An Assessment of the Predictive Performance of Current Machine Learning–Based Breast Cancer Risk Prediction Models: Systematic Review

Abstract

Background Several studies have explored the predictive performance of machine learning–based breast cancer risk prediction models and have shown controversial conclusions. Thus, the performance of the current machine learning–based breast cancer risk prediction models and their benefits and weakness need to be evaluated for the future development of feasible and efficient risk prediction models. Objective The aim of this review was to assess the performance and the clinical feasibility of the currently available machine learning–based breast cancer risk prediction models. Methods We searched for papers published until June 9, 2021, on machine learning–based breast cancer risk prediction models in PubMed, Embase, and Web of Science. Studies describing the development or validation models for predicting future breast cancer risk were included. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used to assess the risk of bias and the clinical applicability of the included studies. The pooled area under the curve (AUC) was calculated using the DerSimonian and Laird random-effects model. Results A total of 8 studies with 10 data sets were included. Neural network was the most common machine learning method for the development of breast cancer risk prediction models. The pooled AUC of the machine learning–based optimal risk prediction model reported in each study was 0.73 (95% CI 0.66-0.80; approximate 95% prediction interval 0.56-0.96), with a high level of heterogeneity between studies (Q=576.07, I2=98.44%; P<.001). The results of head-to-head comparison of the performance difference between the 2 types of models trained by the same data set showed that machine learning models had a slightly higher advantage than traditional risk factor–based models in predicting future breast cancer risk. The pooled AUC of the neural network–based risk prediction model was higher than that of the nonneural network–based optimal risk prediction model (0.71 vs 0.68, respectively). Subgroup analysis showed that the incorporation of imaging features in risk models resulted in a higher pooled AUC than the nonincorporation of imaging features in risk models (0.73 vs 0.61; Pheterogeneity=.001, respectively). The PROBAST analysis indicated that many machine learning models had high risk of bias and poorly reported calibration analysis. Conclusions Our review shows that the current machine learning–based breast cancer risk prediction models have some technical pitfalls and that their clinical feasibility and reliability are unsatisfactory.