JMIR Publications

ABSTRACT

Background:

Artificial intelligence (AI) integrated with point-of-care (POC) imaging has emerged as a promising approach to expand access to diagnostic capabilities in settings with limited specialist access. However, no systematic review has comprehensively evaluated AI-assisted clinical decision support across multiple POC imaging modalities, assessed explainability implementation, or quantified clinical impact evidence gaps.

Objective:

To systematically identify, evaluate, and synthesize evidence on AI-based clinical decision support systems utilizing point-of-care imaging for diagnostic purposes, with particular attention to task-shifting potential, explainability implementation, and clinical outcome evidence.

Methods:

We searched PubMed, Scopus, IEEE Xplore, and Web of Science (January 2018 to November 2025). We included research studies evaluating AI/machine learning systems applied to POC-capable imaging modalities in POC clinical settings with clinical decision support outputs. Two reviewers independently screened studies, extracted data across 15 domains, and assessed methodological quality using the QUADAS-2 tool. We developed novel frameworks to evaluate the implementation of explainability (XAI cascade) and clinical impact evidence (impact pyramid). Narrative synthesis was performed due to substantial heterogeneity in the data.

Results:

Of 2,113 records identified, 20 studies met inclusion criteria, encompassing approximately 78,296 patients across 15 countries. Studies evaluated tuberculosis (n=5), breast cancer (n=3), deep vein thrombosis (n=2), and nine other conditions using ultrasound (n=7, 35%), chest X-ray (n=5, 25%), photography-based and colposcopic imaging (n=3, 15%), fundus photography (n=2, 10%), microscopy (n=2, 10%), and dermoscopy (n=1, 5%). Median sensitivity was 92% (IQR 85.7-98.0%), and median specificity was 90.6% (IQR 70.0-95.7%). Task-shifting was demonstrated in 65% of studies (n=13), with non-specialists achieving specialist-level performance after a median of 1 hour of training. The XAI implementation cascade revealed critical gaps: 15 of 20 studies (75%) did not mention explainability, 2 studies (10%) provided explanations to users, and no studies (0%) evaluated whether clinicians understood the explanations or whether XAI influenced their decisions. The clinical impact pyramid showed 3 studies (15%) reported technical accuracy only, 13 studies (65%) reported process outcomes, 4 studies (20%) documented clinical actions, and no studies (0%) measured patient outcomes. Methodological quality was concerning, as 14 of 20 studies (70%) were at high or very high risk of bias, with verification bias (n=14, 70%) and selection bias (n=10, 50%) being the most common. The overall certainty of evidence was very low (GRADE ⊕◯◯◯), primarily due to risk of bias, heterogeneity, and imprecision.

Conclusions:

AI-assisted POC imaging demonstrates promising diagnostic accuracy and enables meaningful task-shifting with remarkably minimal training requirements. However, the field exhibits critical evidence gaps, including a complete absence of patient outcome measurement, inadequate evaluation of explainability, regulatory misalignment, and a lack of cross-context validation despite claims of global applicability. Gaps must be addressed through implementation research with patient outcome endpoints, rigorous XAI evaluation, and multi-context validation studies before widespread clinical adoption. Review limitations include restrictions to English-language publications, exclusion of grey literature, and substantial heterogeneity that precludes meta-analysis. Clinical Trial: This review was not registered.