Abstract
Background: Laboratory-based frailty indices (FI-Lab) offer potential alternatives to manual assessment in emergency care settings, but how should we select features and time-frames to find the best balance between coverage and performance? We evaluated multiple FI-Lab configurations to determine the optimal configuration requirements for reliable automated frailty assessment.
Methods: We analyzed 74,493 ED visits from 54,075 patients aged ≥70 years across two London hospitals (2017-2021), comparing five FI-Lab configurations and a drug-adjusted version against nurse-assessed Clinical Frailty Scale scores. Configurations varied in observation windows (12-36 months), minimum data requirements (1-10 months of data), and calculation approaches. Outcomes included 90-day mortality, length of stay, and readmission risk.
Results: Nurse assessments consistently showed superior outcome discrimination (c-statistic 0.726 for 90-day mortality), though automated measures performed strongly (best FI-Lab c-statistic 0.718). FI-Lab measures demonstrated effect sizes comparable to age for mortality prediction (HR range 1.37-1.55 per standard deviation), indicating clinical relevance. The mean-type FI-Lab showed the strongest automated performance (HR 1.29, 95% CI 1.22-1.37), but notably, even configurations requiring minimal data maintained similar predictive validity. Information criteria suggested automated measures provided more consistent scoring (drug-adjusted AIC=-45,984 vs nurse assessment AIC=116,715), though with some loss of predictive power.
Conclusions: While nurse assessments predicted outcomes best, the similar performance across FI-Lab configurations suggests that complex data requirements may be unnecessary for effective automated frailty screening. Given their complementary strengths - clinical insight from nurse assessment and scoring consistency from automated measures - a combined approach could enhance frailty screening in emergency care. Additionally, automated frailty screening could help triage nursing assessments. Further work is needed to identify the minimum feature set that ensures maximum population coverage, while maintaining consistent results regardless of the number of features available to provide reliable automated adjuncts to clinical assessment.