Purpose <p>This study aims to synthesize the effectiveness of machine learning (ML) in the diagnosis and prediction of congenital heart disease (CHD), providing evidence for the subsequent development of ML models for CHD.</p> Methods <p>PubMed, EMBASE, Web of Science, and Cochrane were searched for studies on the application of ML in CHD up to April 8, 2024. The risk of bias was assessed using the prediction model risk of bias assessment tool (PROBAST). Subgroup analyses by task types (diagnosis and prediction) and different populations were performed.</p> Results <p>Fifty-two papers covering 850,866 subjects were included. The sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and summary receiver operating characteristic (SROC) curves of ML in the diagnosis of CHD were 0.90 (95%CI: 0.86–0.93), 0.90 (95%CI: 0.85–0.93), 8.6 (95%CI: 5.8–12.6), 0.11 (95%CI: 0.08–0.16), and 0.95 (95%CI: 0.82–0.99), respectively. The sensitivity and specificity of the ML model for predicting whether the fetus would develop CHD based on the clinical characteristics of pregnant women were 0.749 (95% CI: 0.69–0.81) and 0.876 (95% CI: 0.83–0.92), respectively. A mortality risk prediction model based on clinical characteristics had a sensitivity of 0.82 (95% CI: 0.71–0.90) and a specificity of 0.83 (95% CI: 0.77–0.88). The model for predicting coagulation status after CHD surgery had a sensitivity of 0.84 and a specificity of 0.70; the model for predicting malnutrition 1 year after CHD surgery had a sensitivity of 0.85 and a specificity of 0.88.</p> Conclusions <p>ML, especially image-based ML, appears to be an effective tool for the diagnosis and prediction of CHD. Nonetheless, this conclusion is drawn based on limited evidence. More and larger open datasets should be covered in future studies.</p>