Balanced-Reservoir Replay: A Simple Class-Aware Modification for Continual Learning in Highly Imbalanced Settings

Catastrophic forgetting and class imbalance pose major challenges in continual learning for critical care. Predictive models must safely adapt to incoming patient data. In this paper, we introduce Balanced-Reservoir Replay (BRR), a simple modification to reservoir sampling that preserves minority-class examples when the buffer is full. We evaluated BRR on the eICU-CRD dataset (2,520 patients, 8.37% mortality) using only age and gender, in a task-incremental setup with eight tasks and buffer sizes ranging from 0 to 1000 samples.Using a 1000-sample buffer (~ 40% of past data), BRR outperformed standard sequential training, achieving an average AUC of 0.628 ± 0.019 compared to 0.610 ± 0.016 (p < 0.01). It also reduced forgetting by 16% (0.106 ± 0.047 vs. 0.126 ± 0.066), which narrowed the gap with cumulative training by 39%. A clinical composite score, which penalized forgetting twice as much as accuracy gains (λ = 2.0), indicated that a buffer size of 1000 was optimal. We conducted a sensitivity analysis of eight configurations of training duration (40 vs. 80 epochs per task) and forgetting penalties (λ = 0.5, 1.0, 2.0, 5.0) supported the validity of this selection.This work questions the idea of minimal memory in continual learning. It indicates that retaining some past data can improve accuracy, enhance stability, and help meet regulatory requirements for slowly changing, low-data clinical tasks. BRR only needs about 50 KB of data, and it’s easy to understand and use in places where resources are limited. It also protects patient privacy while being used for continuous learning.