Method for analyzing cough sounds using disease signatures to diagnose respiratory diseases - Patent Application 20070122999

Abstract

1. A method for diagnosing one or more diseases of a patient's airway, comprising: obtaining a cough sound from the patient; processing the cough sound to generate a cough sound feature signal representing one or more cough sound features from the cough segment; obtaining one or more disease signatures based on the cough sound feature signal; and classifying the one or more disease signatures to consider the cough segment as indicative of one or more of the diseases, wherein obtaining one or more disease signatures based on the cough sound feature signal includes applying the cough sound features to each of one or more pre-trained disease signature decision machines, each of the decision machines being pre-trained to classify the cough sound features as corresponding to either a specific disease or a non-disease state, or as corresponding to a first specific disease or a second specific disease different from the first specific disease.

Description

Technical Field

[0001] The present invention relates to a method and apparatus for assisting medical staff in diagnosing and treating patients suffering from respiratory diseases.

Background Art

[0002] Any reference to prior art methods, apparatus, or documents should not be construed as constituting any evidence or admission that they form or formed part of common general knowledge.

[0003] In a previous study by one or more of the same inventors, which is the subject of International Patent Application PCT / AU2013 / 000323 incorporated herein by reference in its entirety, the airway sounds of patients were recorded and cough sounds were identified therefrom. Features were extracted from the cough sounds to form test feature vectors, which were then applied to a pre-trained classifier, preferably a logistic regression model, to diagnose the presence of respiratory dysfunction such as pneumonia in the patient.

[0004] The method for diagnosing disease states described in PCT / AU2013 / 000323 is functioning well and has been successfully implemented commercially, but there is still a need for improvement. For example, it would be advantageous if a method were provided that was an improvement in that it could result in a more accurate diagnosis. Also, it would be preferable if the method could incorporate region-specific information into the diagnostic process. Further, it would be desirable if some of the subjective nature of standard clinical diagnoses could be addressed.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Non-Patent Documents

[0006] [Non-Patent Document 1] Abeyratne, UR, et al., Cough sound analysis can rapidly diagnose childhood pneumonia. Annals of biomedical engineering, 2013. 41(11):2448-2462. [Non-patent document 2] Kosasih, K., et al., Wavelet augmented cough analysis for rapid childhood pneumonia diagnosis. IEEE Transactions on Biomedical Engineering, 2015. 62(4): 1185-1194. [Non-patent document 3] Abeyratne, U.Blind reconstruction of non-minimum-phase systems from 1-D oblique slices of bispectrum.1999.IET. [Non-patent document 4] Hinton GE, Salakhutdinov RR.Reducing the dimensionality of data with neural networks science.2006 July 28;313(5786):504-7 Summary of the Invention [Problem to be solved by the invention]

[0007] It is an object of the present invention to provide a method and apparatus for aiding in the diagnosis of disease states of the airways. [Means for solving the problem]

[0008] According to a first aspect of the present invention there is provided a method for diagnosing one or more diseases of the respiratory tract in a patient, the method comprising: obtaining cough sounds from a patient; Steps of processing a cough sound to generate a cough sound feature signal representing one or more cough sound features from the cough segment; Steps of obtaining one or more disease signatures based on the cough sound feature signal; Steps of classifying one or more disease signatures to consider the cough segment as indicative of one or more of the diseases; including The step of obtaining one or more disease signatures based on the cough sound feature signal includes applying the cough sound features to each of one or more pre-trained disease signature determination machines, each of said determination machines being pre-trained to classify the cough sound features as corresponding to either a specific disease or a non-disease state, or as corresponding to a first specific disease or a second specific disease different from the first specific disease.

[0009] In a preferred embodiment of the present invention, each of the one or more disease signature machines comprises a trained Logistic Regression Model (LRM).

[0010] Each trained Logistic Regression Model (LRM) may be trained using a reduced set of training features that are features of the available set of training features determined to be important for the LRM, thereby avoiding overfitting of the LRM.

[0011] The training features may be determined to be training features important for the LRM by calculating an average p-value for all the training features and then selecting the features having an average p-value smaller than a threshold P min More details of this process according to an embodiment of the present invention are described in Appendix F of this specification.

[0012] In this embodiment, the independent variable of each LRM is the value of the cough sound feature, and the output value from the LRM comprises the predicted probability of a cough indicating the first specific disease for the second specific disease or for the non-disease state.

[0013] The disease signature may also be trained to generate a scale related to specific evaluation or measurement results used in the department of respiratory medicine. Example: Wheeze Severity Scores (WSS) by training the disease signature to separate low WSS from high WSS, low FEV1 measured in spirometry from high FEV1, and low FEV1 / FVC measured in spirometry from high FEV1 / FVC.

[0014] In another embodiment of the present invention, one or more disease signature machines may include one or more trained neural networks. Other classifiers or models that provide continuous output (e.g., generalized linear models, hidden Markov models, etc.) may also be used as signal machines.

[0015] The method may further include applying clinical patient measurements as independent variables to the disease signature determination machine in addition to cough characteristics.

[0016] In one embodiment of the present invention, the step of classifying one or more disease signatures to consider a cough segment as indicative of one or more of the diseases includes applying one or more disease signatures to a signal classifier trained to cover all disease groups.

[0017] In another embodiment of the present invention, the step of classifying one or more disease signatures to consider a cough segment as indicative of one or more of the diseases includes applying one or more disease signatures to a number of classifiers each trained to recognize a disease of interest.

[0018] In a preferred embodiment, the classifiers are each of the following diseases: bronchiolitis (S BO ), croup (S C ), asthma / RAD (S A ), pneumonia (S P), lower respiratory tract disease (S LRTD ), initial URTI(S U ) is trained to recognize one of

[0019] Preferably, the d disease signatures are generated by a step of obtaining one or more disease signatures based on the cough sound feature signal, and correspondingly, the artificial neural network (ANN) has a d-dimensional input layer having one input neuron corresponding to each of the disease signatures.

[0020] In a preferred embodiment of the present invention, the ANN has a k-dimensional output layer, with each neuron in the output layer outputting a probability corresponding to a disease.

[0021] In a preferred embodiment of the present invention, the step of classifying one or more disease signatures comprises using a composite effective probability measure P Q 'of, P' Q =P Q (1-P Z ) where P Q has an index of the probability of a patient belonging to disease Q, and P Z provides an indication of the probability of the patient belonging to disease Z. Therefore, the product P Q (1-P Z ) denotes the probability of the composite event that a patient belongs to disease Q and does not belong to disease Z.

[0022] The method may include calculating a cough index for each of the target diseases, wherein a patient's cough index for a target disease is calculated as the ratio of the patient's coughs classified as indicative of the target disease to the total number of coughs analyzed for said patient.

[0023] The method may further include administering a particular treatment, for example a treatment known to be effective for the patient based on the particular disease diagnosed.

[0024] Preferred features, embodiments, and variations of the present invention can be identified from the following detailed description, which provides sufficient information for those skilled in the art to practice the invention. The detailed description should not be construed as limiting the scope of the foregoing summary of the invention in any way. The detailed description makes reference to several drawings, as follows: [Brief explanation of the drawings]

[0025]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

[0026] FIG. 1 shows a flowchart 100 of a method according to a preferred embodiment of the present invention. The method 100 includes four main processing boxes or blocks as follows: a cough feature calculation block 112, a disease signature generation block 114, a disease signature selection block 116, and a disease signature classifier block 118. Each of these blocks can be implemented to provide an apparatus by using, for example, a computer specially programmed to implement the various functions described herein. Alternatively, the various processing blocks can also be implemented using custom integrated circuit chips and discrete logic. More specifically, some blocks, such as the classifier block 118, can implement an artificial neural network (ANN) and can be implemented using a dedicated ANN integrated circuit such as the Synapse chip by IBM, or the Zeroth processor by Qualcomm, or their equivalents.

