Researchers at Sylvester Comprehensive Cancer Center have discovered that artificial intelligence models can predict which cancer survivors are most likely to face emergency department visits, hospitalizations, and worsening symptoms—opening a new pathway to catch problems before they spiral out of control.
The challenge facing cancer care today is a paradox: just as patients finish their grueling treatments and need support most, their contact with care teams often drops off. Many survivors experience new or evolving health struggles months or years after chemotherapy and radiation end, yet without a clear way to identify who is most vulnerable, clinicians struggle to intervene early. This gap between the end of treatment and the start of vigilant monitoring can mean that treatable problems go unaddressed until they become urgent enough to send patients to the emergency room.
A new study published in JCO–Clinical Cancer Informatics suggests AI offers a solution. Led by Akina Natori, M.D., MSPH, an oncologist in the Division of Medical Oncology at Sylvester, the research team analyzed data from more than 25,000 cancer survivors tracked over three years. Rather than relying on clinical records alone—which capture test results, scans, and treatment history—the researchers incorporated patient-reported outcomes, or PROs: the patients' own accounts of how they were feeling, including emotional well-being, fatigue, functional limitations, and practical struggles that traditional medical data often misses.
The machine learning models proved remarkably effective at spotting risk. When researchers identified the highest-risk 10 percent of survivors, this small group accounted for roughly half of all subsequent emergency visits, hospitalizations, and severe symptom episodes. For acute events like ER visits and hospitalizations, recent clinical activity was the strongest signal—what happened in the last few months mattered more than baseline disease characteristics. But for predicting symptom burden, the picture shifted. Longer-term trends told a clearer story, and adding patient-reported outcomes nearly doubled how well the models performed compared with clinical data alone.
Frank J. Penedo, Ph.D., director of Sylvester's Survivorship and Supportive Care Institute and the study's senior author, described the stakes plainly: "For many patients, new or evolving challenges arise after treatment ends, just as routine clinical contact often tapers off, raising a critical question: how can we identify those at higher risk earlier, before these concerns intensify and become harder to address?"
The implications for cancer care are substantial. If AI tools can reliably forecast which survivors are heading toward trouble, care teams can intervene with targeted symptom management, psychosocial support, or closer monitoring—before a patient deteriorates to the point of needing emergency care. This represents a fundamental shift from reactive medicine, where patients seek help only when symptoms become unbearable, to proactive survivorship support grounded in data.
The research highlights why patient voices matter: symptoms, emotional struggles, and functional limitations that patients experience are often invisible in electronic health records. By treating PROs not as retrospective descriptions of what patients have endured, but as prospective indicators of future need, the researchers transformed data that care teams already have access to into actionable intelligence. For the millions of cancer survivors navigating life after treatment, this breakthrough in early identification could mean the difference between managing complications at home and facing an unexpected crisis.