Patent Application
20210287793
The disclosure relates to a medical system for generating a medical suggestion. More particularly, the disclosure relates to an AI-based medical system capable of providing explainable descriptions about the medical suggestion.
Recently the concept of computer-aided medical system has emerged in order to facilitate diagnosis for patients. The computer-aided medical system may request patients to provide some information, and then the computer aided medical system may provide a diagnosis or a recommendation of the potential diseases based on the interactions with those patients. The computer-aided medical system may aid a doctor in diagnosing, or aid a patient in consulting or self-diagnosing.
Most of the computer-aided medical system utilizes an Artificial Intelligence technology (including machine learning and/or neural network model) to predict the potential diseases or give related recommendations. However, the AI-based technology usually provides a result (diagnosis or recommendation) without any explanation. Therefore, it is hard for the patient or the doctor to trust or to understand the result provided by the AI-based technology.
The disclosure provides a control method which includes following operations. A symptom input status and a test result status are collected. A neural network is utilized to generate a test suggestion, a predicted test result distribution and a predicted disease distribution according to the symptom input status and the test result status. The test suggestion includes a candidate test. Information gains of the candidate test relative to diseases are estimated according to the predicted test result distribution and the predicted disease distribution. An explainable description about the test suggestion is generated according to the information gains of the candidate test.
The disclosure provides a control method which includes following operations. A symptom input status and a test result status are collected. The symptom input status includes symptom answers. The test result status includes test results. A neural network is utilized to generate a predicted disease distribution according to the symptom input status and the test result status. A predicted disease list is generated according to the predicted disease distribution. An attention mask is applied to filter the symptom answers and the test results for obtaining an attention input. An explainable description about the predicted disease list is generated according to the attention input.
The disclosure provides a medical system, which includes an interface and a processor. The interface is configured for receiving a symptom input status and a test result status. The symptom input status includes symptom answers. The test result status includes test results. The processor is coupled with the interface.
According to some embodiments, in a test suggestion phase, the processor utilizes a neural network to generate a test suggestion, a predicted test result distribution and a predicted disease distribution according to the symptom input status and the test result status. The test suggestion includes a candidate test. The processor estimates a plurality of information gains of the candidate test relative to diseases according to the predicted test result distribution and the predicted disease distribution. The processor generates a first explainable description about the test suggestion according to the information gains of the candidate test.
According to some embodiments, in a disease prediction phase, the processor generates a predicted disease list according to the predicted disease distribution. The processor applies an attention mask to filter the symptom answers and the test results for obtaining an attention input. The processor generates a second explainable description about the predicted disease list according to the attention input.
In some embodiments, the disclosure can provide AI decision-making explanations for the automatic diagnosis system in the test suggestion phase and the disease prediction phase. In the test suggestion phase, the distribution of test results predicted by the neural network can be used to calculate the information gain of each test in distinguishing different diseases, so as to explain a correspondence between the suggest test and a target disease to distinguish. In the disease prediction phase, an attention mask is utilized to find out critical symptoms and/or critical test results, which are important for disease prediction.
It is to be understood that both the foregoing general description and the following detailed description are demonstrated by examples, and are intended to provide further explanation of the invention as claimed.
Embodiments of the invention will now be described with reference to the attached drawings in which:
Reference will now be made in detail to the present embodiments of the disclosure, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers are used in the drawings and the description to refer to the same or like parts.
Reference is made to
In some embodiments, the processor 140 is communicated with the interface 120. The medical system 100 is configured to interact with the user U1 through the interface 120 (e.g. collecting a symptom input status Ssym from the user U1, providing some symptom inquiries Sqry to the user U1, collecting corresponding symptom responses Sans from the user U1, providing a test suggestion TS to the user U1, collecting test result inputs Str from the user U1). Based on aforesaid interaction history, the medical system 100 is able to analyze, diagnose or predict a potential disease occurring to the user U1.
The medical system 100 is trained with a machine learning algorithm or a reinforcement learning algorithm, such that the medical system 100 is capable to inquire and diagnose based on limited patient data. In some embodiments, the medical system 100 adopts a reinforcement learning (RL) framework to formulate inquiry and diagnosis policies (e.g., Markov decision processes). In some embodiments, a neural network is trained based on the machine learning algorithm or the reinforcement learning algorithm by the processor 140 according to some training data (e.g., known medical records) and the trained neural network is stored in the storage 160.