[0027] Patient sounds 102 are recorded in box 106 from patient 101 via microphone 104. The audio recording is passed to cough identification and segmentation block 108, which processes the audio recording and automatically segments it into cough sounds, e.g., CG1 110a and CG2 110b. The cough sounds 110a, 110b, consisting of digitized signals, e.g., electronic MP3 file segments, are transferred to cough feature calculation block 112. Cough feature calculation block 112 is configured to process each of the cough sounds to extract a cough feature value in the form of a signal, e.g., an electrical signal transmitted on a circuit board or integrated circuit conductor, representing the degree to which each cough possesses each of several characteristic features described above. The output signal from cough feature calculation block 112 is processed by disease signature generation block 114. The disease signature generation block 114 comprises a plurality of pre-trained decision machines 2a-2n, preferably trained logistic regression machines, each pre-trained to classify a cough feature signal as indicative of a first or second particular disease (or group of diseases). One of the pre-trained decision machines 2a-2n may also be trained to classify a cough feature signal as indicative of a particular disease or normal (i.e., non-disease).

[0028] The disease signature selection block 116 is configured to receive the signals output from the disease signature generation block 114 and identify therefrom the set of disease signatures that provide the best diagnostic results for the training / validation set of data at the output of the classifier block 118. The disease signature selection block 116 is configured to consider multiple factors as follows: (i) The ability to provide excellent training and validation performance in the individual disease signature generation process. (ii) Domain-specific knowledge (e.g., if the target of the final classifier is to diagnose asthma / RAD, a WSS signature is expected to be useful since wheezing is a strong indicator of asthma / RAD. If the classifier has to diagnose pneumonia, a disease signature of pneumonia vs. bronchiolitis may be shown, as these are known to present diagnostic dilemmas in clinical practice, etc.). (iii) Maximizing the diagnostic performance in the classifier output on the training / validation set based on a search process for disease signatures.

[0029] The output values from the selected disease signature selection block 116 are passed to a classifier block 118, which preferably consists of multiple pre-trained neural networks, each of which is pre-trained to classify the output of the selected disease signature generator as any one of several predetermined diseases.

[0030] The diagnostic indication block 120 is responsive to the classifier block 118 and presents indications regarding a disease diagnosis that can be used by a clinician to confirm the disease diagnosis and provide an appropriate treatment to the patient. The output box 120 includes the presentation, for example on a visual electronic display, of a diagnostic indication of a single disease or several diseases to the clinician or other caregiver for the patient, who may then apply a particular treatment, for example a treatment known to be effective for the patient based on the particular disease diagnosed.

[0031] Further details are provided in the discussion below.

[0032] Cough characteristics calculation box 112 In box 112, segments from the patient record identified as containing cough events are applied to a processing block configured to extract mathematical features from each cough event in the manner described in PCT / AU2013 / 000323.

[0033] Additionally, according to a preferred embodiment of the present invention, the cough feature calculation block 112 is also configured to extract cepstral coefficients inspired by Mel-Frequency Cepstral Coefficients (MFCC) coefficients based on the higher-order spectral slices. The new coefficients are referred to herein as BFCCs (Bispectral Frequency Cepstral Coefficients) and are described in further detail in Appendix B(i) of the present specification.

[0034] The cough feature calculation box is configured to perform the following processes during use: (i) Let x denote the discrete-time speech signal from any cough event. (ii) Segment x into, for example, three equal-sized non-overlapping sub-segments. The goal is to capture the variation of mathematical features within a single cough. i Let denote the i-th sag segment of x, where i=1, 2, 3. Each sub-segment x i The following features are calculated from the sigma: 8 bispectral coefficients (BC), non-Gaussianity score (NGS), first 4 formant frequencies (FF), logarithmic energy (LogE), zero crossing (ZCR), kurtosis (Kurt), 31 Mel-frequency cepstral coefficients (MFCC), and Shannon entropy (ShE). (iii) Additionally, we compute 13 wavelet features using the entire cough event data [2]. See Appendix B for a brief description of each cough feature. (iv) Total C f = 157 features are extracted from each cough event.

[0035] Disease Signature Generation Box 114 In previous studies that were the subject of PCT / AU2013 / 000323, patients were directly classified by training a logistic regression model on the feature set calculated in box 112.

[0036] In contrast to the approach employed in PCT / AU2013 / 000323, a preferred embodiment of the present invention has several objectives, (a) as a way to inject region-specific information into the diagnostic process, (b) as a way to address the subjective nature of standard clinical diagnoses, and (c) as a way to produce a more accurate diagnosis, and includes a new procedure herein referred to as "disease signature generation".

[0037] Disease signature generation, which takes place in box 114, involves mapping the input features from the cough feature calculation block 112 to a set of axes represented in digital memory that vary continuously between 0 and 1 and define a space in which the classifier can better diagnose individual diseases. Each of these axes is referred to as a disease signature axis. All patients provide a response along each of the disease signature axes, and the set of response values for each disease signature for each patient is then sent to a classifier block 118 that generates a diagnosis presented in the diagnostic instruction block 120. As will be explained, the disease signature axes embed region-specific knowledge.

[0038] In one embodiment, the input features from box 112 to box 114 can be converted into disease signatures by training a logistic regression model (LRM). For example, the signature axis can be achieved by training a set of models such as {(bronchiolitis vs. normal), (pseudomembranous laryngitis vs. normal), (bronchiolitis vs. pseudomembranous laryngitis), (bronchiolitis vs. all diseases), (pseudomembranous laryngitis vs. A / RAD)... etc.}. Each model provides one disease signature. The actual disease signature used depends on the disease group requiring diagnosis. Signatures can also be constructed using models such as (high wheezing severity score WSS = 5, 6, 7 vs. low wheezing severity score WSS = 0, 1, 2) and clinical signature-based models (e.g., fever, runny nose) in which clinical signatures are mapped to continuous variables between 0 and 1. As an alternative to LRM, any other classifier that maps inputs to continuous decision variables, for example, a neural network, can be used in generating the disease signature block.

[0039] A logistic regression (LR) model is a generalized linear model that uses several independent variables to estimate the probability of a categorical event. The relevant independent variables are mathematical features calculated from cough events, and the categorical events are disease subgroups. Thus, in the above example, an LR model can be trained to predict the probability of a cough belonging to bronchiolitis in relation to pseudomembranous laryngitis disease. The LR model is derived using a regression function to estimate the probability Y given the following independent features:

[0040]

number

[0041] z=β0+β1·q1+β2q2+...+β n-1 q F (2) In (2), β0 is called the intercept, and β1 and β2 are the independent variables q1, q2, ... q F is called the regression coefficient of

[0042] In order to generate the required LR models 2a, ..., 2n of the disease signature generation block 114, it is necessary to train them using the acquired cough sounds. The LRM models 2a, ..., 2n are trained to lead to their ability to accurately classify diseases or measurement classes in the relevant patient subgroups. Once trained, they are used to generate a continuous value probability output signal between (0, 1) for all patients in need of diagnosis.

[0043] Cough sound acquisition Cough sounds were recorded from two clinical sites in Perth, Western Australia, the Joondalup Health Campus (JHC) and Princess Margaret Hospital (PMH). The patient population included children aged 0 to 12 years suspected of having respiratory diseases such as pneumonia, asthma / RAD (reactive airway disease), bronchiolitis, croup, and upper respiratory tract infections (URTIs). The Human Ethics Committees of the University of Queensland, the Joondalup Health Campus, and Princess Margaret Hospital approved the research protocol and patient recruitment procedures.

[0044] Patients who met the inclusion criteria (cough, expiratory wheeze, shortness of breath, inspiratory wheeze, URTI) and did not meet the exclusion criteria (requiring respiratory assistance, without consent) were recruited for the study. Healthy children defined as those without any symptoms of respiratory disease at the time of measurement were also recruited.

[0045] Cough sounds were recorded using an Apple iPhone (registered trademark) 6s (see Box 106 in Figure 1). The audio data was recorded at a bit depth of 16 bits per sample and a sampling rate of fs = 44.1k samples / second. The smartphone recorder was placed at an angle of approximately 45° about 50 cm from the mouth.

