Stevens--Johnson syndrome (SJS) and toxic epidermal necrolysis (TEN) are rare but severe diseases. This study aimed to validate the predictive ability of risk models in patients with SJS/TEN and propose possible refinement in China. Patients in the Department of Dermatology of Huashan Hospital from January 2008 to January 2019 were included. Results showed that the severity-of-illness score for TEN (SCORTEN) had a good discrimination (area under the receiver operating characteristic curve (AUC), 0.78), and it was superior to auxiliary score (AS) and ABCD-10, which indicates age, bicarbonate level, cancer, dialysis, and 10% involved body surface area (AUC, 0.69 and 0.68, respectively). The calibration of SCORTEN (Hosmer–Lemeshow goodness-of-fit test, = 0.69) was also better than that of AS ( = 0.25) and ABCD-10 ( = 0.55). SCORTEN and ABCD-10 were similar (Brier score (BS), 0.04 and 0.04) in terms of accuracy of predictions. In addition, the imaging appearance of pulmonary consolidation on computed tomography was associated with high mortality. Refined models were formed using the variables and this imaging appearance. The refined AS and ABCD-10 models were similar in discrimination compared with the original SCORTEN (0.74 vs. 0.78, = 0.23; 0.74 vs. 0.78, = 0.30, respectively). Therefore, SCORTEN showed good discrimination performance, calibration, and accuracy, and refined AS or ABCD-10 model may be an option when SCORTEN variables are not available.