A Predictive Analytics Approach for Proactive Heart Failure Diagnosis Using Machine Learning

Abstract

The early and precise detection of heart failure can lead to the mitigation of disease progression and decrease cardiovascular mortality. However, traditional diagnostic methods utilize subjective judgment of symptoms and isolated clinical factors, which limit the scope for early intervention. This paper suggests a framework of predictive analytics for early detection of heart failure through efficient machine learning models, trained on publicly available clinical data. Experiments were performed on the Heart Failure Clinical Records data set, collected from UCI Machine Learning Repository, consisting of demographic attributes, physical measurements and biochemical markers. We used a detailed preprocessing pipeline including normalization, outlier detection and class imbalance problem management. To increase the discriminative ability, a hybrid feature selection method involving mutual information ranking and evolutionary optimization was devised. Finally, diverse learners, including Random Forest, Support Vector Machine, Gradient Boosting and XGBoost were trained and merged through a weighted ensemble learning scheme to improve robustness and generalization. The learners and the hybrid feature selection method were assessed using accuracy, precision, recall, F1-score, RMSE and AUC metrics. Our proposed ensemble learning system obtained an accuracy of 97.1% and an AUC of 0.986. The proposed ensemble approach successfully outperformed the individual learning models and the standard feature selection technique.