Artificial intelligence-based personalized oncological outcome prediction model for upper urinary tract urothelial carcinoma after radical nephroureterectomy: A development and multicenter validation.
To develop and validate an artificial intelligence (AI)-based personalized outcome prediction model for upper-urinary tract urothelial carcinoma patients undergoing radical nephroureterectomy.
Data from patients who underwent radical nephroureterectomy between 2010 and 2020 across three hospitals were retrospectively analyzed. A model was developed using one tertiary center's data and externally validated with data from two other hospitals. An AI model using XGBoost as risk estimator and bootstrapped Weibull Accelerated Failure Time model for 10-year survival probability was employed. Hyperparameter tuning used Optuna method. Model efficacy was assessed using concordance index, average Brier score, D-calibration, and six-month interval time-dependent area under the curve (AUC).
Of 1,039 patients, 627 qualified after excluding 50 with neoadjuvant chemotherapy. Model development used 564 patients (507 training, 57 test) with 9:1 stratified random split, plus 63 for internal validation and 362 for external validation. Significant parameters included preoperative glomerular filtration rate (p<0.001), hydroureteronephrosis (p=0.013), pathological N stage (p<0.001), concurrent carcinoma in situ (p<0.001), disease progression (p<0.001), and survival rate (p<0.001). Disease-free survival (DFS) model's concordance index: internal validation 0.789, external validations 0.734 and 0.771. Overall survival (OS) model's concordance index: internal validation 0.819, external validations 0.780 and 0.771. Mean time-dependent AUC was 0.66-0.77 for DFS and 0.67-0.80 for OS during 10-year periods.
AI-based model effectively predicts disease-free and OS outcomes for upper-urinary tract urothelial carcinoma patients with post-radical nephroureterectomy, showcasing robust performance across multicenter settings.
Data from patients who underwent radical nephroureterectomy between 2010 and 2020 across three hospitals were retrospectively analyzed. A model was developed using one tertiary center's data and externally validated with data from two other hospitals. An AI model using XGBoost as risk estimator and bootstrapped Weibull Accelerated Failure Time model for 10-year survival probability was employed. Hyperparameter tuning used Optuna method. Model efficacy was assessed using concordance index, average Brier score, D-calibration, and six-month interval time-dependent area under the curve (AUC).
Of 1,039 patients, 627 qualified after excluding 50 with neoadjuvant chemotherapy. Model development used 564 patients (507 training, 57 test) with 9:1 stratified random split, plus 63 for internal validation and 362 for external validation. Significant parameters included preoperative glomerular filtration rate (p<0.001), hydroureteronephrosis (p=0.013), pathological N stage (p<0.001), concurrent carcinoma in situ (p<0.001), disease progression (p<0.001), and survival rate (p<0.001). Disease-free survival (DFS) model's concordance index: internal validation 0.789, external validations 0.734 and 0.771. Overall survival (OS) model's concordance index: internal validation 0.819, external validations 0.780 and 0.771. Mean time-dependent AUC was 0.66-0.77 for DFS and 0.67-0.80 for OS during 10-year periods.
AI-based model effectively predicts disease-free and OS outcomes for upper-urinary tract urothelial carcinoma patients with post-radical nephroureterectomy, showcasing robust performance across multicenter settings.
Authors
Lee Lee, Kim Kim, Lim Lim, You You, Song Song, Jeong Jeong, Hong Hong, Hong Hong, Ahn Ahn, Jeong Jeong, Ku Ku, Suh Suh
View on Pubmed