A METHOD AND APPARATUS FOR DIAGNOSIS OF MALADIES FROM PATIENT SOUNDS

TECHNICAL FIELD

The present invention concerns a method and automated apparatus for diagnosing maladies such as, though not limited to, Obstructive Sleep Apnea (OSA) from patient sounds.

RELATED APPLICATIONS

The present application claims priority from Australian provisional patent application No. 2018903933 filed 17 Oct. 2018, the disclosure of which is hereby incorporated herein by reference.

BACKGROUND ART

Any references to methods, apparatus or documents of the prior art are not to be taken as constituting any evidence or admission that they formed, or form part of the common general knowledge.

One common malady is the sleep disorder of Obstructive Sleep Apnea syndrome (OSA). The prevalence of OSA in adults varies from 17-26% in males and 9-28% in females [2]. At present over 85% of OSA patients remain undiagnosed [3]. OSA is characterized by a repetitive upper airway collapse during sleep. Full closure of the upper airway is termed “apnea” and partial closure is termed “hypopnea”. The average number of apnea and hypopnea events per-hour of sleep is termed the Apnea-Hypopnea Index (AHI). AHI is a major clinical severity measure for OSA.

The current standard for OSA diagnosis is Polysomnography (PSG)[4]. PSG requires continuous monitoring of multiple physiological signals over the course of a night. Physical contact of sensors with the patient is essential for these measurements. The several hours of PSG data are manually reviewed by an expert sleep technician. Reviewing PSG data is a labor intensive, time consuming and expensive process. PSG is also inconvenient to patients, especially the pediatric population, and results are subjective and unsuitable for population screening.

In the past several researchers have attempted to use patient sounds for diagnosis of maladies related to dysfunctions of the respiratory system. For example, the patient sounds may include snoring sounds used to diagnose OSA. Other maladies, such as pneumonia, asthma, bronchitis, croup and chronic obstructive pulmonary disease (COPD), Tracheobronchomalacia (TBM) or cystic fibrosis also cause characteristic patient sounds. Many of the existing methods depend on the identification of segments of the patient sound that are characteristic of the malady in question. For example, in the case of the malady being OSA then snore segments from the overnight sound data are identified. Hence if the snore segmentation algorithm fails to identify any snore segments or if the patient did not snore then results of the test will be indeterminate. Furthermore, procedures for identifying sounds that are characteristic of a malady of interest, such as snore sounds for OSA diagnosis, or a cough sound for pneumonia diagnosis, in a lengthy patient sound recording are computationally expensive and may be inaccurate. Therefore there is a need for an improved method of diagnosing a malady which does not rely on identification of sounds that are characteristic of a malady of interest in segments of the patient sounds.

SUMMARY OF THE INVENTION

According to a first aspect of the present invention there is provided a method for diagnosing a malady of a patient from sounds of the patient including the steps of:

- making a digital recording of the sounds of the patient;
- processing the digital recording to extract one or more features for sub-segments of each of a number epochs of the digital recording;
- determining deviation scores from a probability distribution for each epoch based on said extracted features;
- applying a test vector derived from the deviation scores to a pre-trained decision machine; and
- presenting a diagnosis of the malady based on an output from said decision machine.

For example, the malady may comprise OSA or a disease state such as pneumonia or another malady that causes a change from normal patient sounds, such as, pneumonia, asthma, bronchitis, croup and chronic obstructive pulmonary disease (COPD), Tracheobronchomalacia (TBM) or cystic fibrosis.

The features may be one or more of pitch, entropy, formants, a Gaussianity or other probability distribution measure and higher-order spectra-based features.

An embodiment of the invention may involve computing a Chi-squared test statistic between a MFCC distribution and a target probability distribution and using the computed test statistic directly as a feature to input to the decision machine.

Another embodiment of the invention may involve computing p-values for a Chi-squared test statistic between a MFCC distribution and the target distribution and use the p-value directly as a feature to feed the decision machine.

The target distribution may be a Gaussian distribution.

Alternatively, other embodiments may involve computing a KS test (Kolmogorov-Smirnov) test statistic in the place of the Chi-squared test statistic.

Another embodiment of the invention may make use of a Lilliefors test for normalcy with the Gaussian distribution.

According to a further aspect of the present invention there is provided a method for diagnosing OSA of a patient including the steps of:

- making a digital recording of sounds of the patient;
- processing the digital recording to extract a multiplicity of MFCCs for sub-segments of each of a number epochs of the digital recording;
- determining deviation scores from a probability distribution for each epoch based on the MFCCs;
- applying a test vector derived from the deviation scores to a pre-trained decision machine; and
- presenting a diagnosis of OSA on the basis of an output from said decision machine.

According to another aspect of the present invention there is provided a method of operating one or more electronic processors to diagnose the presence of Obstructive Sleep Apnea (OSA) of a patient comprising:

- acquiring a digital audio signal of sounds of the patient in an electronic storage assembly accessible to said processors;
- identifying a number of epochs of the digital audio signal;
- identifying a plurality of sub-segments for each of the epochs;
- for each sub-segment of each of the epochs determining an associated multiplicity of mel-frequency cepstral coefficients (MFCCs);
- determining deviation scores from a probability distribution for each of the epochs in respect of each of the multiplicity of MFCCs;
- forming a test vector for the patient based upon the deviations scores from the probability distribution of the MFCCs;
- applying the test vector to a pre-trained decision machine stored in said electronic storage assembly to thereby generate an OSA signal indicating OSA or non-OSA for the patient; and
- controlling a display responsive to the one or more electronic processors to display a message corresponding to the OSA signal.

According to a preferred embodiment of the present invention the forming of the test vector based upon the deviations scores of the MFCCs includes applying a comparator to each of the deviation scores. For example, the comparator may comprise a set of instructions executed by the one or more processors to implement a decision routine.

In an embodiment the output of the routine is a “1” signal if the deviation score is above a threshold or a “0” signal if the deviation score is equal to or below the threshold.

Preferably the method further includes forming components of the test vector for each of the MFCCs by producing sums of outputs from the comparator. In an embodiment the the method includes producing the sums of the outputs from the comparator for each MFCC over all of the epochs.

The method may include averaging each of the sums of the outputs over all of the epochs.

In a preferred embodiment of the invention the method includes reducing dimensionality of the test vector. For example, the method may include removing all but a subset of components of the test vector previously adjudged to be statistically significant for production of the OSA signal from the pre-trained decision machine.

Preferably the method includes forming the test vector on the basis of the entire digital audio signal.

In one embodiment of the invention the probability distribution is a Gaussian distribution and the deviation from a probability distribution score is a non-Gaussianity Score (NGS) or non-Gaussianity “Index” though other distributions may also be used and measures of deviation from those distributions may also be used.

For example, other embodiments may involve computing a KS test (Kolmogorov-Smirnov) test statistic in the place of the Chi-squared test statistic.

Another embodiment of the invention may make use of a Lilliefors test for normalcy with the Gaussian distribution.

According to a further aspect of the present invention there is provided an apparatus for diagnosing the presence of Obstructive Sleep Apnea (OSA) of a patient comprising:

- a microphone;
- an audio interface including an analog-to-digital converter (ADC) coupled to the microphone;
- an electronic storage assembly coupled to the ADC and arranged to store a digitized audio file of patient sounds from the audio interface;
- an epoch identification assembly configured to process the digitized audio file and identify a number of epochs therein;
- a sub-segment identification assembly configured to process the digitized audio file and identify a plurality of sub-segments therein for each of the epochs;
- a Mel-Frequency Cepstral Coefficient generator that is responsive to the epoch identification assembly and the sub-segment identification assembly and arranged to process the digitized audio file to produce a multiplicity of mel-frequency cepstral coefficients (MFCCs) signals for each of the sub-segments;
- a deviation from probability distribution score assembly that is responsive to the Mel-Frequency Cepstral Coefficient generator and which is arranged to process the MFCCs signals for each of the sub-segments to produce deviation from probability distribution scores for each of the MFCCs signals for each epoch;
- a test-vector generator assembly that is responsive to the deviation from probability distribution score assembly and which is arranged to store a test vector for the patient in the electronic storage assembly;
- a decision assembly that is coupled to the at least one electronic processor and arranged to process the test vector to produce a OSA diagnosis signal; and
- a human-machine interface that is coupled to the decision assembly and arranged to present the OSA diagnosis to a human.

