through the body of the patient A or through the gap between the skin of the patient A and the chest piece are examples of the stationary noise. When the intensity of the lung sounds in the time-series acoustic signals is small and the intensity of the background noise is large, it is difficult to detect abnormality in the lung sounds. Therefore, the analysis object lung sound acquisition means 162 first uses a bandpass filter to extract time-series acoustic signals in the frequency band of 100 Hz to about 2 kHz from the time-series acoustic signals output from the electronic stethoscope 11. Then, the analysis object lung sound acquisition means 162 calculates the intensity of the lung sounds and the intensity of the background noise in the extracted time-series acoustic signals, and calculates the difference degree thereof as an index value of the quality of the lung sounds. Hereafter, a method of calculating an index value of quality of lung sounds will be described.

FIG. 12 is a schematic diagram illustrating an exemplary waveform of time-series acoustic signals including lung sounds output from the electronic stethoscope 11. In general, lung sounds include tracheal breath sounds, bronchial vesicular breath sounds, bronchial breath sounds, and vesicular breath sounds. FIG. 12 is a schematic diagram illustrating lung sounds that are exemplary vesicular breath sounds heard on almost all chest wall, that is, at all auscultation positions (1) to (12). Referring to FIG. 12, in the time-series acoustic signals including lung sounds, the amplitude largely varies at the start of inspiration. Moreover, at the start of expiration, the amplitude also largely varies although it is less than that at the start of inspiration. Therefore, the analysis object lung sound acquisition means 162 compares the time-series acoustic signals with a threshold T1 with which amplitude variation at the start of inspiration can be determined, and detects the point of time when the amplitude of the time-series acoustic signals becomes larger than the threshold T1 as the inspiration start time. Further, the analysis object lung sound acquisition means 162 sets the start time of inspiration up to the start time of the next inspiration to be a section of one breathing cycle, compares the amplitude of the time-series acoustic signals in the section with a threshold T2 (<T1) with which amplitude variation at the start of expiration can be determined, and detects the point of time when the amplitude of the time-series signals becomes larger than the threshold T2 as the expiration start time. Here, for merely distinguishing a pause phase from the other phases, it is only necessary to detect the start of inspiration. However, in the present invention, a phase other than the pause phase is further divided into an inspiratory phase and an expiratory phase. Therefore, the start of expiration is also detected.

In general, it is known that breathing of a person is configured of an inspiratory phase of about one second and an expiratory phase of about one second, and a pause phase of about one to one and a half seconds until the next inspiration. That is, there is a pause phase during which neither inspiration nor expiration is made, immediately before the inspiration start point of time. The analysis object lung sound acquisition means 162 detects a predetermined period (for example, one second) immediately before the detected inspiration start point of time as a pause phase. Then, the analysis object lung sound acquisition means 162 calculates the intensity of the time-series acoustic signals in the pause phase as the intensity of the background noise. As the intensity of the time-series acoustic signals, a root-mean-square of the amplitude value may be used for example. However, it is not limited thereto, and may be an amplitude or the like. Further, the analysis object lung sound acquisition means 162 calculates a value obtained by subtracting the intensity of the background noise from the intensity of the time-series acoustic signals in the inspiratory phase and/or expiratory phase, as the intensity of the lung sounds. Then, the analysis object lung sound acquisition means 162 uses the ratio of the calculated intensity of the lung sounds to the intensity of the background noise, as an index value of the lung sound quality. Note that an index value of the lung sound quality is not limited to that described above. It is also possible to use an S/N ratio calculated from the intensity of the lung sounds and the intensity of the background noise as an index value.

In the examples described above, a method of detecting a pause phase has been described by using vesicular breath sounds as an example. However, at auscultation positions of the middle lung field and the upper lung field, bronchial vesicular breath sounds are also heard together with the vesicular breath sounds. However, in the bronchial vesicular breath sounds, the amplitude of inspiration is equal to or larger than the amplitude of the expiration. Therefore, even in the case where the bronchial vesicular breath sounds are heard together with the vesicular breath sounds, it is possible to detect inspiration start timing and expiration start timing by the method as described in FIG. 12. However, when the bronchial vesicular breath sounds are similar to the trachea breath sounds, the amplitude may be larger at the time of expiration than that at the time of inspiration. Therefore, when the bronchial vesicular breath sounds are similar to the trachea breath sounds, inspiration and expiration may be reversed in the method described in FIG. 12. Specifically, detection may be performed as described below, for example.