In some embodiments, the medical system 100 is established with a computer, a server or a processing center. The processor 140 can be implemented by a central processing unit (CPU), a graphic processing unit (GPU), a tensor processing unit (TPU), an application-specific integrated circuit (ASIC) or any equivalent computation unit. The interface 120 can include an output interface (e.g., a display panel for display information) and an input device (e.g., a touch panel, a keyboard, a microphone, a scanner or a flash memory reader) for user to type text commands, to give voice commands or to upload some related data (e.g., images, medical records, or personal examination reports). As shown in
As shown in
Reference is further made to
As shown in
In some embodiments, the neural network 142 can be trained with a machine learning algorithm or a reinforcement learning algorithm according to the training data in advance. In some embodiments, the training data includes known medical records. The medical system 100 utilizes the known medical records in the training data to train the neural network 142. In an example, the training data can be obtained from data and statistics information from the Centers for Disease Control and Prevention (www.cdc.gov/datastatistics/index.html). The medical system 100 is capable to inquire and diagnose based on limited patient data. Further details about how to train the neural network 142 will be discussed later in some other embodiments.
After training, the neural network 142 is able to generate output status OUT based on contents of the symptom input status Ssym and the test result status Str of the input status IN. Reference is further made to
As shown in
As shown in
In step S220, the neural network 142 is able to generate output status OUT based on contents of the symptom input status Ssym and the test result status Str. In some embodiments, the output status OUT includes the symptom-query state values Osq, the test-suggestion state values Otest, the predicted disease distribution Odd and the predicted test result distribution Otrd. Reference is further made to
As shown in
In step S230, the processor 140 select the next action according to the output status OUT. In some embodiments, if the output status OUT indicates the symptom information is not enough (e.g., an amount of the answered symptom queries is less than a threshold, test-suggestion state values Otest is not high enough, or a maximum probability of one candidate diseases CD_1 to CD_x in the predicted disease distribution Odd is not high enough) to give a test suggestion or make a disease prediction, the control method will enter a symptom inquiry phase P1 to generate a symptom query Sqry in step S241. In some embodiments, the symptom query Sqry is generated according to a maximum state value among the symptom-query state values Osq. For example, if the symptom-query state values Osq indicating the symptom query Sqry_2 (e.g., “do you feel pain at your stomach?”) with the maximum state value, the symptom query Sqry_2 is generated and presented to the user U1. The user can answer to the symptom query Sqry_2. In step S242, the symptom answer Sans relative to the symptom query Sqry_2 can be collected from the user U1 through the interface 120. The processor 140 can updates the symptom input status Ssym according to the symptom answer Sans. As shown in
In some embodiments, when the medical system 100 and the control method 200 gather enough symptom information, the medical system 100 and the control method 200 will enter a test suggestion phase P2 as shown in
In some embodiments, the test suggestion TS includes at least one candidate test(s). As the embodiment shown in Table 1, the test suggestion TS includes three candidate tests CT1˜CT3, which includes Rapid Influenza diagnostic test (CT1), the Throatology Examination (CT2) and the Thoracic Examination (CT3). These candidate tests CT1˜CT3 are suggested according to three highest state values among the test-suggestion state values Otest. The embodiment shown in Table 1, these three candidate tests CT1˜CT3 in the test suggestion TS are illustrated for demonstration, and the medical system 100 and the control method 200 are not limited to provide a specific amount of candidate test(s). For example, the medical system 100 and the control method 200 in some embodiments can provide one candidate test to ten candidate tests in the test suggestion TS. In some AI systems, a test suggestion similar to Table 1 can be generated and provided to the user, but if the test suggestion TS is provided to the user U1 without any explanations, it is hard for the user to understand, to verify, or to trust the test suggestion TS. Sometimes, even a medical profession is hard to explain the test suggestion TS generated by an AI system. In some embodiments discussed in following paragraphs, the medical system 100 and the control method 200 are able to provide a first explainable description ED1 to explain the test suggestion TS.
In step S252, an explanation module 146 in the processor 140 is configured to estimate information gains of each candidate test relative to different diseases according to the predicted test result distribution Otrd and the predicted disease distribution Odd. For example, if there are 70 diseases (D1˜D70) considered in the medical system 100, the explanation module 146 will estimate one information gain between “CT1” and “D1”, another information gain between “CT1” and “D2” . . . and another information gain between “CT1” and “D70”. Similarly, the explanation module 146 will estimate one information gain between “CT2” and “D1”, another information gain between “CT2” and “D2” . . . and another information gain between “CT2” and “D70”. In some embodiments, the explanation module 146 can be implemented by software instructions executed by the processor 140.
For brevity, the following embodiment discussed how to calculating one information gain between one candidate test and one target disease. An information gain Ĝd(IN, CT) of the candidate test CT on a patient with the current input status IN relative to a target disease “d” can be estimated as below:
Ĝ
d(IN,CT)=Îd(IN)−ΣvϵV
The information gain Ĝd(IN, CT) of the candidate test CT on a patient with the current input status IN relative to a target disease “d” is estimated in reference with a first Gini index Îd(IN) about the target disease “d” in a group with the current input status IN before performing the candidate test CT, a probability {circumflex over (p)}test(IN, CT, v) to get a target result “v” in the candidate test CT, and a second Gini index Îd(IN, CT=v) about the target disease “d” in a group with the target result “v” after performing the candidate test CT. The target result “v” is one of possible results VCT of performing the candidate test CT.
