Diagnostic Accuracy of Machine-Learning Models on Predicting Chemo-Brain in Breast Cancer Survivors Previously Treated with Chemotherapy: A Meta-Analysis

Abstract

We performed a meta-analysis of chemo-brain diagnostic, pooling sensitivities, and specificities in order to assess the accuracy of a machine-learning (ML) algorithm in breast cancer survivors previously treated with chemotherapy. We searched PubMed, Web of Science, and Scopus for eligible articles before 30 September 2022. We identified three eligible studies from which we extracted seven ML algorithms. For our data, the χ² tests demonstrated the homogeneity of the sensitivity's models (χ² = 7.6987, df = 6, p-value = 0.261) and the specificities of the ML models (χ² = 3.0151, df = 6, p-value = 0.807). The pooled area under the curve (AUC) for the overall ML models in this study was 0.914 (95%CI: 0.891-0.939) and partial AUC (restricted to observed false positive rates and normalized) was 0.844 (95%CI: 0.80-0.889). Additionally, the pooled sensitivity and pooled specificity values were 0.81 (95% CI: 0.75-0.86) and 0.82 (95% CI: 0.76-0.86), respectively. From all included ML models, support vector machine demonstrated the best test performance. ML models represent a promising, reliable modality for chemo-brain prediction in breast cancer survivors previously treated with chemotherapy, demonstrating high accuracy.

Keywords: breast cancer; chemo-brain; chemotherapy; diagnostic accuracy; machine learning.

Figure 1

Flow diagram of the study selection process according to PRISMA guidelines.

Figure 2

Methodologic quality of the included studies according to QUADAS-2 tool. Green represents low, yellow unclear and red high risk of bias (Lin 2021 [15], Wang 2022 [16], Chen 2019 [17]).

Figure 3

Sensitivity and specificity of machine learning models in the study. The names of the models are used accordingly to the models detailed in Table 1.

Figure 4

(A) Weighted crosshair plot with arbitrary coloring. (B) Summary plot with the proportional hazard model approach (PHM). The crosshairs displayed the uncertainty in sensitivity and specificity for every model. The solid line represents SROC curve, plotted together with the dotted lines representing confidence interval in the summary. The circle lines represent the 95% confidence regions for the model estimates.

Figure 5

Forest plot for a univariate meta-analysis using the diagnostic odds ratio. The names of the models are used accordingly to the models detailed in Table 1.

Figure 6

Hierarchical summary receiver operating characteristics—HSROC curve for overall machine learning models in the study.