AI approach predicts glomerular filtration rate after nephrectomy
Click Here to Manage Email Alerts
NASHVILLE, Tenn. — A fully automated artificial intelligence approach predicted post-nephrectomy glomerular filtration rate with comparable accuracy as validated clinical prediction models, according to study results.
The findings — presented at International Kidney Cancer Symposium: North America — show AI-based predictions can be incorporated into decision-making without clinical details or measurements, saving considerable clinician time, according to researchers.
“Treatment of patients with renal masses requires a multidisciplinary approach,” Nour Abdallah, MD, postdoctoral research fellow at Cleveland Clinic, told Healio. “Urologists and oncologists are always trying to find a personalized approach for each patient based on their needs. One of these needs is good kidney function after nephrectomy.
“It’s very difficult for clinicians to predict postoperative [glomerular filtration rate]. By creating this fully automated model, we are giving them a very accurate prediction in less time.”
American Urological Association recommends estimating postoperative glomerular filtration rate (GFR) among patients with renal masses, with the goal to prioritize partial nephrectomy over radical nephrectomy for those who have GFR less than 45 mL/min/1.73m2.
There are accurate validated models based on clinical equations or renal volumes from hand or semi-automated segmentations; however, uptake in clinical practice is limited.
Abdallah and colleagues sought to develop an AI GFR prediction calculated automatically on preoperative CT scan, with the goal to predict postoperative GFR with the same accuracy as a validated clinical model.
The analysis included 293 patients (median age, 60 years; interquartile range, 51-68; 40.6% women) who underwent partial nephrectomy (62.1%) or radical nephrectomy (37.1%). Median tumor size was 4.2 cm (range, 2.6-6.1). Most tumors (91.8%) were malignant, with 35.1% being high grade and 25.6% being high stage. About one in five (21.8%) had necrosis.
The closest recorded GFR value prior to surgery served as preoperative GFR, and postoperative GFR was recorded at least 90 days after surgery.
Researchers determined split-renal function in a fully automated way from preoperative CT and their deep learning segmentation model.
Abdallah and colleagues programmed their algorithm to estimate postoperative GFR through the following calculation: 1.24 x preoperative GFR x contralateral split renal function for radical nephrectomy and 89% of preoperative GFR for partial nephrectomy.
Researchers compared GFR estimations from AI and the clinical model with measured postoperative GFR using correlation coefficients. Investigators then used logistic regression and areas under the curves (AUC) to compare the models’ abilities to predict postoperative GFR less than 45.
When Abdallah and colleagues compared measured postoperative GFR, they calculated correlation coefficients of 0.75 for the AI model and 0.77 for the clinical model.
When researchers assessed prediction of postoperative GFR less than 45, results showed similar performance between the AI model (AUC, 0.89) and the clinical model (AUC, 0.9).
The AI model produces an instant result with comparable accuracy. It likely takes at least 5 minutes for clinicians to make this prediction on their own, Abdallah said.
“Because they are so busy, every second we can give back to them is extremely important,” Abdallah told Healio.
Abdallah and colleagues plan to validate the AI model in a larger cohort and eventually implement it into the treatment paradigm at Cleveland Clinic.
“Artificial intelligence is a tremendous asset in kidney cancer management, and we hope this work is a step toward increasing the trust physicians have in it,” Abdallah said. “We also believe it’s very important for clinicians to be active in the AI research world, because they have to ensure its proper and appropriate use in this field.”
Perspective
Back to Top
Sumanta Kumar Pal, MD, FASCO
One of the big challenges in kidney cancer is that when patients are eligible for surgery, we don’t know how to make the decision between radical nephrectomy — which is removal of the entire kidney — or partial nephrectomy, which is removal of a portion of it.
This research team has come up with an AI mechanism to predict what postoperative kidney function will look like based on a CT scan. This is something that my urology colleagues and I try to “guesstimate” before a patient goes to the operating room, but we’re incredibly bad at it. By looking at a CT scan, it’s very challenging to predict what postoperative kidney function will look like.
I am very impressed with these findings. It looks as though — with a very high AUC of 0.89 — this approach has been able to tease out who is going to develop deterioration of kidney function after surgery. This could be an immensely powerful tool to decide who to take to surgery and who not to take to surgery.
This is a huge area of need, and it’s not just relevant to urologists. Medical oncologists face a similar decision. We’re treating patients with upfront systemic therapy. and we really need to determine who might ultimately be a candidate down the line for cytoreductive nephrectomy, which is removal of the kidney in the metastatic setting. This gives us a good sense of whether a patient will be left with sufficient renal function afterward, so this is incredibly powerful.
Collapse
Abdallah N, et al. Abstract 5. Presented at: IKCS: North America; Nov. 9-11, 2023; Nashville, Tenn.
Click Here to Manage Email Alerts