Screening device, method, and system for structural heart disease
A wireless wearable sensor device, system, method, and non-transitory computer readable medium for screening for structural heart disease based on electrocardiogram, phonocardiogram, and/or accelerometer signals on a patient's skin surface.
1 . A wearable sensor device for screening of structural heart disease, comprising:
at least one sensor measuring electrocardiogram (ECG), phonocardiogram (PCG), and accelerometer (ACC) signals on a patient's skin surface and sending the ECG, PCG, and ACC signals to a processor; and
the processor executing a software application stored in a non-transitory computer executable storage medium to perform processes of:
dividing the measured ECG, PCG, and ACC signals into sequences, by dividing the measured ECG, PCG, and ACC signals into respective windows of arbitrary durations to obtain the sequences, based on a measure of body acceleration derived from the ACC signals,
screening decisions on the sequences of the ECG and PCG by feeding each sequence individually into a sequence screening engine, wherein the sequence screening engine screens each sequence based on segmentation specific features and sequence specific features, and classification is performed for the segmentation specific features and the sequence specific features independently using a machine learning classifier trained to generate a screening decision of each sequence indicative of a normal or abnormal heart condition,
combining a plurality of screening decisions from the sequence screening engine over time into a sequence of screening decisions, and
screening for structural heart disease by feeding the sequence of screening decisions into a sequence aggregate model and outputting an indication of a presence or an absence of the screened structural heart disease to a display to guide medical management or surgical treatment planning for structural or congenital heart disease, wherein the sequence aggregate model comprising at least one of a recurrent neural network, a finite state machine, or a voting-based decision rule.
2 . The wearable sensor device of claim 1 , wherein the processes further comprise directly deriving sequence specific features from one of the sequences before any segmentation.
3 . The wearable sensor device of claim 1 , wherein the screening decisions includes applying a screening algorithm on the sequences of the ECG and PCG.
4 . The wearable sensor device of claim 1 , wherein the processes further comprise aligning the ECG, PCG, and ACC signals in time.
5 . The wearable sensor device of claim 1 , wherein the segmentation specific features are derived after segmenting the PCG signals into constituent heart phases including S1, systole, S2, and diastole for each heartbeat.
6 . The wearable sensor device of claim 1 , wherein the screening decisions on the sequences ECG and PCG further includes independently employing single channels of segmentation including physiological segmentation and model based segmentation to determine constituent heart stages.
7 . The wearable sensor device of claim 6 , wherein a weighted sum of single channel are used to obtain a final position of the segmentation.
8 . The wearable sensor device of claim 6 , wherein the physiological segmentation includes:
determining fiducial points in a first sequence including a peak of an R wave, and an end of a T wave for each heartbeat;
determining locations of S1 and systole (Sys) in a second sequence by using a time duration between peak of R wave (T 1 ) and end of T wave (T 2 ) of a beat from the first sequence;
determining locations of S2 and diastole (Dia) in the second sequence by using the time duration between an end of T wave of a beat and peak of R wave of a next beat from the first sequence; and
learning the boundary between the S1 and systole, and the S2 and diastole from a training set.
9 . The wearable sensor device of claim 6 , wherein the model based segmentation includes:
determining fiducial points in the second sequence using Hidden Markov based model (HMM);
modeling an expected duration of T 1 and T 2 using a probabilistic distribution;
dividing T 1 and T 2 into heart phases of S1 and systole, and S2 and diastole, respectively; and
learning boundaries from a training set.
10 . The wearable sensor device of claim 9 , wherein the model based segmentation further includes a subset of the heart phases.
11 . The wearable sensor device of claim 6 , wherein a weighted sum of single channel detections of the physiological and model based segmentation is compared to a threshold to determine a final position of the segments.
12 . The wearable sensor device of claim 11 , wherein features are extracted based on at least one of the output of segmentation or from the second sequence.
13 . The wearable sensor device of claim 12 , wherein the classification is performed for the segmentation specific features and sequence specific features independently using Classifier A and Classifier B, respectively,
wherein the Classifier A and the Classifier B includes machine learning algorithms including at least one of: logistic regression, support vector machines, artificial neural networks, or gradient boosting.
14 . The wearable sensor device of claim 13 , wherein the output of the classifiers A and B are combined at a final sequence detector, based on a linear or non-linear rule to determine a decision.
15 . A non-transitory computer-readable medium storing executable instructions that, in response to execution, cause a computer to perform operations comprising:
receiving measured signals from at least one sensor, the measured signals including electrocardiogram (ECG), phonocardiogram (PCG), and accelerometer (ACC) signals on a patient's skin surface;
dividing the measured signals into sequences, by dividing the measured ECG, PCG, and ACC signals into respective windows of arbitrary durations to obtain the sequences, based on a measure of body acceleration derived from the ACC signals;
screening decisions on the sequences of the ECG and PCG by feeding each sequence individually into a sequence screening engine, wherein the sequence screening engine screens each sequence based on segmentation specific features and sequence specific features, and classification is performed for the segmentation specific features and the sequence specific features independently using a machine learning classifier trained to generate a screening decision of each sequence indicative of a normal or abnormal heart condition;
combining a plurality of screening decisions from the sequence screening engine over time into a sequence of screening decisions; and
screening for structural heart disease by feeding the sequence of screening decisions into a sequence aggregate model and outputting an indication of a presence or an absence of the screened structural heart disease to a display to guide medical management or surgical treatment planning for structural or congenital heart disease, wherein the sequence aggregate model comprising at least one of a recurrent neural network, a finite state machine, or a voting-based decision rule.
16 . A wearable sensor system for screening of structural heart disease, comprising:
a wearable sensor device measuring electrocardiogram (ECG), phonocardiogram (PCG), and accelerometer (ACC) signals on a patient's skin surface and sending the ECG, PCG, and ACC signals to a processor; and
the processor executing a software application stored in a non-transitory computer executable storage medium to perform the process of:
dividing the measured signals into sequences, by dividing the measured ECG, PCG, and ACC signals into respective windows of arbitrary durations to obtain the sequences, based on a measure of body acceleration derived from the ACC signals,
screening decisions on the sequences of the ECG and PCG by feeding each sequence individually into a sequence screening engine, wherein the sequence screening engine screens each sequence based on segmentation specific features and sequence specific features, and classification is performed for the segmentation specific features and the sequence specific features independently using a machine learning classifier trained to generate a screening decision of each sequence indicative of a normal or abnormal heart condition,
combining a plurality of screening decisions by feeding the sequences to screening decision into a sequence aggregate model from the sequence screening engine over time into a sequence of screening decisions, and
screening for structural heart disease by feeding the sequence of screening decisions into a sequence aggregate model and outputting an indication of a presence or an absence of the screened structural heart disease to a display to guide medical management or surgical treatment planning for structural or congenital heart disease, wherein the sequence aggregate model comprising at least one of a recurrent neural network, a finite state machine, or a voting-based decision rule.