A Contrastive Learning-Based Interpretable Prediction Model for Patients with Heart Failure

Heart failure is a group of complex clinical syndromes due to any structural or dysfunctional abnormality of the heart that results in impaired filling or ejection capacity of the ventricles. Using historical Electronic Health Records (EHRs) to forecast the risk of critical events in heart failure (HF) patients is an important area of research in the field of personalized medicine. However, it is difficult for some machine learning models to predict the risk of critical events owing to data imbalance and poor feature performance in the EHR data of HF patients. While time series-based deep neural networks have achieved excellent results, they lack interpretability. To solve these problems, this study focuses on proposing a deep neural network prediction model of critical events in heart failure patients based on Contrastive learning and Attention mechanism (CLANet). We evaluate our model on a real-world medical dataset, and the experimental results demonstrate that CLANet improves by 2–10 \(\%\) over the conventional methods.