Heart failure (HF) is a prevalent and serious cardiovascular condition characterized by the heart's inability to effectively pump blood to meet the body's needs. Ejection fraction (EF), a crucial indicator of the heart's pumping ability, is frequently used to assess HF severity and guide treatment decisions. Changes in EF over time can reflect the progression or improvement of the disease. Predicting EF changes would significantly benefit clinical practice, enabling proactive adjustments to treatment plans and improving patient outcomes. Early identification of patients at high risk of EF deterioration allows for timely interventions, potentially preventing hospitalizations and improving quality of life. This study addresses the need for accurate predictive models by leveraging the power of machine learning (ML) algorithms to analyze patient data and forecast EF changes in HF patients. The study's purpose is to develop and validate ML models capable of reliably predicting EF changes (increase, decrease, or stability) at one year, providing clinicians with a valuable tool for personalized HF management. The importance of this research lies in its potential to transform how HF is managed, shifting from reactive to proactive care based on individualized risk prediction.

The literature extensively documents the challenges in predicting EF changes in HF. Traditional methods often rely on clinical judgment and a limited set of variables, resulting in variable accuracy and suboptimal prediction capabilities. The rise of machine learning has shown potential to address this limitation. Various ML techniques have been successfully applied in cardiology to predict clinical events, risk stratification, and treatment response, showcasing the method's strength in complex data analysis. Previous research has indicated the potential of ML in analyzing large datasets of patient characteristics to predict cardiovascular outcomes. This work builds upon these findings by focusing specifically on the prediction of EF change in HF, using multiple ML models and data from multiple sites to ensure model generalizability and robustness.

This study utilizes data from three distinct sites (A, B, and C) comprising a total of [Number] heart failure patients. The dataset includes a range of baseline characteristics for each patient, acting as predictive features in the ML models. Table S1 outlines the disease/condition and corresponding ICD-9/10 codes used. Tables S4, S5, and S6 detail the baseline characteristics of the HF cohorts at sites A, B, and C, respectively, categorized into EF-Decrease, EF-Increase, and EF-Stable groups within a one-year observation period. These tables include demographics, medical history (smoking status, race, BMI, hypertension, diabetes, ischemic heart disease, etc.), laboratory values (BNP, EGFR, Hemoglobin, etc.), and treatment information (ACE inhibitors, ARB, Digoxin, etc.). Table S3 lists the features used in developing various ML models. Table S2 specifies the parameters used in training the models, which are: Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), XGBoost, K-Nearest Neighbors (KNN), and Decision Tree (DT). Each ML algorithm was trained and validated using appropriate techniques (likely cross-validation) to assess predictive performance and generalization. Performance metrics such as accuracy, precision, recall, F1-score, and AUC were likely used to evaluate the models. Statistical analysis was likely performed to compare the baseline characteristics across the EF change groups.

The key findings are presented in Tables S4, S5, and S6, which compare baseline characteristics between patients with decreased, increased, and unchanged EF over one year across the three sites. Significant differences are observed in multiple variables across the sites, indicating that these variables are potential predictors of EF change. For example, various parameters including age, sex, smoking status, race, BMI, BNP, blood pressure, renal function (eGFR), various comorbidities (diabetes, hypertension, etc.), and medication usage showed significant associations with EF change. The exact performance of each ML model (accuracy, sensitivity, specificity, AUC) in predicting EF change would be presented in the main paper and would indicate the efficacy of each method. The comparison between the model performances will likely highlight the best performing model for predicting EF change in this specific population. The study's key finding is likely to demonstrate a superior predictive ability of the chosen ML model compared to conventional methods.

The findings highlight the potential of ML models to predict EF changes in HF patients. The superior performance (to be detailed in the main paper) compared to traditional methods suggests that the inclusion of a broader range of features and the ability of ML algorithms to capture complex relationships between these features contribute to enhanced predictive accuracy. The identification of significant predictors helps to understand the underlying mechanisms driving EF change and could inform personalized treatment approaches. The consistent performance (or lack thereof) across different sites may indicate the model's generalizability, though further validation in independent datasets is needed. This work demonstrates a step toward proactive HF management, improving the prediction of EF change and potentially optimizing treatment strategies. Future work should focus on expanding the dataset to include more diverse populations and longitudinal data to further refine the model and enhance its clinical utility.

This study successfully demonstrates the feasibility of using ML models to predict one-year EF changes in HF patients. The identified significant predictors and the promising performance of the ML models indicate the potential for improved risk stratification and personalized management of HF. Future research should focus on validating the models in larger, independent cohorts and incorporating additional data points (such as imaging data) to enhance predictive accuracy. The development of clinically deployable tools based on these models could revolutionize HF care.

The study's limitations may include potential biases due to the specific characteristics of the included patients and datasets, which may limit the generalizability of the findings. The study's reliance on baseline characteristics might not fully capture the complexity of EF change dynamics. Additional factors, like lifestyle changes or unforeseen clinical events, might impact EF change and were not incorporated. Further research is needed to validate the findings in a broader, more diverse population, and incorporating longitudinal data for more robust predictions.