FFMHNet: Drug Recommendation Based on Feedforward Neural Network and Multi-head Self-attention Mechanism

Nowadays, Deep learning has gained increasing prominence in drug recommendation systems; however, critical challenges persist in feature fine-grained extraction and crossmodal information fusion. To address these limitations, we propose FFMHNet - a novel drug recommendation model integrating feedforward neural networks (FNN) with multi-head attention mechanisms to enhance recommendation accuracy while mitigating drug-drug interaction (DDI) risks. Leveraging electronic health records (EHRs), our framework processes multidimensional temporal data including patient diagnoses and procedures. The patient representation module employs FNNs to fuse temporal features from diagnostic and procedural information, coupled with bidirectional GRU networks to capture longitudinal treatment dependencies. Concurrently, the drug representation module synergizes graph convolutional networks with multi-head self-attention mechanisms to analyze structural drug features and dynamically model complex pharmacological associations. Experimental results demonstrate FFMHNet’s superiority over state-of-the-art baselines: $\mathbf{2. 3 \%}$ absolute reduction in DDI rate. Besides, performance gains of $1.65 \%$ (Jaccard: 0.5304), 1.85% (PRAUC: 0.7799), and 1.09% (F1-score: 0.6839) Ablation studies confirm the complementary effects between FNN and attention components. This research establishes an effective framework balancing recommendation precision and medication safety, advancing intelligent therapeutic decision-making through multimodal feature integration.