Wearable sensor-based multi-disease detection continual learning framework

A method for continually learning to detect new diseases with a single neural network model, comprising:

• generating data for a current mission by receiving data from a wearable medical sensor and preprocessing the received data;
• receiving data from previous missions from a exemplar-replay-style continuous learning (CL) algorithm;
• analyzing the data for the current mission to determine if more output neurons and/or detection heads are needed in a multi-headed deep neural network (DNN) model for the current mission, and if so, expanding the DNN model include more output neurons and/or detection heads;
• train the DNN model jointly with substantially equal data from the current mission and previous missions;
• and
• preserving or generating data by either:
• obtaining the average training loss of each data instance in the current mission after training, computing a threshold value for data preservation, and preserving data instances whose average training loss values are above the threshold value for future replays;
• or
• modeling a joint multivariate probability distributions of real training data through both parametric and non-parametric density estimation methods, and sampling as much synthetic data as required from the joint multivariate probability distributions.