Diagnostic accuracy of machine-learning models for detecting pulmonary embolism in patients with pneumonia: a systematic review and meta-analysis

Pulmonary embolism (PE) in patients with pneumonia presents a major diagnostic challenge because of overlapping clinical features and the limited specificity of conventional tools. Machine learning (ML)–based models have been proposed to improve diagnostic accuracy by integrating clinical, laboratory, and imaging data. Five databases were searched from inception to May 2026. Fifteen studies met inclusion criteria, of which ten provided sufficient data for quantitative synthesis. The primary meta-analysis was restricted to studies reporting patient- or examination-level binary PE detection against an imaging or adjudicated reference standard; clinical prediction, segmentation/localization, thrombosis, and mortality outcomes were analyzed separately, and overlapping public-dataset studies were reduced to one estimate per dataset. Pooled sensitivity, specificity, and summary receiver operating characteristic estimates were calculated using a bivariate random-effects model. Between-study heterogeneity was assessed using the I² statistic. Meta-regression and subgroup analyses explored sources of heterogeneity. Study quality was evaluated using the PROBAST + AI tool, and certainty of evidence was assessed using the GRADE framework. Across included studies, ML-based models showed high overall diagnostic discrimination for PE detection in pneumonia. The primary analysis of six patient/examination-level studies yielded a pooled sensitivity of 0.91 (95% CI 0.76–0.97; 95% prediction interval 0.16–0.99), specificity of 0.92 (95% CI 0.68–0.98; 95% prediction interval 0.02–1.00), and a summary AUROC of 0.96 (95% CI 0.94–0.98). However, substantial heterogeneity (I 2 = 77.8% overall; sensitivity I² = 76.0%, specificity I² = 85.3%) was observed across pooled estimates. Image-based deep-learning models applied to CTPA generally showed higher accuracy than clinical or hybrid approaches. According to GRADE, the overall certainty of evidence was rated as very low. Prediction intervals were wide, and the pooled operating point should not be read as the expected performance of a deployable model. Differences by reference standard and clinical setting were examined descriptively and are hypothesis-generating only. ML-based models show promise as adjunctive tools for PE detection in pneumonia, but current evidence is limited by heterogeneity and low certainty. Prospective, multicenter validation and standardized evaluation are required before clinical implementation. The current evidence is heterogeneous, indirect, and insufficient to establish the performance of a deployable tool in any individual clinical setting.