Predicting Population Health with Machine Learning: A Scoping Review Study Data
Description
Objective: To determine how machine learning has been applied to prediction applications in population health contexts. Specifically, to describe which outcomes have been studied, the data sources most widely used and whether reporting of machine learning predictive models aligns with established reporting guidelines. Design: A scoping review. Data sources: MEDLINE, EMBASE, CINAHL, ProQuest, Scopus, Web of Science, Cochrane Library, INSPEC and ACM Digital Library were searched on 18 July 2018. Eligibility criteria: We included English articles published between 1980 and 2018 that used machine learning to predict population-health-related outcomes. We excluded studies that only used logistic regression or were restricted to a clinical context. Data extraction and synthesis: We summarised findings extracted from published reports, which included general study characteristics, aspects of model development, reporting of results and model discussion items. Results: Of 22 618 articles found by our search, 231 were included in the review. The USA (n=71, 30.74%) and China (n=40, 17.32%) produced the most studies. Cardiovascular disease (n=22, 9.52%) was the most studied outcome. The median number of observations was 5414 (IQR=16 543.5) and the median number of features was 17 (IQR=31). Health records (n=126, 54.5%) and investigator-generated data (n=86, 37.2%) were the most common data sources. Many studies did not incorporate recommended guidelines on machine learning and predictive modelling. Predictive discrimination was commonly assessed using area under the receiver operator curve (n=98, 42.42%) and calibration was rarely assessed (n=22, 9.52%). Conclusions: Machine learning applications in population health have concentrated on regions and diseases well represented in traditional data sources, infrequently using big data. Important aspects of model development were under-reported. Greater use of big data and reporting guidelines for predictive modelling could improve machine learning applications in population health. Registration number: Registered on the Open Science Framework on 17 July 2018 (available at https://osf.io/rnqe6/).