Development and validation of risk scores for all-cause mortality for the purposes of a smartphone-based ‘general health score’ application: a prospective cohort study using the UK Biobank

Background: Even though established links exist between individuals behaviours and potentially adverse health outcomes, to date either univariate, simpler models or multivariate, yet difficult to employ ones, have been developed. Such models are unlikely to be successful at capturing the wider determinants of health in the broader population. Hence, there is a need for a multidimensional, yet widely employable and accessible, way to obtain a comprehensive health metric. Objective: To develop and validate a novel, easily interpretable points-based health score (C-Score) derived from metrics measurable using smartphone components, and iterations thereof that utilise statistical modelling and machine learning approaches. Methods: Comprehensive literature review to identify suitable predictor variables for inclusion in a first iteration points-based model. This was followed by a prospective cohort study in a UK Biobank population for the purposes of validating the C-Score, and developing and comparatively validating variations of the score using statistical/machine learning models to assess the balance between expediency and ease of interpretability versus model complexity. Primary and secondary outcome measures: Discrimination of a points-based score for all-cause mortality within 10 years (Harrell s c-statistic). Discrimination and calibration of Cox proportional hazards models and machine learning models that incorporate C-Score values (or raw data inputs) and other predictors to predict risk of all-cause mortality within 10 years. Results: The cohort comprised 420,560 individuals. During a cohort follow-up of 4,526,452 person-years, there were 16,188 deaths from any cause (3.85%). The points-based model had good discrimination (c-statistic = 0.66). There was a 31% relative reduction in risk of all-cause mortality per decile of increasing C-Score (hazard ratio: 0.69, 95% CI: 0.663 to 0.675). A Cox model integrating age and C-Score had improved discrimination (8% percentage points, c-statistic = 0.74) and good calibration. Machine learning approaches did not offer improved discrimination over statistical modelling. Conclusions: The novel health metric (C-Score) has good predictive capabilities for all-cause mortality within 10 years. Embedding C-Score within a smartphone application may represent a useful tool for democratised, individualised health risk prediction. A simple Cox model using C-Score and age optimally balances parsimony and accuracy of risk predictions and could be used to produce absolute risk estimations for application users.