Differentiated knowledge distillation: Patient-specific single-sample personalization for electrocardiogram diagnostic models.

To achieve optimal performance in practical applications, the electrocardiogram (ECG) diagnosis models have to be personalized using the ECG data of specific patients. Most current research on personalized ECG diagnosis centres around large-sample transfer learning aimed at specific populations or collection conditions. However, the time and economic costs of collecting such a large amount of data are unacceptable for specific individuals. Moreover, the characteristics of ECG signals mean that most data augmentation methods may not produce positive effects. To address these issues, we propose a novel differentiated knowledge distillation (DKD) method, which purposefully transfers specific knowledge in knowledge distillation by inputting differentiated data to the teacher and student models. Specifically, this method constructs differences between the input data of student and teacher models using one single ECG record of a specific patient, so as to transfer individual knowledge of the patient into the student model while maintaining the diagnostic ability of the baseline teacher model. For other ECG data of the target patient, the model personalized by this method shows a significant improvement of up to 10% in accuracy, micro-F1-score, and micro-sensitivity compared to the baseline model. Additionally, this method demonstrates excellent efficacy for various existing networks in the field of ECG diagnosis. [ABSTRACT FROM AUTHOR]