For decades, oncologists have worked with fragmented information. Imaging tests, biopsy results and clinical records often sit in separate databases, making it difficult to see how one piece of data affects another. The 2025 wave of studies suggests that artificial intelligence (AI) can close that gap by merging data from multiple sources into a single model that captures the full complexity of a patient’s disease.
Seeing the Full Picture of Cancer
A large study published in Nature Cancer used real-world data from more than 15,000 patients across 38 tumor types to test how well multimodal AI could predict outcomes. The model, trained on a mix of medical images, clinical notes and tumor biology, successfully identified the key factors that influenced survival and treatment response. Researchers said the results show how AI can detect subtle relationships between genetics, body composition and therapy success patterns that might be too complex for humans to spot.
A related study focused on stage II colorectal cancer, where doctors often struggle to decide whether patients need chemotherapy after surgery, multimodal AI usage proved to be helpful. By combining imaging, molecular and clinical data, the AI system improved prediction accuracy compared to current risk tools. The findings suggest that some patients could safely avoid additional treatment, while others who are at higher risk could be flagged for closer monitoring.
At the American Society of Clinical Oncology (ASCO) conference earlier this year, researchers presented another model that integrates pathology images, genetic markers, and patient histories for high-risk prostate cancer. The system helped identify which patients would benefit most from second-generation hormone therapies, pointing to how multimodal AI can make personalized treatment more precise.
From the Lab to the Clinic
While many multimodal AI tools are still being tested, several companies are starting to bring them closer to real-world use. BostonGene showcased its multimodal analytics platform, demonstrating how AI-driven data integration can reveal tumor behavior and immune interactions that traditional analyses miss.
The company said its platform combines genomic, transcriptomic, proteomic and digital pathology data to generate personalized “molecular portraits” of each tumor, enabling physicians to tailor therapies more effectively and researchers to accelerate biomarker discovery.
Flatiron Health presented new methods using large language models to extract key medical data from patient records, making it easier to build the integrated datasets multimodal systems rely on.
Practical Barriers
Experts say multimodal AI could make cancer care more personalized and cost-effective by helping doctors match patients to treatments more accurately and by improving how clinical trials are designed. AI models could one day create synthetic control groups reducing the need for large control populations in trials or continuously learn from real-world data to adjust treatment recommendations over time.
Still, large-scale deployment faces real challenges. Medical data are often stored in incompatible systems, making it difficult to integrate. Privacy regulations can also limit data sharing, even when anonymized.
Regulatory agencies are sharpening their approaches to AI in healthcare. The U.S. Food and Drug Administration published draft guidance in January 2025 that outlines how AI- and machine language-enabled devices should be evaluated across their full life cycle, including design, validation, marketing submission and post-market monitoring.
At the same time, the European Medicines Agency released a “Reflection Paper” on the use of AI throughout the medicinal-product life cycle covering discovery, clinical trials, manufacturing and post-authorization activities. Despite these moves, global harmonization is still lacking, which leaves many hospitals, AI developers and device makers unsure about how to validate, update and govern AI systems once deployed.
Source: https://www.pymnts.com/