System and method for physiological monitoring and feature set optimization for classification of physiological signal

A system for physiological monitoring, the system comprising:

• one or more memories storing instructions;
• and one or more hardware processors coupled to the one or more memories, wherein said one or more hardware processors are configured by said instructions to:
• identify a clean physiological signal training set from an input physiological signal based on a Dynamic Time Warping (DTW) of a plurality of segments associated with the physiological signal, wherein the input physiological signal is received from one or more sensors that are capable of monitoring subject's physiological health including smartphone inbuild sensors, a digital stethoscope, and an acoustic sensor, wherein the input physiological signal comprises a physiological audio signal obtained from a subject and the physiological audio signal includes a phonocardiogram (PCG) signal, wherein to identify the clean physiological signal training set, the one or more hardware processors are further configured by the instructions to:
• segment the input physiological signal into a plurality of segments;
• identify a clean template segment from the plurality of segments, the clean template segment associated with a noise level less than a pre-computed threshold level of noise in the input physiological signal:
• apply a DTW algorithm on the plurality of segments to calculate DTW distance between each of the plurality of segments and the clean template segment;
• and
• select at least one segment from the plurality of segments as a clean physiological signal segment of the clean physiological signal training set based on a comparison of the DTW distance of the plurality of segments with a pre-computed DTW threshold distance, wherein the DTW distance measures an intrinsic dissimilarity between the clean physiological signal segment and the noisy physiological signal segment as a semi-supervised technique, wherein a segment from the plurality of segments having the DTW distance greater than the pre-computed DTW threshold distance is categorized as a noisy physiological signal segment and discards the noisy physiological signal segment, wherein the differentiation between the noisy physiological signal segment and the clean physiological signal segment facilitates denoising the PCG signal before performing classification and results in eliminating artefacts and false-positives from the PCG signal, and wherein a segment from the plurality of segments having the DTW distance less than or equal to the precomputed DTW threshold distance is categorized as the clean physiological signal segment;
• extract an optimal feature set from the clean physiological signal training set based on a Maximum Consistency and Maximum Dominance (MCMD) property associated with the optimal feature set, using an MCMD feature selection module, the MCMD property is maximally supreme in their consistency property to clearly identify a target class C={normal;
• abnormal}, and the MCMD property strictly optimizes on an objective function associated with conditional likelihood maximization over a plurality of selection parameters so that diverse properties of the plurality of feature selection parameters are captured, wherein the optimal feature set is a subset of the plurality of features, a set of maximum consistent features and the set of maximum dominance features, and present in both the set of maximum consistent features and the set of maximum dominance features;
• and
• classify the input physiological signal into normal signal components and abnormal signal components using the optimal features set by (i) training a classifier with training data set, and (ii) classifying the input physiological signal as normal or abnormal using the trained classifier, and further the optimal feature set is input to a non-linear Support Vector Machine (SVM) classifier with radial bias function (RBF) kernel, wherein the SVM classifier classifies the abnormalities in the PCG signal automatically without human intervention, wherein the classifier is utilized for identification of cardiac abnormality of the subject from the PCG signal with an accuracy of approximately more than 85%.

A non-transitory computer-readable medium having embodied thereon a computer program for executing a method for physiological monitoring, the method comprising:

• identifying a clean physiological signal training set from an input physiological signal based on a Dynamic Time Warping (DTW) of a plurality of segments associated with the physiological signal, via one or more hardware processors, wherein the input physiological signal is received from one or more sensors that are capable of monitoring subject's physiological health including smartphone inbuild sensors, a digital stethoscope, and an acoustic sensor, wherein the input physiological signal comprises a physiological audio signal obtained from a subject and the physiological audio signal includes a phonocardiogram (PCG) signal, wherein identifying the clean physiological signal training set comprises:
• segmenting the input physiological signal into a plurality of segments:
• identifying a clean template segment from the plurality of segments, the clean template segment associated with a noise level less than a pre-computed threshold level of noise in the input physiological signal;
• applying a DTW algorithm on the plurality of segments to calculate DTW distance between each of the plurality of segments and the clean template segment;
• and
• selecting at least one segment from the plurality of segments as a clean physiological signal segment of the clean physiological signal training set based on a comparison of the DTW distance of the plurality of segments with a pre-computed DTW threshold distance, wherein the DTW distance measures an intrinsic dissimilarity between the clean physiological signal segment and the noisy physiological signal segment as a semi-supervised technique, wherein a segment from the plurality of segments having the DTW distance greater than the pre-computed DTW threshold distance is categorized as a noisy physiological signal segment and discards the noisy physiological signal segment, wherein the differentiation between the noisy physiological signal segment and the clean physiological signal segment facilitates denoising the PCG signal before performing classification and results in eliminating artefacts and false-positives from the PCG signal, and wherein a segment from the plurality of segments having the DTW distance less than or equal to the pre-computed DTW threshold distance is categorized as the clean physiological signal segment;
• extracting an optimal feature set from the clean physiological signal training set based on a Maximum Consistency and Maximum Dominance (MCMD) property associated with the optimal feature set, via the one or more hardware processors using an MCMD feature selection module, the MCMD property is maximally supreme in their consistency property to clearly identify a target class C={normal;
• abnormal}, and the MCMD property strictly optimizes on an objective function associated with conditional likelihood maximization over a plurality of feature selection parameters so that diverse properties of the plurality of feature selection parameters are captured, wherein the optimal feature set is a subset of the plurality of features, a set of maximum consistent features and the set of maximum dominance features, and present in both the set of maximum consistent features and the set of maximum dominance features;
• and
• classifying the input physiological signal into normal signal components and abnormal signal components using the optimal features set by (i) training a classifier with training data set, and (ii) classifying the input physiological signal as normal or abnormal using the trained classifier, and further the optimal feature set is input to a non-linear Support Vector Machine (SVM) classifier with radial bias function (RBF) kernel, wherein the SVM classifier classifies the abnormalities in the PCG signal automatically without human intervention, wherein the classifier is utilized for identification of cardiac abnormality of the subject from the PCG signal with an accuracy of approximately more than 85%.