Session: PD43: Kidney Cancer: Localized: Surgical Therapy V
PD43-06: Fully Automated AI-generated models predict post-operative glomerular filtration rate after renal surgery with similar accuracy to a validated clinical model.
Introduction: The American Urologic Association recommends estimating postoperative (postop) glomerular filtration rate (GFR) in patients with renal mass to prioritize partial (PN) over radical nephrectomy (RN) when postop GFR < 45 mL/min/1.73m2. Validated models based on clinical equations or renal volumes from hand or semi-automated segmentations are quite accurate but have limited uptake in clinical practice. We aimed to develop an artificial intelligence (AI)-GFR prediction calculated automatically on a preoperative (preop) computed tomography (CT) scan to predict a postop GFR as accurately as a validated clinical model. Methods: Three hundred patients having PN or RN for renal tumor from the KiTS21 challenge were analyzed. We excluded 7 patients having bilateral tumors. Preop GFR was the closest recorded value, and postop GFR was =90 days. Split-renal function (SRF) was determined in a fully automated way from preop CT and our deep learning segmentation model. We programmed the algorithm to estimate postop GFR as 1.24×preop GFR×contralateral SRF for RN and 89% of preop GFR for PN. We compared AI-predicted GFR to a validated clinical model (GFR = 35 + preop GFR (x0.65) -18 (if RN) -age (x0.25) +3 (if tumor size > 7 cm) - 2 (if diabetes)). We compared AI and clinical model estimations of GFR to the measured postop GFR using correlation coefficients (R) and compared their ability to predict postop GFR < 45 using logistic regression and areas under the curves (AUC). Results: In 293 patients, median age was 60 years ((IQR) 51-68), 40.6% were female, and 62.1% had PN. Median tumor size was 4.2 (2.6-6.1), and 91.8% of the tumors were malignant, of which 35.1% were high-grade, 25.6% high-stage, and 21.8% had necrosis. When comparing measured postop GFR, the correlation coefficients were 0.75 and 0.77 for AI and clinical models, respectively. For the prediction of postop GFR < 45, the AI and clinical models performed similarly (AUC 0.89 and 0.9, respectively). Conclusions: Our study introduces a fully automated prediction of postop GFR based on CT and baseline GFR with comparable predictive accuracy to validated clinical prediction models. These AI-based predictions can be implemented for decision-making without clinical details, clinician time, or measurements needed. SOURCE OF Funding: None
