Method, electronic apparatus, and computer readable medium of constructing classifier for disease detection

1. A method of constructing a classifier for disease detection, adapted to an electronic apparatus, comprising:

constructing a codebook of representative features based on a plurality of disease-irrelevant data, wherein the disease-irrelevant data is non-medical data;

extracting a plurality of transfer-learned disease features from a plurality of disease-relevant bio-signals according to the codebook, wherein both the disease-irrelevant data and the disease-relevant bio-signals are time-series data, and the disease-irrelevant data comprises a plurality of audio signals; and

performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection,

wherein the steps of extracting a plurality of transfer-learned disease features from the plurality of disease-relevant bio-signals according to the codebook and performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection comprise:

extracting a plurality of first transfer-learned disease features from a plurality of heart sound signals according to the codebook;

extracting a plurality of second transfer-learned disease features from a plurality of electrocardiac (ECG) signals according to the codebook;

concatenating the first transfer-learned disease features and the second transfer-learned disease features to form a plurality of fused feature vectors; and

performing supervised learning on the fused feature vectors to train the classifier for disease detection.

2. The method according to claim 1 , wherein the step of constructing the codebook of the representative features based on the disease-irrelevant data comprises:

constructing the codebook of the representative features based on the disease-irrelevant data by using a first representation learning model, wherein the first representation learning model comprises a plurality of layers of neurons, wherein the layers comprise at least one inner layer, and wherein activation values of the inner layers corresponds to the representative features of the disease-irrelevant data.

3. The method according to claim 2 , wherein the step of extracting the transfer-learned disease features from the disease-relevant bio-signals according to the codebook comprises:

for each of the bio-signals:

inputting the disease-relevant bio-signal into the first representation learning model;

propagating information in the disease-relevant bio-signal through the layers; and

obtaining the transfer-learned disease features of the disease-relevant bio-signal according to at least one of the inner layers.

4. The method according to claim 1 , wherein the step of constructing the codebook of the representative features based on the disease-irrelevant data comprises:

constructing the codebook of the representative features based on the disease-irrelevant data by using a second representation learning model, wherein the second representation learning model comprises a plurality of layers of units, wherein the layers comprise at least one hidden layer having a plurality of hidden units, and wherein probabilistic values of the hidden units correspond to the representative features of the disease-irrelevant data.

5. The method according to claim 4 , wherein the step of extracting the transfer-learned disease features from the disease-relevant bio-signals according to the codebook comprises:

for each of the bio-signals:

inputting the disease-relevant bio-signal into the second representation learning model;

propagating information in the disease-relevant bio-signal through the layers; and

obtaining the transfer-learned disease features of the disease-relevant bio-signal according to at least one of the hidden units.

6. The method according to claim 1 , further comprises:

pre-processing a plurality of raw disease-relevant bio-signals by one or a combination of filtering, segmentation, frequency-domain transformation and time-frequency transformation to obtain pre-processed bio-signals; and

extracting a plurality of heuristic features from the pre-processed bio-signals.

7. The method according to claim 6 , wherein the step of performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection, the method further comprises:

concatenating the transfer-learned disease features and the heuristic features to form a plurality of fused feature vectors corresponding to the disease-relevant bio-signals; and

performing supervised learning on the fused feature vectors to train the classifier for disease detection.

8. The method according to claim 6 , the step of performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection further comprises:

performing supervised learning on the transfer-learned disease features to train a first classifier;

performing supervised learning on the heuristic features to train a second classifier;

inputting the transfer-learned disease features and the heuristic features of the disease-relevant bio-signals respectively into the first classifier and the second classifier; and

concatenating outputs of the first classifier and the second classifier corresponding to each of the disease-relevant bio-signals to generate a corresponding concatenated output; and

performing supervised learning on the concatenated outputs to train the classifier for disease detection.

9. The method according to claim 1 , wherein after the step of performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection, the method further comprises:

obtaining a new disease-relevant bio-signal; and

classifying the new disease-relevant bio-signal by using the classifier for disease detection to generate a classification result; and

outputting the classification result.

10. An electronic apparatus, comprising:

a communication device;

a storage device, recording a plurality of modules; and

a processor, coupled to the communication device and the storage device, and accessing and executing the modules stored in the storage device, wherein the modules comprise:

a codebook construction module, constructing a codebook of representative features based on a plurality of disease-irrelevant data obtained via the communication device, wherein the disease-irrelevant data is non-medical data;

a feature extraction module, extracting a plurality of transfer-learned disease features from a plurality of disease-relevant bio-signals obtained from at least one bio-sensing device via the communication device according to the codebook, wherein both the disease-irrelevant data and the disease-relevant bio-signals are time-series data, and the disease-irrelevant data comprises a plurality of audio signals, wherein the feature extraction module extracts a plurality of first transfer-learned disease features from a plurality of heart sound signals according to the codebook and extracts a plurality of second transfer-learned disease features from a plurality of electrocardiac (ECG) signals according to the codebook;

a feature classification module, performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection; and

a feature fusing module, concatenating the first transfer-learned disease features and the second transfer-learned disease features to form a plurality of fused feature vectors, wherein the feature classification module performs supervised learning on the fused feature vectors to train the classifier for disease detection.

