Open Access

ARTICLE

DRAGON-MINE: Deep Reinforcement Adaptive Gradient Optimization Network for Mining Rare Events in Healthcare

Mohammed Abdullah Alsuwaiket^*

Department of Computer Science and Engineering Technology, University of Hafr Al Batin, Hafr Al Batin, Saudi Arabia

* Corresponding Author: Mohammed Abdullah Alsuwaiket. Email:

Computer Modeling in Engineering & Sciences 2026, 146(3), 35 https://doi.org/10.32604/cmes.2026.078169

Received 25 December 2025; Accepted 23 February 2026; Issue published 30 March 2026

Abstract

The healthcare field is fraught with challenges associated with severe class imbalance, wherein such critical conditions like sepsis, cardiac arrest, and drug adverse reactions are rare but have dire clinical consequences. This paper presents a new framework, Deep Reinforcement Adaptive Gradient Optimization Network to Mining Rare Events (DRAGON-MINE), to demonstrate how deep reinforcement learning can be used synergistically with adaptive gradient optimization and address the inherent weaknesses of current methods in the prediction of rare health events. The suggested architecture uses a dual-pathway consisting of a reinforcement learning agent to dynamically reweigh samples and an adaptive gradient optimizer to follow novel learning rates. With extensive experiments on the MIMIC-IV and eICU-CRD datasets, DRAGON-MINE consistently outperforms recent state-of-the-art methods for sepsis, cardiac arrest, and adverse drug reaction prediction, achieving AUROC values of 92.3% and 91.6% for sepsis prediction on MIMIC-IV and eICU-CRD, respectively, while consistently outperforming Transformer-, CNN-RNN-, and Fed-Ensemble-based methods across all evaluated tasks and datasets, with particularly strong gains observed in precision–recall performance under severe class imbalance. With its high sensitivity (88.4%) and specificity (90.2%), DRAGON-MINE enables reliable early warning of rare clinical events in critical care settings while minimizing false alarms, supporting safer clinical decision support systems, and demonstrating strong potential for scalable deployment across multi-institutional intensive care environments through federated learning.

---

Keywords

---