Predicting 30-day Mortality for Patient with Heart Failures: A Machine Learning Approach

Abstract

Title of Thesis : Predicting 30-day Mortality for Patient with Heart Failures: A Machine Learning Approach
Author : Muhammad Hanif Amiruddin
Thesis Advisor : Hsu, Min-Huei, Ph.D

Background
Heart failure (HF) is a leading cause of hospital readmissions and short-term mortality worldwide. Accurate early prediction of outcomes, such as 30-day mortality, could improve clinical decision-making and enable proactive management for high-risk patients. However, real-time clinical data such as vital signs are not always consistently available, particularly across different hospital systems. To address this challenge, we aimed to develop a practical and interpretable machine learning model using only routinely available laboratory and demographic features collected within the first 48 hours of admission. This approach reflects real-world data limitations and supports early risk assessment that could be realistically deployed in diverse hospital settings.
Methods
We developed a LightGBM-based classification model using retrospective data from two sources: an internal cohort (MIMIC-T) and an external cohort (MIMIC-III). The primary outcome was 30-day in-hospital mortality among HF patients. Patients were included if they were ≥18 years old, admitted with a diagnosis of heart failure, and had a hospital stay between 1 and 30 days. Only first hospitalizations were considered to ensure patient-level independence. Exclusion criteria included missing values in key demographic or laboratory variables and admissions with implausible age or length of stay. Laboratory and demographic variables collected within the first 48 hours of admission were used as predictors. All features were pre-processed using a standardized pipeline: numerical variables were imputed using iterative imputation and scaled with standardization; categorical variables were imputed with the most frequent value and one-hot encoded. The models were evaluated using a combination of discrimination and calibration metrics, including AUC, F1-score, precision, recall, and calibration curves. To assess generalizability, we tested two validation scenarios: (1) training on internal data and testing on external data, and (2) reversing this setup. SHAP values were used to interpret model predictions and identify the most influential features across settings.
Result
The LightGBM model trained on the internal validation with MIMIC-T and MIMIC III achieved an AUC of 0.80 and AUC of 0.78 demonstrating strong discriminative performance in the development cohort. However, external validation on MIMIC-III showed a notable drop in performance (AUC = 0.71), reflecting case-mix differences and limited generalizability. Calibration also revealed risk overestimation in the mid-probability range. In contrast, the reverse scenario with training on MIMIC-III and testing on MIMIC-T cohort resulted in improved balance with AUC = 0.76 and better calibration. SHAP analysis consistently highlighted age, hemoglobin, platelet count, and BUN as dominant predictors across both setups. These findings suggest that the external dataset represent with more stable model.
Conclusion
This study shows the feasibility of predicting 30-day mortality in HF patients using laboratory-based machine learning models without relying on real-time vital signs. While the model performed well on internal validation, its generalizability was limited when applied directly to a different population. Reversing the training strategy improved performance, emphasizing the role of population context and data consistency in model development. Interpretability through SHAP and model calibration analysis further strengthened the model's clinical relevance. Although limited by feature availability and dataset harmonization, this work provides a foundation for developing scalable, transparent, and adaptable predictive tools in clinical practice. These findings highlight the potential of machine learning models to support early risk stratification in heart failure patients, with age and BUN emerging as key predictors. Such tools can aid clinicians in prioritizing monitoring and care, particularly for older adults, male patients, and those with elevated BUN levels.

Keywords: Heart failure, mortality prediction, machine learning, LightGBM, SHAP, external validation