High-accuracy ECG image interpretation using parameter-efficient Low-Rank Adaptation (LoRA) fine-tuning with multimodal LLaMA V.3.2

Objective To develop and evaluate a high-accuracy ECG image interpretation model using parameter-efficient Low-Rank Adaptation (LoRA) fine-tuning with the multimodal LLaMA V.3.2 model. Methods and analysis We fine-tuned the multimodal LLaMA V.3.2 model (ECG-LLaMA) using parameter-efficient LoRA with a rank of 64 on the ECGInstruct dataset, which contains 1 million samples of ECG images paired with expert annotations. The model was evaluated using area under the curve (AUC), Macro F1 score, Hamming loss and a report generation quality score evaluated by GPT-4O using criteria of medical accuracy, completeness and clinical utility. Ablation studies were conducted to assess the impact of LoRA rank and compare parameter-efficient tuning with partial parameter tuning. Results The fine-tuned model achieved significant improvements over the baseline LLaMA V.3.2 model across all metrics, with an AUC of 0.98 (vs 0.51), Macro F1 of 0.74 (vs 0.33), Hamming loss of 0.11 (vs 0.49) and report generation quality score of 85.4 (vs 47.8). Parameter-efficient LoRA outperformed partial parameter tuning, with optimal performance at a LoRA rank of 64. Error analysis revealed higher accuracy for common arrhythmias (>90%) with more challenges in detecting subtle abnormalities. Conclusion Parameter-efficient LoRA fine-tuning of the multimodal LLaMA V.3.2 model significantly improves ECG image interpretation accuracy, offering a promising approach for developing artificial intelligence systems that can assist in clinical ECG interpretation, particularly in settings where expert cardiologists may not be readily available.