According to another aspect of the present invention there is provided a computer readable medium bearing tangible, non-transitory machine readable instructions for execution by one or more electronic microprocessors including instructions for:

- acquiring a digital audio signal of sounds of the patient in an electronic storage assembly accessible to said processors;
- identifying a number of epochs of the digital audio signal;
- identifying a plurality of sub-segments for each of the epochs;
- for each sub-segment of each of the epochs determining an associated multiplicity of mel-frequency cepstral coefficients (MFCCs);
- determining a deviation from probability distribution score for each of the epochs in respect of each of the multiplicity of MFCCs;
- forming a test vector for the patient based upon the deviation from probability distribution score of the MFCCs;
- applying the test vector to a pre-trained decision machine stored in said electronic storage assembly to thereby generate an OSA signal indicating OSA or non-OSA for the patient; and
- controlling a display responsive to the one or more electronic processors to display a message corresponding to the OSA signal.

In one embodiment of the invention the distribution is a Gaussian distribution and the deviation from probability distribution score assembly is a non-Gaussianity score (NGS) assembly and the deviation score is a non-Gaussianity Score or “index”. It will be realized that other distributions are also useable and encompassed by embodiments of the present invention and some of these other distributions are described toward the end of this specification.

According to a further aspect of the present invention there is provided a method for diagnosing OSA of a patient including the steps of:

- making a digital recording of sounds of the patient;
- processing the digital recording to extract a multiplicity of MFCCs for sub-segments of each of a number epochs of the digital recording;
- determining deviation from a probability distribution score for each epoch based on the MFCCs;
- applying a test vector derived from the deviation from probability distribution score to a pre-trained decision machine; and
- presenting a diagnosis of OSA on the basis of an output from said decision machine.

BRIEF DESCRIPTION OF THE DRAWINGS

Preferred features, embodiments and variations of the invention may be discerned from the following Detailed Description which provides sufficient information for those skilled in the art to perform the invention. The Detailed Description is not to be regarded as limiting the scope of the preceding Summary of the Invention in any way. The Detailed Description will make reference to a number of drawings as follows:

FIG. 1. is a block diagram of an example of an example of a specially configured diagnostic device for diagnosing a malady such as OSA, according to a preferred embodiment of the present invention.

FIG. 2: Is a physical view of the diagnostic device of FIG. 1 displaying a recording duration and sampling rate screen on its LCD touch screen interface.

FIG. 3: Is a physical view of the diagnostic device of FIG. 1 displaying a recording-in-progress screen on its LCD touch screen interface.

FIG. 4 Is a first portion of a flowchart of a method coded as instructions in a diagnostic App stored in a digital memory of the diagnostic device.

FIG. 5 is a second portion of the flowchart of FIG. 4.

FIG. 6 depicts a portion of a digital waveform of a patient sound generated by an analog-to-digital converter of the diagnostic device.

FIG. 7 depicts a longer duration of the digital waveform showing a number of epochs identified therealong.

FIG. 8 depicts a single epoch of the digital waveform with pre-emphasis applied and showing sub-segments of the epoch identified therealong.

FIG. 9 is a flowchart of a procedure that is implemented by the diagnostic device during its operation for generating MFCCs for each subsegment.

FIG. 10 graphically illustrates MFCCs for each sub-segment of each pre-emphasized epoch.

FIG. 11 graphically illustrates the values of three exemplary MFCCs each over a single pre-emphasized epoch.

FIG. 12 comprises three data plots of sample distributions that are referred to for explanation of non-Gaussianity score.

FIG. 13 graphically illustrates the non-Gaussianity of each of the three data plots and presents a Non-Gaussianity Score (NGS) for each.

FIG. 14 graphically illustrates non-Gaussianity Scores for each of three MFCCs for each of N pre-emphasized epochs.

FIG. 14A Is a physical view of the diagnostic device of FIG. 1 displaying a diagnosis on its LCD touch screen interface.

FIG. 15: Comprises three charts of Mean (a) Age (b) Body Mass Index (BMI) and (c) Neck Circumference (NC) with 95% confidence interval across the 4 OSA subject groups.

FIG. 16: Is a boxplot showing classification performance of the LRM when trained using features computed with different audio file format. The tops and bottoms of each “box” are the 25th and 75th percentiles of the samples, respectively. Error bar indicates interquartile ranges. The line in the middle of each box is the sample median.

FIG. 17: Is a chart of test classification results of the LRM with change in sampling rate of the audio data. Only selected features were used to train and test the models.

FIG. 18: Illustrates Mean classification Sensitivity and Specificity of LRM with 95% confidence interval, when trained using features computed with data sampled at different sampling rate (Fs). There is no significance difference in LRM sensitivity at different FS. Only specificity of LRM at Fs=2000 Hz is significantly lower from that at Fs=44100.

FIG. 19 is a block diagram of a dedicated diagnosis apparatus according to a further embodiment of the present invention.

DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS

Referring initially to FIG. 1 there is shown a diagnostic device 1 in the form of a unique combination being a computational device in the form of a smart phone in combination with a diagnostic application software product. The diagnostic device 1 includes at least one microprocessor 3 that accesses an electronic memory 5. The electronic memory 5 includes an operating system 8 such as the Android operating system or the Apple iOS operating system, for example, for execution by the microprocessor 3. The electronic memory 5 also includes the diagnostic application software product or “App” 6 according to a preferred embodiment of the present invention. The diagnostic App 6 includes instructions that are executable by the microprocessor 3 in order for the diagnostic device 1 to process sounds from a patient 2 and present a diagnosis of a malady such as OSA to a clinician 4 by means of LCD touch screen interface 11. In the exemplary embodiment that will be primarily discussed reference will be made to a malady being OSA and thus the device will be referred to as OSA diagnostic device 1. However in other embodiments the device 1 may be configured by App 6 to diagnose other maladies of the respiratory system such as pneumonia, asthma, bronchitis, croup and chronic obstructive pulmonary disease (COPD), Tracheobronchomalacia (TBM) or cystic fibrosis. The App 6 includes instructions for the microprocessor 3 to implement a trained predictor or decision machine, which in the presently described preferred embodiment of the invention comprises a specially trained Logistic Regression Model 20. It will be realised that in other embodiments of the invention other suitable decision machines may be used, such as an artificial neural network or a Bayesian decision machine and thus the invention is not limited to the use of an LRM only.

The microprocessor 3 is in data communication with a plurality of peripheral assemblies 9 to 23, as indicated in FIG. 1, via a data bus 7. Consequently, if required the diagnostic device 1 is able to establish voice and data communication with a voice and/or data communications network 31 via WAN/WLAN assembly 23 and radio frequency antenna 29.

Although the OSA diagnostic device 1 that is illustrated in FIG. 1 is provided in the form of a smartphone it might equally be some other computational device such as a laptop, or tablet in combination with a software product containing instructions to implement a method according to an embodiment of the invention such as will be described.

In a preferred embodiment the OSA diagnostic device 1 is programmed with App 6 so that it operates as a decision device that requires no external sensors, physical contact with patient 2 or communication network 31.

In use the nominal distance from the microphone 25 of device 1 to the face of patient 2 is set to about 50 cm, but may vary between 40 cm to 70 cm due to patient movements.

Referring now to FIGS. 4 and 5, there is shown a flowchart of a method according to a preferred embodiment of the present invention, which the OSA diagnosis device 1 implements under the control of the instructions that are coded into the OSA diagnostic App 6 in order to make a diagnosis of whether or not patient 2 is suffering from a malady, being OSA in the present exemplary embodiment. The health carer 4 can then use the diagnosis to provide appropriate therapy, for example a positive pressure airway device or other suitable therapy to alleviate the OSA.

At box 41 of FIG. 4, the microprocessor 3 operates the LCD screen 11 to display a prompt for a user, e.g. clinician 4, to commence recording the in-air sounds 39 of patient 2.

The breathing sound 39 of patient 2 is recorded by the diagnostic device 1 and FIG. 2 shows the diagnostic device 1 displaying a recording commencement screen 59 on LCD touch screen 11 for the clinician 4 to enter the recording parameters. The recording parameters are a patient ID number, and the “Timeout”, i.e. the duration of the recording that is to be made and also the analogue to digital sample rate to be used. In the present instance the duration that has been selected is 10 hours and the sample rate that is to be used is 44.1 kHz. FIG. 3 shows the screen 61 that is displayed once clinician 4 presses the “Record” button in screen 59 of FIG. 2.

As the recording proceeds an audio file is stored in an electronic storage assembly such as either memory 5 or secondary memory 14, which is typically a Secure Digital (SD) memory card. The audio file may be stored in a compressed format such as MP3 or in a non-compressed format such as a WAV or FLAC file. The pros and cons of using a compressed format as opposed to an uncompressed format will be discussed later in this specification. Depending on the hardware configuration the selection of the sample rate may alter a sample rate parameter in Audio Interface 21 or alternatively the analog-to-digital conversion may be made at 44.1 kHz in the audio interface 21 and then down-sampled by the microprocessor 3 in accordance with instructions in OSA Application 6.

