AACR 2026: Real-time AI may help clinicians catch sepsis earlier
Industry Buzz
Early recognition of sepsis in oncology patients is difficult because their baseline physiology is often already altered. Immunosuppression can blunt the typical inflammatory response, so patients may not present with classic signs like fever or elevated white blood cell count.
—Elizabeth Rubin, MD
[It] would have to be vetted over and over and reviewed retrospectively in research studies, with consistent validation to ensure accuracy and reliability across different patient populations.
—Jeffrey Velotta, MD
Sepsis detection in oncology remains difficult. Symptoms overlap with treatment toxicity, and delays are common, contributing to disease-related mortality. Cancer patients face a higher sepsis risk and worse outcomes, with immunosuppression, treatment toxicity, and atypical presentations often complicating the diagnosis.
In an attempt to improve early identification of sepsis, investigators studied the use of real-time AI models trained on multimodal EHR data. Their findings, to be presented at the upcoming 2026 AACR annual meeting, showed high accuracy in detecting sepsis in oncology populations.[]
Detection can be elusive
Jeffrey Velotta, MD, thoracic surgeon at Kaiser Permanente Northern California and expert contributor for The Mesothelioma Center at Asbestos.com, said, “Sepsis looks a lot like many other etiologies, so figuring out what is sepsis vs hypovolemia vs other things can be difficult, especially in complex oncology patients, where multiple conditions may be present at once.”
He further adds, “The other thing is, sepsis often does not grow out microorganisms, which makes it tough to find the source and can delay confirmation of the diagnosis.”
Traditional early warning systems rely on static thresholds and generate high false-positive rates. As explained by Elizabeth Rubin MD, a board-certified physician and Clinical Advisor at Embers Recovery, “Early recognition of sepsis in oncology patients is difficult because their baseline physiology is often already altered. Immunosuppression can blunt the typical inflammatory response, so patients may not present with classic signs like fever or elevated white blood cell count. This limits the effectiveness of traditional screening tools, which are largely based on these markers. In practice, clinicians often rely on clinical judgment, which introduces variability and can delay intervention.”
Can AI close the gap?
The investigators who will be presenting at AACR developed a real-time AI platform that integrated structured and unstructured EHR data, including vitals, labs, medications, and clinical notes.[]
They used it to analyze multimodal EHR data from 7,480 hospitalized patients, including cancer subgroups, to evaluate its ability to detect sepsis earlier. The model identified sepsis with high sensitivity, flagging about 86% of cancer patients with sepsis within four days of onset.
Patients flagged by the system had significantly higher ICU transfer and mortality rates, indicating the model captured clinically meaningful deterioration.
This aligns with prior literature. A meta-analysis of 30 studies,[] including approximately 5.5 to 6.0 million patients, found that machine learning models using EHR data have consistently shown higher diagnostic performance for early sepsis detection compared to traditional scoring systems. Some studies reported substantially higher sensitivity and earlier identification of sepsis onset.
Real-world use remains to be proven
Dr. Velotta expressed his opinion: “I think it has promise, particularly for analyzing large volumes of patient data in real time, though its effectiveness will depend on how well it performs in real-world clinical settings.”
In real-world implementation, AI-based sepsis early warning systems have been associated with improved outcomes, including reduced mortality when clinicians respond promptly to alerts.[]
Alert fatigue remains a barrier, with override rates exceeding 60% in some studies of clinical decision support systems.[]
Although specific data in oncology patients is limited, diagnostic overlap between infection, malignancy, and treatment toxicity further reduces specificity, increasing the burden of non-actionable alerts. AI models attempt to address this by prioritizing context and longitudinal data rather than isolated thresholds.
As explained by Dr. Rubin, “Regarding alert fatigue, AI has the potential to improve specificity and reduce unnecessary alerts. However, if poorly implemented, it can contribute to cognitive overload. The goal should be fewer, more actionable alerts that support, not replace, clinical decision-making.”
Clinicians are cautious. Dr. Velotta says, “[It] would have to be vetted over and over and reviewed retrospectively in research studies, with consistent validation to ensure accuracy and reliability across different patient populations.” He observes that these systems could add “another layer of complexity,” but he says he is “still open to it, especially if it can demonstrate clear benefits without adding unnecessary burden to clinical workflows."
Dr. Rubin echoes the sentiment, adding, “To build trust, clinicians need strong evidence that these tools improve outcomes, not just predictive accuracy. External validation, clear performance metrics, and usability at the bedside are key factors. Without that, adoption will remain limited.“