Introduction: PC-RPLND for metastatic non-seminoma is a challenging procedure. In up to 50% of patients undergoing PC-RPLND, the histopathology results reveal necrosis/fibrosis, and the surgery was hence not necessary.
Several different prediction models have been introduced with the aim of selecting the right patients for surgery. The aim of this study was to try to improve the most used model (by Vergouwe) by adding new variables and to test it on our Swedish Norwegian Testicular Cancer Group (SWENOTECA) RETROP data. RETROP is a population-based dataset of patients with non-seminoma that underwent PC-RPLND between 2007 and 2014.
Methods: Patients with Non-Seminoma Germ Cell Tumour (NSGCT) in Sweden and Norway that underwent PC-RPLND between 1st September 2007 and 1st September 2014 were included. Information for the study regarding teratoma in orchiectomy specimen, lymph node size pre and post chemotherapy, AFP, HCG and LDH levels before chemotherapy, chemotherapy given, Royal Marsden clinical stage, prognostic group according to the IGCCCG and histopathology results from PC-RPLND were obtained at time of surgery, from SWENOTECA register and chart review.
Statistical analysis
Discrimination and calibration analyses were used to validate Vergouwes results. Calibration plots were created and Hosmer–Lemeshow test was calculated. Clinical utility expressed as Net benefit were analyzed using Decision curve analysis.
The original algoritm was trained with Random Forest, a machine learning program. Additional data such as IGCCCG prognostic group, Royal Marsden clinical stage, number of chemotherapy courses, AFP, HCG and lymph node shrinkage as continuous variables with non-linear restricted cubic spines were added.
Results: In total 284 patients met the criteria to be included in this study.
Discrimination analysis showed good reproducibility with AUC of 0.819 (95% CI 0.765 – 0.863) compared to Vergouwes original study in 2007 with AUC between 0.77 and 0.84. The calibration plot, as well as Hosmer-Lemeshow test (p=0.44) showed good calibration.
For patients with post-chemo lymph nodes between 10-19 mm, 14 % would be classified as false negatives using the model with a 70% prediction level.
Machine learning did not improve the model.
Conclusions: The model was validated in this material with good reproducibility. For clinical use the model needs further development with new variables.
.jpg)
.jpg)