Deep reinforcement learning-based multi-modal framework for explainable drug recommendation in healthcare

• Introduces a deep reinforcement learning (DRL)-based drug recommendation system that integrates multimodal data for accurate and effective decision support. • Applies an improved adaptive attention-based BERT model along with a hybrid stacked LSTM–GRU architecture with dual sparse attention (SLD-SA) to process unstructured, sequential, and heterogeneous medical data. • Proposes an optimized ANN–DQN (OANN-DQN) model incorporating Chebyshev Grey Lag Goose Optimization (C-GLGO) to improve medication mapping and reinforcement learning performance. • Achieves state-of-the-art performance on three datasets—MIMIC-III, MIMIC-IV, and eICU—with accuracy exceeding 98%, while integrating LIME for model explainability and transparency. Drug recommendation systems have been developed to assist medical practitioners in accurate decision-making, thereby enhancing the treatment process and assist clinicians by reducing their workload. On the other hand, several researchers have proposed drug recommendation systems based on clinical guidelines, howerver this process is time-consuming and sometimes leads to inaccurate recommendations. Hence, to mitigate these limitations, this proposed approach introduces a novel deep reinforcement learning (DRL) based recommendation system leveraging multi-modal data features. The clinical datasets were collected from multiple healthcare platforms, and pre-processed for further processing. In order to extract unstructured and sequential information, an improved adaptive attention based BERT and Stacked LSTM-assisted GRU with dual sparse attention (SLD-SA) models are employed, which effectively capture the clinical data features, including demographics, prescriptions, diagnosis, lab results and vital signs. Furthermore, structured data are encoded using entity embedding, and the resulting feature representations are fused via gated cross-attention to enhance feature learning. For drug recommendation, a deep reinforcement learning-based Optimized Artificial Neural Network integrated with a Deep Q-Network (OANN-DQN) is proposed, which learns the mapping between patient conditions and suitable medications. The policy optimization is performed using the Chebyshev Grey Lag Goose Optimization (C-GLGO) technique, improving the learning efficiency of the DRL framework. To ensure interpretability, Local Interpretable Model-agnostic Explanations (LIME) are applied. The proposed model is evaluated on MIMIC-III, MIMIC-IV, and eICU benchmark datasets, achieving accuracies of 98.96%, 98.36%, and 98.42%, respectively.