AI Model Predicts Patient Response to Bladder Cancer Treatment

Muscle-invasive bladder cancer (MIBC) is a severe form of bladder cancer where tumors invade the detrusor muscle of the bladder. The standard treatment for MIBC has been neoadjuvant chemotherapy (NAC) followed by radical cystectomy (RC), but this approach results in significant challenges. Radical cystectomy carries a mortality rate of 0.3–5.7%, along with considerable surgical morbidity, as 64% of patients experience postoperative complications within 90 days. Although around 35% of MIBC patients achieve a complete pathologic response (pCR), meaning no residual tumor remains after NAC treatment, predicting which patients will benefit from this treatment has been difficult due to the tumor's heterogeneity. Despite numerous studies in this area, developing accurate prediction models and identifying biomarkers that can reliably indicate how patients will respond to treatment have proven challenging. Now, researchers have developed a more effective model to predict the response of MIBC patients to chemotherapy.

This predictive model, created at Weill Cornell Medicine (New York, NY, USA), utilizes the power of artificial intelligence (AI) and machine learning. It combines whole-slide tumor imaging data with gene expression analyses, improving upon previous models that relied on only one data type. The study, published in npj Digital Medicine, highlights key genes and tumor characteristics that could determine how well patients respond to treatment. By accurately predicting an individual’s response to the standard treatment for MIBC, this model could help clinicians personalize care and potentially spare patients who respond well from undergoing bladder removal surgery.

To enhance the predictive capabilities of the model, the researchers used data from the SWOG Cancer Research Network, which designs and conducts multi-center clinical trials for adult cancers. They specifically integrated data from tumor sample images and gene expression profiles, which show which genes are activated or suppressed. For image analysis, they employed graph neural networks, a specialized AI technique that captures the arrangement and interaction of cancer cells, immune cells, and fibroblasts within the tumor. Automated image analysis was also used to identify the various cell types present in the tumor site. By combining the image-based data with gene expression profiles, the AI-driven model significantly outperformed models using either data type alone in predicting clinical response.

Looking ahead, the researchers plan to incorporate additional data, such as mutational analyses of tumor DNA, which can be detected in blood or urine, as well as spatial analyses to identify the precise cell types present in the bladder. The model also presented new hypotheses for further testing, such as the idea that the ratio of tumor cells to normal cells, like fibroblasts, can impact chemotherapy response predictions. Moving forward, the team aims to validate their findings with other clinical trial cohorts and is open to expanding their collaboration to determine whether their model can predict therapeutic response in a wider range of patients.

“We want to identify the right treatment for the right patient at the right time,” said co-lead Dr. Bishoy Morris Faltas. “The dream is that patients would walk into my office, and I could integrate all of their data into the AI framework and give them a score that predicts how they would respond to a particular therapy. It’s going to happen. But physicians like me will have to learn how to interpret these AI predictions and know that I can trust them—and to be able to explain them to my patients in a way they can also trust.”