Validation and Comparison of Various Models for Cancer Screening

The receiver operating characteristic (ROC) curve was used to evaluate the performance of the SVM-, KNN-, LR-based cancer screening models and the single tumour markers in panels. ROC curves for all machine learning methods and tumour markers were generated using SPSS (Version 20; SPSS Inc.). Furthermore, the area under the curve (AUC) was calculated to compare the discrimination abilities of machine learning methods and single tumour markers. Moreover, the performance of these machine learning methods was tested using the validation set. The performance of the combined test was also evaluated. The algorithm of the combined test was based on the threshold of each tumour marker. The thresholds of the tumour markers used in this study were 15 ng/mL for AFP, 5 ng/mL for CEA, 37 U/mL for CA19-9, 3.3 ng/mL for CYFRA21-1, 2.5 ng/mL for SCC, 4 ng/mL for PSA, 35 U/mL for CA125, and 30 U/mL for CA15-3. The sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Youden index of the cancer screening models were calculated for the machine learning methods and the combined tests. The 95% confidence interval (CI) of the Youden indices of each method was calculated for further analysis, as detailed in previous studies [17, 18]. Moreover, for clinical consideration, the relative risk reduction (RRR), absolute risk reduction (ARR: cancer screened), and absolute risk increase (ARI: false diagnosis) were evaluated.