The accuracy of an electronic-Surprise-Question defining end-of-life cohorts in a whole adult population by algorithmic digital risk stratification: the Proactive Risk-Based and Data-Driven Assessment of Patients at the End of Life (PRADA).

Abstract

Current methods for identifying end-of-life cannot be applied systematically to large populations. We have developed, tested, validated a mortality probability algorithm with that level of scalability. This was a prospective whole adult population cohort study in Wolverhampton, a high deprivation, multiethnic city in the UK. Integrated hospital, community and primary care data spanned 2.5 years on 236,321 adults (age ≥18 years) including 6153 who had died. A binary logistic regression model (p < 0.001) generated mortality probability. This was triaged in a 2-step algorithm, based on care process measures and probability cut points. This digital enquiry, termed the e-Surprise-Question (e-SQ), allocated prognostic categories of e-SQ-Yes and e-SQ-No (>1, ≤1 year survival respectively). Those alive at baseline were followed prospectively (n = 230,168, e-SQ-Yes (n = 217,625), e-SQ-No (n = 12,543). At 12 months, mortality was 2753 (1.2%), with 1366 (0.6%) in e-SQ-Yes vs 1377 e-SQ-No (11.0%, 50% of all deaths, OR 19.4 (17.9-20.9), p < 0.001 (binary logistic regression)). The model's ROC c-statistic for 1-year mortality was 0.73 (0.72-0.74) (p < 0.001) and sensitivity, specificity, positive and negative predictive values 50.0%, 95.1%, 11.0%, and 99.4% respectively. This methodology is applicable at scale, anticipating mortality prognosis with statistical significance and clinically meaningful accuracy. The prognostic findings can be presented to clinicians for validation, further assessment and care planning for improved outcomes. South Staffordshire Medical Centre Charitable Trust Rotha Abraham Bequest (Charity number 509324) and the Royal Wolverhampton NHS Trust Charity (Charity number 1059467).