Analysis of three models to predict pathohistology in patients undergoing postchemotherapy RPLND for (pcRPLND) advanced nonseminomatous germ cell tumors (NSGCT)
Sie können Bookmarks mittels Listen verwalten, loggen Sie sich dafür bitte in Ihr SLUB Benutzerkonto ein.
Medientyp:
E-Artikel
Titel:
Analysis of three models to predict pathohistology in patients undergoing postchemotherapy RPLND for (pcRPLND) advanced nonseminomatous germ cell tumors (NSGCT)
Beschreibung:
<jats:p> e16053 </jats:p><jats:p> Background: We aim to validate the two best performing prediction models (Vergouwe and Leao) for final pathohistology in NSGCT patients undergoing pcRPLND and we introduce a new radiomics approach. Methods: A cohort of 496 patients treated between 2008 and 2018 was used to validate the 2 prediction modelsusing published formulas and thresholds.For group comparisons, we used t-test or chi-square test. ROC were plotted (sensitivity against 1-specificity) and AUC was calculated. We determined the optimal cut point and used bootstrapping (1,000 replications) to estimate its variability. A p-value of < 0.05 was considered statistically significant. For radiomics, lymph nodes identified on CT images, were semiautomatically segmented with 93 radiographic features (pyRadiomics package). A linear support vector machine algorithm was applied to analyze reproducible radiomics features. A continuous reduction of features analyzed was performed using Random Forest algorithms and ROC analysis. Results: The Vergouwe model had a significantly better AUC compared to Leao model (0.749 [CI 0.706-0.792] vs. 0.689 [0.642-0.736], p = 0.004) to predict benign histology. At a threshold of > 70% for the probability of benign disease, the Leao model would have avoided RLND in 8.6% with benign disease with an error rate of 5.6% for viable tumor. The Vergouwe model would avoid pcRPLND in 23.4% with benign disease with an error rate of 12.7% for viable tumor/teratoma. Of the 93 radiomic features analyzed, 51 features were reproducible. Applying the trained algorithm resulted in a AUC of 82% with a diagnostic sensitivity of 81% and a specificity of 83%. Conclusions: The discriminatory accuracy of both models is not sufficient to safely select patients for surveillance strategy instead of pcRPLND. The radiomics model is promising but needs prospective validation. Further studies including new biomarkers are needed to optimize the accuracy of potential prediction models. </jats:p>