An Efficient Machine Learning Procedure for Selection of Core Predictive Variables in the Presence of Multicollinearity for Acute Heart Failure Detection: Methodological Study

Objective: Application of machine learning techniques in comprehending heart failure conditions has been noted in recent research on acute HF (AHF). Nevertheless, the complexities of selecting important features and handling multicollinearity inherently pose significant challenges. Hence, this study aims to identify vital AHF risk factors using machine learning when multicollinearity is present.Material and Methods: Investigations were conducted on patients managed for AHF or non-acute HF (NHF) at University College Hospital (UCH), Ibadan, Nigeria. The population of the study included AHF patients admitted into the Cardiology Unit of Department of Medicine, UCH, Ibadan. It involved 160 previously collected data on AHF cases by Adebayo et al. and 136 NHF patients from UCH Cardiology Department database involving 43 variables. After data cleaning, 256 patients were eventually sampled to assess the consistency among machine learning techniques. Bias reduction generalized linear model, Least Absolute Shrinkage and Selection Operator, ridge regression, random forest (RF), and support vector machine algorithms were employed. Results: The study revealed RF as a robust tool for precise AHF feature selection. This model consistently demonstrated exceptional performance metrics (over 95%), highlighting key predictors, such as ''elevated jugular venous pressure'', ''third heart sound'' and ''paroxysmal nocturnal dyspnea'', offering critical clinical insights. Conclusion: This study can be considered pioneering research in exploring machine learning for the assessment of AHF, particularly with regard to Nigeria's real-life data, thereby contributing significantly to the mitigation of the existing literature gap on AHF.