To test if our artificial intelligence (AI)-postoperative glomerular filtration rate (GFR) prediction is as accurate as a validated clinical model. The American Urologic Association recommends estimating postoperative GFR in patients with renal masses and prioritizing partial nephrectomy (PN) when GFR would be <45 ml/minutes/1.73 m² if radical nephrectomy (RN) was performed. Previously validated models have limited clinical uptake.

We included 300 patients undergoing nephrectomy for renal tumors from the KiTS19 challenge. Preoperative GFR was collected just before surgery, and new baseline GFR 3-12 months postoperatively. Split-renal function (SRF) was determined in a fully automated way from preoperative computed tomography, combining our deep learning segmentation model, then using those segmentation masks to estimate postoperative GFR = 1.24 × GFR_Pre-RN × SRF_{Contralateral} for RN and 89% of GFR_preoperative for PN. A clinical model estimated postoperative GFR = 35 + GFR_preoperative x 0.65–18 (if RN)–age x 0.25 + 3 (if tumor>7 cm)⁻² (if diabetes). We compared the AI and clinical model GFR estimations to the measured postoperative GFR using correlation coefficients and their ability to predict GFR < 45 using logistic regression.

Median age was 60 years, 41% were female, and 62% had PN. Median tumor size was 4.2 cm, and 92% were malignant. Compared to the measured postoperative GFR, correlation coefficients were 0.75 and 0.77 for the AI and clinical models, respectively. The AI and clinical models performed similarly for predicting GFR < 45 (areas under the curve 0.89 and 0.9, respectively).

Our fully automated prediction of new baseline renal function is as accurate as a validated clinical model without needing clinical details, clinician time, or measurements.