First, the frequency in which the amplitude of the frequency spectrum of auscultated lung sounds becomes maximum is compared with a predetermined threshold frequency. Then, when the frequency in which the amplitude of the frequency spectrum of the auscultated lung sounds becomes maximum is equal to or higher than the threshold frequency, it is determined that the bronchial vesicular breath sounds included in the lung sounds are similar to the trachea breath sounds, and the start timing of inspiration and the start timing of expiration are detected by reversing inspiration and expiration in the method described in FIG. 12. On the contrary, when the frequency in which the amplitude of the frequency spectrum of the auscultated lung sounds becomes maximum is lower than the threshold frequency, it is determined that the bronchial vesicular breath sounds included in the lung sounds are not similar to the trachea respiration sounds, and the start timing of inspiration and the start timing of expiration are detected by the method described in FIG. 12. The threshold frequency is a threshold with which whether or not the bronchial vesicular breath sounds included in the lung sounds are similar to the trachea breath sounds can be distinguished. For example, the threshold frequency can be determined previously from the frequency band between the frequency in which the amplitude of the frequency spectrum of the vesicular breath sounds becomes maximum and the frequency in which the amplitude of the frequency spectrum of the trachea breath sounds becomes maximum that is higher than that. Further, instead of the “frequency in which the amplitude of the frequency spectrum becomes maximum”, “spectrum center of gravity” used as a measure of representing the spectrum shape may be used.

Further, in the example described above, the start time of expiration and the start time of inspiration are detected from the time-series acoustic signals output from the electronic stethoscope 11, and a predetermined period of time immediately before the detected start point of inspiration is detected as a pause phase. However, the method of detecting an inspiratory phase, an expiratory phase, and a pause phase is not limited to that described above. For example, the analysis object lung sound acquisition means 162 may be configured to acquire estimated probabilities of an inspiratory phase, an expiratory phase, and a pause phase for each section from a learning model, by inputting time-series acoustic signals including the lung sounds of the patient A into the learning model having been learned through machine learning for estimating which section of the time-series acoustic signals including the lung sounds output from the electronic stethoscope is an inspiratory phase, an expiratory phase, or a pause phase. The learning model can be generated in advance through machine learning using a machine learning algorism such as a neural network by using time-series acoustic signals including various lung sounds as teacher data. Further, the analysis object lung sound acquisition means 162 may detect breath timing such as the start of inspiration and the start of expiration of the patient A from those other than the time-series acoustic signals output from the electronic stethoscope. For example, the analysis object lung sound acquisition means 162 may detect the breath timing of the patient A by using a breath amount sensor such as a lung tachograph or a breath band for detecting a shape change of the chest or abdominal region due to the breathing action by a sensor.

Then, the analysis object lung sound acquisition means 162 compares the index value of the quality of the lung sounds with a quality threshold set in advance (step S26). Then, when the index value of the quality of the lung sounds is smaller than the threshold, the analysis object lung sound acquisition means 162 displays, on the screen display unit 14, warning indicating that the quality of the lung sounds at the auscultation position auscultated by the electronic stethoscope 11 is bad (step S27). The operator who recognizes the warning performs an operation to obtain the lung sounds of the focused auscultation position by the electronic stethoscope 11 again, after taking measures to decrease the background noise or increase the lung sounds (step S28). As measures to reduce the background noise, it is considerable to close the window to make the room silent, put on the chest piece closely to the skin of the patient A so as to prevent environmental sounds from entering from the gap between the skin of the patient A and the chest piece, and the like. Further, as measures to increase the lung sounds, it is considerable to instruct the patient A to breathe more largely. At that time, it is also possible to instruct the breath timing to the patient A by the method as described in Patent Literature 9 for example. Then, the analysis object lung sound acquisition means 162 returns to the processing of step S25 and repeats the same processing as that described above.

On the contrary, when the index value of the quality of the lung sounds is equal to or larger than the threshold, the analysis object lung sound acquisition means 162 removes the period of pause phase and the background noise from the digital time-series acoustic signals including the lung sounds of the focused auscultation position, and stores, in the analysis object lung sound information 153, the digital time-series acoustic signals after the removal of the period of the pause phase and the background noise, in association with the focused auscultation position (step S29). Removal of the period of pause phase and the background noise is performed as described below.