The procedure that microprocessor 3 uses to make a diagnosis of a malady, which in the present example is OSA, and which comprises instructions that make up App 6 is illustrated in the flowchart of FIGS. 4 and 5 which will now be further described.

- 1. At box 41 the nocturnal breathing sound signal 39 of the patient 2 is recorded by microphone 25 and digitized by the audio interface 21 to produce a digitized signal “x” (indicated as item 31 of FIG. 6) that is conveyed to microprocessor 3 along bus 7. The digitized patient sound data is recorded from patient 2 at sampling frequency Fs which is typically 44.1 kHz or less.
  - FIG. 7, which shows a far longer portion of the waveform x on a much compressed timescale (horizontal axis) than in FIG. 6. At box 43 (FIG. 4) the microprocessor 3 segments signal x into non-overlapping blocks of size k₁=[10, 15, 20, 30] seconds. In a clinical setting these non-overlapping blocks are referred as ‘epochs’ and thus from here onward the same terminology will be followed. There are a total of N epochs as indicated in FIG. 7 and the symbol x(i) represents the i^thepoch, where i=1, 2, . . . , N.
- 2. In the present example the OSA diagnosis device 1 has been set to record the patient sound 39 at Fs=44100 Hz. If at box 45 a lower sampling frequency is desired, for example to decrease the size of the resulting file, then microprocessor 3 can execute instructions in OSA App 6 to resample x(i) by choosing sampling frequency from one of (for example) Fs=[22050; 11025; 8000; 6000; 4000; 2000; 1000; 500; 200;] Hz.
- 3. At box 47 the microprocessor filters each of the epochs x(1), . . . , x(N) using a Butterworth band-pass filter with lower cut-off frequencies (LCF)=20 Hz (to remove low frequency rumble) and higher cut-off frequency (HCF)=Fs/2.
- 4. At box 48 microprocessor 3 applies a pre-emphasis filter, as set out in equation (1) below to produce pre-emphasized epochs y(i). If a[p] represents signal x(i) in i^thepoch with p=1, 2 . . . , P as sample numbers then a pre-emphasis filter can be implemented as shown in (1). The pre-emphasis filtering boosts the overall energy of the audio signal and improves the signal-to-noise ratio at higher frequencies. The parameter a in (1) can be adjusted to set the filter to the desired cut-off frequency using Fc in (2). The App 6 can include instructions for the microprocessor 3 to present controls on the touchscreen of device 1 for the user to make adjustments if required.

$\begin{matrix} b [p] = a [p] - α a [p - 1] & (1) \\ α = e^{- 2 π \frac{F_{C}}{F_{S}}} & (2) \end{matrix}$

- 5. At box 49 microprocessor 3 sub-segments each of the pre-emphasis filtered epochs y(i) using a non-overlapping rectangular window of length k₂=100 ms. Let y(i,j) represent j^thsub-segment of the i^thepoch data. j=1, 2, . . . , J as illustrated in FIG. 8. For example, if the Epoch length k₁is 10 seconds then there will be 10,000 subsegments in each epoch and thus J will equal 10,000.
- 6. At box 51 (FIG. 5) of the method, the Microprocessor 3 computes twelve Mel-Frequency Cepstral Coefficients (MFCC) for each of the j=1, 2, . . . , J data sub-segments y(i,j) as illustrated in FIG. 9. It will be realized that in other embodiments of the invention more or fewer than 12 MFCCs may be calculated and that the number of MFCCs does not have to be exactly 12 for the method to work. MFCCs are modeled on the basis of the human perception of speech via the anatomical auditory system. They provide some resilience to the non-linguistic sources of variance in speech signal. The computation of MFCCs for each sub-segment y(i,j) is illustrated in FIG. 9 in block diagram form and is well known in the prior art. Computing the MFCCs involves the estimation of short-term power spectra for each of a number of filters of a mel-frequency scale filter bank. That is done by applying a Discrete Fourier Transform 70 to the pre-emphasized sub-segments, then applying a log function 74 to the output from each of the filters of the mel-filter bank 72 and then applying a Discrete Cosine Transform (DCT) 76 to produce the MFCCs. Since the log power spectrum is real and symmetric an Inverse DFT reduces to the Discrete Cosine Transform (DCT) 76. The output from the DCT 76 comprises spectral mel-frequency cepstral coefficients (MFCCs) z_c(i,j) being the c^thcoefficient of the j^thsub-segment of the ith pre-emphasized epoch. FIG. 10 graphically represents the MFCCs 1 to 12 vertically for each segment j=1, . . . , J of each epoch i=1, . . . , N. FIG. 11 shows a plot of three of the MFFCs c_l, c_iand C₁₂for each sub-segment across a single pre-emphasized epoch y(i). These plots are only for illustrative purposes to assist in understanding the procedure that is being applied by the microprocessor 3 to the digitized sound wave. It will be realized that the actual plots of MFCCs that are determined may have substantially different appearances.
- 7. At box 53 of FIG. 5, microprocessor 3 then calculates a measure or “score” of deviation from a known statistical distribution. A number of distributions that may be used are described at the end of this specification. In the present example the statistical distribution that is used is a Gaussian distribution and the measure of deviation from Gaussian distribution is a non-Gaussianity score (NGS). The NGS ξ_c(i) for each of the MFCC components over each epoch is determined. ξ_c(i) represents the NGS of the c^thMFCC component in the i^thpre-emphasized epoch. For example, the value of 3^rdMFCC over the sub-segments in a given epoch may be close to a Gaussian distribution, in which case the NGS for the 3^rdMFCC will be a low value for that epoch, or it may be very unlike a Gaussian distribution, in which case the NGS for the 3^rdMFCC will be a high value for that epoch. The following description will mainly refer to the use of a Gaussian distribution however other statistical distributions may also be used and as previously mentioned, these are described toward the end of this specification.
  - Three exemplary sample distributions are shown in FIG. 12 parts (a), (b) and (c). It will be visually observed that the plot of part (a) appears to match a Gaussian “bell curve” whereas that of part (c) does not and that of part (b) is intermediate. The normal probability plot is a plot of the midpoint probability positions of a given data segment versus the theoretical quantiles of a normal distribution. If the distribution of the data under consideration is normal, the plot will be linear. Other probability distributions will lead to plots that deviate from linearity, with the particular nature and amount of deviation depending on the actual distribution itself. FIG. 13 parts (a), (b), (c) show a normal (Gaussian) distribution as a linear dashed line on which corresponding plots from each of FIGS. 11 (a), (b) and (c) have been superimposed showing the increasing deviation from Gaussianity.
- 8. The increase in the NGS score is therefore a quantitative measure of the deviation from Gaussianity, of MFCC component values in an epoch. As previously mentioned methods of NGS computation are known in the prior art and are centered on computing the normal probability plot for each of the MFCCs for each epoch. Detailed methods of NGS computation can be found in H. Ghaemmaghami, U. Abeyratne, and C. Hukins, “Normal probability testing of snore signals for diagnosis of obstructive sleep apnea,” in Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE, 2009, pp. 5551-5554, the contents of which is hereby incorporated by reference in its entirety. FIG. 13 graphically represents NGS scores, normalized on a vertical scale of zero to one for each of three MFCCs (C1, Ci and C12) for each epoch 1, . . . , N.
- 9. At box 55 the microprocessor 3 implements a comparator and compares each NGS ξ_c(i) that was computed in box 53 against a threshold η to define L_c(i) using (3).

$\begin{matrix} L_{c} (i) = {\begin{matrix} 1 & if ξ_{c} (i) > η \\ 0 & if ξ_{c} (i) \leq η \end{matrix} & (3) \end{matrix}$

- 10. At box 61 microprocessor 3 computes an MFCC-Index vector, Ψ_cfor all MFCC components using (4).

$Ψ_{c} = \frac{\sum_{i}^{N} L_{c} (i)}{N}$

- 11. At box 63 microprocessor 3 produces a reduced dimension test vector Ψ_sc-test by removing some components from Ψ_c. As will be described, the components of Ψ_cthat are removed have been previously judged to have little, or no, influence on the diagnosis.
- 12. At box 65 the test vector is applied to the pre-trained LRM 20. It will be realised that other types of decision machines or “classifiers” can be used as well. The output of the LRM is a signal that represents a number that is “1” or very close thereto and so indicates a diagnosis of “OSA present” or “0” or very close thereto and so indicates a diagnosis of “no OSA present”.
- 13. At box 67 the microprocessor 3 operates LCD touch screen interface 11 to present the diagnosis screen 63 in respect of patient 2 to the carer 4 as shown in FIG. 14A. Diagnosis screen 63 includes a message 65 indicating the diagnosis of OSA in the particular patient 2.