In the equation (1) above, the probability {circumflex over (p)}test (IN, CT, v) to get a target result “v” in the candidate test CT can be acknowledged according to the predicted test result distribution Otrd generated by the neural network 142.
In aforesaid equation (1), the first Gini index Îd(IN) can be estimated as below:
Î
d(IN)=1−πdis(IN,d)2−(1−πdis(IN,d))2 (2)
In aforesaid equation (2), πdis(IN, d) is an estimated probability that a patient with the current input status IN is diagnosed to have the target disease “d”. The πdis(IN, d) can be acknowledged according to the predicted disease distribution Odd generated by the neural network 142.
In aforesaid equation (1), the second Gini index Îd(IN, CT=v) can be estimated as below:
Î
d(IN,CT=v)=1−πdis(INv,d)2−(1−πdis(IN,d))2 (3)
In aforesaid equation (3), πdis(INv, d) is an estimated probability that a patient (with the current input status IN and expected to have the target result “v” in the candidate test CT) is diagnosed to have the target disease “d”. The πdis(INv, d) can be acknowledged by updating the input status IN into another input status INv (filling the target result “v” into the result of the candidate test CT) and utilize the neural network 142 to re-calculate the predicted disease distribution Odd according to the input status INv.
When the information gain Ĝd(IN, CT) of the candidate test CT is higher, it means the candidate test CT has a higher significance in distinguishing whether the patient with the current input status IN has the target disease “d” or not. When the information gain Ĝd(IN, CT) of the candidate test CT is lower, it means the candidate test CT is not helpful in distinguishing whether the patient with the current input status IN has the target disease “d” or not. By calculating the information gains Ĝd(IN, CT) of the candidate test CT on each one of all candidate diseases, the medical system 100 and the control method 200 can acknowledge that the candidate test CT is important in distinguishing which one among all candidate diseases.
In step S253, the explanation module 146 is configured to generate a first explainable description ED1 about the test suggestion TS according to the information gains of the candidate test(s).
The following Table 2 is an example about a first explainable description ED1, which is generated by the medical system 100 and the control method 200 to explain the test suggestion TS.
As shown in Table 2, the first explainable description ED1 is helpful for the user U1 to understand the test suggestion TS given by the medical system 100 and the control method 200. In some embodiments, the confidence levels of the suspected disease shown in the first explainable description ED1 can be determined by the predicted disease distribution Odd according to the current input status IN. In some embodiments, target distinguish diseases of the candidate tests shown in the first explainable description ED1 can be determined by the information gains of each candidate test relative to each disease.
After the suggest tests are performed on the user U1, the test results about these suggested tests can be collected through the interface 120 in step S254. In step S255, the processor 140 is able to update the test result status Str in the input status IN according to these collected test results. The control method 200 will return to S220, to re-generate the output status OUT by the neural network 142 according to the input status IN after updating. In the case, the predicted disease distribution Odd in the output status OUT will also consider the test results in these suggested tests.
After the symptom inquiry phase P1, the test suggestion phase P2, the medical system 100 and the control method 200 may gather enough information (about symptom queries and test results) to predict a disease of the user U1. The control method 200 will enter a disease prediction phase P3, and in Step S261, the processor 140 generates a predicted disease list DP (and/or a medical department recommendation corresponding to the predicted disease list DP correspondingly) according to the predicted disease distribution Odd generated by the neural network 142 according to the input status IN. As shown in
The following Table 3 shows an example of the predicted disease list DP. In the example shown in the following Table 3, the predicted disease list DP includes three different diseases for demonstration, and the disclosure is not limited thereto.
In some embodiments, the predicted disease list DP includes at least one predicted disease(s). As shown in Table 3, the predicted disease list DP includes three predicted diseases, which includes Acute gastroenteritis, Norovirus infection and Cholera. These predicted diseases are suggested according to three highest state values among the predicted disease distribution Odd. The embodiment shown in Table 3, these predicted diseases in the predicted disease list DP are illustrated for demonstration, and the medical system 100 and the control method 200 are not limited to provide a specific amount of predicted disease(s). For example, the medical system 100 and the control method 200 in some embodiments can provide one predicted disease to ten predicted diseases in the predicted disease list DP.
