TY - JOUR T1 - Machine Learning Approaches to Determine Feature Importance for Predicting Infant Autopsy Outcome JF - medRxiv DO - 10.1101/2020.05.21.20105221 SP - 2020.05.21.20105221 AU - J Booth AU - B Margetts AU - W Bryant AU - R Issitt AU - C Hutchinson AU - N Martin AU - NJ Sebire Y1 - 2020/01/01 UR - http://medrxiv.org/content/early/2020/05/22/2020.05.21.20105221.abstract N2 - Introduction Sudden unexpected death in infancy (SUDI) represents the commonest presentation of postneonatal death, yet despite full postmortem examination (autopsy), the cause of death is only determined in around 45% of cases, the majority remaining unexplained. In order to aid counselling and understand how to improve the investigation, we explored whether machine learning could be used to derive data driven insights for prediction of infant autopsy outcome.Methods A paediatric autopsy database containing >7,000 cases in total with >300 variables per case, was analysed with cases categorised both by stage of examination (external, internal and internal with histology), and autopsy outcome classified as ‘explained (medical cause of death identified)’ or ‘unexplained’. For the purposes of this study only cases from infant and child deaths aged <= 2 years were included (N=3100). Following this, decision tree, random forest, and gradient boosting models were iteratively trained and evaluated for each stage of the post-mortem examination and compared using predictive accuracy metrics.Results Data from 3,100 infant and young child autopsies were included. The naïve decision tree model using initial external examination data had a predictive performance of 68% for determining whether a medical cause of death could be identified. Model performance increased when internal examination data was included and a core set of data items were identified using model feature importance as key variables for determining autopsy outcome. The most effective model was the XG Boost, with overall predictive performance of 80%, demonstrating age at death, and cardiovascular or respiratory histological findings as the most important variables associated with determining cause of death.Conclusion This study demonstrates the feasibility of using machine learning models to objectively determine component importance of complex medical procedures, in this case infant autopsy, to inform clinical practice. It further highlights the value of collecting routine clinical procedural data according to defined standards. This approach can be applied to a wide range of clinical and operational healthcare scenarios providing objective, evidence-based information for uses such counselling, decision making and policy development.Competing Interest StatementThe authors have declared no competing interest.Funding StatementNo direct funding. GOSH research is indirectly supported by the NIHR GOSH BRCAuthor DeclarationsI confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.YesThe details of the IRB/oversight body that provided approval or exemption for the research described are given below:The database allows controlled use of deidentified information and its use for research has been approved by the appropriate Research Ethics Committee / IRB (REC approval 16LO1910, London, Bloomsbury).All necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived.YesI understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).YesI have followed all appropriate research reporting guidelines and uploaded the relevant EQUATOR Network research reporting checklist(s) and other pertinent material as supplementary files, if applicable.YesNonidentifiable aggregate data can be obtained from the authors ER -