Producing the Lrm

In order to create the trained Logistic Regression Machine (LRM) 20 the Inventors initially recorded sounds from Q=41 patients including individuals with symptoms such as daytime sleepiness, snoring, tiredness lethargy etc. and who were suspected of OSA. It will be realised that a similar procedure is followed in order to train the LRM for detection of other maladies and that in that case sounds would be recorded from patients suffering from the malady in question.

The steps that have previously been described in relation to boxes 43 to 61 of FIGS. 4 and 5 were then performed in respect of each of the Q patients in the database and a feature matrix M of the size Q×Ψ_cwas formed. Q represents the total number of patients and Ψ_crepresents a feature vector from each patient.

Pattern Classifier

As previously discussed, App 6 includes instructions for implementation of a logistic-regression model (LRM) as the “pattern classifier” or “decision machine” for classifying test patient sounds as suffering from a malady being OSA in the exemplary embodiment. It will be realized that in other embodiments of the invention other types of decision machine may also be used such as trained neural nets, Bayesian decision machines and support vector machines and that other maladies, such as those that have previously been referred to may be the subject of the training of the pattern classifier or decision machine.

The LRM that is implemented by App 6 in the present embodiment of the invention is the best LRM that could be determined by the methodology that the Inventors have devised and which will now be described.

An LRM is a generalized linear model, which uses several independent features to estimate the probability of a categorical event (dependent variable). In the present case, the dependent variable Y is assumed to be equal to ‘one’ (Y=1) for ‘OSA’ subjects and ‘zero’ for ‘non-OSA subjects. OSA and non-OSA subjects were defined using 3 different AHI thresholds, AHI=[5; 15; 30;]. These AHI thresholds are routinely used in the clinical practice to define the severity of OSA as follows:

AHI < 5
no OSA

5 ≤ AHI < 15
mild OSA

15 ≤ AHI < 30
moderate OSA

AHI ≥ 30
severe OSA

As is known in the prior art, an LRM model is derived using a regression function to estimate the probability Y given the independent features in Ψ_cas follows:

$\begin{matrix} (Y = 1 |_{Ψ1, Ψ2, \dots, Ψ c}) = \frac{e^{w}}{e^{w} + 1} & (5) \\ w = β_{0} + β_{1} \cdot Ψ_{1} + β_{2} \cdot Ψ_{2} + \dots ++ β_{c} \cdot Ψ_{c} & (6) \end{matrix}$

In (6), β₀, is called the intercept and β₁, β₂and so on are called the regression coefficients of independent variables. To select the optimal decision threshold A from Y (that subject is OSA if Y>λ; non-OSA otherwise) the Receiver-Operating Curve (ROC) analysis was used.

The Inventors used a K-fold cross validation (KCV) technique for the LRM design, setting K=10. In KCV technique, subject population in the database is randomly partitioned into K-equal size non-overlapping subsamples. Then of the K subsamples, data from subjects in K−1 subsamples are used to train the LRM model and data from subjects in the remaining one subsample is used to test the model. This process is systematically repeated K times such that each patient in the database is used to test the model exactly one time. At the end of this process, we end up with κ different LRM models. To evaluate the performance of the designed κ LRMS, performance measures such as Sensitivity (Sn), Specificity (Sp), Accuracy (Ac), Positive Predicted Value (PPV) and Negative Predicted Value (NPV) were computed.

Feature Selection

Feature selection is a technique of selecting a subset of features for building a robust classifier. Optimal feature selection requires the exhaustive search of all possible subsets of features. However, it is impractical to do so when large numbers of features are used as candidate features. Therefore, an alternative approach was used based on p-value to determine significant features. During LRM design, a p-value can be computed for each feature to indicate how significant that feature is to the model. Important features have low p-value. The Inventors used this property of an LRM to select a reasonable combination of features that facilitate the classification, in the model during the training phase. The technique that was used consisted of computing the mean p-value associated with Ψ_cfor κ LRM models. Then selecting the features with mean p-value less than a threshold p_ths. Let Ψ_scbe the feature vector with subset of the selected MFCC component index and M_fs(of size Q×Ψ_sc) be the feature matrix computed from selected features.

Once the significant features were known and selected they were used to build a new set of LRMs, following K-fold cross validation (K=10) as previously described. At the end of this process, κ_fsnumber of LRMs were produced using the selected features.

As previously mentioned, the Inventors used breathing sound data from Q=41 subjects. According to AHI severity these subjects were divided into four groups namely:

(i) Group 1, non-OSA subjects with RDI<5

(ii) Group 2, 5≤AHI<15, mild OSA,

(iii) Group 3, 15≤AHI<30, moderate OSA and

(iv) Group 4, AHI≥30, Severe OSA.

Table 1 sets out the demographic details of the subjects in the database for four subject groups.

TABLE 1

Demographic details of the subjects.

Group 1
Group 2
Group 3
Group 4

Non-OSA
Mild OSA
Moderate OSA
Sever OSA

(RDI < 5)
(5 ≤ RDI < 15)
(15 ≤ RDI < 30)
(RDI ≥ 30)

N
7
16
5
13

Age
50 ± 14
53 ± 12
63 ± 10
56 ± 12

M:F
4:3
9:7
5:0
12:1

BMI
27 ± 3
36 ± 10
30 ± 3.7
40 ± 8

AHI
1.76 ± 1.2
8.68 ± 2.5
23.08 ± 3
62.58 ± 24

FIG. 15 shows plots for mean (a) Age; (b) Body Mass Index (BMI) and (c) Neck Circumference (NC) with 95% confidence interval for subject groups. One way analysis of variance (ANOVA) statistical test showed no significant difference between the mean Age of subjects among the groups. Quite interestingly mean BMI of the severe OSA subjects was significantly higher than for non-OSA subjects. Similarly mean NC of the non-OSA subjects is significantly lower than mean NC of mild OSA and severe OSA.

Comparison Between Different File Formats

One of the Inventors' objectives was to evaluate the effect of data compression on the classifier performance. For this the nocturnal breathing sound audio data was recorded from subjects in raw audio data format, WAV format. Then using Adobe Audition™ the data was converted into FLAC (loss-less audio format) and Mp3 (lossy audio data format).

The average length of the audio data recordings from Q=41 subjects were 7 hours and 4 minutes with standard deviation of 1 hour and 38 minutes. The average size of an audio data recording with Fs=44100 Hz, were, WAV file=2.25±0.24 Giga bytes, FLAC file=0.95±0.11 Giga bytes and that of MP3 file=0.61±0.06 Giga bytes. On average size of a FLAC audio data file with Fs=44100 Hz was 58±5% smaller than that of WAV file and Mp3 audio data file was 73±0.04% smaller than that of WAV file.

The Inventors investigated a snore sound waveform and its spectrogram using different audio file formats and at different sampling rates. They found no difference between the WAV file format and the FLAC file format and no difference in the time domain or in the frequency domain at all the sampling rates. With respect to the Mp3 audio file, no obvious changes could be seen in the time domain signal however a clear attenuation of the higher frequencies could be seen in the spectrogram. However high frequency attenuation could only be seen at Fs=44100 Hz and was not present at Fs=8000 Hz or 2000 Hz.

Classification Results—Comparison Between File Format at Fs=44100

As previously discussed the LRM were trained using Ψ_cfeature vectors which were derived from MFCC and NGS following a K-fold cross validation technique to classify patients into OSA and non-OSA. The LRM were trained to classify patients into OSA and non-OSA at different AHI thresholds of [5; 15; 30;]. The LRM were initially trained using all features and then the LRM models were retrained using a selected sub-set of features.

Table 2 gives the test classification results for OSA diagnosis at different AHI thresholds optimized for epoch lengths. These results are for audio data sampled at Fs=44,100 Hz.

TABLE 2

Cross validation results for OSA and non-OSA classification at

different AHI thresholds.

EPL
Sensitivity
Specificity
PPV
NPV
Accuracy
Features

AHI = 5

WAV
10
85
43
88
38
78
All

10
94
86
97
75
93
[8, 10,

11, 12]

FLAC
10
85
43
88
38
78
All

10
94
86
97
75
93
[8, 10,

11, 12]

MP3
30
88
29
86
33
78
All

30
88
86
97
60
88
[2, 5,

8, 10]

AHI = 15

WAV
30
61
78
69
72
71
All

30
83
91
88
88
88
[3, 9]

FLAC
30
61
78
69
72
71
All

30
83
91
88
88
88
[3, 9]

MP3
20
72
78
72
78
76
All

20
83
87
83
87
85
[1, 2,

3, 7, 9]

AHI = 30

WAV
10
100
75
65
100
83
All

10
100
93
87
100
95
[1, 2, 3,

6, 10, 11]

FLAC
10
100
75
65
100
83
All

10
100
93
87
100
95
[1, 2, 3,

6, 10, 11]

MP3
30
92
79
67
96
83
All

30
92
89
80
96
90
3

It will be observed from Table 2 that there is no difference in classification accuracy between WAV and FLAC audio data at all the AHI thresholds. When selected features are used for model training, WAV and FLAC audio format have classification sensitivities/specificities of 94/86%, 83/91% and 100/93% respectively at AHI=5, 15 and 30. Classification results using Mp3 audio data format were slightly lower than WAV/FLAC audio data format. The sensitivities/specificities of the Mp3 data was 88/86%, 83/87 and 92/89% respectively at AHI=5, 15 and 30.