First, the analysis object lung sound acquisition means 162 divides the digital time-series acoustic signals including the lung sounds of the focused auscultation position into a section configured of an inspiratory phase and an expiratory phase immediately thereafter (hereinafter referred to as an inspiratory/expiratory section) and a section of a pause phase (hereinafter referred to as a pause section). Then, the analysis object lung sound acquisition means 162 calculates the frequency spectrum of the inspiratory/expiratory section and the pause section by applying fast Fourier transform (FFT) to the digital time-series acoustic signals in each of the inspiratory/expiratory section and the pause section. Then, the analysis object lung sound acquisition means 162 subtracts the frequency spectrum of the pause section from the frequency spectrum of the inspiratory/expiratory section. By the subtraction, the background noise included in the inspiratory phase and the expiratory phase is suppressed. Then, the analysis object lung sound acquisition means 162 applies inverse frequency transform to the frequency spectrum of the inspiratory/expiratory section to thereby generate digital time-series acoustic signals after the removal of the noise in the inspiratory/expiratory section. Then, the analysis object lung sound acquisition means 162 records the generated digital time-series acoustic signals after the removal of the noise in the inspiratory/expiratory section, in the analysis object lung sound information 153 in association with the focused auscultation position. Note that the analysis object lung sound acquisition means 162 may remove the period of pause phase from the digital time-series acoustic signals including the lung sounds of the focused auscultation position and not remove the background noise. In that case, the analysis object lung sound acquisition means 162 divides the digital time-series acoustic signals including the lung sounds of the focused auscultation position into two, that is, the inspiratory/expiratory section and the pause section, and records the digital time-series acoustic signals of the inspiratory/expiratory section in the analysis object lung sound information 153 in association with the focused auscultation position.

Then, the lung sound abnormality detection means 163 detects abnormality in the lung sounds from the lung sound data recorded in the analysis object lung sound information 153 in association with the focused auscultation position, and records the detection result in the analysis object lung sound information 153 in association with the focused auscultation position (step S30). Detection of abnormality in the lung sounds is performed by the abnormality detection method using supervised learning as described above.

Each time abnormality detection of lung sound data of the focused auscultation position is performed by the lung sound abnormality detection means 163, the analysis result output means 164 displays the abnormality detection result on the screen display unit 14 (step S31). Thereby, the operator can immediately recognize whether or not the lung sound data of the auscultation position is abnormal lung sounds, at the time of auscultation.

Upon completion of acquisition and analysis of the lung sound data of the focused auscultation position, the analysis object lung sound acquisition means 162 determines whether or not acquisition and analysis of lung sound data have been completed for all auscultation positions (step S32). When there remains any auscultation position in which acquisition has not been completed, the analysis object lung sound acquisition means 162 moves the focus to the next auscultation position in the sequence (step S33), and returns to step S24 and repeats the same processing as that described above.

When acquisition and analysis of lung sound data of all auscultation positions have been completed, the analysis object lung sound acquisition means 162 ends the processing of FIG. 10. The analysis object lung sound acquisition means 162 may end the processing of FIG. 10 before acquisition and analysis of lung sound data of all auscultation positions have been completed due to the circumstances or the like of the patient A. When the processing of FIG. 10 is terminated, the lung sound information 1534 in the analysis object lung sound information 153 corresponding to the auscultation position in which acquisition and analysis of lung sound data have not been performed still has a Null value.

Next, the details of step 813 of FIG. 9 to calculate the emergency level 1535 will be described.

The lung sound abnormality detection means 163 determines the severity of the heart failure of the patient A on the basis of the analysis result of the lung sound data of each auscultation position, and calculates the emergency level 1535 based on the determined severity. When determining the severity of the heart failure, the lung sound abnormality detection means 163 determines the severity of the heart failure with reference to a determination table for determining the severity of the heart failure from the analysis result of the lung sound data of each auscultation position.

FIG. 13 illustrates an example of the determination table. The determination table illustrated in FIG. 13 includes a column corresponding to each of the auscultation positions (1) to (12) one to one, and a row corresponding to a degree of severity one to one, and at an intersection between a column and a row, a + sign indicating that there is abnormality in the lung sounds and a − sign indicating that there is no abnormality in the lung sounds are set. Referring to FIG. 13, in the determination table, when there is no abnormality in the lung sounds of any auscultation position, it is determined that the severity is 0. Further, in the determination table, when there is abnormality in the lung sounds at at least one of the auscultation positions (11) and (12) set in the lower lung field in the anterior side of the chest and there is no abnormality in the lung sounds at the other auscultation positions (1) to (10), it is determined that the severity is 1.

