Presentation Authors: Andrew Hung, Jian Chen, Los Angeles, CA, Zequn Liu, Beijing, China, People's Republic of, Jessica Nguyen*, Los Angeles, CA, Sanjay Purushotham, Baltimore, MD, Yan Liu, Los Angeles, CA
Introduction: Urinary continence is an important outcome after a robot-assisted radical prostatectomy (RARP). Survival analysis (time-to-event analysis) models predicting time to urinary continence recovery have not yet been trained with validated and objective measures of surgeon performance, such as automated performance metrics (APMs). Our study provides three predictive models (conventional regression analysis, machine learning, and deep learning algorithms) with APMs and clinicopathological data (CPD) to predict time to urinary continence recovery after RARPs.
Methods: APMs (instrument motion tracking and system events data) and CPD were collected for 161 RARPs. The RARP was broken down into 12 steps, and for each step, 41 developed and validated APMs were reported. The predictive models were trained with three data sets: 1) APMs only; 2) CPD (i.e. prostate volume, Gleason score, PSA, pathological stage) only; 3) APMs + CPD. We utilized Cox Proportional Hazards (CPH), Random Survival Forests (RSF), and Deep learning Models based Survival analysis (DeepSurv) to compare conventional regression analysis, machine learning, and deep learning, respectively. Concordance Index (CI) and Mean Absolute Error (MAE) measured prediction performance. For the top-performing model, data inputs (APMs and CPD) were ranked based on importance for urinary continence (no pads or 1 safety pad) prediction.
Results: Out of the 161 RARP patients, 100 had at least three months follow-up data. 79% of patients achieved urinary continence with a median time of 126 days (16-553 days). DeepSurv, constructed on APMs and CPD, achieved the greatest prediction performance, with a high CI (0.60) and low MAE (85.9). Trained on the same data set, the CPH model had a CI 0.54 and MAE 134.7, while the RSF model had CI 0.58 and MAE 101.2. For DeepSurv, as the superior survival analysis model, only APMs (no CPD) were ranked in the top 25 data inputs for urinary continence prediction. Furthermore, three of the five top-ranked features were APMs measuring right instrument articulation during the vescio-urethral anastomosis and prostatic apical dissection steps of the RARP.
Conclusions: When trained with APMs and CPD, deep learning algorithms (DeepSurv) proved to be the superior survival analysis model to predict time to urinary continence recovery after a RARP. This model showed that APMs have a greater influence on time to urinary continence prediction.
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 also supported in part by the National Institute Of Biomedical Imaging And Bioengineer