[0046] Database and experimental plan The database used to train the various models required to implement the device in Figure 1 consisted of cough records and detailed clinical diagnostic information for each patient, including final diagnosis, clinical examination findings, and laboratory and imaging results. Demographic information was also available in a patient-anonymized format.

[0047] Diagnostic Group (case definitions used in diagnosing the disease are given in Appendix A).

[0048] Normal group (Nr): Healthy volunteers with no identifiable respiratory disease at the time of measurement.

[0049] Early URTI group (U): Patients with only upper respiratory tract infection (URTI) without medically identifiable lower respiratory tract complications or other respiratory diseases at the time of measurement.

[0050] Pseudomembranous laryngitis group (C): Patients with a diagnostic classification of pseudomembranous laryngitis alone or comorbid URTI.

[0051] Asthma / Reactive Airways Disease Group (A): Patients with a diagnostic classification of asthma or reactive airways disease with or without URTI as a comorbidity.

[0052] Clinical pneumonia group (P): patients with a diagnostic classification of clinical pneumonia with or without URTI as a comorbidity.

[0053] Bronchiolitis group (Bo): Patients with a diagnostic classification of bronchiolitis with or without URTI as a comorbidity.

[0054] Bronchitis group (Bc): Patients with a diagnostic classification of bronchitis with or without URTI as a comorbidity.

[0055] The total subjects were divided into two mutually exclusive sets for training, validation, and testing the classifier model. The two mutually exclusive sets were (1) the training-validation set (TrV) and (2) the prospective test set (PT). Each subject belonged to only one set. Subjects with diagnostic uncertainty (as indicated by the clinical team) and comorbid conditions other than URTIs were excluded from the TrV.

[0056] The training-validation set TrV is used to train and validate the model according to leave-one-out validation (LOOV) or K-fold cross-validation techniques. The LOOV method involves using data from all but one patient to train the model and cough events from the remaining patients to validate the model. This process was systematically repeated so that each patient in the TrV was used to validate the model exactly once. In K-fold cross-validation, the original samples were randomly divided into K equal-sized subsamples. A single subsample was retained as validation data for training the model. The remaining (K-1) samples were used to train the model. The process was repeated K times until all data in the TrV had been used once in training the method. Note that LOOV is a special case of K-fold cross-validation, where K is set to the total number of data in the set TrV (N).

[0057] Table 1 (overleaf) lists the various LR models 2a, ..., 2n trained in the disease signature generation block 114 of FIG.

[0058] [Table 1]

[0059] The data in the TrV dataset was used to train and validate all LR models listed in Table 1.

[0060] In the simplest form, only cough-based features are used to train the LR model. However, the inventors recognized the existence of several simple clinical measurements that can be used to improve the performance of the LR model with minimal complexity and without additional cost. Triggered by this, the inventors added simple clinical features to the cough-based features and trained a second set of the LR model list in Table 1 (Table 1). Table 2 (Table 2) shows the simple clinical features with the cough-based features added.

[0061]

Table 2

[0062] Feature Selection Feature selection is a technique for selecting relevant features for designing an optimal disease signature model. For example, when constructing an LR model as a disease signature using the TrV set, a p-value that captures how important a particular feature is to the model is calculated for each input feature. Important features have low p-values. This property of the LR model was used across the entire TrV set to select an appropriate combination of features. When a subset of important features is known, it was used to retrain the LR model based on leave-one-out-of-K (LOOV) training / validation in the TrV dataset. Further details of the feature selection process can be found in the previous PCT application referenced previously, and Abeyratne, U.R., et al., Cough sound analysis can rapidly diagnose childhood pneumonia. Annals of biomedical engineering, 2013. 41(11):2448 - 2462.

[0063] Selecting a Good LR Model The leave-one-out-of-K (LOOV) training / validation process results in N k LR models, where N k is the number of patients in the TrV set. N kThe value of N varies for the different LR models listed in Table 1 due to the different numbers of patients in different disease groups within the TrV set. k From the LR models, one of the best models was selected based on the k-mean clustering algorithm. Further details on using the k-mean clustering algorithm for model selection can be found in [1].

[0064] R c,j Let denote the selected disease signature based on the LR model j trained using only cough features, and R cf,j Let,denote the selected LR model trained using simple clinical features in addition to,cough.,Once the LR model is selected, it is run on all patients in the,TrV dataset to generate a disease signature.

[0065] These disease signatures are then used to calculate the “response” ρ of a given patient i to all hypotheses tested via different disease signature values. c,ij =R c,j (x i ) and ρ cf,ij =R cf,j (x i ) to provide the vector x i is the Model R c,j (x i ) and Model R cf,j (x i ) represents the set of information such as cough sounds and clinical signs of the ith patient.

[0066] As an example, if a patient being tested has bronchiolitis, the patient's response to a model such as {bronchiolitis vs. all other diseases} should give a strong value close to 1 ("complete response"), while {A / RAD vs. all other diseases} should give a lower response ("partial response"). However, patients will give a particular response to each disease signature axis (LRM model), and a vector V representing the collection of such responses for each patient "i" will be created. i,c and V i,cfcharacterizes the disease from which the patient is suffering. V i,c =[{ρ c,ij}|j = all selected disease signatures (LRM models)] (3) V i,cf =[{ρ c,ij}|j = all selected disease signatures (LRM models)] (4) Not all models 2a,...,2n available in signal generation block 114 are necessarily utilized in final disease signature classifier block 118. The selection of disease signatures is performed in disease signature selection block 116, which is configured to take into account domain-specific knowledge and observations about how much a particular disease signature contributes to the final disease diagnostic performance as well as information redundancy between disease signatures. The entries in Table 1 provide a preferred list of such disease signatures selected from a larger collection.

[0067] Classifier Block 118 The function of the classifier block 118 is to generate a vector V i,c and V i,cf The goal is to use a classification algorithm to label the likelihood of each cough recorded from each patient belonging to a particular disease group. An overall diagnosis is then determined for each patient using several different methods. Multiple classifier blocks can be constructed to cover disease groups of interest or to build a single classifier that covers all disease groups. While any classifier can be used in the disease signature classifier block 118, a softmax artificial neural network layer (softmax ANN) was used as the classifier.

[0068] Figure 2 shows a typical structure 200 of the softmax ANN used in block 118 of FIG. 1. For the training and validation of the ANN, data from the DTrV database was used and the LOOV (or k-fold) cross-validation process was continued. The advantage of using the softmax function is that the value of each neuron is in the range between 0 and 1, and the sum of all neurons in the output layer is 1. This makes it a useful function for modeling and predicting the probabilities of disease subgroups. For example, if it is desired to build a model for diagnosing disease Q and it is known that disease Z appears as a confounding disease (e.g., there is some overlap in symptoms), the diagnostic performance can be improved by compiling the composite effective probability measure P Q ' as P' Q = P Q (1 - P Z ) (5) where P Q is treated as an indicator of the probability that a patient belongs to disease Q, and P Z can be treated as an indicator of the probability that a patient belongs to disease Z. Thus, the product P Q (1 - P Z ) indicates the probability of the composite event that the patient belongs to {disease Q and does not belong to disease Z).

[0069] The ANN shown in FIG. 2 represents an artificial neural network architecture having a d-dimensional input layer 201 composed of neurons 203 (inputs received from the disease signature block 114 in FIG. 1) and a k-dimensional output layer 205. Each neuron 207 in the output layer 205 corresponds to a disease subgroup.