Further, in the determination table, when there is abnormality in the lung sounds at both the auscultation positions (11) and (12), there is abnormality in the lung sounds at either one of the auscultation positions (5) and (6) set in the lower lung field of the posterior side of the chest, and there is no abnormality in the lung sounds in the other auscultation positions (1) to (4) and (7) to (10), it is determined that the severity is 2. The severity N set in the last row means that there is abnormality in the lung sounds at all auscultation positions (1) to (12). In FIG. 13, although description of one or more degrees of severity are omitted between the severity 2 and the severity N, auscultation positions having abnormality in the lung sounds and auscultation position not having abnormality in the lung sounds are also set for them. In one or more degrees of severity between the severity 2 and the severity N, the number of auscultation positions at which there is abnormality in the lung sounds is four or larger and less than twelve, and the number increases as closer to the severity N.

In the determination table illustrated in FIG. 13, the severity of heart failure is classified into N+1 classes from severity 0 to severity N, depending on the combination of presence or absence of abnormality in the lung sounds at the auscultation positions (1) to (12). Here, the severity 0 is a state where no abnormal lung sound is heard. Therefore, it can be said that heart failure is remitted. The severity 1 is a condition where abnormal lung sounds are heard only in the lower lung field of the anterior side of the chest. Therefore, heart failure is mild although not remitted, and is a condition in which some patients are discharged from hospital in such a condition. The severity 2 is a condition where abnormal lung sounds are heard in one of the lower lung fields of the posterior side of the chest in addition to the lower lung field of the anterior side of the chest. Therefore, it can be said that it is severe than the severity 1. However, it still belongs to the mild case, so there is a high possibility of preventing re-hospitalization if appropriate treatment is taken at this point.

The determination table in which the severity of heart failure is determined from the analysis result of the auscultation position is not limited to that illustrated in FIG. 13. For example, it is known that when rales are heard only at the end of inspiration, it is mild, and when the rales are heard immediately after the start of inspiration, it is severe. Therefore, in addition to presence or absence of abnormal lung sounds at each auscultation position, the type of abnormal lung sounds and the timing that the abnormal lung sounds are heard may be added to the determination table, and the severity of the heart failure may be determined according to a combination of an auscultation position, presence or absence of abnormal lung sounds, the type of abnormal lung sounds, and the timing that the abnormal lung sounds are heard.

Further, the lung sound abnormality detection means 163 may determine the severity of the heart failure of the patient A from the number of auscultation positions where the abnormal lung sounds are heard, regardless of the auscultation position. For example, the lung sound abnormality detection means 163 may determine the severity to be 0, 1, 2, 3, or 4 (maximum) when the number of auscultation positions where abnormal lung sounds are heard is 0, 1 to 2, 3 to 4, 5 to 8, or 9 or more, respectively.

Further, when the processing illustrated in FIG. 10 is terminated due to circumstances or the like of the patient A so that at least part of the analysis results in the lung sound information 1534 of the respective auscultation positions has a NULL value, the lung sound abnormality detection means 163 performs the processing as described below. First, the lung sound abnormality detection means 163 determines whether or not the following condition is satisfied: the number of auscultation positions in which the analysis result has a Null value, that is, lung sound data is not acquired and analysis of whether or not the lung sounds are abnormal lung sounds has not been performed, is less than a first threshold set in advance. In other words, the lung sound abnormality detection means 163 determines whether or not the following condition is satisfied: the number of auscultation positions in which the lung sound data is acquired and analysis of whether or not the lung sounds are abnormal lung sounds has been performed, is equal to or larger than a second threshold set in advance. Here, the first threshold and the second threshold may be fixed values or variable values corresponding to the latest condition of the patient. In the case of fixed values, the first threshold may be 4 or smaller and the second threshold may be 8 or larger, for example. In the case of variable values, for the patient who has no abnormal lung sound in any auscultation position in the latest condition, the first threshold may be 10 or smaller and the second threshold may be 2 or larger for example, and for the other patients, the values may be the same as the fixed values. Then, when the condition is not satisfied, the lung sound abnormality detection means 163 does not calculate the severity (and does not calculate the emergency level accordingly), and ends the current lung sound analysis in error and displays the fact on the screen display unit 14. This is because not to provide erroneous information to the operator and the like.

