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)

Abstract

AbstractObjectiveTo 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 Cox’s 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.