Development of a digitally-obtainable 10-year all-cause mortality risk score based on data from 497,712 UK Biobank participants

Background: All-cause mortality (ACM) scores are a useful tool for identifying individuals with decreased life expectancy. An interpretable score consisting of smartphone-obtainable variables could allow for long-term management of individual health and support the next generation of healthcare monitoring and preventative practices. The aim of this study was to develop a 10-year ACM risk score using the UK Biobank dataset, using only digitally-obtainable variables. Methods The models were developed using the full UK Biobank cohort comprising nearly 500,000 individuals. We extracted 399 features from the dataset and, through a data-driven feature selection process with subsequent clinical review, identified 34 features for the final model. As part of the study, we compared two survival analysis approaches: Cox proportional hazards model and DeepSurv, a deep learning-based survival analysis algorithm. Results Before feature selection, Cox performed similarly to DeepSurv, achieving a c-index of 0.771 (95% CI 0.770–0.772) and 0.774 (95% CI 0.772–0.775) on the test dataset, respectively. Using the selected 34 features, the c-index of Cox decreased slightly to 0.770 (95% CI 0.769–0.770) and DeepSurv to 0.758 (95% CI 0.755–0.762). The models show excellent calibration at 10 years. Discussion This study describes a smartphone-compatible ACM score, with excellent predictive ability. The score described here improves on a previous smartphone-compatible score, C-Score, by incorporating non-modifiable factors in addition to variables which can be actively modified to reduce risk. Conclusion We have developed a smartphone-compatible ACM score that is comprehensive, easily interpretable and actionable, and as such, could provide a powerful tool for preventative healthcare.