[0070] In a preferred embodiment of the present invention, five different softmax ANN models were trained to identify disease subgroups, bronchiolitis, croup, asthma / RAD, pneumonia, and lower respiratory tract disease (LRTD). Details of these disease-specific softmax ANNs are provided below. 1. Softmax ANN for bronchiolitis (S BO): In this softmax ANN model, a disease signature centered on bronchiolitis was used. The sigmoid LR model used to train the ANN was {R BO,R ,R A,BO ,R P,BO ,R BO,BC ,R BO,C ,R BO,U ,R HW,LW}. In the output layer, the dimension of k was set to 6 neurons, and each neuron corresponded to one disease subgroup, bronchiolitis, asthma / RAD, pneumonia, bronchitis, pURTI, croup. In Equation (5), P Q is the output of the bronchiolitis neuron, and P Z is the output of the RAD neuron. 2. Softmax ANN for croup (S C ): In this ANN model, a disease signature centered on croup was used. The sigmoid LR model used was {R C,R ,R A,C ,R P,C ,R B0,C ,R BC,C ,R C,U}. In the output layer, the dimension of k was set to 6 neurons, and each neuron corresponded to one disease subgroup, bronchiolitis, asthma / RAD, pneumonia, bronchitis, pURTI, croup. In Equation (5), P Q is the output of the croup neuron, and P Z is the output of the pURTI neuron. 3. Softmax ANN for asthma / RAD (S A ): In this ANN model, a disease signature centered on asthma / RAD was used. The sigmoid LR model used was {R A,R ,R A,C ,R A,p ,R A,BO ,R A,BC ,R A,U ,R HW,LW}. In the output layer, the dimension of k was set to 6 neurons, each neuron corresponding to one disease subgroup, bronchiolitis, asthma / RAD, pneumonia, bronchitis, pURTI, croup. In Equation (5), P Q is the output of the asthma / RAD neuron, and P Z is the output of the pURTI neuron. 4. Softmax ANN for pneumonia (S P ): In the softmax ANN model for pneumonia, the signature LR model used was {R A,P , R A,BO , R A,C , R P,R , R P,U , R P,C , R P,BO , R P,BC , R BO,R , R C,R , R HW,LW}. In the output layer, the dimension of k was set to 4 neurons, one neuron for each of bronchiolitis, pneumonia, and croup, and one neuron for the diseases of asthma / RAD, bronchitis, pURTI. In Equation (5), P Q is the output of the pneumonia neuron, and P Z is the output of the neuron for the diseases of asthma / RAD, bronchitis, pURTI. 5. Softmax ANN for lower respiratory tract diseases (S LRTD ): Lower respiratory tract disease (LRTD) is an umbrella term used to represent diseases with lower respiratory tract lesions. The LRTD subgroup combines patients from the diseases of pneumonia, asthma / RAD, bronchiolitis, and bronchitis. The LRTD soft ANN model was trained to distinguish LRTD diseases from croup and pURTI. In this ANN, a disease signature centered on LRTD was used. The signature LR model used was {R A,U , R A,C , R P,C , R P,U , R BO,U , R BO,C , R BC,C , R C,R , R C,U,R A,N ,R P,N ,R BO,N ,R BC,N ,R HW,LW In the output layer, the dimension of k was set to three neurons: one neuron each for URTI and pseudomembranous laryngitis, and one neuron for the LRTD disease group. In equation (5), P Q is the output of the LRTD neuron, and P Z is the output of the neuron for pURTI. 6. Softmax ANN for initial URTI (S U ): In this ANN model, a disease signature centered on pURIT was used. The signature LR model used was {R A,U ,R P,U ,R BO,R ,R BO,U ,R BC,U ,R C,R ,R C,U ,R HW,LW ,R P,Nr ,R U,Nr In the output layer, the dimension of k is set to 6 neurons, and each neuron corresponds to one disease subgroup: bronchiolitis, asthma / RAD, pneumonia, bronchitis, pURTI, and pseudomembranous laryngitis. In equation (5), P Q is the output of the pURTI neuron, and P Z is the output of the pneumonia neuron.

[0071] To train the cough-only softmax ANN, an LR signature model developed using only cough features was used. To train the cough plus simple clinical features, an LR signature model developed using only cough features and / or cough plus simple clinical features was used.

[0072] Selecting a preferred softmax ANN model The LOOV cross-validation process results in N softmax models, where N is the number of patients in the TrV dataset. From the N models, the inventors selected one of the best models according to the k-means clustering algorithm again. Further details of using the k-means clustering algorithm for model selection can be found in [1]. S S represents the selected softmax ANN, and λ S is assumed to be the corresponding probability decision threshold for a particular disease subgroup. S S Once is selected, all the parameters of the model are determined and the training process is completely terminated. Model S S is then used as the best model for further testing. In the K-fold cross-validation approach, a similar approach was used.

[0073] Other factors investigated by the inventors when training the softmax ANN were the network size, training epochs, training rate, stopping criteria, and the difference between training and validation errors. These factors were investigated with the aim of minimizing overfitting of the dataset accessed by the inventors and maximizing generalization to a statistically similar but previously unseen population.

[0074] The following describes some preferred embodiments of the method used to arrive at a diagnostic decision for each patient.

[0075] i). Cough index and patient-based classification In this approach, the output of the classifier (e.g., softmax ANN) is directly (e.g., P Q , P Z etc.), or P' QA probability composite measure such as [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 12

[0076] A cough index was then calculated for each of the target diseases as follows: C T is the total number of coughs analyzed from patient i, and C S Let be the number of coughs labeled as "successful" by the softmax ANN model. Then, let cough index Cl for patient i for target disease subgroup j be the number of coughs labeled as "successful" by the softmax ANN model. i,j is Cl i,j =C S / C T It is calculated as:

[0077] ii) Classification based on deep learning strategies The raw output of the softmax layer (which is a real number that varies continuously between 0 and 1) can be further processed in the spirit of deep learning techniques. Some preferred embodiments are given below. a) The raw output of the softmax ANN in block 1183 may be fed to another classifier, such as a neural network, that simultaneously receives input from other similar networks trained with different philosophies. For example, the other network could be based entirely on clinical signatures that parents can easily observe and report to a clinician. The clinical signature network could have a parallel architecture to the disease signature generation block 114 in FIG. 1, or a simpler variant such as a set of selected LRM classifiers (or ANN classifiers, etc.) that map clinical signatures to desired disease groups. b) The LRM-based disease signature generator models 2a, ..., 2n in block 114 of Figure 1 can be replaced by other classifiers, such as ANN layers, making the entire network ANN-based, following the philosophy of deep learning networks. The inventors have constructed and tested such models. Clinical signature models can be incorporated as described above.

[0078] In one preferred embodiment, we used an encoder approach to train a deep learning network. The final classifier itself was a neural network with one neuron representing each disease class of interest.

[0079] D. Test the process In this section, we describe the performance of the final diagnostic model in a prospective data set. Prior to using the prospective data set, we completely freeze our diagnostic models and no further training or parameter adjustments or protocol changes are permitted.

[0080] result A. Dataset Table 3 (Table 3) details the subject population used in this study. For this work, we used coughing data from a total of N=1151 subjects (982 patients and 169 healthy subjects) to develop, validate and test our model. These patients were divided into two non-overlapping data sets: (1) training-validation set (TrV) and (2) prospective test set (PT). Patients were assigned to each set based on the order of presentation to the hospital.

[0081] [Table 3]

[0082] Training-validation dataset: The inventors froze their dataset for model training and validation, having a total of 1011 subjects (852 patients and 159 healthy individuals) recruited from two sites. 600 patients and 134 healthy individuals.

[0083] Prospective test dataset: There were records of a total of 130 patients and 10 healthy individuals from all recording sites JHC (9 from JHC and 1 from PMH).

[0084] Following clinical adjudication and consultation with the clinical team, subjects within the dataset were classified into various diagnostic subgroups, a normal group (Nr), an early URTI group (U), a croup group (C), an asthma / responsive airway disease group (A), a clinical pneumonia group (P), a bronchiolitis group (Bo), and a bronchitis group (Bc).

[0085] B. Performance of the disease signature model in the training-validation dataset Out of the 1011 subjects, 725 subjects (602 patients and 123 healthy individuals) were ultimately used for model training and validation.

[0086]

Table 4

[0087] Table 4 shows the number of patients within each disease subgroup used for the training-validation model.

[0088] LR model for normal vs. disease: First, the inventors investigated the performance of the LR signature model in classifying coughs from healthy individuals and subjects in any disease subgroup. Table 5 shows the leave-one-out validation results of this investigation.