On the contrary, when the condition is satisfied, the lung sound abnormality detection means 163 assumes that no abnormal lung sound is detected at auscultation positions in which analysis of whether or not abnormal lung sounds are heard has not been performed, and calculates the severity. Then, the lung sound abnormality detection means 163 holds the calculated severity as the most optimistic value. That is, when the calculated severity is the severity 1, it is not held as “severity 1” but held as “severity 1 or higher” or “at least severity 1”. For example, it is assumed that with respect to the patient who has no abnormal lung sound in any auscultation position in the latest condition, acquisition and analysis of lung sound data is performed on only two positions, that is, the auscultation positions (11) and (12), and abnormal lung sounds are detected at at least one of the auscultation positions. In that case, the lung sound abnormality detection means 163 assumes that no abnormal lung sound is detected at the other auscultation positions (1) to (10), determines the severity to be the severity 1 based on the determination table of FIG. 13, and accordingly determines to be “severity 1 or higher”.

When determining the severity of the heart failure on the analysis result of the lung sound data of each auscultation position as described above, the lung sound abnormality detection means 163 determines the emergency level 1535 from the determined severity. For example, the lung sound abnormality detection means 163 may determine the emergency level 1535 only based on the severity 0 to N of the heart failure. That is, the lung sound abnormality detection means 163 may set the range that can be taken by the emergency level 1535 to be N+1 classes from the emergency level 0 to the emergency level N, and determine an emergency level i that corresponds to the determined severity i (i=0 to N) of the heart failure one to one.

Further, the lung sound abnormality detection means 163 may determine the emergency level 1535 on the basis of the severity 0 to N of the heart failure and the condition of the patient A. For example, as the condition of the patient A, whether or not the weight is increased by a certain quantity in a unit period (for example, 3 kg or more in a week), presence or absence of subjective symptoms such as edema, cough, anorexia, or the like, whether or not the pulse exceeds a prescribed number, and the like may be considered. Then, the lung sound abnormality detection means 163 may set the emergency level that is obtained by correcting the emergency level determined based on the severity of the heart failure to be higher according to the condition of the patient A, as a final emergency level. For example, when a weight increase is observed although the emergency level determined from the severity of the heart failure is the emergency level 0 or 1, the lung sound abnormality detection means 163 may increase the emergency level to be 1 or 2. However, the upper limit of the emergency level after the correction is N.

Next, description will be given on an operation after transmission of an email, to which the analysis object lung sound information 153 is attached as a file, to a terminal device of a medical specialist in heart failure by the analysis result output means 164.

In the terminal device of a medical specialist in heart failure that received the email, the analysis object lung sound information 153 stored in the attached file is analyzed by the medical specialist in heart failure. The analysis object lung sound information 153 is not limited to be attached as a file but may be in a form to be shared with a medical specialist in heart failure in SaaS format by posting a link or the like. For example, a medical specialist replays the lung sound data of each auscultation position recorded in the analysis object lung sound information 153 by a personal computer or the like, and diagnoses whether or not adventitious sounds such as rales are heard from the lung sounds of the patient A. Then, the medical specialist creates auscultation observations on the lung sound data of respective auscultation positions, and records them in the analysis object lung sound information 153 as illustrated in FIG. 6. The analysis object lung sound information 153 in which the auscultation observations of the medical specialist are recorded is returned to the lung sound analysis device 10, that is, the transmission source, by means of a communication means such as an email.

The learning data generation means 165 updates the original analysis object lung sound information 153 recorded in the storage unit 15, according to the analysis object lung sound information 153 with auscultation observations received via the communication I/F unit 12 of the lung sound analysis device 10. Then, for each auscultation position of the analysis object lung sound information 153, the learning data generation means 165 creates learning data for each set of lung sound data and auscultation observation recorded corresponding thereto. For example, from a set of auscultation observation indicating abnormal lung sounds and lung sound data, the learning data generation means 165 creates learning data including a label indicating abnormal lung sounds and the lung sound data. Further, from a set of auscultation observation indicating normal lung sounds and lung sound data, the learning data generation means 165 creates learning data including a label indicating normal lung sounds and the lung sound data. At that time, personal information of the patient may be added to the label. The learning data generation means 165 also records the created learning data in the learning data DB 154 in association with the auscultation position. At that time, the learning data creation means 165 may add a time stamp of record date/time and the like to the learning data so as to distinguish the data from other learning data.

As described above, the learning data added to the learning data DB 154 is to be used for re-learning the model for detecting abnormality in lung sounds by the learning means 166 from the next time. In this way, it is possible to gradually improve the accuracy of the model for detecting abnormality in lung sounds.

