Autonomous Self-Refining Neural Systems For Continuous Learning In Real-Time Healthcare Monitoring Applications

Abstract

Round-the-clock monitoring of various physiological parameters like ECG, heart rate, blood oxygen saturation levels, and respiratory rate has come to be the cornerstone of today’s critical-care and remote patient monitoring systems. But most of the current deep learning models that have been developed for such purposes are trained only once using a particular static dataset and their performance starts deteriorating gradually due to drifting of physiology of patients, sensor calibration, and demographic changes, which is further aggravated with the occurrence of catastrophic forgetting each time the neural network is retrained. This paper suggests an autonomous self-refining neural system which consists of a lightweight convolutional recurrent (CNN-LSTM) inference model along with an online drift detection layer, selective experience replay buffer, and EWC-based regularized. Evaluations are performed on a synthetic stream of vital sign data, which is modelled from the PhysioNet MIT-BIH Arrhythmia and MIMIC-III datasets, with six deployment periods of sequential use and two concept-drift events simulating sensor recalibration and population shift, respectively. With respect to a static baseline model and an ablation experiment involving the use of only the replay strategy, our self-refining architecture maintains performance above 90 percent accuracy throughout deployment and recovers completely following each drift event, whereas the static model degrades from 91.4 percent to 74.1 percent accuracy during the same period. These results highlight the feasibility of using autonomous self-refinement, along with forgetting-resistant regularization, to achieve trustworthy adaptive and continuously learning health monitoring systems deployable at the edge.