[0089]

Table 5

[0090] It can be seen from Table 5 (Table 5) that after feature selection, all LR models were able to separate diseased coughs from normal coughs with very high accuracy.

[0091] LR models between disease subgroups: The next goal of the inventors is to investigate the performance of the LR signature model when classifying coughs from two different disease groups. This investigation helps to determine how well the LR model captured the disease signatures of the diseases. Table 6(A) (Table 6) shows the leave-one-out validation results of this investigation when all features were used for model training.

[0092] [Table 6]

[0093] The following Table 6(B) (Table 7) shows the results after feature selection.

[0094] [Table 7]

[0095] According to Table 6(B) (Table 7), after feature selection, most of the LR signature models achieved moderate to high accuracy (70 - 90%) when separating coughs from two classes. The highest accuracy was achieved when discriminating croup or bronchiolitis coughs from any other disease or disease group. The least accurate LR signature models were for pneumonia vs. bronchitis and bronchitis vs. pURTI (accuracy ~65%).

[0096] C. Performance of the Softmax Model in the Training-Validation Dataset

[0097] [Table 8]

[0098] The results from Section 3(B) showed that the LR signature model was quite successful in capturing disease-specific disease signatures. Using these disease signatures in Step 3 of Section 2(C), we trained a softmax neural network model to separate the target disease cough from other diseases. Then, using the cough index and applying an optimal threshold, we achieved the ultimate goal of classifying diseases at the patient level.

[0099] Table 7(B) (Table 9) shows the leave-one-out validation results using a softmax ANN model to separate one disease from the rest. According to these results, all models except for pneumonia are able to predict the target disease with very high sensitivity and specificity. The best validation results were obtained for the pseudomembranous laryngitis model, with 100% sensitivity and 96% specificity (using a simple clinical signature model in addition to cough) and 95% sensitivity and 92% specificity (cough-only model). The second-best results were obtained for the bronchiolitis model, followed by the early URTI, asthma, and LRTD models. All models trained using simple clinical features in addition to cough clearly outperformed models trained using cough-only features.

[0100] [Table 9]

[0101] The leave-one-out validation process results in N softmax models, where N is the number of patients in the TrV dataset. From the N models, we again selected one of the best models according to the k-means clustering algorithm. Table 8 shows the performance of the selected ANN models.

[0102] [Table 10]

[0103] From Table 7(B) (Table 9) and Table 8 (Table 10), it can be seen that the pseudomembranous laryngitis, bronchiolitis, URTI, and asthma models are all highly sensitive and specific at the same time. On the other hand, the pneumonia model has moderate specificity and high sensitivity. Therefore, we hypothesize that it is possible to use other disease models within the pneumonia model to eliminate false-positive cases and improve its specificity. To test this hypothesis, we sequentially applied the asthma, early URTI, pseudomembranous laryngitis, and bronchiolitis models as postscreeners in the pneumonia model. To avoid true-positive cases of pneumonia being screened as other diseases, we used a threshold applied to the cough index of the screening model. This threshold indicates how confidently the screening model indicates that the subject does not have pneumonia. The screening threshold was optimized using the training-validation dataset.

[0104] [Table 11]

[0105] Table 9 (above) shows the results of sequentially applying various disease postscreeners to the pneumonia model. It can be seen that increased specificity of the ANN pneumonia model is achieved with a modest decrease in sensitivity. The decrease in sensitivity was -9% for the cough-only model and -4% for the cough plus simple clinical features model. The increase in specificity was 18% for the cough-only model and 16% for the cough plus simple clinical features model. Although not required, postscreeners and prescreeners may be used in embodiments of the invention.

[0106] Encouraged by the positive effect of the post-screeners in the pneumonia model, the inventors investigated their application to other disease models. The inventors' analysis showed that the performance of the specificity of the bronchiolitis and LRTD models could be improved by 1-3% by using the croup screener. In the asthma and early URTI models, no improvement in performance was seen. Since the performance of the croup model is already very high, the screener was not tried. Table 10 (Table 12) shows the results of this investigation.

[0107]

Table 12

[0108] D. Performance of the Softmax Model in the Prospective Test Dataset Figure 3 shows the flowchart 300 of the diagnostic algorithm after all training is completed, all parameters are fixed, and the selected model is ready for the prospective test. These models were tested in a prospective dataset that was completely independent from the training-validation dataset. All models except pneumonia achieved high accuracy in predicting the target disease. Among all the models, croup and bronchiolitis were found to be the best, achieving sensitivity and specificity in the range of 86% - 100% for both the cough-only model and the model with simple clinical features in addition to cough across the entire prospective study list. The performance of the asthma / RAD model was moderate with cough-only features but was significantly improved by the addition of simple clinical features to the cough features (sensitivity of approximately 93% and specificity of approximately 90%). See Appendix E for details of the prescreener and post-screener blocks.

[0109] 1. Appendix A 1. Clinical Pneumonia US Clinical Pneumonia Case Definition - Used when labeling the dataset from JHC and PMH 2. WHO Radiation Main Endpoint Pneumonia (PEP) US WHO Pneumonia Case Definition - Used when labeling X-ray datasets from JHC and PMH 3. Pseudomembranous Laryngitis US Clinical Pseudomembranous Laryngitis Case Definition - Used when labeling datasets from JHC and PMH 4. Bronchiolitis US Bronchiolitis Case Definition - Used when labeling datasets from JHC and PMH 5. Asthma (A) / Reactive Airway Disease (RAD) - A / RAD US A / RAD Case Definition - Used when labeling datasets from JHC and PMH 6. Bronchitis US Bronchitis Case Definition - Used when labeling datasets from JHC and PMH 7. Upper Respiratory Tract Infection US URTI Case Definition - Used when labeling datasets from JHC and PMH 8. Lower Respiratory Tract Disease (LRTD) US LRTD Case Definition - Used when labeling datasets from JHC and PMH

[0110] 2. Appendix B Calculate characteristics from cough Cough characteristics The method of the present inventors requires the calculation of several mathematical characteristics from the cough sound. This section describes the characteristics calculated from each sub-segment x of the cough sound x recorded by the present inventors i where i = 1, 2, 3. i) Bispectral Frequency Cepstral Coefficients (BFCC, a total of 24 characteristics, 8 from each part of the cough segment) - The third-order spectrum of the signal is known as the bispectrum [3]. Unlike the power spectrum (second-order spectrum based on autocorrelation), the bispectrum preserves Fourier phase information. The bispectrum B i of segment x xi (ω1, ω2) is from (6)

[0111]

number

[0112] where w(τ1,τ2) is a bispectral window function, such as the minimum bispectral bias upper window used in this paper, and Cx i (τ1,τ2) is the x estimated using (2) i where ω1 and ω2 indicate the digital frequencies.

[0113]

number

[0114] In (7), Q is the length of the considered third-order correlation delay, and x i is a zero-mean signal.

[0115] The bispectrum is a 2D signal. However, for linear signals, any 1D tilted slice of the bispectrum other than the slices parallel to the axis, ω1=0, ω2=0, and ω1+ω2=0, can be proven to carry sufficient information to characterize the entire 2D bispectrum in phase factors. In this work, we consider the diagonal slice P(ω) defined by ω1=ω2=ω, i.e., P(ω)=Bx i We capture the information available in the bispectrum via (ω,ω).

[0116] Then, the filter operator is applied to the diagonal slice P(ω) and we compute the bispectral frequency cepstral coefficients using (8).

[0117]

number

[0118] In (8), ζ represents a filter operator having lower and upper cut-off frequencies ω l and ω h and g is the gain constant of the filter. In (8), ζ can be a triangular filter, a rectangular filter, a trapezoidal filter, or a filter of a more complex shape. For the work of this document, the inventors used a rectangular filter with g = 1 and, as shown in Table 11 (Table 13), calculated eight BFCC coefficients using the following values of ω l and ω h .