As described above, according to the present embodiment, it is possible to collect learning data for detecting abnormality in lung sounds in a clinic with no medical specialist or at home of a patient. This is because the lung sound analysis device 10 acquires time-series acoustic signals including the lung sounds of a heart failure patient, detects abnormal lung sounds from the acquired time-series acoustic signals, transmits analysis object lung sound information in which the acquired time-series acoustic signals and the detection result are associated with each other to a terminal device of a medical specialist, and when receiving analysis object lung sound information in which observations on the time-series acoustic signals by the medical specialist are added, creates learning data for detecting abnormality in lung sounds based on the analysis object lung sound information to which the observations by the medical specialist are added.

Second Exemplary Embodiment

FIG. 14 is a block diagram of a lung sound analysis system 20 according to a second exemplary embodiment of the present invention. Referring to FIG. 14, the lung sound analysis system 20 is configured of a plurality of lung sound analysis devices 21, a server device 22, and a terminal device 24. The lung sound analysis devices 21, the server device 22, and the terminal device 24 are communicably connected with each other over a network such as the Internet.

The lung sound analysis device 21 is an information processing device that acquires and analyzes lung sounds from a patient who received treatment for heart failure and was discharged from hospital. The terminal device 24 is a terminal device used by a medical specialist in heart failure. The lung sound analysis device 21 and the terminal device 24 may be a smartphone, a tablet terminal, a PDA, a laptop personal computer, or the like, but is not limited thereto. The lung sound analysis device 21 includes an electronic stethoscope, a communication IN unit, an operation input unit, a screen display unit, a storage unit, and an arithmetic processing unit that are not illustrated. The terminal device 24 includes a communication I/F unit, an operation input unit, a screen display unit, a storage unit, and an arithmetic processing unit that are not illustrated.

The server device 22 is a computer that provides, to the lung sound analysis devices 21, various services required for lung sound analysis over the network 23. For example, the server device 22 stores therein at least part of the lung sound record 152, the analysis object lung sound information 153, the learning data DB 154, and the program 151 illustrated in FIG. 1, and provides the lung sound analysis devices 21 with them over the network 23. Therefore, the lung sound analysis device 21 is not needed to store at least part of the lung sound record 152, the analysis object lung sound information 153, the learning data DB 154, and the program 151 in the storage unit 15 as compared with the lung sound analysis device 10 of FIG. 1, so that the memory capacity can be reduced.

The server device 22 also provides the lung sound analysis device 21 with at least part of the functions of the lung sound record acquisition means 161, the analysis object lung sound acquisition means 162, the lung sound abnormality detection means 163, the analysis result output means 164, the learning data generation means 165, and the learning means 166 illustrated in FIG. 1, over the network 23. For example, the server device 22 executes at least part of the processing of steps S1 to S2 of FIG. 7, steps S11 to S14 of FIG. 9, and steps S21 to S33 of FIG. 10, or the learning data generation process, on behalf of the lung sound analysis device 21. Therefore, in the lung sound analysis device 21, the configuration of the arithmetic processing unit 16 can be simplified as compared with the lung sound analysis device 10 of FIG. 1.

Further, the terminal device 24 has a function of replaying lung sound data of each auscultation position recorded in the analysis object lung sound information received by email or the like from the lung sound analysis devices 21, a function of inputting auscultation observations by a medical specialist in heart failure, and a function of transmitting analysis object lung sound information with auscultation observations to the lung sound analysis device 21 by email or the like. While FIG. 14 illustrates only one terminal device 24, a plurality of terminal devices 24 corresponding to the number of medical specialists in heart failure may exist.

Third Exemplary Embodiment

FIG. 15 is a block diagram of a lung sound analysis system 30 according to a third exemplary embodiment of the present invention. Referring to FIG. 14, the lung sound analysis system 30 is configured of an acquisition means 31, a detection means 32, an observation acquisition means 33, and a generation means 34.

The acquisition means 31 is configured to acquire time-series acoustic signals including lung sounds of a heart failure patient. The acquisition means 31 may be configured as similar to step S24 of FIG. 10 for example, but is not limited thereto.

The detection means 32 is configured to detect abnormal lung sounds from the time-series acoustic signals acquired by the acquisition means 31. The detection means 32 may be configured as similar to step S30 of FIG. 10 for example, but is not limited thereto.

