International Journal of Science, Engineering and TechnologyAuthors: Anshika Saini, Radha Kumari, Tanvi Verma, Vanshika Gupta, Dr. Amit Kumar
Abstract: A silent but critical challenge confronts every hospital’s AI system after deployment: disease patterns shift, patient demographics evolve, and new outbreaks emerge, yet the underlying model remains frozen at its training snapshot. This temporal mismatch causes accuracy to erode steadily without any visible warning signals to clinical staff. The present work constructs and validates a Continual Learning (CL) pipeline specifically engineered to eliminate this vulnerability. Three coordinated mechanisms form the backbone of the approach. An Elastic Weight Consolidation (EWC) regularizer identifies which model parameters are most responsible for previously learned disease distinctions and applies a Fisher-Information-weighted penalty to prevent those parameters from drifting during future updates. A fixed-capacity Experience Replay buffer stores five hundred past patient records that are blended with each new data batch at training time, ensuring earlier disease representations are continuously refreshed. An ensemble of three independent drift sensors—ADWIN, DDM, and the Page-Hinkley test—runs concurrently on the incoming patient stream; the moment any sensor signals a statistically significant distributional change, an incremental retraining cycle is triggered automatically. Four clinical scenarios are simulated to exercise the system: a routine disease baseline, a sudden COVID-19 variant surge, a monsoon-driven dengue outbreak, and a post-outbreak mixed-prevalence recovery. Evaluated across these phases, the pipeline attains a mean classification accuracy of 88.1% over five diseases using ten routine clinical measurements. Most significantly, when Phase 1 data is re-evaluated after three complete learning cycles, accuracy falls by only 1.0%, confirming that knowledge acquired during initial training survives subsequent adaptation intact. Drift is detected at every phase boundary with no false alarms, and post-update accuracy recovers within a single retraining cycle. A Streamlit monitoring application delivers live accuracy trends, configurable alert thresholds, and drift event histories to clinical administrators without requiring any machine learning expertise.
