Abstract Real-time, proactive detection has become a necessity as health systems face converging pressures that legacy tools cannot solve: rising acuity, workforce strain, and persistent quality gaps that erode margins, capacity, and reputation. Clinical decision support has historically lagged because rudimentary rules-based systems drown clinicians in false alerts, eroding trust without reliably changing outcomes at the bedside. Bayesian Health’s real-time clinical intelligence platform was built to close this gap by continuously synthesizing multimodal EHR data (vitals, labs, medications, procedures, and longitudinal history) to learn patient-specific baselines and surface only the most actionable, high-precision signals to care teams. Across diverse health systems, this platform has been used to enable earlier recognition and intervention for a range of high-impact conditions, improving both bedside care and system performance. This talk will focus on concrete, real-world applications of the platform for sepsis, clinical deterioration, and palliative care needs, illustrating how earlier detection and workflow-integrated interventions drive measurable gains in clinical outcomes and margins, and why the value of early, precise detection is amplified in high-risk populations such as oncology patients, whose higher baseline acuity and treatment-related toxicities both elevate their risk of harm and make emerging deterioration harder to recognize. Case Study: Detecting Sepsis within a Oncology Patient Subpopulation. Background: Sepsis remains one of the leading causes of in-hospital mortality, and traditional rules-based alerts have struggled to deliver timely, actionable detection without overwhelming clinicians with false alarms. Oncology patients experience an elevated risk for sepsis and septic shock, and those who experience sepsis are more likely to suffer in-hospital mortality than non-cancer patients with sepsis. Early recognition of sepsis onset is challenging due to the overlap between signs of sepsis and inflammation and tumor- and treatment-related toxicities. We investigated the ability of Bayesian Health’s real-time artificial intelligence-based computational algorithm to perform timely detection of sepsis onset in hospitalized patients with cancer. This algorithm ingests data directly from the electronic health record, such as vital signs, laboratory results, medications, procedures, and patient-specific historical results, processing it in a continuous manner to flag presence of sepsis as early as possible. Methods: Analysis included 7480 adult emergency and inpatient encounters from three tertiary and quaternary academic centers (Cleveland Clinic Main Campus, Sinai Hospital of Baltimore, and Strong Memorial Hospital) and one community hospital (Cleveland Clinic Fairview Hospital) with sufficient data for inclusion (age ≥ 18, presence in an emergency or inpatient care unit, and at least one measurement of a panel of vital signs). Patients with cancer (including subgroups of solid tumor and hematologic malignancy) were identified by current or past recorded diagnoses (ICD-10 codes). All included patient records underwent a multi-tiered retrospective adjudication process. An automated screen labeled records with no evidence of infection or organ dysfunction as non-sepsis, removing them from need for manual review. Two independent physicians then adjudicated a sample of the eliminated records in addition to each remaining record for both presence of sepsis and time of onset, with a third physician reconciling disagreements. Evaluation of the algorithm included all discrete flags generated by the system during a period of time it was not in clinical use and not visible to care teams. Results: In the overall population, 3. 1% of all patients and 6. 0% of cancer patients were adjudicated to have had sepsis, including zero from a sample of 200 of the records marked as non-sepsis in automated screening. The detection algorithm identified sepsis with high and consistent accuracy across all patients and across cancer patients, including those with solid tumors and with hematologic malignancies (Table). A flag for sepsis occurred in 85. 8% of all sepsis patients and 86. 1% of cancer patients with sepsis within 4 days of sepsis onset. Sepsis was present in 24. 2% of all cancer patients flagged (22. 9% of solid tumor patients and 27. 9% of hematologic malignancy patients). Compared to patients who were never flagged, all patients who were flagged, whether they had sepsis or not, were more likely to experience in-hospital mortality (7. 9% vs 0. 3%, p 0. 001) or ICU transfer (38. 2% vs 7. 9%, p 0. 001). This was also true within oncology patients (in-hospital mortality: 8. 8% vs 0. 4%, p 0. 001; ICU transfer: 34. 7% vs 9. 7%, p 0. 001). Implementation of an automated sepsis detection system with high accuracy in populations such as cancer patients, who have elevated risk of poor outcomes following sepsis, provides an opportunity to meaningfully improve timely delivery of care and reduce subsequent morbidity and mortality while minimizing alert fatigue for clinicians. Citation Format: Suchi Saria, Varesh Prasad, Nicasia Beebe-Wang, Ellie Zhang, Neri Cohen, David Graham, Conrad Gleber, James Morrison, Jacob Strand, Alexandra Tallman. Real-time multimodal AI enables individualized care in challenging subpopulations abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts) ; 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86 (8Suppl): Abstract nr PL03-03.
Saria et al. (Fri,) studied this question.