Machine learning enhanced acute heart failure phenotype prediction using natural language processing and random forest

Figure 1. The distribution of HF phenotype classification displayed from final study participants with LVEF information available.

Figure 2. Performance of random forests for HF phenotype classification using different data configurations. (A) illustrates the effectiveness of random forest models using textual data with 47 features, (B) depicts models using laboratory test data with 53 features, and (C) shows models combining textual and laboratory test data with 100 features. Each subfigure includes ROC curves on the left and PRC curves on the right.

Figure 3. Performance of random forests for HF phenotype classification with varying feature quantities using combined data. (A) illustrates the effectiveness of random forest models using top 30 features, (B) depicts models using top 20 features, and (C) shows models using top 10 features. Each subfigure includes ROC curves on the left and PRC curves on the right.

Figure 4. SHAP (SHapley Additive exPlanations) value distributions for top features in predicting HF phenotype. The figure presents summary plots for HFpEF (left), HFmrEF (middle), and HFrEF (right) phenotypes. Each row represents a feature, and each point represents a patient. The x-axis shows the SHAP value, indicating the impact of the feature on the model output. Red points denote high feature values, while blue points indicate low values. Features are ranked by their overall importance in predicting each phenotype. This visualization reveals the complex relationships between features and their impact on heart failure phenotype classification, highlighting key predictors such as hypokinesis, cardiomyopathy, and severe global dysfunction across different phenotype.

Figure 5. Mean SHAP (SHapley Additive exPlanations) values for top features across HF phenotype. This bar plot illustrates the average impact of key features on model output for HFrEF (blue), HFpEF (pink), and HFmrEF (olive) classifications. Features are ranked by their overall importance, with hypokinesis, cardiomyopathy, and severe global dysfunction showing the highest mean SHAP values. The x-axis represents the mean absolute SHAP value, indicating the magnitude of each feature’s impact on model predictions. This visualization provides a comparative view of feature importance across the three heart failure phenotypes, highlighting the differential influence of clinical and physiological factors in predicting each phenotype.