Multi-Task ECG Model Training with Related Abnormality Data

The patent applies preliminary action by pre-training the deep learning model on abundant related abnormality data before fine-tuning on the target abnormality data. This preliminary training on related tasks (e.g., other ECG abnormalities) establishes initial feature extraction capabilities and parameter configurations that can be transferred to the target task, reducing the amount of target-specific training data needed.

The patent utilizes parameter changes by adjusting model parameters during transfer learning from related tasks to the target task. The model parameters learned from related abnormality classification are modified and adapted through fine-tuning on the target abnormality data, allowing the model to leverage previously learned features while adapting to the specific characteristics of the target abnormality type.

The patent applies segmentation by dividing the multi-classification problem into multiple single-task classification models, each dedicated to recognizing a specific type of electrocardio abnormality. Instead of training one large multi-class model that requires balanced data across all classes, the approach segments the task into multiple specialized models that can be trained independently on their respective abnormality types, reducing the data balance requirement for each individual model.

The patent implements universality through a shared feature extraction backbone that serves multiple classification tasks. The deep learning model uses common layers for feature extraction that are shared across different abnormality types, while having separate classification heads for each specific abnormality. This allows the model to achieve multi-classification coverage while training each task independently, reducing the need for perfectly balanced multi-class training data.