Published Paper Details:

Cardiovascular disease (CVD) is one of the serious health issues in the world that causes the death of many people annually. Early CVD diagnosis is important in enhancing survival. The proposed research will contribute to creating a novel system to detect CVDs based on the synthesis of innovative hybrid techniques to impute missing data, detect outliers, balance the classes, select features, and classify them. It is also aimed at improving accuracy and reliability using a new combination of techniques, hyperparameter optimization and prevention of Type II errors. This framework applies a real-time dataset that belongs to the Salem Private Hospital and a benchmark dataset that is found in the UCI repository to identify CVDs. The Heuristic-SHAP Adaptive MissForest (HSAMF) approach addressed missing data by adopting a hybrid technique of imputation that involves the combination of MissForest, SHAP and heuristic rules. Removal of outliers was done through the Hybrid Global-Local-Structural Outlier Detection (HGLS-OD) method, which is a combination of Local Outlier Factor (LOF), Isolation Forest (IF) and Multi-Model Outlier Detection (MMOD) algorithm. Target encoding was implemented on categorical data. Class imbalance was addressed using optimized K-Means and Synthetic Minority Oversampling TEchnique (OKSMOTE), which includes K-Means, SMOTE, optuna and RF. The Min-MaxScaler was used to perform data normalization. The features were selected using the Cluster-Weighted Mutual Information - Genetic Algorithm (CWMI-GA) methodology that comprised the application of Mutual Information (MI), clustering and Genetic Algorithm (GA). Various classification methods had been used, such as traditional methods, such as Optimized Random Forest (ORF), Optimized XGBoost (OXGB) and ensemble techniques, such as Optimized Bagging-Boosting Stacked Ensemble (OBBSE), Optimized Heterogeneous Soft Voting Ensemble (OHSVE), Optimized Feature-Augmented Heterogeneous Stacking (OFAHS), Optimized Heterogeneous Bootstrap-Ensemble (OHBE) and Optimized Heterogeneous Sequential Boosting (OHSB). Optuna was used to maximize the classification and threshold tuning procedures. The OFAHS model was superior to all the other models with the default threshold of 99.1% when applied to real-time data and 97% when applied to benchmark data. The ideal threshold of 0.47 greatly minimized Type II errors. With this threshold, accuracy increased further to 99.4 and 98.5 on real-time and benchmark data, respectively. Thus, not only were the Type II errors lowered by modifying the threshold, but also the reliability of the forecasts was enhanced. This study demonstrates significant improvement in cardiovascular disease detection through advanced methods that led to improved data handling, accuracy, and performance. The findings are the precursors of the prospective advancements in cardiovascular diagnostics and health management.