FIG. 16 shows the boxplot of test classification Sensitivity and Specificity for the three types of audio data file format. To generate this graph, results of LRM from a file format, at different AHI thresholds and at different Fs were pooled together. According to FIG. 17, the LRM show no significance (p=0.47) difference in classification performance when trained using MFCC features computed with different file format.

Classification Result—Effect of Sampling Frequency

As previously discussed, the patient sounds may be resampled with different sampling frequencies Fs=[22050; 11025; 8000; 6000; 4000; 2000; 1000; 500; 200;] Hz. Note that audio data is initially recorded at Fs=44100 Hz. MFCC features were then computed with resampled data. FIG. 17 illustrates the variation in model test classification performance (sensitivity and specificity) with different Fs. Results in FIG. 17 are from selected MFCC features. FIG. 18 shows the mean test classification Sensitivity and Specificity with 95% confidence interval, achieved for audio data with different sampling rate Fs. To generate the graph in FIG. 18, results at specific Fs from three file formats at different AHI thresholds were pooled together. According to FIG. 17 and FIG. 18 a gross variation in sensitivity and specificity with change in data sampling rate can be seen. In general, Sensitivity initially increases with decrease in Fs, reaching at its peak at Fs=11025 Hz. It then starts decreasing, reaching its lowest value at Fs=2000 Hz across all file formats. Note that though a decrease in Sensitivity can be seen at lower Fs, however this decrease is insignificant. In Specificity generally remained stable when Fs decreased from 44100 to 22050, then it decreased at Fs=11025 Hz. Then from Fs=11025 Hz it starts increasing up-to Fs=6000 Hz and then gain starts decreasing reaching lowest at Fs=2000 Hz. The decrease in Specificity was significant but only with respect to Fs=44100.

The results indicate that methods according to embodiments of the present invention can classify patients into OSA and non-OSA at different AHI threshold with a high accuracy.

In the past several researchers [10-16] have attempted to use snoring sounds to diagnose OSA and many of the existing methods [10-12, 14] have depended on the identification of snore segments from the overnight sound data. Hence if the snore segmentation algorithm fails to identify any snore segments or if the patient did not snore then results of the test will be indeterminate.

In contrast to those previous methods that have relied on detection of snore segments in the patient sound for subsequent diagnosis of OSA, preferred embodiments of the invention described herein capture the instantaneous characteristics of the upper airway present in continuous recordings of the breath sound.

Furthermore, preferred embodiments of the invention make use of MFCC features for the diagnosis of OSA via measuring the amount of deviation of MFCC features from Gaussianity in a given sound segment (“epoch”). This approach has the advantage of better performance, robustness against AHI variation and low computational complexity as it does not depend on identifying snore segments from breath sound data.

The Inventors' results also illustrate that it is possible to record the patient sounds, i.e. the sounds of the patient breathing, with a compressed audio format and at a low sampling rate without compromising on classification accuracies. The results show that it is possible to achieve a sensitivity/specificity of 97/86%, 94/83% and 92/89% respectively at AHI threshold of 5, 15 and 30, with breath sound data recorded using Mp3 file format at Fs=6000 Hz (FIG. 17). With these settings the memory space required to record breath sound data of 8 hours duration will be <28 megabytes which is important where it is desired to store lengthy sounds from a number of patients. The automated, non-contact and mobile technology according to preferred embodiments of the present invention provides an excellent tool for population screening.

Previously in FIG. 1 a block diagram of an OSA diagnostic device 1 according to a preferred embodiment of the present invention was provided and discussed. It is also possible in another embodiment of the invention to provide a dedicated OSA diagnostic device rather one that is comprised of a specially programmed microprocessor based apparatus such as a specially programmed smartphone.

A dedicated OSA diagnostic apparatus 100 for diagnosing the presence of Obstructive Sleep Apnea (OSA) of a patient 2 is illustrated in FIG. 19. The apparatus includes a system clock 123 for synchronizing the various modules of the apparatus. A microphone 120 is coupled via an anti-aliasing filter 121 to an analog-to-digital converter (ADC) 122. The output from the ADC 122 is received by an electronic storage assembly 124, which stores a digitized audio file of patient sounds from the audio interface ADC 122. A pre-emphasis assembly 126 is coupled to the output of the data storage assembly 124 for applying pre-emphasis to the digitized audio signal.

The apparatus 100 includes an epoch identification assembly 128 that is coupled to an output side of the pre-emphasis assembly 126 to process the digitized audio file and identify a number of epochs in the audio file. A sub-segment identification assembly 130 is provided that is arranged to process the digitized audio file and identify a plurality of sub-segments therein for each of the epochs.

The sub-segment ID assembly 130 and the Epoch ID Assembly 128 provide respective outputs to the Mel-Frequency Cepstral Coefficient generator 132 which processes the digitized audio file from the pre-emphasis assembly 126 to produce a multiplicity of mel-frequency cepstral coefficients (MFCCs) signals for each of the sub-segments.

A non-Gaussianity Score calculation assembly 134 is provided that is responsive to the Mel-Frequency Cepstral Coefficient generator and which is arranged to process the MFCC signals from the MFCC generator 132 for each of the sub-segments to produce NGS scores for each of the MFCCs signals for each epoch as identified by the Epoch ID Assembly 128. In other embodiments of the invention a deviation from probability distribution score calculation assembly may be used to calculate a score for deviation from another distribution other than Gaussian.

The output from the NGS calculator 134 is passed to a comparator 136 which compares each of the MFCCs to a threshold value and respectively outputs a “0” or a “1” if the MFCC value is below or above threshold.

The output from the comparator is summed and averaged by Sum-and-Average block 138 to produce an initial test-vector which is subsequently reduced in dimension by Component Reduction assembly 140 to produce a reduced MFCC feature test vector. The reduced MFCC feature test vector is then passed to a decision machine block 142 which generates an OSA/non-OSA signal in response to the reduced MFCC feature test vector.

The apparatus 100 includes a human-machine interface including diagnostic display 146 that is coupled to the decision machine block 142 and which is arranged to present the OSA diagnosis to a human.

Whilst the previous discussion focused on a method and apparatus according to a preferred embodiment of the invention that uses deviation from Gaussian distribution, other measures of deviation from a known statistical distribution may also be used in other embodiments of the present invention and some of these are listed below. In other embodiments App 6 may include instructions for microprocessor 5 to implement each of the following statistical techniques as an alternative to determining deviation from Gaussian distribution.

- 1. Compute the Chi-squared test statistic between the MFCC distribution and the target distribution (e.g. Gaussian) and use it directly as a feature to feed the classifier.
- 2. Compute p-values for the Chi-squared test statistic between the MFCC distribution and the target distribution (e.g. Gaussian) and use the p-value directly as a feature to feed the classifier.
- 3. Use the KS test (Kolmogorov-Smirnov) test statistic in the place of Chi-squared described in (1) and (2) above
- 4. Use the Lilliefors test for normalcy as described in (1) and (2) above, with the Gaussian distribution.
- 5. Use a combination of deviation measures, including the NGS measure.

Results on the above methods 1-5 are set forth below.

