ClinMamba: Leveraging state–space modeling for medication prediction in Electronic Medical Records

With the growth of Electronic Medical Records (EMRs), extracting clinical knowledge for precise diagnosis and treatment is crucial in smart healthcare. However, drug prescription prediction models struggle with multi-modal information fusion, temporal dependency, and class imbalance, limiting their effectiveness in capturing the complexities of patient care. To address these challenges, we introduce ClinMamba, a multi-modal drug recommendation model that combines state–space modeling, sequence learning, and channel attention mechanisms. ClinMamba enhances the joint modeling of patients’ historical behaviors and static attributes. The model features a state–space-based LocalMamba module to capture the local dynamic features within medication time series, while leveraging ResNet and LSTM architectures to encode static information. Additionally, the Cross-Fusion module is incorporated to bolster inter-modal interactions, and the ECA attention mechanism selectively amplifies crucial channels, culminating in multi-label prescription predictions. We evaluate the model on the extensive ICU dataset MIMIC-III, which includes ten typical critical care diagnostic tasks. The experimental results demonstrate that ClinMamba significantly outperforms existing state-of-the-art models across several key metrics, including AUROC and AUPR, showcasing superior predictive power, robustness, and generalization capacity.