In some other embodiments, in Step S261, the processor 140 can generate a medical department recommendation according to the predicted disease list DP, and the medical system 100 and the control method 200 can provide the medical department recommendation to the user U1 instead of the predicted disease list DP. For example, the medical department recommendation may include “Emergency Department” corresponding to the predicted disease “acute gastroenteritis”.
In some AI systems, a test suggestion similar to Table 3 can be generated and provided to the user, but if the predicted disease list DP is provided to the user U1 without any explanations, it is hard for the user to understand, to verify, or to trust the predicted disease list DP. Sometimes, even a medical profession is hard to explain the predicted disease list DP generated by an AI system. In some embodiments discussed in following paragraphs, the medical system 100 and the control method 200 are able to provide a second explainable description ED2 to explain the predicted disease list DP.
In a demonstrational example, information in the input status IN (the symptom input status Ssym and the test result status Str) corresponding to the predicted disease list DP shown in Table 3 may include:
(feature 1) Have fever;
(feature 2) No trauma;
(feature 3) Have cough symptoms;
(feature 4) No symptoms of runny nose;
(feature 5) Symptom of vomiting;
(feature 6) Symptom of diarrhea;
(feature 7) No symptom of general muscle ache; and
(feature 8) Vibrio cholerae culturing test: negative.
In Step S262, referring to
In step S263, the attention mask MSK is applied to the input status IN to obtain the attention input INm. For example, the attention mask MSK can allow the features 1, 6 and 7 to pass, such that the attention input INm includes (feature 1) fever; (feature 6) diarrhea and (feature 7) no general muscle ache.
In step S264, the explanation module 146 is configured to generate a second explainable description ED2 about the predicted disease list DP according to the attention input INm.
The following Table 4 is an example about a second explainable description ED2, which is generated by the medical system 100 and the control method 200 to explain the predicted disease list DP in Table 3.
As shown in Table 4, the second explainable description ED2 is helpful for the user U1 to understand the predicted disease list DP given by the medical system 100 and the control method 200.
In some embodiments, the attention module 144 is trained (about how to generate the attention mask MSK according to the input status IN) in a training phase according to some training data. This attention module 144 can be implemented by several fully-connected layers, which will generate a set of attention masks based on the input status IN (can be feed according to the known medical records in the training data), selectively block some of the input features, so that the following neural network 142 will only perform calculations based on some of the features. After training, the neural network 142 can determine which features are important according to the input, and the attention module 144 generates the corresponding attention mask to retain the important features in the input and block the unimportant features, so that the entire model can still generate the correct features based on the filtered input status. By utilizing the attention mask MSK, when the predicted disease list DP is generated, the medical system 100 can identify which input features the predicted disease list DP is based on, according to the attention mask MSK generated by the attention module 144.
In some embodiments, the neural network 142 is trained in advance according to some training data (e.g., known medical records). The processor 140 utilizes the neural network 142 to generate the output status OUT and accordingly selects the sequential actions from a set of candidate actions. In some embodiments, the sequential actions include some symptom inquiry actions, one or more medical test actions (suitable for providing extra information for predicting or diagnosing the disease) and a disease prediction action.
When the processor 140 selects proper actions (e.g., some proper symptom inquiries, some proper medical test actions or a correct disease prediction action matching with the medical records in the training data), corresponding rewards will be provided to the neural network 142. In some embodiments, the neural network 142 is trained to maximize cumulative rewards in response to the sequential actions. In some embodiments, the cumulative rewards can be calculated by a sum of a symptom abnormality reward, a test abnormality reward, a test cost penalty and/or a positive/negative prediction reward. In other words, the neural network model 142 is trained to ask proper symptom inquiries, suggest proper medical tests and make the correct disease prediction at its best.
Based on embodiments above, the disclosure can provide AI decision-making explanations for the automatic diagnosis system in the test suggestion phase and the disease prediction phase. In the test suggestion phase, the distribution of test results predicted by the neural network can be used to calculate the information gain of each test in distinguishing different diseases, so as to explain a correspondence between the suggest test and a target disease to distinguish. In the disease prediction phase, an attention mask is utilized to find out critical symptoms and/or critical test results, which are important for disease prediction.
Although the present invention has been described in considerable detail with reference to certain embodiments thereof, other embodiments are possible. Therefore, the spirit and scope of the appended claims should not be limited to the description of the embodiments contained herein.
It will be apparent to those skilled in the art that various modifications and variations can be made to the structure of the present invention without departing from the scope or spirit of the invention. In view of the foregoing, it is intended that the present invention cover modifications and variations of this invention provided they fall within the scope of the following claims.
This application claims priority to U.S. Provisional Application Ser. No. 62/987,881, filed Mar. 11, 2020, which is herein incorporated by reference.
| Number | Date | Country | |
|---|---|---|---|
| 62987881 | Mar 2020 | US |