Non-Segmentation Based OSA Classification Results (Scored Using 2007 Alternate Criteria):

Data statistics for 73 usable recordings:

- RDI<15=39
- RDI>=15=34
- RDI=>30=22
- Male=42
- Female=31

Training and Leave-One-Out Validation Results

Results Summary (RDI threshold=15):

Leave One Out

Training Results
Validation Results

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
91.26
92.38
91.18
92.31

Chi Sq Test of Mfcc
88.52
89.99
88.24
84.62

(T-stat value)

KS test of MFCC
88.07
89.60
88.24
84.62

(T-stat value)

Lilliefors test of MFCC
91.17
92.31
91.18
87.18

(T-stat value)

Chi Sq Test of Mfcc
87.91
89.46
85.29
84.62

(p-value)

KS test of MFCC
88.07
89.60
88.24
87.18

(p-value)

Lilliefors test of MFCC
88.44
89.92
85.29
84.62

(p-value)

NGS + Lilliefors test
95.65
97.40
91.18
97.44

Leave One Out

Training Results
Validation Results

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
91.13
92.27
91.18
92.31

Chi Sq Test of Mfcc
85.05
86.96
85.29
84.62

(T-stat value)

KS test of MFCC (T-stat
90.97
92.13
85.29
87.18

value)

Lilliefors test of MFCC
93.95
94.73
88.24
87.18

(T-stat value)

Chi Sq Test of Mfcc (p-
84.60
86.57
85.29
84.62

value)

KS test of MFCC
88.07
89.60
88.24
87.18

(p-value)

Lilliefors test of MFCC
85.09
87.00
85.29
84.62

(p- value)

NGS + Lilliefors test
91.13
92.27
91.18
92.31

Dividing Data into Training and Testing & Objectively Removing Noisy Recordings

Training Set
Test Set

Number of Subjects
53
20

Male
31
11

Female
22
9

RDI < 15
29
10

RDI ≥ 15
24
10

Set 1 [Train and LOV=53 and Independent Test=20]:

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
87.99
89.29
88.00
89.29
70
100

Chi Sq Test of
87.92
89.21
88.00
82.14
70
60

Mfcc (T-stat

value)

KS test of
91.84
92.72
92.00
85.71
50
80

MFCC (T-stat

value)

Lilliefors test

of MFCC (T-
85.76
87.29
88.00
85.71
80
90

stat value)

Chi Sq Test of
83.99
85.71
84.00
82.14
70
40

Mfcc (p-value)

KS test of
87.76
89.08
88.00
85.71
90
80

MFCC (p-

value)

Lilliefors test
87.99
89.28
84.00
89.29
80
80

of MFCC (p-

value)

NGS +
91.06
92.03
92
89.29
70
100

Lilliefors test p

value

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
95.53
96.02
92.00
89.29
80
70

Chi Sq Test of Mfcc
87.76
89.08
88.00
85.71
80
70

(T-stat value)

KS test of MFCC (T-
91.92
92.78
92.00
85.71
50
70

stat value)

Lilliefors test of
88.38
89.63
88.00
89.29
80
80

MFCC (T-stat value)

Chi Sq Test of Mfcc
91.76
92.65
88.00
89.29
60
70

(p-value)

KS test of MFCC (p-
87.53
88.87
88.00
85.71
60
90

value)

Lilliefors test of
88.07
89.35
88.00
85.71
80
70

MFCC (p-value)

NGS + Lilliefors test
95.84
96.29
96
92.86
80
90

p value

Set 2 [Created after shuffling the training and test data; Train and LOV=53 and Independent Test=20]:

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
92.08
92.92
92.00
89.29
90
80

Chi Sq Test of Mfcc
87.92
89.22
88.00
89.29
60
70

(T-stat value)

KS test of MFCC (T-
88.15
89.42
88.00
85.71
80
50

stat value)

Lilliefors test of
88.07
89.35
88.00
89.29
60
60

MFCC (T-stat value)

Chi Sq Test of Mfcc
87.84
89.15
88.00
89.29
80
60

(p-value)

KS test of MFCC (p-
88.38
89.63
88.00
85.71
80
80

value)

Lilliefors test of
88.00
89.28
88.00
85.71
80
80

MFCC (p-value)

NGS + Lilliefors test
91.76
92.65
92
89.28
80
60

p value

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
91.76
92.65
92.00
89.29
80
90

Chi Sq Test of Mfcc
87.76
89.08
88.00
85.71
80
70

(T-stat value)

KS test of MFCC (T-
87.92
89.22
88.00
89.29
80
80

stat value)

Lilliefors test of
91.69
92.58
92.00
89.29
90
80

MFCC (T-stat value)

Chi Sq Test of Mfcc
87.99
89.29
88.00
89.29
70
60

(p-value)

KS test of MFCC (p-
88.46
89.69
88.00
89.29
80
80

value)

Lilliefors test of
87.99
89.29
88.00
89.29
50
80

MFCC (p-value)

NGS + Lilliefors test
100
100
96
92.86
90
60

p value

Set 3 [Created after shuffling the training and test data; Train and LOV=53 and Independent Test=20]:

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
91.84
92.72
92.00
89.29
80
80

Chi Sq Test of Mfcc
87.84
89.14
88.00
85.71
80
70

(T-stat value)

KS test of MFCC (T-
87.99
89.29
88.00
89.29
70
70

stat value)

Lilliefors test of
91.84
92.72
92.00
89.29
80
70

MFCC (T-stat value)

Chi Sq Test of Mfcc
87.84
89.14
88.00
85.71
60
60

(p-value)

KS test of MFCC
92.00
92.85
92.00
85.71
30
60

(p-value)

Lilliefors test of
92.00
92.85
92.00
89.29
80
70

MFCC (p-value)

NGS + Lilliefors test
92
92.85
92
89.28
90
80

p value

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
91.84
92.72
92.00
89.29
90
80

Chi Sq Test of Mfcc
92.00
92.85
92.00
89.29
60
60

(T-stat value)

KS test of MFCC (T-
91.76
92.65
92.00
89.29
90
60

stat value)

Lilliefors test of
88.31
89.56
88.00
89.29
70
60

MFCC (T-stat value)

Chi Sq Test of Mfcc
91.46
92.37
88.00
89.29
50
70

(p-value)

KS test of MFCC (p-
91.69
92.58
88.00
89.29
40
50

value)

Lilliefors test of
87.99
89.29
88.00
89.29
80
80

MFCC (p-value)

NGS + Lilliefors test
91.84
92.72
92
89.29
90
80

p value

Iphone Data Analysis:

Total Iphone dataset available=83

Scored with 2007 Alternate=81 [data recorded between 2010 and 2014]

Scored with 2012 Recommended=2 [data recorded from 2015 onward]

Non-Segmentation based analysis (Scored using 2007 Alternate criteria):

Total dataset available=81

- RDI Missing=4
- CPAP study=2
- Audio recording corrupt=3 (4/11/2014; 7/11/2014P1; 24/1/2013P2)
- Objective Noise detection algorithm rejected 2 recordings based on excessive noise=2 (19/11/2012P1; 31/01/2013P1)

Data statistics for 70 usable recordings:

- RDI<15=38
- RDI>=15=32
- RDI=>30=20
- Male=40
- Female=30

Dividing Data into Training and Testing & objectively removing Noisy recordings

Buffer Size=8;

Set 1 [Train and LOV=50 and Independent Test=20]:

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
84.44
86.30
86.36
85.71
70
80

Chi Sq Test of Mfcc
84.35
86.16
86.36
85.71
70
70

(T-stat value)

KS test of MFCC
83.99
85.86
86.36
85.71
60
80

(T-stat value)

Lilliefors test of
91.74
93.51
86.36
89.29
60
80

MFCC (T-stat

value)

Chi Sq Test of Mfcc
86.35
89.28
86.36
85.71
70
80

(p-value)

KS test of MFCC
83.99
85.86
86.36
85.71
60
80

(p-value)

Lilliefors test of
86.45
88.29
86.36
85.71
60
80

MFCC (p-value)

NGS + Lilliefors
100.00
100.00
90.91
89.29
60
50

test p value

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 30
Specificity
Sensitivity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
91.00
92.93
86.36
89.29
70
80

Chi Sq Test of Mfcc
86.36
87.92
86.36
85.71
80
80

(T-stat value)

KS test of MFCC
90.91
92.85
90.91
85.71
60
60

(T-stat value)

Lilliefors test of
92.02
93.73
86.36
89.29
70
80

MFCC (T-stat

value)

Chi Sq Test of Mfcc
90.81
92.78
90.91
89.29
80
80

(p-value)

KS test of MFCC
90.91
92.85
90.91
85.71
60
60

(p-value)

Lilliefors test of
86.64
89.50
86.36
85.71
60
80

MFCC (p-value)

NGS + Lilliefors
100.00
100.00
95.45
89.29
80
50

test p value

Set 2 [Created after shuffling the training and test data; Train and LOV=50 and Independent Test=20]:

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
86.08
87.71
86.36
85.71
100
50

Chi Sq Test of Mfcc
83.90
85.71
86.36
85.71
80
20

(T-stat value)

KS test of MFCC (T-
83.71
85.57
86.36
85.71
80
50

stat value)

Lilliefors test of
91.00
92.93
90.91
92.86
70
50

MFCC (T-stat value)

Chi Sq Test of Mfcc
90.91
92.85
86.36
85.71
70
60

(p-value)

KS test of MFCC (p-
83.71
85.57
86.36
85.71
80
50

value)

Lilliefors test of
86.08
87.71
86.36
85.71
80
30

MFCC (p-value)

