JMIR Publications

ABSTRACT

Background:

Population ageing, multimorbidity, and polypharmacy are increasing the burden of adverse events among community-dwelling older adults managed in primary care. Primary care nurses are well positioned to detect early clinical deterioration and trigger proactive interventions; however, the scope, characteristics, and real-world readiness of predictive models applicable to primary care nursing remain unclear.

Objective:

To identify, describe, and synthesise predictive models published between 2010 and 2025 that are applicable to primary care nursing for the early detection of adverse events in community-dwelling adults, and to summarise evidence on model performance, external validation, implementation, and key methodological gaps.

Methods:

A scoping review was conducted between September and November 2025 following PRISMA-ScR guidance. Searches were performed in PubMed/MEDLINE, Scopus, and Web of Science using MeSH/DeCS terms related to predictive modelling, machine learning/artificial intelligence, adverse events, primary care/community care, and nursing. Peer-reviewed studies published in English or Spanish were eligible if they developed, validated, evaluated, or implemented predictive models for adverse events relevant to primary care or community settings and potentially applicable to nursing practice. Two reviewers independently screened titles/abstracts and full texts, with disagreements resolved by consensus involving a third reviewer. Data were extracted using a standardised form and synthesised narratively due to heterogeneity.

Results:

Seven primary studies (reported in eight articles, 2013–2025) met the inclusion criteria, spanning five countries (the Netherlands, the United Kingdom, Denmark, Singapore, and Spain). Identified models ranged from structured nurse-administered or clinician-judgement tools (EASY-Care TOS; RISC) to risk equations derived from primary care electronic health records (QMortality/QFrailty), telemonitoring-based time-series approaches (including an LSTM model for heart failure readmissions), and attention-based deep learning models trained on longitudinal primary care data (SIDIAP). Reported discrimination was generally moderate to high (AUROC ≈ 0.70–0.88, depending on outcome and prediction horizon). Robust multicentre external validation was clearly documented only for QMortality/QFrailty, whereas most other models lacked formal external validation and provided limited calibration reporting. Operational deployment was described in a minority of studies, mainly in telemonitoring programmes and population-level risk stratification initiatives. One additional systematic review was identified to contextualise evidence on implementation and risk-stratification-linked interventions but was not counted as a primary modelling study.

Conclusions:

The evidence highlights a complexity–implementation paradox: more sophisticated models tend to achieve higher statistical performance but face greater barriers to integration into nursing workflows, whereas structured clinical instruments are more feasible but demonstrate more modest performance and limited impact evaluation. Across studies, there was a near-complete absence of equity assessment (e.g. stratified performance by sex, ethnicity, or socioeconomic status), limited geographic and temporal validation, and scarce evidence of clinical impact from controlled studies. Future progress is likely to depend less on incremental algorithmic advances and more on rigorous external and temporal validation, calibration and monitoring for drift, explicit equity evaluation, and the linkage of risk predictions to clearly defined, nurse-led protocols and implementation strategies.