[0119] [Table 13]

[0120] ii) Non-Gaussianity score (NGS, a total of three features, one from each part of the cough segment) - The NGS score is a numerical measure of the non-Gaussianity of a given segment of the data x i . The normal probability plot can be used to obtain a visual measure of the Gaussianity of a set of data, and the NGS score is a method of quantifying non-Gaussianity based on regression analysis. The inventors used (9) to estimate the NGS score, where p and q represent the normal probability plots of the reference normal data and the analysis data (x i ). The symbol N is the number of data points used in the probability plot.

[0121] [Number]

[0122] iii) Formant Frequencies (12 features in total, 4 from each part of the cough segment) - In speech analysis, formant frequencies (FF) are referred to as the resonances of the vocal tract. In the analysis of a cough, it is reasonable to expect that the resonances of the entire airway that contribute to the production of the cough sound will be represented in the formant structure. One classic example for this is expiratory wheezing. The presence of mucus can also change the acoustic properties of the airway. We included the first four formants (F1, F2, F3, F4) in our candidate feature set. We found that the cough segment x i F1–F4 were calculated by peak-picking the linear predictive coding (LPC) spectrum of the signal. For this work, we used a 14th-order LPC model with parameters determined via the Levinson-Durbin recursive procedure.

[0123] iv) Logarithmic Energy (LogE, 3 features in total, one from each part of the cough segment) - for all sub-segments x i The logarithmic energy for was calculated using (10).

[0124]

number

[0125] In (4), ε is an arbitrary small positive constant added to prevent inadvertent calculation of the logarithm of 0.

[0126] v) Zero Crossings (Zcr, 3 features in total, one from each part of the cough segment) - the number of zero crossings is determined for each sub-segment x i was counted.

[0127] vi) Kurtosis (Kurt, 3 features in total, one from each part of the cough segment) - Kurtosis is the x i is a measure of how peaky the probability density distribution of x is. iis the fourth central moment and can be calculated using (11), where μ and σ denote the mean and standard deviation of x i respectively.

[0128] [Number]

[0129] vii) Mel-frequency cepstral coefficients (MFCCs, a total of 93 features, 31 from each part of the cough segment) - MFCCs are widely used in speech recognition systems. MFCCs provide some resilience to non-linguistic sources of variance in the speech signal. They also provide the orthogonality feature that facilitates the training of classifiers. The calculation of MFCCs involves estimating the short-term power spectrum, mapping it to the Mel-frequency scale, and then calculating the cepstral coefficients. In the work of the present inventors, the present inventors included 31 MFCC coefficients in their feature set.

[0130] viii) Shannon entropy (ShE, a total of 3 features, 1 from each part of the cough segment): Cough sounds are complex signals representing contributions from various substructures of the airway. Some of these components exhibit quasi-periodic structures, while others exhibit random probabilistic characteristics. In this work, the present inventors calculated the Shannon entropy to capture these features. The Shannon entropy (ShE) of all subsegments x i was calculated using (12).

[0131] [Number]

[0132] ix) Wavelet features (WvL, a total of 13 features from each cough): The inventors' research has shown the usefulness of wavelet features in the diagnosis of pneumonia [see IEEE conference papers]. For this work, the inventors calculated 13 wavelet features from each cough segment. For details, see [2].

[0133] 3. Appendix C 1. Expiratory wheeze signature generator Expiratory wheeze in children is a common symptom of many respiratory diseases. Expiratory wheeze is defined as a high-pitched whistling sound generated during breathing. Expiratory wheeze is most commonly associated with asthma, but also exists in other respiratory diseases such as bronchiolitis, bronchitis, pneumonia, cystic fibrosis, and foreign body aspiration. It is often used in differential diagnosis and in separating lower respiratory tract diseases from upper respiratory tract infections. More details are available in Appendix A: Case Definitions.

[0134] The presence or absence of expiratory wheeze is an important decision node in the clinical decision tree, as practiced by the clinical community. However, clinically detecting it is not always an easy task. Expiratory wheeze is a qualitative phenomenon and a secondary effect of fundamental changes in physiology / pathology. The ability of a physician to detect expiratory wheeze at a specific time during an examination depends on many factors, including that expiratory wheeze is present at that time, the physician places the stethoscope at the appropriate location over the lungs, the expiratory wheeze sound is generated with sufficient intensity to overcome energy losses during its propagation from the lungs to the body surface, and the clinician has the skill to perceive and detect the sound. The underlying physiological reason is airway narrowing due to various reasons, and expiratory wheeze is a surrogate measure of that phenomenon. In some situations, expiratory wheeze may not occur even in severe diseases due to airflow limitation according to the severity of the disease (e.g., "silent chest" in severe asthma / RAD).

[0135] To capture the severity of expiratory wheezing, clinicians have defined many different versions of the Wheezing Severity Score (WSS). The version of our clinical collaborators uses three different subscores to calculate the WSS. These are the presence of expiratory wheezing and the respiratory phase in which it occurs, respiratory rate, and use of accessory muscles. We developed one disease signature to capture the WSS using only cough or cough augmented with simple signs that a parent can observe. Our WSS model was trained with the goal of separating high WSS (5 - 9) from low WSS scores (0, 1) using a continuous signature scale (LRM output) that varies between 0 and 1.

[0136] 2. Lung Function Signature Generator Lung function testing techniques, particularly spirometry, are used in the definitive diagnosis of some respiratory diseases such as asthma and chronic obstructive pulmonary disease (COPD) when available. Spirometry provides numerical measurements such as FEV1 and FVC. Using coughs collected during spirometry, we built cough-based disease signature models such as high FEV1 vs low FEV1, high FEV1 / FVC vs low FEV1 / FVC. Once trained, those disease signature models are then used in all patients as disease signature generators that provide an output between (0, 1).

[0137] 4. Appendix D Clinical Sign-Based Disease Signatures and Diagnostic Models Clinicians rely heavily on clinical signs (observed by them or reported by parents) when diagnosing some respiratory diseases. Previously, we investigated and built models to identify the best clinical features (Indonesia study) for diagnosing pneumonia based on clinical signs alone. We also investigated what happens when a small number of coughs are added to the clinical sign model.

[0138] The inventors constructed a model of only clinical signs for diagnosing respiratory diseases (using signs that can be reported by parents) according to the architecture shown in FIG. B. In certain embodiments, signs such as fever, expiratory wheeze, rhinorrhea, and age, gender, etc. are used to construct a disease signature based on LRM modeling. The LRM model converts the signs by category into continuous outputs at the disease signature generator level, and these continuous outputs are then classified at the classifier block level. In other embodiments, other classifier methods such as ANN can be used for the same purpose. The inventors can also add cough to the clinical sign model to improve performance, as the inventors did in the patent of clinical signs submitted in the Indonesia study. Further, the inventors can use the clinical sign signature to improve classification using deep learning techniques.

[0139] The process of the neural deep learning architecture is separated into two stages discussed here.

[0140] Stage 1 - Training layer by layer - In the first stage, the inventors trained three types of neural networks individually to perform a specific task.

[0141] Stage 1 - Neural network type 1. Feature Encoding Neural Network (FeNN): Here, the inventors implement the concept of an autoencoder. An autoencoder is a feed - forward neural network trained to reproduce its input at the output. The hidden layer in the autoencoder encodes a code that can be used to represent the input data. After training, the encoder stage output is used in the next step. FIG. 5 shows an example of an autoencoder trained for feature mapping. It has 144 input sizes representing the size of the input feature vector, the hidden layer (encoder) size is 10, and the output layer (decoder) size is 144, which is the same size as the input layer.

[0142] Stage 1 - Neural Net Type 2. Signature Neural Net (SgNN): The output from the encoder is used to train a feed-forward neural network to generate a disease signature as described in the section "Signature Generator Block". This signature ANN does not have softmax neurons in the hidden and output layers. The advantage of using softmax neurons is that it results in a probability function that varies between 0 and 1, similar to the LR model, but here the mapping from input to output is non-linear. Figure 6 shows an example of a disease signature neural net. The input to this network is from the encoder. This neural network is trained to generate a disease-specific disease signature. For all 26 disease signatures, the neural net was trained for all the disease signature models listed in Table 1. Table 12 shows the leave-one-out validation results when classifying a disease cough from a normal cough using the signature neural net model. Table 13 shows the leave-one-out validation performance of the signature neural net model.

