Neophytos Stylianou1, Artur Akbarov2, Evangelos Kontopantelis2, Iain Buchan2, Ken W Dunn3. 1. Centre for Health Informatics, Institute of Population Health, University of Manchester, UK. Electronic address: neophytos.stylianou@postgrad.manchester.ac.uk. 2. Centre for Health Informatics, Institute of Population Health, University of Manchester, UK. 3. University Hospital South Manchester, Greater Manchester, UK.
Abstract
INTRODUCTION: Predicting mortality from burn injury has traditionally employed logistic regression models. Alternative machine learning methods have been introduced in some areas of clinical prediction as the necessary software and computational facilities have become accessible. Here we compare logistic regression and machine learning predictions of mortality from burn. METHODS: An established logistic mortality model was compared to machine learning methods (artificial neural network, support vector machine, random forests and naïve Bayes) using a population-based (England & Wales) case-cohort registry. Predictive evaluation used: area under the receiver operating characteristic curve; sensitivity; specificity; positive predictive value and Youden's index. RESULTS: All methods had comparable discriminatory abilities, similar sensitivities, specificities and positive predictive values. Although some machine learning methods performed marginally better than logistic regression the differences were seldom statistically significant and clinically insubstantial. Random forests were marginally better for high positive predictive value and reasonable sensitivity. Neural networks yielded slightly better prediction overall. Logistic regression gives an optimal mix of performance and interpretability. DISCUSSION: The established logistic regression model of burn mortality performs well against more complex alternatives. Clinical prediction with a small set of strong, stable, independent predictors is unlikely to gain much from machine learning outside specialist research contexts.
INTRODUCTION: Predicting mortality from burn injury has traditionally employed logistic regression models. Alternative machine learning methods have been introduced in some areas of clinical prediction as the necessary software and computational facilities have become accessible. Here we compare logistic regression and machine learning predictions of mortality from burn. METHODS: An established logistic mortality model was compared to machine learning methods (artificial neural network, support vector machine, random forests and naïve Bayes) using a population-based (England & Wales) case-cohort registry. Predictive evaluation used: area under the receiver operating characteristic curve; sensitivity; specificity; positive predictive value and Youden's index. RESULTS: All methods had comparable discriminatory abilities, similar sensitivities, specificities and positive predictive values. Although some machine learning methods performed marginally better than logistic regression the differences were seldom statistically significant and clinically insubstantial. Random forests were marginally better for high positive predictive value and reasonable sensitivity. Neural networks yielded slightly better prediction overall. Logistic regression gives an optimal mix of performance and interpretability. DISCUSSION: The established logistic regression model of burn mortality performs well against more complex alternatives. Clinical prediction with a small set of strong, stable, independent predictors is unlikely to gain much from machine learning outside specialist research contexts.
Authors: Leah L Zullig; Shelley A Jazowski; Tracy Y Wang; Anne Hellkamp; Daniel Wojdyla; Laine Thomas; Lisa Egbuonu-Davis; Anne Beal; Hayden B Bosworth Journal: Health Serv Res Date: 2019-08-20 Impact factor: 3.402
Authors: Sukrit S Jain; Indra Neil Sarkar; Paul C Stey; Rajsavi S Anand; Dustin R Biron; Elizabeth S Chen Journal: AMIA Annu Symp Proc Date: 2018-12-05