Feed

Electrocardiogram analysis is vital for cardiovascular diagnosis, yet traditional methods struggle with inter-patient variability and complex multi label correlations. Existing Large Language Model integrations often freeze internal parameters, which limits adaptation to idiosyncratic signal patterns across diverse subjects.
Approach. To overcome these challenges, this study proposes a framework that integrates Large Language Models fine-tuned via Low Rank Adaptation with a multi-label dynamic graph convolutional network. The application of Low Rank Adaptation allows the model to encode specific signal features and mitigate inter-patient variability while retaining essential clinical semantic knowledge. Furthermore, the dynamic graph convolutional module is designed to capture the complex correlations between diverse cardiac diseases, thereby enhancing the modeling of multi-label dependencies. The final inference strategy leverages refined semantic embeddings to achieve accurate classification of diagnostic categories without requiring extensive labeled training data for every new patient population.
Main results. Experimental results on the PTB-XL dataset yield F1-scores of 0.781 for the five class task and 0.370 for the twenty class task, representing improvements of 12.4% and 12.9% respectively compared to state of the art methods. Evaluation on the CPSC2018 dataset further confirms robust cross-dataset generalization with an F1-score of 0.606, which corresponds to an 18.8% relative improvement over existing approaches.
Significance. By effectively resolving inter patient variability and modeling disease correlations, the proposed framework provides a scalable and accurate solution for diagnosing cardiac conditions in complex clinical environments where annotated data is limited.
&#xD.