[0143] Stage 1 - Neural Net Type 3. Classification Neural Net (CaNN): The output of the disease signature neural net was then used to train a final classification softmax neural net. This neural network is similar to the one described with reference to classifier block 118.

[0144] In a preferred embodiment, the inventors trained a single softmax ANN model to identify all target disease subgroups, bronchiolitis, croup, asthma / RAD, pneumonia, and lower respiratory tract diseases. The signature neural net used to train the CaNN was {R Hw,Lw 、R A,R 、R A,C 、R A,P 、RA,Bo , R A,Bc , R A,U , R A,Nr , R P,Nr , R Bo,Nr , R BC,Nr , R C,Nr , R U,Nr} It was. In the output layer, the dimension of k was set to 7 neurons, and each neuron corresponded to one disease subgroup, bronchiolitis, asthma / RAD, pneumonia, bronchitis, pURTI, croup, and normal.

[0145] Stage 2 - Fine - tuning stage. In this stage, the individually trained neural networks from Stage 1 were connected to each other to create the stacked deep neural network (DNN) shown in FIG. 7. FIG. 7 shows a DNN where the first layer represents a feature - encoding neural net, the second layer represents a disease - signature neural net, and the third output layer represents a classification neural net.

[0146] The DNN was then retrained using the training - validation dataset with a limited number of training epochs to fine - tune the DNN network parameters. The fine - tuning of the DNN was performed according to the leave - one - out validation technique.

[0147] Table 16 (Table 18) represents the leave - one - out validation results for classifying patients using the DNN model.

[0148] 5. Appendix E Prescreener and Post - screener For a further embodiment of the present invention, reference is made to the block diagram of the diagnostic model 300 shown over two pages in FIGS. 4 and 5. The diagnostic model 300 includes a disease - signature generation block 114 and a classifier block 118 of the first embodiment 100 of FIG. 1 (see FIG. 1), which may be called a primary model. However, the second embodiment of the diagnostic model 300 also includes a prescreener block 111 and a post - screener block 121 to improve the diagnostic performance of the entire diagnostic method.

[0149] i) Pre - screener block 111 The function of the pre - screener block 111 is to screen out subjects who are not intended for analysis by the primary model.

[0150] As an example, consider a situation where the task of the primary model is defined as diagnosing a specific disease, for example, bronchiolitis, from a mixture of other given diseases in a population of subjects who visit a medical facility. The result of the entire diagnostic algorithm is either "Bronchiolitis yes / no? = no" or "Bronchiolitis yes / no? = yes". The pre - screener 111 in this situation can be designed to separate healthy individuals from bronchiolitis cases and report the result as "Bronchiolitis yes / no? = no" without the need to send the cases to the primary model for further analysis.

[0151] Each of the screener models 4a,..., 4n in the screener block 111 has a high - set decision threshold to ensure that a person with the actual disease targeted in the primary model is not wrongly screened out of further analysis by obtaining a non - disease label.

[0152] ii) Post - screener block 121 The function of the post - screener block 121 is to target the detection of dominant false positives in the primary model and improve the diagnostic performance of the primary model by making corrections.

[0153] As an example, consider a situation where the task of the primary model is defined as diagnosing a specific disease, e.g., bronchiolitis, from a mixture of other given diseases in a population of subjects who visit a medical facility. The result of the entire diagnostic algorithm is either "Bronchiolitis yes / no? = no" or "Bronchiolitis yes / no? = yes". Assume that the inventors know that croup patients are present as a dominant false positive group in the primary model. That is, in the "Bronchiolitis yes / no? = yes" group, the inventors find a significant number of subjects with a clinical diagnosis of croup. In this scenario, the inventors' approach is to construct a post - screener model {croup vs. bronchiolitis} that is trained to select croup patients from a mixture of croup and bronchiolitis subjects. The inventors then process the "Bronchiolitis yes / no? = yes" group with their {croup vs. bronchiolitis} model and move the detected croup patients to the "Bronchiolitis yes / no? = no" side of the primary model output.

[0154] Depending on the need and effectiveness, it is possible to apply multiple post - screener models 6a,..., 6m to a given primary model. There may be cases where a post - screener is not useful or necessary in a given primary model. When it is necessary, a decision threshold is set high to ensure that people with the actual disease targeted in the primary model do not accidentally move to the opposite side of the diagnostic decision, and the screener model is used conservatively.

[0155] 6. Appendix F Feature Reduction in Model Development The diagnostic models discussed herein were developed using cross-validation (CV) methods on available clinical datasets. For available data, K-fold cross-validation (K=10) and leave-one-out validation (LOOV) methods were used. Both of these methods have their advantages and disadvantages. 10-fold CV tends to lead to models with low variance and high bias for new datasets (generalization performance in previously unseen datasets), while LOOV tends to lead to models with high variance and low bias for new datasets.

[0156] Due to the smaller size of the available datasets, the LOOV method was often preferred over 10-fold splitting when developing diagnostic models. To compensate for the higher generalization variance in the models, a feature reduction process was developed that aims to build models that are as small as possible, leading to a lower risk of overfitting the model. The procedure is described below.

[0157] Feature optimization / reduction is a technique for selecting a subset of relevant features to build a robust classifier. Optimal feature selection requires exhaustive exploration of all possible subsets of features. However, doing so for the large number of features we use as candidate features is impractical. Therefore, an alternative approach based on p-values was used to determine disease signature features in the disease signature model design phase using a logistic regression model (LRM). In LRM design, a p-value is calculated for each feature, indicating how important that feature is to the model. Important features have low p-values, and this property of LRM was used to determine the optimal combination of features that facilitate classification in the model during the training phase.

[0158] The approach taken is to calculate the average p-value for all features across the entire dataset and then apply a threshold P minIt is composed of selecting features with an average p-value smaller than [a certain value]. The detailed method will be described in the following steps. 1.F N =[F1,F2,F3,...,F n represents the initial set of all N features. 2.F N is used to train a logistic regression model (LRM) according to the leave-one-out validation (or K-fold) process. Calculate the average p-value for all features. P N =[P f1 ,P f2 ,P f3 ,...,P fn (1) (1) In this case, P N represents the set of average values related to the initial set of all N features. 3. By selecting features whose average p-value is less than P ths =P O , a new subset F' N of features is created from F N . 4. According to the leave-one-out validation (or K-fold) process, use the set of features F' N to train the LRM. Calculate the average p-value for the set of features F' N . 5. By selecting features whose average p-value is less than P ths , a new subset F" N of features is created from F' N . 6. If the size of F" N is equal to that of F' N , that is, if all features in F' N have an average p-value less than P ths , use equation (2) to change P ths . P ths =P ths -c (2) Here, c << P ths . 7.F" NSet size is F' N Repeat step 6 until you have less than the set. 8.F" N Feature size is F min until it is smaller than ths Reset to its original value and repeat steps 4-7. 9. For every subset of the selected feature model, the performance (sensitivity, specificity, and kappa value) is calculated. To select the optimal subset of features, we followed the following scheme. a. Select the subset of features that has the highest kappa value in model performance. This subset is called F a Let's say. b. where F is the function whose performance (in terms of sensitivity / specificity) is characterized. a Identify all subsets of features that are within q% of the model performance by the subset. c. From this pool of identified subsets of features, select the subset whose size is smallest but larger than Z and whose size is F a If this is not possible, select a subset of F as the optimal feature subset. a Select .

[0159] In one particular embodiment, the following parameter values were used in the algorithm given above: F for the cough-only model N The size of F is 157 (i.e., 157 input features to the LRM were started). N The size of the sigma is 157 or more, depending on the number of clinical features used. P ths (in step 3) = PO = 0.20 c (in step 6) = 0.001 F min (in step 8) = 10 Z (in step 9) = 10

[0160] 7. In one particular embodiment, the following parameter values were used in the algorithm given above: The results obtained with the method are shown in Table 12 (Table 14) and Table 13 (Table 15) below.

