Validation of a prediction model for post‐chemotherapy fibrosis in nonseminoma patients

Abstract

ObjectiveTo validate Vergouwe's prediction model using the Swedish and Norwegian Testicular Cancer Group (SWENOTECA) RETROP database and to define its clinical utility.Materials and methodsVergouwe's prediction model for benign histopathology in post‐chemotherapy retroperitoneal lymph node dissection (PC‐RPLND) uses the following variables: presence of teratoma in orchiectomy specimen; pre‐chemotherapy level of alpha‐fetoprotein; β‐Human chorionic gonadotropin and lactate dehydrogenase; and lymph node size pre‐ and post‐chemotherapy. Our validation cohort consisted of patients included in RETROP, a prospective population‐based database of patients in Sweden and Norway with metastatic nonseminoma, who underwent PC‐RPLND in the period 2007–2014. Discrimination and calibration analyses were used to validate Vergouwe's prediction model results. Calibration plots were created and a Hosmer–Lemeshow test was calculated. Clinical utility, expressed as opt‐out net benefit (NBopt‐out), was analysed using decision curve analysis.ResultsOverall, 284 patients were included in the analysis, of whom 130 (46%) had benign histology after PC‐RPLND. Discrimination analysis showed good reproducibility, with an area under the receiver‐operating characteristic curve (AUC) of 0.82 (95% confidence interval 0.77–0.87) compared to Vergouwe's prediction model (AUC between 0.77 and 0.84). Calibration was acceptable with no recalibration. Using a prediction threshold of 70% for benign histopathology, NBopt‐out was 0.098. Using the model and this threshold, 61 patients would have been spared surgery. However, only 51 of 61 were correctly classified as benign.ConclusionsThe model was externally validated with good reproducibility. In a clinical setting, the model may identify patients with a high chance of benign histopathology, thereby sparing patients of surgery. However, meticulous follow‐up is required.