The observation acquisition means 33 is configured to transmit analysis object lung sound information in which the time-series acoustic signals acquired by the acquisition means 21 and a detection result by the detection means 32 are associated with each other to a terminal device of a medical specialist, and receive analysis object lung sound information to which observations by the medical specialist on the time-series acoustic signals are added, from the terminal device. The observation acquisition means 33 may be configured as similar to the analysis result output means 164 of FIG. 1 for example, but is not limited thereto.

The generation means 34 is configured to generate learning data for detecting abnormal lung sounds, on the basis of the analysis object lung sound information to which observations by the medical specialist are added, received by the observation acquisition means 33. The generation means 34 may be configured as similar to the learning data generation means 165 of FIG. 1 for example, but is not limited thereto.

The lung sound analysis system 30 configured as described above functions as described below. First, the acquisition means 31 acquires time-series acoustic signals including lung sounds of a heart failure patient. Then, the detection means 32 detects abnormal lung sounds from the acquired time-series acoustic signals. Then, the observation acquisition means 33 transmits analysis object lung sound information in which the time-series acoustic signals acquired by the acquisition means 31 and a detection result by the detection means 32 are associated with each other to a terminal device of a medical specialist, and receives analysis object lung sound information to which observations by the medical specialist on the time-series acoustic signals are added, from the terminal device. Then, the generation means 34 generates learning data for detecting abnormal lung sounds, on the basis of the analysis object lung sound information to which the observations by the medical specialist are added, received by the observation acquisition means 33.

According to the lung sound analysis system 30 that is configured and operates as described above, it is possible to efficiently collect learning data for detecting abnormal lung sounds, in a clinic without a medical specialist or at home of a patient. This is because the lung sound analysis system 30 acquires time-series acoustic signals including lung sounds of a heart failure patient, detects abnormal lung sounds from the acquired time-series acoustic signals, transmits analysis object lung sound information in which the acquired time-series acoustic signals and a detection result are associated with each other, to a terminal device of a medical specialist, and upon receipt of analysis object lung sound information to which observations by the medical specialist on the time-series acoustic signals are added, generates learning data for detecting abnormal lung sounds, on the basis of the analysis object lung sound information to which the observations by the medical specialist are added.

While the present invention has been described with reference to the exemplary embodiments described above, the present invention is not limited to the above-described embodiments. The form and details of the present invention can be changed within the scope of the present invention in various manners that can be understood by those skilled in the art. For example, configurations as described below are also included in the present invention.

For example, in the lung sound analysis device 10 illustrated in FIG. 1, the lung sound abnormality detection means 163 uses an abnormality detection method using the supervised learning to detect abnormality in the lung sound data of a patient acquired by the electronic stethoscope. However, the lung sound abnormality detection means 163 may detect abnormal lung sounds by using another abnormality detection method such as an abnormality detection method using non-supervised learning. As an abnormality detection method using non-supervised learning, a method in which normal lung sound data is learned without using abnormal sounds for learning and lung sounds deviated from the learned normal sounds are detected as abnormal lung sounds, may be used.

Further, the analysis object lung sound acquisition means may instruct the subject to breathe larger, when it is determined that the lung sounds are not recorded correctly, for example. Further, the analysis object lung sound acquisition means may instruct the operator on the auscultation position by means of augmented reality (AR) display. Further, the analysis object lung sound acquisition means may change the auscultation position on the basis of the previously registered information such as sex and the like of the subject. Further, the analysis object lung sound acquisition means may start breath instruction when it is detected that the stethoscope is put on, that is, the chest piece comes into contact with, the body of the subject. Further, the analysis object lung sound acquisition means may perform breath instruction by avatar display or voice designated by the subject. Further, when acquisition of lung sounds is not performed within a predetermined period, the analysis object lung sound acquisition means may urge acquisition of lung sounds by avatar display or voice designated by the subject. Further, the analysis result output means may display an analysis result, a lung sound record used for the analysis, and storage information including the analysis object lung sound information, on the screen display unit or the like in the time-series manner. Further, the analysis result output means may transmit information to the server even when abnormality is not detected.

INDUSTRIAL APPLICABILITY

The present invention is applicable to a system for analyzing lung sounds of a person, and in particular, applicable to a system for creating learning data for detecting abnormality in lung sounds.

The whole or part of the exemplary embodiments disclosed above can be described as, but not limited to, the following supplementary notes.