[0161] [Table 14]

[0162] [Table 15]

[0163] [Table 16]

[0164] [Table 17]

[0165] [Table 18]

[0166] References The following documents are incorporated herein by reference in their entirety: 1. Abeyratne, UR, et al. Cough sound analysis can rapidly diagnose childhood pneumonia. Annals of biomedical engineering, 2013. 41(11): 2448-2462. 2. Kosasih, K., et al., Wavelet augmented cough analysis for rapid childhood pneumonia diagnosis. IEEE Transactions on Biomedical Engineering, 2015. 62(4): pp. 1185 - 1194 3. Abeyratne, U. Blind reconstruction of non - minimum - phase systems from 1 - D oblique slices of bispectrum. 1999. IET. 4. Hinton GE, Salakhutdinov RR. Reducing the dimensionality of data with neural networks science. July 28, 2006; 313(5786): 504 - 7

[0167] In accordance with the law, the present invention has been described in a language that is more or less specific to structural or methodological features. The term "comprising", and its variants such as "comprises" and "consisting of", are used throughout in an inclusive sense and not for the exclusion of any additional features. Since the means described herein comprise preferred forms of carrying out the present invention, it should be understood that this specification is not limited to the specific features illustrated or described. Accordingly, this specification claims in either that form or a modification thereof within the proper scope of the appended claims as appropriately construed by those skilled in the art.

[0168] Throughout the specification and the claims (if any), unless the context requires otherwise, the terms "substantially" or "about" are to be understood not to be limited to the values of the ranges modified by the terms.

[0169] Features, integers, characteristics, compounds, chemical moieties, or chemical groups described in connection with a particular aspect, embodiment, or example of the present invention are to be understood as applicable to any other aspect, embodiment, or example described herein, unless inconsistent therewith.

[0170] Any embodiment of the present invention is meant to be merely illustrative and not meant to be a limitation on the present invention. Thus, it should be understood that various other changes and modifications can be made to any of the embodiments described without departing from the spirit and scope of the present invention.

Explanation of Reference Numerals

[0171] 2a to 2n Decision Machine, LR Model, LRM Model, Model 100 Flowchart 101 Patient 102 Patient Voice 104 Microphone 106 Box 108 Segmentation Block 110a CG1 Cough Sound 110b CG2 Cough Sound 111 Prescreener Block, Prescreener, Screener Block 112 Cough Feature Calculation Block, Cough Feature Calculation Box, Box, 114 Disease Signature Generation, Disease Signature Generation Block, Box, Signal Generator Block, Disease Signature Block, Block 116 Disease Signature Selection Block 118 Disease Signature Classifier Block 120 Diagnosis Instruction Block, Output Box 121 Post - screener Block 200 Structure 201 d - dimensional Input Layer 203 Neuron, Output Layer 205 k - dimensional Output Layer, Output Layer 207 Neuron

Claims

1. A method for diagnosing one or more diseases of a patient's airway by a computer, comprising: obtaining, by the computer, a recording of the patient's cough sound; processing, by the computer, the recording of the cough sound to generate cough segments, and obtaining a cough sound feature signal representing one or more cough sound features from the cough segments; outputting, by the computer, a plurality of signals representing disease signatures generated by applying the cough sound feature signal to each of one or more pre-trained disease signature determination machines; selecting, by the computer, one or more disease signatures, and outputting a signal indicating the selected one or more disease signatures from the plurality of signals; inputting, by the computer, a signal indicating the selected one or more disease signatures, classifying the cough segments into diseases corresponding to the selected one or more disease signatures, and determining that the cough segments indicate one or more of the diseases; wherein each of the disease signature determination machines is pre-trained to identify the cough sound feature signal as corresponding to either a specific disease or a non-disease state, and is pre-trained to identify the cough sound feature signal as corresponding to a first specific disease or a second specific disease different from the first specific disease.

2. The method according to claim 1, wherein each of the one or more disease signature determination machines comprises a trained logistic regression model (LRM).

3. The method according to claim 2, wherein each trained logistic regression model (LRM) is a set of training features that are features determined to be important for the LRM, and is trained using a reduced set, thereby avoiding overfitting of the LRM.

4. The method according to claim 3, wherein the training features are determined to be important training features for the LRM by calculating an average p-value for all training features and then selecting the training features having an average p-value smaller than a threshold P min ​

5. The variables of each LRM are values of the cough sound feature signal, and the output value from the LRM is when the cough sound feature signal corresponds to either the specific disease or the non-disease state, and when it corresponds to the non-disease state, a predicted probability of a cough indicating the first specific disease. In the case where the cough sound feature signal corresponds to the first specific disease or the second specific disease, and in the case where it corresponds to the second specific disease, it includes the predicted probability of cough indicating the first specific disease. The method according to any one of claims 2 to 4.

6. The method according to claim 1, wherein the one or more disease signature determination machines include one or more trained neural networks.

7. The method according to any one of claims 1 to 6, including the step of applying clinical patient measurement values as variables to the disease signature determination machine in addition to the cough sound feature signal.

8. The step of outputting the plurality of signals representing the disease signatures generated by applying the cough sound feature signal to each of the one or more pre-trained disease signature determination machines by the computer includes applying the one or more cough sound feature signals to a single classifier trained to cover all disease groups. The method according to any one of claims 1 to 7.

9. The step of outputting the plurality of signals representing the disease signatures generated by applying the cough sound feature signal to each of the one or more pre-trained disease signature determination machines by the computer includes applying the one or more cough sound feature signals to a number of classifiers each trained to recognize a disease of interest. The method according to any one of claims 1 to 7. **Claim 10**: The step of inputting, by the computer, a signal indicating the selected one or more disease signatures, classifying the cough segment as indicating one or more of the diseases corresponding to the selected one or more disease signatures, and determining that the cough segment indicates one or more of the diseases, wherein the step is bronchiolitis (S BO ), croup (S C ), asthma / RAD (S A ), pneumonia (S P ), lower respiratory tract disease (S LRTD ), early URTI (S u ), and classifying the cough sound feature signal as one of the diseases, the method according to claim 8 or 9.

11. d disease signatures are selected by the computer by the step of selecting one or more disease signatures and outputting the signals indicating the selected one or more disease signatures from the plurality of signals, and correspondingly, the artificial neural network (ANN) used as the number of classifiers has a d-dimensional input layer having one input neuron corresponding to each of the disease signatures. The method according to claim 9.

12. The method according to claim 11, wherein the ANN has a k-dimensional output layer, and each neuron in the k-dimensional output layer outputs the probability corresponding to the disease.

13. The step of inputting, by the computer, a signal indicating the selected one or more disease signatures, classifying the cough segment into a disease corresponding to the selected one or more disease signatures, and determining that the cough segment indicates one or more of the diseases, is a composite effective probability measure P Q ' being P' Q =P Q (1 - P Z ) including the step of compiling as, where P Q comprises an indicator of the probability that the patient belongs to disease Q, P Z comprises an indicator of the probability that the patient belongs to disease Z, whereby the product P Q (1 - P Z ) indicates the probability of the compound event that the patient belongs to disease Q and does not belong to disease Z, the method according to any one of claims 1 to 12.

14. The method according to claim 10, comprising the step of calculating a cough index for each of the target diseases, wherein the cough index of the patient for the target disease is calculated as a ratio of the coughs of the patient classified as indicative of the target disease to the total number of coughs analyzed for the patient.

15. The method according to any one of claims 1 to 14, comprising the step of applying one or more post-screening tests to the selection from the step of selecting the one or more disease signatures to detect dominant false positives.

16. The method according to claim 15, further comprising the step of adjusting the step of selecting the one or more disease signatures based on the detected false positives.

17. The method according to any one of claims 1 to 16, further comprising the step of presenting an instruction regarding disease diagnosis based on the diagnosed specific disease.

Images