11. The electronic apparatus according to claim 10 , wherein the codebook construction module constructs the codebook of the representative features based on the disease-irrelevant data by using a first representation learning model, wherein the first representation learning model comprises a plurality of layers of neurons, wherein the layers comprise at least one inner layer, and wherein activation values of the inner layers corresponds to the representative features of the disease-irrelevant data.

12. The electronic apparatus according to claim 11 , wherein for each of the bio-signals, the feature extraction module inputs inputting the disease-relevant bio-signal into the first representation learning model, propagates information in the bio-signal through the layers, and obtains the transfer-learned disease features of the disease-relevant bio-signal according to at least one of the inner layers.

13. The electronic apparatus according to claim 10 , wherein the codebook construction module constructs the codebook of the representative features based on the disease-irrelevant data by using a second representation learning model, wherein the second representation learning model comprises a plurality of layers of units, wherein the layers comprise at least one hidden layer having a plurality of hidden units, and wherein probabilistic values of the hidden units correspond to the representative features of the disease-irrelevant data, wherein

for each of the bio-signals, the feature extraction module inputs the disease-relevant bio-signal into the second representation learning model, propagates information in the disease-relevant bio-signal through the layers, and obtains the transfer-learned disease features of the bio-signal according to at least one of the hidden units.

14. The electronic apparatus according to claim 10 , wherein the modules further comprise:

a pre-processing module, performing one or a combination of filtering, segmentation, frequency-domain transformation and time-frequency transformation on a plurality of raw disease-irrelevant data and raw disease-relevant bio-signals to obtain pre-processed disease-irrelevant data and pre-processed bio-signals, wherein

the codebook construction module constructs the codebook of representative features based on the pre-processed disease-irrelevant data; and

the feature extraction module extracts a plurality of heuristic features from the pre-processed bio-signals.

15. The electronic apparatus according to claim 14 , wherein the modules further comprise:

a feature fusing module, concatenating the transfer-learned disease features and the heuristic features to form a plurality of fused feature vectors corresponding to the disease-relevant bio-signals, wherein the feature classification module performs supervised learning on the fused feature vectors to train the classifier for disease detection.

16. The electronic apparatus according to claim 14 , wherein

the feature classification module performs supervised learning on the transfer-learned disease features to train a first classifier, performs supervised learning on the heuristic features to train a second classifier, and inputs the transfer-learned disease features and the heuristic features of the disease-relevant bio-signals respectively into the first classifier and the second classifier,

wherein the modules further comprise:

a feature fusing module, concatenating outputs of the first classifier and the second classifier corresponding to each of the disease-relevant bio-signals to generate a corresponding concatenated output, wherein the feature classification module performs supervised learning on the concatenated outputs to train the classifier for disease detection.

17. The electronic apparatus according to claim 10 , wherein the feature classification module obtains a new disease-relevant bio-signal from the bio-sensing device by the communication device, classifies the new disease-relevant bio-signal by using the classifier for disease detection to generate a classification result, and outputs the classification result.

18. A non-transitory computer readable medium, storing programs to be loaded into an electronic apparatus to perform steps of:

constructing a codebook of representative features based on a plurality of disease-irrelevant data, wherein the disease-irrelevant data is non-medical data;

extracting a plurality of transfer-learned disease features from a plurality of disease-relevant bio-signals according to the codebook, wherein both the disease-irrelevant data and the disease-relevant bio-signals are time-series data, and the disease-irrelevant data comprises a plurality of audio signals; and

performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection,

wherein the steps of extracting a plurality of transfer-learned disease features from the plurality of disease-relevant bio-signals according to the codebook and performing supervised learning based on the transfer-learned disease features to train the classifier for disease detection comprise:

extracting a plurality of first transfer-learned disease features from a plurality of heart sound signals according to the codebook;

extracting a plurality of second transfer-learned disease features from a plurality of electrocardiac (ECG) signals according to the codebook;

concatenating the first transfer-learned disease features and the second transfer-learned disease features to form a plurality of fused feature vectors; and

performing supervised learning on the fused feature vectors to train the classifier for disease detection.