32 Identifying feature importance in post-mortem outcome with gradient boosting machines. (22nd November 2019)

• Record Type: Journal Article
• Authors: Booth, John; Margetts, Ben; Sebire, Neil
• Is Part Of: Archives of disease in childhood. Volume 104:(2019)Supplement 4
• Page Start: A13
• Page End: A13
• Publication Date: 2019-11-22
• Subjects: Children -- Diseases -- Periodicals; Infants -- Diseases -- Periodicals; 618.920005
• DOI: 10.1136/archdischild-2019-gosh.32
• Languages: English
• ISSNs: 0003-9888

Abstract

Abstract : Post-mortems are complex procedures that utilise a significant amount of hospital resources, yet despite this, cause of death is only determined in 45% of cases. The event itself can be very traumatic for the parents of the child yet is essential for providing further clinical understanding of the patient's cause of death. Given this, there is an imperative to extract the greatest possible value from the data. Here, we investigated whether machine learning could be used to derive novel insights from the prediction of post-mortem outcomes. A post-mortem database containing 7000 events across 300 attributes was analysed and categorised into stage of examination (external and internal). The outcome of the examination was summarised as either 'cause of death determined' or 'not determined'. From these summarised data, cases were filtered by children aged ≤ 2 years, resulting in a dataset of 3, 100 events. Following this, within a GOSH Digital Research Environment (DRE) workspace, decision tree, random forest, and gradient boosting machine models were iteratively built for each stage of the post-mortem examination and compared using their accuracy metrics. The naïve decision tree model using external examination data had a predictive performance of 67%. Model performance notably increased when trained on internal examination data. At each stage of the examination, a core set of data items, of which the final set included age, BMI, and heart weight were highlighted using model feature importance as key variables for determining post-mortem outcome. The use of increasingly complex modelling techniques was able to boost the predictive performance of the model by as much as 10%. This project clearly shows the value of collecting clinical procedural data which can then be modelled using machine learning techniques to inform clinical practice. With more time, further modelling, including unsupervised clustering could be undertaken to derive further insights.