ECG based physiological event prediction
ECG based physiological event prediction is a project centered around the use of machine learning to detect and predict physiological events on a macro scale, from ECG readings. Objectives of this project are twofold. Initially, developing algorithms to predict a future cardiac arrest event from a live stream of ECG data. Next, detect and predict stress and fatigue from a live stream ECG data.
These algorithms will be trained in a weakly supervised manner. In both of these objectives, the localization of the occurrence of each respective event is a very involved problem. In order to solve this problem we will take advantage of several techniques developed with the intent that they be used for time series analysis through the lens of machine learning. The first of these techniques is Personalization. By using Personalization in our event detection framework we will be able to achieve a much higher degree of accuracy on new test subjects that were not included in the training process. By changing our classifier so that it is specifically tailored to the new test subject we will be able to more effectively detect events that occur within the physiology of that person. Another technique we use is Early Event Detection. By analyzing a monotonically increasing indicator function we are able to detect events before they have completed. We will apply this to the ECG event localization problem by treating the entire sequence as an event with a very shallow and gradually increasing monotonic indicator function. In this way we will be able to predict events such as cardiac arrest or, in a more temporally localized manner, stress and fatigue.
Finally, we will implement Segment-SVM to perform the actual event detection. Segment-SVM is a modification of the primal formulation for SVM classifiers. Segment-SVMs are used to perform SVM based machine learning on time series with a focus on event detection. This makes them ideal for our application in localizing the occurrence of physiological events from ECG time series data.