Predicting readmission rates in critically ill heart failure patients during a 90-day vulnerable phase using interpretable machine learning models

OBJECTIVE: The readmission rate of critically ill Heart Failure (HF) patients remains high during the vulnerable phase. However, predictive models based on more comprehensive comorbidities and medication histories are lacking. This study aims to extract these factors to develop interpretable models for predicting readmission risk. METHODS: The authors recruited critically ill HF patients from the MIMIC-IV database (as training and internal validation cohorts) as well as from the MIMIC-III database (as external validation cohorts). Four models, including Neural Multitasking Logistic Regression (NMTLR) were constructed and evaluated. Furthermore, Shapley Additive Explanations (SHAP) interpreted feature importance, simplifying the optimal model based on variable importance. RESULTS: A total of 12,126 patients were included in this study. Among four predictive models, the NMTLR model demonstrated the best predictive performance with the Area Under the Curve (AUC) of 0.752 (internal) and 0.785 (external), with the C-index of 0.7408 (internal) and 0.7724 (external), and with the mean cumulative/dynamic Area Under the Curve (mean AUC) score of 0.747 (internal) and 0.763 (external). Moreover, the Integrated Brier Score (IBS) of the NMTLR model was 0.062 (internal) and 0.043 (external). The SHAP analysis showed over one-third of the top 20 features in the NMTLR model were comorbidities and medications, including a newly relevant drug named psychoanaleptics. Furthermore, the compact NMTLR model also performed well. CONCLUSION: Despite the limited representation of modern HF medications, both the full and compact NMTLR models are useful for predicting HF readmission. Additionally, psychoanaleptics might be a new predictor of readmission, warranting increased clinical attention.