This meta-analysis aimed to evaluate the diagnostic capability of machine learning (ML) models in predicting stone-free status following percutaneous nephrolithotomy (PCNL).

A comprehensive literature search was conducted across MEDLINE, Embase, Scopus, Cochrane, Google Scholar and supplementary databases was undertaken until June 2023. Inclusion criteria were English publications assessing the sensitivity and specificity of ML in predicting post PCNL stone-free status. Studies on non-human subjects or with incomplete data sets were excluded. Quality assessment utilized the Cochrane Risk of Bias Tool. Pooled sensitivity, specificity, and other diagnostic metrics were calculated using Meta-Disc 1.4 software.

Of the 65 initial articles, 5 met the inclusion criteria, representing a total of 1,773 participants. The accuracy of ML models ranged from 44% to 94.8%. The pooled sensitivity and specificity were 0.60 (95% CI [0.57, 0.63]) and 0.87 (95% CI [0.84, 0.89]), respectively. The pooled positive likelihood ratio was 4.69 (95% CI [3.82, 5.77]) and the negative likelihood ratio was 0.45 (95% CI [0.41, 0.48]). The diagnostic odds ratio was 10.93 (95% CI [8.35, 14.33]). The area under the curve (AUC) stood at 0.9372, signifying an excellent diagnostic performance.

Machine learning models demonstrate significant potential in accurately predicting stone-free status post-PCNL. However, the small number of included studies, retrospective designs, and heterogeneity in ML approaches limit generalizability. Standardized definitions, larger multicenter datasets, and prospective validation are required before routine clinical adoption.