JMIR Publications

ABSTRACT

Background:

Health resource planning in urban agglomerations is complicated by spatial heterogeneity of demand, unequal infrastructure accessibility, and varying institutional workload. While artificial intelligence offers potential for forecasting personnel, finances, and material needs while accounting for territorial accessibility, the evidence base on artificial intelligence applicability across different resource types remains fragmented.

Objective:

To systematically identify, appraise, and synthesize studies on the application of artificial intelligence, machine learning, and quantitative algorithmic approaches to forecasting and allocating human, financial, and material health care resources in urban agglomeration settings.

Methods:

Publications from 2016 to July 2026 on the application of artificial intelligence, machine learning, and deep learning, predictive and simulation modeling, and optimization methods to forecast, plan, allocate, or assess accessibility of human, financial, or material health care resources in urban or territorial settings. PubMed/MEDLINE, Scopus, Web of Science, and Google Scholar were searched from January 2016 to July 2026, supplemented by manual screening of reference lists. PROBAST, ROBINS-I, and the JBI Critical Appraisal Tools were applied according to study design by two independent reviewers, with a third expert in cases of disagreement.

Results:

Of 1021 records identified, 75 studies were included across four domains: human resources (n=18), financial resources (n=13), material/technical resources (n=32), and geographic accessibility (n=12). We found that: 1) across China the population-adjusted equity in health care worker distribution improved (Gini coefficient declined from 0.312 to 0.147) while the geographic equity remained poor (Gini 0.700-0.750). 2) decision-aware machine learning in Sierra Leone increased medicine consumption by an estimated 19%; 3) in the Netherlands, the number of radiology departments using artificial intelligence rose from 14 (2020) to 23 (2022), with a dedicated artificial intelligence budget in only 13 centers; 4) a surgical duration prediction model showed higher accuracy in identifying the risk of an overrun one hour in advance: 89% versus 80% for the Bayesian comparator. The heterogeneity of designs, resource types, outcomes, and performance metrics precluded meta-analysis. The predominance of studies from East Asia, particularly China, limits generalizability of the findings.

Conclusions:

High predictive performance does not guarantee territorial equity. For implementation at the agglomeration level, exemplified by the Almaty agglomeration, artificial intelligence applications require local model validation, parallel assessment of population-adjusted provision and geographic accessibility, and prespecified measurable performance indicators. Artificial intelligence should serve as an analytical decision-support component rather than a replacement for health planning mechanisms.