Research on Heart Failure Mortality Prediction Model Based on Imbalanced Medical Dataset

This study utilized a clinical dataset from Zigong Hospital, accessible through the PhysioNet website, to investigate six-month mortality prediction in patients with heart failure using machine learning and deep learning approaches. Both algorithm-driven and data-driven strategies were compared to address the challenge of data imbalance. The potential of large language models (LLMs) in this domain was also explored. The analysis employed Python for feature selection and data preprocessing, constructing nine predictive models: logistic regression, support vector machine (SVM), decision tree, random forest, XGBoost, LightGBM, CatBoost, multi-layer perceptron (MLP), and one-dimensional convolutional neural network (1D-CNN). Stratified 5-fold cross-validation was utilized for comparative experiments, with the area under the receiver operating characteristic curve (AUC) and Recall as evaluation metrics. The results demonstrated that the data-driven Borderline-SMOTE method substantially improved Recall, though its effect on AUC was minimal. Furthermore, the algorithm-driven threshold adjustment method achieved a more pronounced improvement in Recall compared to the data-driven approach. The LLM-based model attained an AUC of 0.796 and a Recall of 0.358. Among the nine models with threshold adjustment, the LLM's AUC outperformed only the decision tree model, while its Recall exceeded those of XGBoost and LightGBM. In summary, the findings suggest that threshold adjustment is a more effective method for addressing data imbalance. While the LLM-based model exhibited lower performance than the top-performing models, this represents a pioneering effort in leveraging such models for predicting heart failure mortality, offering valuable insights and a foundation for future research.