Introduction: Previously, our group utilized survival analyses to predict continence recovery after robot-assisted radical prostatectomy (RARP) by applying summary automated performance metrics (APMs) reported over whole RARP steps. This preceding work highlighted vesicourethral (VUA) APMs as important for continence prediction. Herein, we update our prediction results with the addition of detailed VUA APMs (metrics reported by sub-stitch increments) and technical assessment utilizing the Robotic Anastomosis Competency Evaluation (RACE).
Methods: APMs (instrument motion tracking/system events) and patient clinicopathological data (CPD) were collected for RARPs at our center. Two evaluators assigned RACE scores. The RARP was broken into 12 steps, each with 41 reported summary APMs. The VUA was further broken down into stitches and sub-stitches for detailed APMs. Deep learning-based predictive models (SVM, Survival Forests, Neural Multitask model and Deep Net with a Mixture of Experts) were trained with various combinations of three data sources: 1) summary vs detailed APMs, 2) CPD, and 3) RACE score to predict urinary continence recovery. Datasets were partitioned into five iterations of training (70%), validation (15%) and test (15%), using different seeds to split the data. Prediction performance reported as mean concordance index (CI) ± standard deviation. Continence recovery was defined as no pads or 1 safety pad.
Results: 91 RARP patients recovered continence in median 178.5 days (range 16-553). Median RACE score was 28.85/30 (25-30). Prediction with detailed VUA APMs combined with summary APMs of other RARP steps performed better than summary APMs alone (mean CI = 0.817 ± 0.056 vs 0.628 ± 0.045). CPD alone produced a mean CI = 0.662 ± 0.017. RACE alone produced a mean CI = 0.726 ± 0.083. Combined surgeon factors (detailed APMs + summary APMs + RACE) had a prediction performance of mean CI = 0.828 ± 0.018. Notably, RACE with detailed APMs only had mean CI = 0.830 ± 0.0049. Similarly, while the full dataset (detailed VUA + summary APMs + CPD + RACE) decreased performance to a mean CI = 0.79 ± 0.03, the full dataset without the summary APMs (detailed APMs + CPD + RACE) had the best performance with a mean CI = 0.837 ± 0.042.
Conclusions: Surgeon performance, as measured by APMs reported on the sub-stitch level during the VUA, outperformed summary APMs across all RARP steps to predict urinary continence recovery. Best performance by the deep learning models utilized surgeon metrics, while patient factors added little additional value. Source of
Funding: This study was funded in part by an Intuitive Surgical Clinical Grant; Intuitive Surgical provided the systems data recorder. Research reported in this publication was supported in part by the National Institute Of Biomedical Imaging And Bioengineering of the National Institutes of Health under Award Number K23EB026493.