NGS + Lilliefors test
87.38
90.01
86.36
85.71
70
60

p value

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
85.81
87.41
86.36
85.71
90
70

Chi Sq Test of Mfcc
90.72
92.71
86.36
89.29
90
60

(T-stat value)

KS test of MFCC (T-
83.81
85.64
86.36
85.71
80
60

stat value)

Lilliefors test of
90.72
92.71
86.36
89.29
70
60

MFCC (T-stat value)

Chi Sq Test of Mfcc
90.72
92.71
86.36
89.29
80
60

(p-value)

KS test of MFCC (p-
83.90
85.71
86.36
85.71
60
40

value)

Lilliefors test of
83.80
85.64
81.82
85.71
90
50

MFCC (p-value)

NGS + Lilliefors test
85.81
87.41
86.36
85.71
90
70

p value

Set 3 [Created after shuffling the training and test data; Train and LOV=50 and Independent Test=20]:

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH=15
Specificity
Sensitivity
Specificity
Sensitivity
Sensitivity
Specificity

NGS of MFCC
90.72
92.71
90.91
85.71
80
70

Chi Sq Test of
90.90
92.86
90.91
89.29
80
40

Mfcc (T-stat

value)

KS test of
83.99
85.79
86.36
85.71
80
80

MFCC (T-stat

value)

Lilliefors test
90.90
92.85
90.91
89.29
80
60

of MFCC (T-

stat value)

Chi Sq Test of
90.81
92.78
86.36
89.29
60
60

Mfcc (p-value)

KS test of
83.99
85.86
86.36
85.71
70
80

MFCC (p-

value)

Lilliefors test
79.53
82.29
81.82
82.14
50
60

of MFCC (p-

value)

NGS +
86.17
89.07
86.36
89.29
80
50

Lilliefors test p

value

Leave One Out

EPOCH
Training Results
Validation Results
Test Results

LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
83.90
85.71
86.36
85.71
60
50

Chi Sq Test of
85.90
87.49
86.36
85.71
70
70

Mfcc (T-stat

value)

KS test of
86.73
89.57
86.36
89.29
60
80

MFCC (T-stat

value)

Lilliefors test
100.00
100.00
100.00
89.29
70
50

of MFCC (T-

stat value)

Chi Sq Test of
86.08
87.63
86.36
85.71
40
60

Mfcc (p-value)

KS test of
86.82
89.65
86.36
89.29
60
80

MFCC (p-

value)

Lilliefors test
84.26
86.08
86.36
85.71
70
70

of MFCC (p-

value)

NGS +
100.00
100.00
86.36
85.71
60
60

Lilliefors test p

value

Android Trained Model Tested on Iphone Dataset

Set 1

Training Set Patient
Test Set Patient

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
64
64
80
80

Chi Sq Test of Mfcc
68
54
90
60

(T-stat value)

KS test of MFCC
55
75
50
80

(T-stat value)

Lilliefors test of MFCC
64
64
80
80

(T-stat value)

Chi Sq Test of Mfcc
77
39
80
40

(p-value)

KS test of MFCC
55
61
40
70

(p-value)

Lilliefors test of MFCC
68
64
90
80

(p-value)

NGS + Lilliefors test
59
75
80
70

p value

Training Set Patient
Test Set Patient

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
59
64
90
70

Chi Sq Test of Mfcc
59
64
80
70

(T-stat value)

KS test of MFCC
64
75
60
70

(T-stat value)

Lilliefors test of MFCC
64
68
90
80

(T-stat value)

Chi Sq Test of Mfcc
77
64
80
70

(p-value)

KS test of MFCC
77
54
70
70

(p-value)

Lilliefors test of MFCC
64
64
90
70

(p-value)

NGS + Lilliefors test
68
86
90
60

p value

Set 2

Training Set Patient
Test Set Patient

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
82
54
70
60

Chi Sq Test of Mfcc
77
71
80
50

(T-stat value)

KS test of MFCC (T-
68
71
60
60

stat value)

Lilliefors test of
77
75
80
60

MFCC (T-stat value)

Chi Sq Test of Mfcc
73
82
90
50

(p-value)

KS test of MFCC (p-
68
68
50
50

value)

Lilliefors test of
77
54
70
40

MFCC (p-value)

NGS + Lilliefors test
86
61
70
30

p value

Training Set Patient
Test Set Patient

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
77
68
60
60

Chi Sq Test of Mfcc
68
64
80
50

(T-stat value)

KS test of MFCC (T-
64
71
50
50

stat value)

Lilliefors test of
73
61
70
50

MFCC (T-stat value)

Chi Sq Test of Mfcc
68
75
70
60

(p-value)

KS test of MFCC (p-
64
71
50
50

value)

Lilliefors test of
64
79
80
50

MFCC (p- value)

NGS + Lilliefors test
95
46
90
80

p value

Set 3

Training Set Patient
Test Set Patient

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
77
68
70
80

Chi Sq Test of Mfcc
77
61
80
70

(T-stat value)

KS test of MFCC (T-
50
61
50
80

stat value)

Lilliefors test of
77
61
70
70

MFCC (T-stat value)

Chi Sq Test of Mfcc
68
79
60
80

(p-value)

KS test of MFCC (p-
45
71
40
60

value)

Lilliefors test of
73
64
70
70

MFCC (p-value)

NGS + Lilliefors test
82
61
60
90

p value

Training Set Patient
Test Set Patient

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
68
64
80
60

Chi Sq Test of Mfcc
50
61
80
70

(T-stat value)

KS test of MFCC (T-
50
64
70
90

stat value)

Lilliefors test of
73
71
70
70

MFCC (T-stat value)

Chi Sq Test of Mfcc
45
68
60
80

(p-value)

KS test of MFCC (p-
68
68
40
50

value)

Lilliefors test of
73
64
60
80

MFCC (p-value)

NGS + Lilliefors test
68
64
80
60

p value

Iphone Trained Model Tested on Android Dataset:

Set 1

Training Set Patient
Test Set Patient

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
64
75
50
50

Chi Sq Test of Mfcc
64
54
30
70

(T-stat value)

KS test of MFCC
52
89
40
90

(T-stat value)

Lilliefors test of MFCC
56
82
60
70

(T-stat value)

Chi Sq Test of Mfcc
60
75
70
60

(p-value)

KS test of MFCC
52
89
40
90

(p-value)

Lilliefors test of MFCC
76
68
30
90

(p-value)

NGS + Lilliefors test
64
61
60
50

p value

Training Set Patient
Test Set Patient

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
52
71
70
70

Chi Sq Test of Mfcc
68
71
50
80

(T-stat value)

KS test of MFCC
68
75
70
50

(T-stat value)

Lilliefors test of MFCC
60
71
50
70

(T-stat value)

Chi Sq Test of Mfcc
72
68
70
60

(p-value)

KS test of MFCC
60
82
60
60

(p-value)

Lilliefors test of MFCC
56
64
40
60

(p-value)

NGS + Lilliefors test
64
61
60
50

p value

Set 2

Training Set Patient
Test Set Patient

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
72
64
60
70

Chi Sq Test of Mfcc
84
75
70
50

(T-stat value)

KS test of MFCC (T-
52
68
40
80

stat value)

Lilliefors test of
68
64
50
70

MFCC (T-stat value)

Chi Sq Test of Mfcc
84
79
60
50

(p-value)

KS test of MFCC (p-
52
68
30
80

value)

Lilliefors test of
96
82
80
40

MFCC (p-value)

NGS + Lilliefors test
72
68
50
60

p value

Training Set Patient
Test Set Patient

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
68
89
60
80

Chi Sq Test of Mfcc
84
79
60
50

(T-stat value)

KS test of MFCC (T-
60
68
50
80

stat value)

Lilliefors test of
84
79
70
50

MFCC (T-stat value)

Chi Sq Test of Mfcc
84
79
60
40

(p-value)

KS test of MFCC (p-
32
54
20
70

value)

Lilliefors test of
76
68
90
60

MFCC (p-value)

NGS + Lilliefors test
68
89
60
80

p value

Set 3

Training Set Patient
Test Set Patient

EPOCH LENGTH = 15
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
52
64
80
60

Chi Sq Test of Mfcc
72
54
60
70

(T-stat value)

KS test of MFCC (T-
24
54
50
70

stat value)

Lilliefors test of
52
71
90
70

MFCC (T-stat value)

Chi Sq Test of Mfcc
44
71
70
50

(p-value)

KS test of MFCC (p-
16
54
50
70

value)

Lilliefors test of
56
82
50
60

MFCC (p-value)

