SLM-MEEF: ECG Report Generation based on Small Language Model and Multi-Expert Ensemble Framework

Research on electrocardiogram (ECG) report generation aims to provide physicians with diagnostically interpretable results, thereby enhancing the acceptability of AI-based ECG diagnostic algorithms in clinical settings. Existing methods face challenges in feature extraction and semantic fusion, making it difficult to balance report accuracy with comprehensiveness, while also requiring high computational resources. This paper proposes a novel approach for ECG report generation based on the Small Language Model and Multi-Expert Ensemble Framework (SLM-MEEF), which integrates Small Language Model (SLM) and Multi-Expert Ensemble (MOE) architectures. Specifically, the ECG-MOE framework is designed to fuse fine-grained and global features using load-balancing loss and shared expert mechanisms. Fine-grained features, crucial for diagnosis accuracy, are extracted through a supervised classification model. Global features, which contain rich semantic information from ECG signals, are derived via a self-supervised learning method to enhance the comprehensiveness of the reports. Experimental results demonstrate that the proposed method significantly improves the generation quality and inference performance of ECG reports.