Multimodal AI Approach May Help Identify Key Elements to Predict Outcomes in RCC

Christopher Weight, MD, MS

Segmentation and partial volumetric analysis demonstrated an improved chance of predicting post-operative kidney function in patients with renal cell carcinoma, and if quantified, this could be leveraged to personalize treatment, according to a presentation from the 2024 Kidney Cancer Research Summit.

“We are going to discover new features. Our next step is to compare this AI [tool] to ECOG. This will become available to all of our patients, and they will all receive a CT scan….This will offer us valuable information and allow us to personalize health care,” Christopher Weight, MD, MS, center director of urologic oncology at the Cleveland Clinic, said during the presentation.

Investigators had 2 aims for researching this multimodal AI approach. The first was to develop an accurate and non-invasive AI tool to evaluate patient information, laboratory tests, and characteristics. These capabilities may allow investigators to differentiate between benign or indolent kidney masses and identify potentially aggressive tumors in the preoperative setting. The second aim was to establish an AI tool using preexisting data to estimate post-operative kidney function through nephron-sparing surgery or total nephrectomy.

Initially, the goal was to evaluate approximately 1000 to 1500 patients. However, with additional technology like pathology and the radiology pipeline, investigators aim to evaluate an approximate total of 3000 patients. An additional 300 to 400 patients are being evaluated in a subset population for gene expression.

Investigators also recently submitted data regarding the age of patients. The hypothesis first focused on what it meant if the AI model incorrectly diagnosed a patient’s age based on their CT scans. Results found that if the AI model thought the patient was younger than they were, they would be discharged earlier from the hospital than those who were predicted to be older than their actual age.

This was also observed in survival, where if the AI tool predicted an age that did not correlate with the patient’s actual age, it was predictive of better or worse outcomes. Weight believes there are additional data to be extrapolated from these results, as there hasn’t been a way to truly analyze these outcomes before the development of this AI tool.

Weight said this trial was originally created based on an observation that an estimated $182 million dollars is spent on overdiagnosis and over-treatment of kidney cancers. Of those who proceed to surgery, it is estimated that 10% to 20% of their tumors are benign, 20% to 30% are indolent, and 30% to 50% are potentially aggressive.

To collect and analyze patient data, Weight outlined a 6-step process:

• The patient is diagnosed
• Patient data are collected
• Data points are analog
• Data points are digitized
• Data are analyzed
• A diagnosis or prognosis is made

However, between retrieving patient data and the analog stage, interobserver error or variability can occur. Between the process of organizing analog data points and digitizing them, human time and capital are wasted. Between digitization and analysis, data may become lost.

The multimodal AI system may reduce half of these steps and potentially limit human error or data loss. With this new procedure, the patient gives their data, which are then analyzed before a diagnosis or prognosis occurs.

Reference

1. Weight C. Multimodal AI-based renal cancer patient care. Presented at the 2024 Kidney Cancer Research Summit (KCRS); Boston, MA, July 11-12, 2024.