Dados do Trabalho

Título

PREDICTION OF URINARY CONTINENCE RECOVERY USING ARTIFICIAL INTELLIGENCE/COMPUTER VISION ANALYSIS OF INTRAOPERATIVE ROBOTIC-ASSISTED RADICAL PROSTATECTOMY (RARP) VIDEO RECORDINGS.

Resumo

INTRODUCTION AND OBJECTIVE: The emergence of advanced video analytic tools such as Artificial Intelligence (AI) and computer vision may make it possible to objectively evaluate intraoperative surgical technique and patient-specific anatomical factors that may impact clinical outcomes. The objective of this study was to develop an AI algorithm to predict post-RARP surgery continence recovery using only intraoperative video. METHODS:

We selected a group of 50 patients with localized prostate cancer who underwent RARP surgery from a single surgeon (VP). 25 patients had “early” recovery of urinary continence outcomes, defined as not requiring urinary protection pads six weeks post-surgery and beyond. The remaining 25 patients had "prolonged" urinary incontinence, defined as still needing four or more urinary pads per day 12 months after the surgical procedure.

For each patient, a set of 7 video segments were extracted from their intraoperative video. Those segments were defined as follows;

1. Before anterior bladder-neck dissection

2. Following posterior bladder-neck dissection

3. Following incision of Denonvillier’s fascia and posterior dissection/nerve sparing

4. Following apical dissection

5. Following bladder-neck reconstruction

6. Following urethral posterior reconstruction

7. Following urethrovesical anastomosis

We utilized the segmented video recordings to develop an AI pipeline based on Transformer architecture and a weakly-supervised learning framework for continence outcome prediction in real-time.


RESULTS:

We evaluated the performance of the developed system using well-established evaluation metrics in a K-fold cross-validation setting based on patient stratification. Our method demonstrates an 84% accuracy, 83% F1 score, 88.0% specificity, and 80% sensitivity for predicting a binary outcome of whether a patient achieves “early” continence or “prolonged” incontinence.


CONCLUSIONS: These initial results demonstrate that using AI algorithms and surgical video to objectively evaluate and predict continence recovery outcomes following RARP is feasible and promising.