Artificial Intelligence-guided Personalized Surgical Planning: AI-physician Avatar Assistance in Partial Nephrectomy

Artificial intelligence (AI) has the potential to transform surgical practice, particularly through predictive algorithms for postoperative outcomes. However, these algorithms often lack user- friendly interfaces for clinician interaction. This study aimed to evaluate an AI-driven model designed to predict the optimal surgical approach for maximizing functional outcomes in patients undergoing robot-assisted partial nephrectomy (RAPN) integrated with a graphical interface that enables direct verbal physician-algorithm interaction. Methods We integrated our machine learning (ML) algorithm, based on random forest regression and trained on preoperative and intraoperative data to predict postoperative three-month eGFR, with a generative AI system. This system enables direct verbal interaction with physicians through a virtual physician avatar (VPA). Preoperative characteristics of patients scheduled for live surgery RAPNs during an international meeting were analyzed by the ML algorithm, which identified the optimal patient-specific combination of intraoperative maneuvers (clamping strategy, resection technique, and suturing approach) to maximize renal function preservation. Prior to the live surgery, surgeons participated in real-time verbal communication with the VPA. If they disagreed with the algorithm’s suggested intraoperative strategy, they stated their preferred approach. In turn, the VPA provided the estimated three-month eGFR drop for the surgeon’s selected strategy. Technical issues encountered during live interaction with the VPA were recorded, along with VPA response delay.

Additionally, the concordance among the VPA’s proposed surgical plan, the surgeon’s plan, and the strategy actually performed intraoperatively was assessed. Finally, we evaluated the accuracy of the ML algorithm in predicting the three-month eGFR drop for the executed intraoperative combination. Results Twelve RAPN preoperative plannings were discussed with the VPA during the meeting. No technical issues were observed. Mean (SD) VPA response delay was 2.2 (0.3) seconds. The strategy recommended by the VPA was accepted by the surgeon in 7/12 (58.3%) cases. In cases where the surgeon disagreed with the algorithm’s recommended approach, the VPA provided an estimated three-month eGFR drop prediction for the surgeon’s selected combination in all instances. In every case, either the VPA’s proposed strategy or the surgeon’s preferred strategy was successfully executed intraoperatively. The accuracy of the three-month eGFR drop prediction, whether the VPA’s or the surgeon’s strategy was followed, was 88.3%. Conclusion The integration of a ML algorithm with a generative AI system, enabling direct oral interaction with a VPA, demonstrates significant potential for improving surgical planning in RAPN. The system showed high accuracy (88.3%) in predicting three-month eGFR outcomes, even when the surgeon chose an alternative strategy.