Diagnostic Accuracy of Machine Learning Models on Mammography in Breast Cancer Classification: A Meta-Analysis

In this meta-analysis, we aimed to estimate the diagnostic accuracy of machine learning models on digital mammograms and tomosynthesis in breast cancer classification and to assess the factors affecting its diagnostic accuracy. We searched for related studies in Web of Science, Scopus, PubMed, Google Scholar and Embase. The studies were screened in two stages to exclude the unrelated studies and duplicates. Finally, 36 studies containing 68 machine learning models were included in this meta-analysis. The area under the curve (AUC), hierarchical summary receiver operating characteristics (HSROC) curve, pooled sensitivity and pooled specificity were estimated using a bivariate Reitsma model. Overall AUC, pooled sensitivity and pooled specificity were 0.90 (95% CI: 0.85-0.90), 0.83 (95% CI: 0.78-0.87) and 0.84 (95% CI: 0.81-0.87), respectively. Additionally, the three significant covariates identified in this study were country (p = 0.003), source (p = 0.002) and classifier (p = 0.016). The type of data covariate was not statistically significant (p = 0.121). Additionally, Deeks' linear regression test indicated that there exists a publication bias in the included studies (p = 0.002). Thus, the results should be interpreted with caution.