Kidney Cancer: An AI Model Estimates Risk Before Surgery

An artificial intelligence model developed by Vita-Salute San Raffaele University (UniSR) and IRCCS San Raffaele Hospital, in collaboration with the University of Florence, estimates cancer-specific mortality risk in patients with non-metastatic kidney cancer before surgery. The study, published in Nature Communications, draws on eight clinical parameters already collected in routine care. Researchers developed the model on a cohort of 2,536 patients treated at San Raffaele Hospital, then validated it on an independent cohort of 580 patients from the Careggi University Hospital in Florence. Built according to the principles of Explainable AI, the model shows the contribution of each variable to the risk estimate, rather than returning a result as a “black box”. 

Renal cell carcinoma is the most common form of kidney cancer. Surgery remains the standard treatment for localised disease, but roughly one patient in three experiences recurrence or progression. Identifying high-risk patients before surgery matters because it shapes the entire course of treatment: the surgical strategy, post-operative surveillance, and any eventual systemic therapy. 

The model uses eight variables, all already available in clinical practice before surgery: tumour size, lymph node involvement, haemoglobin, platelet count, kidney function, age, body mass index, and performance status, a standardised clinical measure of a patient's overall condition. None of these require additional tests beyond the diagnostic work-up already in use. 

The model also follows the principles of Explainable AI, an approach that makes the contribution of each clinical variable to the risk estimate visible and understandable, rather than returning a “black box” result. This feature supports its future integration into clinical practice, since doctors and patients can check which factors weigh most heavily on the prediction. 

In clinical practice, we see patients every day who have very different prognoses even when their tumours look similar. Having a tool that can estimate risk before surgery means adding an objective element to clinical evaluation, in support of increasingly personalised decisions», he adds. «The model doesn't replace the doctor's judgement, t's designed to support it.

, says Dr Alessandro Larcher, urologist at the Urology Unit of IRCCS San Raffaele Hospital.

The algorithm was developed on a cohort of 2,536 patients treated at San Raffaele Hospital in recent years, then validated on an independent cohort of 580 patients from the Careggi University Hospital in Florence. External validation strengthens the reliability of the results beyond the setting where the model originated. Among studies proposing prognostic models based solely on preoperative variables, this is currently the one conducted on the largest patient cohort, with performance that outperforms the leading prognostic models currently available for non-metastatic kidney cancer. 

The study is a new application of S-RACE, the platform developed by UniSR and IRCCS San Raffaele Hospital to transform data collected in everyday clinical practice, known as Real World Data, into tools that support research and clinical decisions, according to the principles of responsible, interpretable artificial intelligence. 

Through a dedicated pipeline, an automated sequence of procedures that organises, quality-checks and prepares clinical data for analysis, researchers turned thousands of records collected in routine care into a dataset ready for model development. Compared with the traditional process of manual data selection curated by an expert clinician, the automated pipeline showed comparable performance. It can therefore reliably support specialists in preparing data for research. The algorithm identified two new variables, in addition to six predictors already present in the manually curated clinical dataset. The finding opens the door to biomarkers not considered by current prognostic models from urological societies, identified in an agnostic way by the AI based solely on the data collected. 

This study is concrete proof of what S-RACE can do. Our platform lets us integrate large volumes of clinical data collected in everyday practice, developing models that are accurate as well as transparent and understandable.This is a crucial step if these technologies are to be adopted with confidence, first in research and, in time, in clinical practice.

, says Dr Alberto Traverso, Scientific Lead of the S-Race Data Science group, Centre of Excellence for Artificial Intelligence.  

Reaching this result depended on a multidisciplinary effort: urologists, radiologists, AI experts, data scientists and bioinformaticians worked together on the model's development and validation. 

This study grew out of the meeting between established clinical experience in managing kidney cancer and advanced expertise in artificial intelligence and data science. Developing models like this means putting to use the wealth of clinical data gathered over more than thirty years of activity», he adds, «turning it into knowledge that improves patient care. 

, notes Professor Andrea Salonia, urologist, andrologist, Director of the Urological Research Institute (URI) and Full Professor of Urology at Vita-Salute San Raffaele University. 

The theme of responsibility also runs through the words of Professor Carlo Tacchetti, Professor of Human Anatomy and Director of the AI Strategic Programme at Vita-Salute San Raffaele University:

Artificial intelligence can create real value for medicine only if it is developed according to criteria of transparency, robustness and verifiability. Our goal is not to build algorithms as an end in themselves, but reliable tools able to support doctors in clinical decisions. That's why the model has also been made available through a web application, which will make it easier to validate in other centres and could, in future, support its integration into clinical research pathways.

, Tacchetti concludes. 

The study was funded by the Italian Ministry of University and Research as part of the D³4Health project – Digital Driven Diagnostics, Prognostics and Therapeutics for Sustainable Healthcare. 

Read the full paper here.