Feed

Gastric cancer (GC) remains a significant global health challenge with high mortality rates, often due to late-stage diagnosis. We hypothesize that Raman spectroscopy (RS) (a modern minimally invasive technique that uses light to analyze the molecular composition of tissue, generating a unique "fingerprint" that reveals biochemical details, distinguishing between normal and diseased tissues.) when combined with Machine learning (ML) would provide accurate and expedite approach of detecting GC. We aim to meta-analyze the diagnostic accuracy of ML-enhanced RS in differentiating GC component from normal tissue. This study was conducted following PRISMA-DTA guidelines. We searched PubMed, Scopus, Web of Science, VHL, and Google Scholar up to the end of February 2025. with an updated search conducted on 14 July 2025. We included any peer-reviewed manuscript that assessed ML-based RS technique for detecting GC components against normal control during endoscopy and reported sufficient data to construct 2 × 2 contingency table for assessing basic diagnostic metrics such as the sensitivity and specificity were included. Methodological quality of studies deemed eligible was assessed using QUADAS-2 risk of bias tool. Data on true positives, true negatives, false positives, and false negatives were extracted to calculate pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), diagnostic odds ratio (DOR), and area under the summary receiver operating characteristic curve (AUC) using R software. Heterogeneity was assessed with I² statistics and Deeks' funnel plot was employed to examine potential publication bias. Moreover, we further subgrouped individual study metrics based on source of sample, RS technique, AI model, and the experimental context to assess their role in solidify results by controlling several confounders for heterogeneity. A total of 28 studies were enrolled comprising 2,392 patients and 8861 gastric spectra. Twenty-one studies (75%) applied per-spectra approach to analyze the diagnostic utility for GC tissue detection from non-pathological tissue. On the other hand, seven studies (25%) approached analysis as of per-patient stratification evaluating GC patients from healthy subjects. The pooled estimates of the sensitivity and specificity of per spectra approach yielded 92% (95% CI: 88-95%) and 93% (95% CI: 89-96%), respectively, and the AUC was 0.955. On the other hand, the pooled analysis of studies implemented per patient assessment approach yielded excellent sensitivity, specificity, and AUC as well with 95% (95% CI: 87-98%), 93% (95% CI: 89-95%), 0.928, respectively. Subgroup analyses showed that studies using the KNN model demonstrated the highest diagnostic accuracy. Conventional Raman spectroscopy also achieved superior performance across most metrics. Serum-based samples yielded higher sensitivity and specificity than tissue samples, though the limited number of serum studies warrants cautious interpretation. In vitro studies showed slightly better diagnostic accuracy than in vivo studies, although the difference was not statistically significant. Substantial heterogeneity was observed across per-spectra studies (I² = 82.1% for sensitivity and 91.2% for specificity), but no significant between-study variation was observed in per-patient analyses (I² = 27.3% for sensitivity and I² = 0% for specificity). No substantial publication bias was detected based on Deeks' funnel plot asymmetry test, with p = 0.394 for the per-spectra analysis and p = 0.858 for the per-patient analysis. Our meta-analyses' results provide strong evidence that ML-enabled RS from different body sources and across various ML algorithms subtypes significantly improve the GC detection rate and is superior at differentiating GC from healthy tissue during upper GI endoscopy. This approach led to a more precise and real-time decision-making regarding biopsy and excision, given our excellent diagnostic accuracy and low between-study heterogeneity that we obtained, integrating ML-enhanced detection of significant spectra into clinical workflows as valuable diagnostic adjunct particularly during the endoscopy will optimize false negative rate and overall patient outcomes.