FRISTS: Interpretable Time Series-Based Heart Failure Risk Prediction

Abstract

Heart failure is an incurable medical condition that affects millions of people globally. Developing prediction models is crucial to prevent patients from progressing to heart failure. Current heart failure prediction models struggle with achieving interpretability, a mandatory trait that allows the model to be applied in healthcare, while possessing high real-world accuracy. We introduce FRISTS, a novel HF prediction approach leveraging sequential times series-based modeling, such as long short-term memory (LSTM) networks, and feature selection. Our study utilized an extensive electronic health record (EHR) database and found that FRISTS outperformed traditional AI models (e.g., Random Forest Classifier, Logistic Regression, and Decision Tree) and more complex machine learning techniques (e.g., XGBoost, LSTM), yielding an average F1 score of 0.805 and receiver operating characteristic area under the curve (ROC AUC) value of 0.990. Our SHAP-inspired permutation method enables interpretation of the feature ranking, enhancing the result transparency and paving the way for a new class of interpretable models. This approach holds promise in enhancing clinical decision-making and patient care in the context of heart failure prevention.