NGS + Lilliefors test
68
71
80
60

p value

Training Set Patient
Test Set Patient

EPOCH LENGTH = 30
Sensitivity
Specificity
Sensitivity
Specificity

NGS of MFCC
64
86
80
70

Chi Sq Test of Mfcc
60
68
50
70

(T-stat value)

KS test of MFCC (T-
28
57
50
70

stat value)

Lilliefors test of
68
82
70
90

MFCC (T-stat value)

Chi Sq Test of Mfcc
68
79
50
50

(p-value)

KS test of MFCC (p-
20
57
50
70

value)

Lilliefors test of
52
86
80
80

MFCC (p-value)

NGS + Lilliefors test
68
71
60
80

p value

In general terms, a method according to an embodiment of an aspect of the present invention comprises a method for diagnosing a malady of a patient from sounds of the patient. The malady may be OSA or a respiratory disease such as pneumonia or some other impairment from normal health that results in changes to the sounds that a patient produces. The method includes the steps of initially making a digital recording of the sounds of the patient and that may be done with a contactless microphone as previously discussed. The digital recording is processed by one or more suitably programmed electronic processors to extract a multiplicity of features for sub-segments of each of a number epochs of the digital recording. Features comprising MFCCs have been discussed in detail but other features can also be used in other embodiments such as pitch, entropy, formants, NGS and higher-order spectra-based features. The features are suitably stored in an electronic data storage apparatus such as an electronic or magnetic storage device or server or network accessible storage. The method then involves operating the processors for determining deviation scores from a probability distribution for each epoch based on the extracted multiplicity of features which are retrieved from the storage. In the preferred embodiment the probability distribution that is used is the Gaussian distribution but other distributions can also be used and have been previously mentioned in the results tabled above. The one or more processors then generate a test vector derived from the deviation scores which is then applied to a pre-trained decision machine which is implemented by the processors or on another data network accessible hardware platform. The decision machine that has primarily been discussed is a LRM but other decisions machines such as artificial neural networks, Bayesian decision machines, support vector machines, might also be used.

Finally a diagnosis of malady on the basis of the output from the decision machine is presented on a display under control of the processors, for example to a clinician in order that suitable therapy can be applied to the patient if a malady has been found to be present. For example, therapy may involve administration of antibiotics (for patients suffering from pneumonia), application of controlled air pressure (for patients suffering from OSA) and other appropriate therapies based upon the diagnosis.

The following references are each incorporated herein in their entireties by cross-reference.

[1] B. M. Altevogt and H. R. Colten, Sleep Disorders and Sleep Deprivation: An Unmet Public Health Problem: National Academies Press, 2006.
[2] Wake Up Australia: the value of healthy sleep: Access Economics: Australian Sleep Association, 2004.
[3] T. Young, L. Evans, L. Finn, and M. Palta, “Estimation of the clinically diagnosed proportion of sleep apnea syndrome in middle-aged men and women,” Sleep, vol. 20, pp. 705-706, 1997.
[4] K. E. Bloch, “Polysomnography: a systematic review,” Technol Health Care, vol. 5, pp. 285-305, 1997.
[5] E. Lugaresi, S. Mondini, M. Zucconi, P. Montagna, and F. Cirignotta, “Staging of heavy snorers' disease. A proposal,” Bulletin europeen de physiopathologie respiratoire, vol. 19, pp. 590-594, 1982.
[6] H. Michael, S. Andreas, B. Thomas, H. Beatrice, H. Werner, and K. Holger, “Analysed snoring sounds correlate to obstructive sleep disordered breathing,” European Archives of Oto-Rhino-Laryngology, vol. 265, pp. 105-113, 2008.
[7] W. Whitelaw, “Characteristics of the snoring noise in patients with and without occlusive sleep apnea,” Am Rev Respir Dis, vol. 147, pp. 635-644, 1993.
[8] A. K. Ng, T. S. Koh, E. Baey, T. H. Lee, U. R. Abeyratne, and K. Puvanendran, “Could formant frequencies of snore signals be an alternative means for the diagnosis of obstructive sleep apnea?,” Sleep medicine, vol. 9, pp. 894-898, 2008.
[9] T. Emoto, U. R. Abeyratne, M. Akutagawa, S. Konaka, and Y. Kinouchi, “High frequency region of the snore spectra carry important information on the disease of sleep apnoea,” Journal of medical engineering & technology, vol. 35, pp. 425-431, 2011.
[10] A. Azarbarzin and Z. Moussavi, “Snoring sounds variability as a signature of obstructive sleep apnea,” Medical engineering & physics, vol. 35, pp. 479-485, 2013.
[11] J. Sola-Soler, J. A. Fiz, J. Morera, and R. Jané, “Multiclass classification of subjects with sleep apnoea-hypopnoea syndrome through snoring analysis,” Medical engineering & physics, vol. 34, pp. 1213-1220, 2012.
[12] N. Ben-Israel, A. Tarasiuk, and Y. Zigel, “Obstructive apnea hypopnea index estimation by analysis of nocturnal snoring signals in adults,” Sleep, vol. 35, pp. 1299-3050, 2012.
[13] H. Alshaer, G. R. Fernie, E. Maki, and T. Douglas Bradley, “Validation of an automated algorithm for detecting apneas and hypopneas by acoustic analysis of breath sounds,” Sleep medicine, vol. 14, pp. 562-571, 2013.
[14] U. Abeyratne, S. de Silva, C. Hukins, and B. Duce, “Obstructive sleep apnea screening by integrating snore feature classes,” Physiological Measurement, vol. 34, p. 99, 2013.
[15] U. R. Abeyratne, A. S. Karunajeewa, and C. Hukins, “Mixed-phase modeling in snore sound analysis,” Medical and Biological Engineering and Computing, vol. 45, pp. 791-806, 2007.
[16] A. S. Karunajeewa, U. R. Abeyratne, and C. Hukins, “Multi-feature snore sound analysis in obstructive sleep apnea-hypopnea syndrome,” Physiological measurement, vol. 32, p. 83, 2011.
[17] N. Oliver and F. Flores-Mangas, “HealthGear: automatic sleep apnea detection and monitoring with a mobile phone,” Journal of Communications, vol. 2, pp. 1-9, 2007.
[18] S. Alqassim, M. Ganesh, S. Khoja, M. Zaidi, F. Aloul, and A. Sagahyroon, “Sleep apnea monitoring using mobile phones,” in e-Health Networking, Applications and Services (Healthcom), 2012 IEEE 14th International Conference on, 2012, pp. 443-446.
[19] T. Hao, G. Xing, and G. Zhou, “iSleep: unobtrusive sleep quality monitoring using smartphones,” in Proceedings of the 11th ACM Conference on Embedded Networked Sensor Systems, 2013, p. 4.
[20] H. Nakano, K. Hirayama, Y. Sadamitsu, A. Toshimitsu, H. Fujita, S. Shin, et al., “Monitoring sound to quantify snoring and sleep apnea severity using a smartphone: proof of concept,” Journal of clinical sleep medicine: JCSM: official publication of the American Academy of Sleep Medicine, vol. 10, pp. 73-78, 2014.
[21] D. Salomon, Data compression: the complete reference: Springer, 2004.
[22] K. Sayood, Introduction to data compression: Newnes, 2012.
[23] H. Teager and S. Teager, “Evidence for nonlinear sound production mechanisms in the vocal tract,” in Speech production and speech modelling, ed: Springer, 1990, pp. 241-261.
[24] H. Ghaemmaghami, U. Abeyratne, and C. Hukins, “Normal probability testing of snore signals for diagnosis of obstructive sleep apnea,” in Engineering in Medicine and Biology Society, 2009. EMBC 2009. Annual International Conference of the IEEE, 2009, pp. 5551-5554.

In compliance with the statute, the invention has been described in language more or less specific to structural or methodical features. The term “comprises” and its variations, such as “comprising” and “comprised of” is used throughout in an inclusive sense and not to the exclusion of any additional features. It is to be understood that the invention is not limited to specific features shown or described since the means herein described herein comprises preferred forms of putting the invention into effect. The invention is, therefore, claimed in any of its forms or modifications within the proper scope of the appended claims appropriately interpreted by those skilled in the art.

Throughout the specification and claims (if present), unless the context requires otherwise, the term “substantially” or “about” will be understood to not be limited to the value for the range qualified by the terms.

Any embodiment of the invention is meant to be illustrative only and is not meant to be limiting to the invention. Therefore, it should be appreciated that various other changes and modifications can be made to any embodiment described without departing from the scope of the invention.

A METHOD AND APPARATUS FOR DIAGNOSIS OF MALADIES FROM PATIENT SOUNDS

Information

Publication Number

Date Filed

Date Published

Inventors

Original Assignees

CPC

International Classifications

Abstract

Description

Claims

Priority Claims (1)

PCT Information