Two clinical factors (BMI, WOMAC total score), 5 shape fea- tures, and 6 radiomic texture features were finally selected. The detailed feature names and their averaged model coefficients were listed in Table 1. Specifically, shape features weighted more than half (58.9%) in the proposed model, and texture features took 32.3% of the weightings, while the clinical features weighted only 8.8%. A higher discriminability of the radiomic-clinical model (testing AUC 1⁄4 0.85 (95CI: 0.76-0.93)) was observed than the baseline clinical only model (testing AUC 1⁄4 0.73 (95CI: 0.61-0.82)), as demonstrated by the separated receiver operating characteristic curves in Figure 1(a). Moreover, the proposed model (testing accuracy 1⁄4 0.81 (95CI: 0.73-0.88)) showed a higher binary prediction correctness than the baseline model (testing accuracy 1⁄4 0.64 (95CI: 0.57-0.73)). The prediction correctness of each testing patient is visualized by the probability distributions (Figure 1 (c)(d)), and bal- anced prediction accuracies can be observed for both models. Notably, 19 out of 25 PFOA positive testing cases were correctly identified by the proposed model, while 16 for the baseline model. The value of the model correctness metrics for both training and testing were listed in Table 2. Both models showed a good calibration demonstrated by the testing calibration curve in Figure 1 (b) and low Brier scores in Table 2.