Machine learning has emerged as a crucial technique for intelligent analysis of electrocardiogram (ECG) signals. Subjective reasons of doctors may lead to different annotation results for the same ECG signal, which means that the label result is not a single value, but a series of numerical outputs. We propose a multi-label classification method based on label distribution. This approach allows for a more robust understanding of ECG patterns, as it assigns probabilities to different disease labels rather than assigning a single categorical label. This paper proposes a Broyden-Fletcher-Goldfarb-Shanno Learning from Label Distribution algorithm (BFGS-LLD-ECG) tailored for ECG multi-label classification. This algorithm leverages the relationships between similar ECG morphologies and rhythms, constructs a maximum entropy model, and employs the efficient quasi-Newton method for maximum likelihood estimation. The proposed algorithm was validated on the CPSC2018 dataset, which contains three ordered disease labels defining the label distribution matrix. The results demonstrated superior performance of label distribution learning, achieving an accuracy of 87.89% and a one-error rate of 10.36%. This method exhibits excellent performance in ECG multi-label classification tasks.