JMIR Publications

ABSTRACT

Background:

Falls and fall-related injuries are major causes of morbidity, mortality, loss of independence, and healthcare use among older adults. Although traditional fall-risk assessment tools are widely used, their predictive performance is often modest and inconsistent across populations and care settings. Digital prediction models based on wearable sensors and electronic health records (EHRs) may support risk stratification, but their predictive performance and readiness for clinical implementation remain uncertain.

Objective:

Within the Italian Digital Lifelong Prevention (DARE) project, we systematically reviewed evidence on digital prediction tools for falls and fall-related fractures in selected populations.

Methods:

We conducted a systematic review guidance for reviews of prediction model studies and reported in accordance with PRISMA 2020. Building on NICE NG249 evidence review, we updated and expanded the search from May 2024 to 6 October 2025 in MEDLINE, Embase, CENTRAL, and Epistemonikos. We included cohort studies evaluating EHR-based, AI/ML-based, rule-based, or wearable sensor-based tools in adults aged ≥65 years, or adults aged 50–64 years with conditions associated with increased fall risk. We extracted the point estimate of the area under the curve (AUC) and its corresponding 95% confidence interval. Risk of bias and applicability were assessed using PROBAST+AI, and certainty of evidence using a modified GRADE approach. Findings were synthesized narratively because of clinical and methodological heterogeneity.

Results:

Of 3,380 unique records screened after removal of duplicates, 6 new studies met inclusion criteria. Combined with 17 from the NICE review, 23 studies were analyzed: 15 EHR-based, 8 wearable-based. EHR-based models showed AUCs from 0.69 to 0.82 in community settings, 0.57 to 0.85 in hospitals, and up to 0.74 in externally validated residential care. Wearable devices showed AUCs 0.60-0.84 in internally validated studies The highest wearable estimate arose from a small internally validated emergency-department cohort. Calibration and clinical utility were infrequently assessed. Risk of bias was predominantly high, primarily due to limitations in the analysis domain. Certainty of evidence was low or very low across all evidence groups. No study provided separately extractable fracture-specific discrimination and calibration estimates.

Conclusions:

Some EHR-based models reported moderate-to-good discrimination, but direct comparison with wearable-based models was not possible because populations, outcomes, settings, and validation strategies differed. Limited external validation, heterogeneous outcome definitions, sparse calibration reporting, and frequent methodological limitations preclude conclusions about readiness for routine clinical implementation. Evidence for fracture-specific prediction remains absent.