A prediction algorithm to improve the accuracy of the Gold Standard Framework Surprise Question end-of-life prognostic categories in an acute hospital admission cohort-controlled study. The Proactive Risk-Based and Data-Driven Assessment of Patients at the End of Life (PRADA).

ObjectiveTo determine the accuracy of a clinical data algorithm allocated end-of-life prognosis amongst hospital inpatients. MethodThe model allocated a predicted Gold Standard Framework end-of-life prognosis to all acute medical patients admitted over a 2-year period. Mortality was determined at 1 year. ResultsOf 18,838 patients, end-of-life prognosis was unknown in 67.9%. A binary logistic regression model calculated 1-year mortality probability (X2=6650.2, p<0.001, r2 = 0.43). Probability cut off points were used to triage those with unknown prognosis using the GSF Surprise Question "Yes" or "No" survival categories (> or < 1 year respectively), with subsidiary classification of "No" to Green (months), Amber (weeks) or Red (days). This digitally driven prognosis allocation (100% vs baseline 32.1%) yielded cohorts of GSFSQ-Yes 15,264 (81%), GSFSQ-No Green 1,771 (9.4%) and GSFSQ-No Amber or Red 1,803 (9.6%). There were 5,043 (26.8%) deaths at 1 year. In Coxs survival, model allocated cohorts were discrete for mortality (GSFSQ-Yes 16.4% v GSFSQ-No 71.0% (p<0.001). For the GSFSQ-No classification, the mortality Odds Ratio was 12.4 (11.4 - 13.5) (p<0.001) vs GSFSQ-Yes (c-statistic of 0.71 (0.70 - 0.73), p<0.001; accuracy, positive and negative predictive values of 81.2%, 83.6%, 83.6% respectively. If this tool had been utilised at the time of admission, the potential to reduce subsequent hospital admissions, death-in-hospital, and bed days was all p<0.001. ConclusionsThe defined model successfully allocated end-of-life prognosis in cohorts of hospitalised patients with strong performance metrics for prospective 1 year mortality, yielding the potential to provide anticipatory care and improve outcomes. What is already known about this topic?End-of-life care is fragmented with excessive hospital admission and death in hospital. Current processes to determine end-of-life prognosis and promote anticipatory care for better outcomes are of limited utility. What this paper adds?A patient centric data integration model permitted the development of a digital health care system (PRADA) which allows the use of advanced analytics to accurately determine end-of-life prognosis among those where it was otherwise unknown. This paper demonstrates the potential benefit of integrating this prediction tool into routine care, at scale, in large population-level cohorts. Implications for practice, theory, or policyIn an era of advancing opportunity from informatics driven heath care, NHS policy, through commissioning to direct care, must now actively deploy such evidence-based digital systems into direct care, most specifically in data sharing across provider boundaries. We particularly hope the research community might consider testing and validating this approach.