A Drift-Aware Adaptive Machine Learning Framework for Reliable Clinical Prediction in Dynamic Healthcare Environments

Abstract

The rise of Machine Learning techniques in healthcare has greatly enhanced the prediction and decision-making tasks. But a key problem is the drift in the statistical characteristics of health care data, causing deterioration of the predictive models. Current solutions either rely on drift detection or retraining strategies that are unaware of the reliability and trustworthiness of the model predictions, posing a significant risk in safety-critical domains. To overcome this challenge, we propose a Drift-Aware Adaptive Machine Learning Framework for reliable prediction in clinical domains that experience drift. This approach combines a hybrid drift detection of statistical divergence and continuous monitoring, alongside a new reliability score model. This score builds on model confidence, error patterns and drift severity for adaptive learning. A smart update strategy is adopted to conduct selectively retraining for severe drift. The proposed approach on benchmark healthcare datasets achieves improved model performance from 72% to 89% accuracy (and improved precision, recall and F1-score) and more reliable predictions. Most significantly, it maintains prediction accuracy despite the changing data environment. The proposed method increases clinical safety by minimising the chance of wrong predictions and providing accurate and reliable decision support. It is very effectively applicable in real-world dynamic healthcare settings.