Prediction of In-Hospital Mortality Among Heart Failure Patients: An Automated Machine Learning Analysis of Mimic-III Database

Abstract

BackgroundMIMIC-III database has been explored to develop predictive risk models for heart failure outcome.1 With Artificial intelligence in general and machine learning in particular exhibiting superior predictive and prognositicative capabilities, when compared with conventional statistical techniques, we explored the current state-of-the-art for automated Machine Learning to develop mortality prediction risk models among those heart failure patients who are being managed in ICU with the aim to optimise risk stratification. MethodologyThe study population comprised 1,177 patients suffering from heart failure being managed intensively. Variables included BMI, comorbidities along with cardiac, renal and blood parameters. The current state of the art (SOTA) for automated Machine Learning (aML) was adopted to develop predictive models using algorithms including Neural Network, eXtreme Gradient Boosting and CatBoost with employment of hyperparameter tuning. Ensemble approach, which is the amalgamation of two or more than two algorithmic models to develop such a model which is better than either of its computive components, was superimposed. Macro weighted average Area Under the Receiver Operating Curve (mWA-AUROC) and log loss, along with other parameters, were adopted to assess the predictive discriminative ability of the developed models. ResultsA stacked ensemble of Neigherest Neighbours, eXtreme Gradiant Boosting, Light Gradiant Boosting Machine, Random Forest, CatBoost, Extra Trees and Neural Network algorithmic models predicted in-hospital mortality in HF patients with an mWA-AUROC of 0.85 and a log loss of 0.28. A specificity of 98.43% (95% CI: 97.28%-99.19%) and a positive likelihood ratio of 19.81 (95% CI: 10.62-36.95) was achieved. Anion gap, along with renal failure and atrial fibrillation, was recognized as the most influential predictor. (Figure 1) ConclusionsOur novel approach to developing in-hospital mortality predictive models for HF patients by exploring automated machine learning provides optimal predictions which, when incorporated into the respective prognosticative protocols, shall translate into a decrease in the morbidity and mortality associated with this ailment by assisting in risk stratification and complication triaging.These authors contributed equally to data curation, development of methodology, statistical analyses and writing with subsequent reviewing and editing of the abstract draft.These authors, indicated alphabetically, contributed equally to statistical analyses and validation.