Mortality prediction models for community-dwelling older adults: A systematic review.

As complexity and comorbidities increase with age, the increasing number of community-dwelling older adults poses a challenge to healthcare professionals in making trade-offs between beneficial and harmful treatments, monitoring deteriorating patients and resource allocation. Mortality predictions may help inform these decisions. So far, a systematic overview on the characteristics of currently existing mortality prediction models, is lacking. To provide a systematic overview and assessment of mortality prediction models for the community-dwelling older population. A systematic search of terms related to predictive modelling and older adults was performed until March 1st, 2024, in four databases. We included studies developing multivariable all-cause mortality prediction models for community-dwelling older adults (aged ≥65 years). Data extraction followed the CHARMS Checklist and Quality assessment was performed with the PROBAST tool. A total of 22 studies involving 38 unique mortality prediction models were included, of which 14 models were based on a cumulative deficit-based frailty index and 9 on machine learning. C-statistics of the models ranged from 0.60 to 0.93 for all studies versus 0.61–0.78 when a frailty index was used. Eight models reached c-statistics higher than 0.8 and reported calibration. The most used variables in all models were demographics, symptoms, diagnoses and physical functioning. Five studies accounting for eleven models had a high risk of bias. Some mortality prediction models showed promising results for use in practice and most studies were of sufficient quality. However, more uniform methodology and validation studies are needed for clinical implementation. [Display omitted] • Multiple mortality prediction models exist for community-dwelling older adults. • Diagnoses and demographics were the most common predictors. • Models showed varying discrimination and calibration, some had promising results. • Transparent reporting of prediction model development was often incomplete. [ABSTRACT FROM AUTHOR]