- (Supplementary Note 1)

A lung sound analysis system comprising:

- acquisition means for acquiring time-series acoustic signals including lung sounds of a heart failure patient;
- detection means for detecting abnormal lung sounds from the acquired time-series acoustic signals;
- observation acquisition means for transmitting, to a terminal device of a medical specialist, analysis object lung sound information in which the acquired time-series acoustic signals and a result of the detection are associated with each other, and receiving, from the terminal device, analysis object lung sound information to which an observation by the medial specialist on the time-series acoustic signals is added; and
- generation means for generating learning data for detecting abnormal lung sounds on a basis of the analysis object lung sound information to which the observation by the medial specialist is added.
- (Supplementary Note 2)

The lung sound analysis system according to supplementary note 1, wherein

- before performing the transmitting, the observation acquisition means determines whether or not abnormal lung sounds are detected in the result of the detection, and performs the transmitting only when abnormal lung sounds are detected.
- (Supplementary Note 3)

The lung sound analysis system according to supplementary note 1 or 2, wherein before performing the transmitting, the observation acquisition means determines whether or not consent to use of the time-series acoustic signals for the learning data is given by the heart failure patient, and performs the transmitting only when the consent is given.

- (Supplementary Note 4)

The lung sound analysis system according to any of supplementary notes 1 to 3, wherein

- before performing the transmitting, the observation acquisition means determines whether or not personal information of the heart failure patient matches personal information of a target for collecting lung sound data, and performs the transmitting only when the personal information matches.
- (Supplementary Note 5)

The lung sound analysis system according to any of supplementary notes 1 to 4, wherein

- the acquisition means determines a pause phase of respiration of the heart failure patient, and divides the time-series acoustic signals into time-series acoustic signals in a period of the pause phase and time-series acoustic signals in a period other than the pause phase according to a result of the determination.
- (Supplementary Note 6)

The lung sound analysis system according to supplementary note 5, wherein the detection means detects abnormality in the lung sounds from the time-series acoustic signals in the period other than the pause phase after the division.

- (Supplementary Note 7)

A lung sound analysis method comprising:

- acquiring time-series acoustic signals including lung sounds of a heart failure patient;
- detecting abnormal lung sounds from the acquired time-series acoustic signals;
- transmitting, to a terminal device of a medical specialist, analysis object lung sound information in which the acquired time-series acoustic signals and a result of the detection are associated with each other, and receiving, from the terminal device, analysis object lung sound information to which an observation by the medial specialist on the time-series acoustic signals is added; and
- generating learning data for detecting abnormal lung sounds on a basis of the analysis object lung sound information to which the observation by the medial specialist is added.
- (Supplementary Note 8)

The lung sound analysis method according to supplementary note 7, wherein

- the transmitting includes, before performing the transmitting, determining whether or not abnormal lung sounds are detected in the result of the detection, and performing the transmitting only when abnormal lung sounds are detected.
- (Supplementary Note 9)

The lung sound analysis method according to supplementary note 7 or 8, wherein

- the transmitting includes, before performing the transmitting, determining whether or not consent to use of the time-series acoustic signals for the learning data is given by the heart failure patient, and performing the transmitting only when the consent is given.
- (Supplementary Note 10)

The lung sound analysis method according to any of supplementary notes 7 to 9, wherein

- the transmitting includes, before performing the transmitting, determining whether or not personal information of the heart failure patient matches personal information of a target for collecting lung sound data, and performing the transmitting only when the personal information matches.
- (Supplementary Note 11)

A computer-readable medium storing thereon a program for causing a computer to execute processing to:

- acquire time-series acoustic signals including lung sounds of a heart failure patient;
- detect abnormal lung sounds from the acquired time-series acoustic signals;
- transmit, to a terminal device of a medical specialist, analysis object lung sound information in which the acquired time-series acoustic signals and a result of the detection are associated with each other, and receive, from the terminal device, analysis object lung sound information to which an observation by the medial specialist on the time-series acoustic signals is added; and
- generate learning data for detecting abnormal lung sounds on a basis of the analysis object lung sound information to which the observation by the medial specialist is added.

REFERENCE SIGNS LIST

10 lung sound analysis device

11 electronic stethoscope

12 communication IN unit

13 operation input unit

14 screen display unit

15 storage unit

16 arithmetic processing unit

151 program

152 lung sound record

153 analysis object lung sound information

154 learning data DB

161 lung sound record acquisition means

162 analysis object lung sound acquisition means

163 lung sound abnormality detection means

164 analysis result output means

165 learning data generation means

166 learning means

LUNG SOUND ANALYSIS SYSTEM

Information

Publication Number

Date Filed

Date Published

Inventors

Original Assignees

CPC

International Classifications

Abstract

Description

Claims

PCT Information