Automated development of clinical prediction models enables real-time risk stratification with exemplar application to hypoxic-ischaemic encephalopathy

Abstract

AbstractReal-time updated risk prediction of disease outcomes could lead to improvements in patient care and better resource management. Established monitoring during pregnancy at antenatal and intrapartum periods could be particularly amenable to benefits of this approach. This proof-of-concept study compared automated and manual prediction modelling approaches using data from the Collaborative Perinatal Project with exemplar application to hypoxic-ischaemic encephalopathy (HIE). Using manually selected predictors identified from previously published studies we obtained high HIE discrimination with logistic regression applied to antenatal only (0.71 AUC [95% CI 0.64-0.77]), antenatal and intrapartum (0.70 AUC [95% CI 0.64-0.77]), and antenatal, intrapartum and birthweight (0.73 AUC [95% CI 0.67-0.79]) data. In parallel, we applied a range of automated modelling methods and found penalised logistic regression had best discrimination and was equivalent to the manual approach but required little human input giving 0.75 AUC for antenatal only (95% CI 0.69, 0.81), 0.70 AUC for antenatal and intrapartum (95% CI 0.63, 0.78), and 0.74 AUC using antenatal, intrapartum, and infant birthweight (95% CI 0.65, 0.81). These results demonstrate the feasibility of developing automated prediction models which could be applied to produce disease risk estimates in real-time. This approach may be especially useful in pregnancy care but could be applied to any disease.