Two stage fine-tuned multimodal generative AI for automated ECG based cardiovascular report generation

The increasing prevalence of cardiovascular diseases (CVDs) calls for innovative diagnostic solutions that are both accurate and scalable. ElectroCardioGrams (ECGs) remain central to cardiac assessment: However, manual interpretation is time consuming and error-prone. To address this challenge, we propose a lightweight multimodal generative AI framework capable of automatically interpreting ECG images and producing structured clinical reports. The framework builds upon the SmolVLM-500M-Instruct model, fine-tuned via Quantized Low-Rank Adaptation (QLoRA) to enable efficient deployment on standard hardware. A custom multimodal ECG dataset ,comprising image report pairs curated from authoritative clinical sources and augmented to mitigate class imbalance, served as the foundation for training. The proposed architecture integrates a vision encoder, a cross-modal fusion mechanism, and a language decoder to effectively align visual ECG representations with diagnostic narratives. Experimental evaluations demonstrate significant improvements in BLEU, ROUGE-L, and BERTScore metrics through a two-phase fine-tuning strategy, highlighting the model's ability to generate clinically coherent and semantically rich reports. Overall, this work contributes a scalable, interpretable, and resource efficient AI framework for cardiac diagnostics, bridging the gap between state of the art deep learning research and real-world clinical practice.