JMIR Publications

ABSTRACT

Background:

Inpatient fall prevention is a frontline nursing responsibility that requires early risk screening, continuous monitoring, and timely delivery of preventive actions under constrained staffing. Existing bedside scales and many EHR-centric prediction models often provide limited linkage from risk identification to actionable, nurse-led prevention pathways and may incompletely report calibration, temporal validation, or decision-analytic utility. Comprehensive geriatric assessment (CGA) is routinely collected in geriatric wards and captures modifiable domains relevant to falls, but it is rarely leveraged for implementation-oriented prediction.

Objective:

To develop and temporally validate a CGA-derived multidomain prediction model to identify hospitalized older adults at high fall risk and to evaluate discrimination, calibration, and clinical utility for nurse-led screening.

Methods:

We conducted a single-center retrospective cohort study in a geriatric inpatient unit (The First Affiliated Hospital of Ningbo University), including patients aged ≥65 years admitted from January 2017 to January 2025 (N=355). High fall risk was defined as Morse Fall Scale (MFS) ≥50. Prespecified predictors were routinely collected CGA-domain measures, including age, sex, BMI, functional status (ADL, IADL), cognition (MMSE), sleep quality (PSQI), nutrition (MNA-SF), depressive symptoms (GDS-15), urinary incontinence (ICIQ-UI SF), mobility/balance (Tinetti/POMA), handgrip strength, and inability to complete a 4-meter walk test. We fitted an elastic-net regularized logistic regression model and evaluated internal performance using 5-fold stratified cross-validation with out-of-fold predictions. Discrimination (AUC), overall error (Brier score), calibration (intercept, slope), and decision curve analysis (DCA) were assessed; uncertainty was quantified via bootstrap resampling (2,000 resamples). Temporal validation used an a priori split (training: 2017–2022; test: 2023–2025).

Results:

Of 355 hospitalizations, 187 (52.7%) were classified as high fall risk (MFS ≥50). Internal validation showed moderate discrimination (AUC=0.694) and acceptable overall error (Brier=0.224), with good calibration (intercept=−0.123; slope=0.987). Using the Youden-index threshold (0.512), sensitivity was 0.722 and specificity 0.527 (PPV=0.659; NPV=0.602). Bootstrap internal validation yielded an AUC 95% percentile interval of 0.639–0.749. DCA indicated higher net benefit than “treat-all” and “treat-none” strategies across threshold probabilities of approximately 0.25–0.70. In temporal validation (training n=196; test n=159; test event rate 66.7%), discrimination remained stable (AUC=0.726; Brier=0.250) while calibration shifted (intercept=1.063; slope=0.895).

Conclusions:

The CGA-derived model achieved stable discrimination and favorable decision-analytic utility, while temporal calibration drift suggests the need for recalibration over time. The model’s domain-based structure is intended to support nurse-led, tiered fall-prevention workflows. Clinical Trial: Medical Research Registration (medicalresearch.org.cn); MR-33-26-012211; https://www.medicalresearch.org.cn/