Real-Time Multimodal AI for Proactive, Individualized Care

At the American Association for Cancer Research (AACR) Annual Meeting 2026 (Abstract PL03-03), Suchi Saria, PhD, positioned real-time multimodal AI as a means to address a core limitation in clinical practice: care is “reactive and random,” resulting in variability in quality across patients and settings. The opportunity, noted Dr. Saria, the John C. Malone Endowed Chair and Associate Professor of Computer Science at Johns Hopkins University, Baltimore, and CEO and Founder of Bayesian Health, is to enable proactive, evidence-based cancer care through systems that are easy to adopt, safe, reliable, and responsible, and that think like a clinical team and act as a trusted partner.

“What we want to enable, and what I think is possible [with AI is] bringing high-quality, consistent care to every visit," she said. 

Limitations of Current Approaches

Dr. Saria noted that existing tools, particularly rules-based or analytic alerting systems, were not built to achieve this goal. These systems can miss approximately 40% of cases, often generating alerts after treatment has already started, producing roughly 50 alerts per true positive, and are therefore adopted at a low rate of 15%. Care quality stays inconsistent when alerts are largely ignored, she emphasized.

She attributed these limitations to several challenges: “needle-in-a-haystack” conditions in many patient safety and quality areas, alarm fatigue from simple rules-based alerts, variability and subtypes that cause generic models to fail, and the risk of overfitting multimodal data. The importance of clinically phenotyped labels, as well as lead time to achieve outcomes, was also highlighted.

Real-Time Multimodal AI Architecture

To address these barriers, Dr. Saria described the four-part architecture of the Bayesian Health real-time clinical intelligence platform:

• Multimodal data integration: longitudinal integration of electronic health record (EHR) data beyond text-based clinical notes, incorporating event-based and continuous time-series data across labs, vitals, medications, and treatments, to build patient-specific baselines

• Mixture-of-experts models: real-time interpretation of data, routing patients to specialized models based on phenotype and evolving signals

• Agentic clinical workflows: modular agents embedded in the EHR that individualize risk, explain why a patient is flagged, “tee up” orders and documentation automatically to reduce friction, and enable reporting to track what happens in response and evaluate system performance

• Monitoring and governance: continuous oversight and validation of population-level performance and detection of drift or shifts in clinical practice

A key realization, she said, was that “just to drop a flag or alert isn’t as helpful to clinicians as it is to be able to have full workflows, [because] workflows that make it possible to act on that signal…are just as critical.”

Translating Performance Into Outcomes

Dr. Saria presented a sepsis case study as a proof point for this approach. She presented an analysis of 7,480 patients from across four health systems using the Bayesian Health platform; in this group, 1,307 patients had cancer or a history of any cancer. Among this subpopulation, 6% had sepsis, as confirmed by expert adjudication; the rate was higher (10%) among patients with current or a history of a hematologic malignancy (n = 231). 

Compared with existing tools, the system improved sensitivity (53% to 85%), increased earlier identification (ie, before treatment initiation; 14% to 69%, with a 3.5-hour earlier lead time), and reduced alert volume nearly 10-fold. As an “exciting” result, she explained that these gains translated into clinical impact, with adoption rates of 80% to 90% and improvements including reduced delays to treatment (−58%), shorter length of stay (−23%), and lower mortality (−33%).

Beyond sepsis, Dr. Saria described a similar approach in palliative care, where reliable, earlier, and appropriate identification of its need increased consultations (44% improved conversion), reduced hospital readmissions (−28%), and lowered costs ($3,700 per case).

“The ability to deploy these tools to not just have insight but drive action is ultimately what impacts patient outcomes,” she emphasized. “This opportunity is not just limited to these two use cases, but [extends across] many different clinical condition areas,” with implications for oncology. 

Enabling Adoption

A recurring theme was that impact depends on adoption. This requires adaptive AI that understands context, collaborative change management, tech-enabled process improvement, and workflow design that makes “the right thing the easy thing to do,” Dr. Saria stated. 

Across deployments, these approaches have resulted in 80% to 95% clinical adoption of the Bayesian Health system.

In this context, Dr. Saria noted that the work of her team was adopted across a variety of systems, stating, “We are not just thinking of AI as the kind of thing that is implemented for the wealthiest or most pristine hospitals, but we really want to bring this value to very diverse populations and very diverse settings.”

DISCLOSURES: Dr. Saria is Founder and CEO of Bayesian Health and holds equity in the company. Johns Hopkins University also holds equity in Bayesian Health. Under a license agreement between Bayesian Health and Johns Hopkins University, Dr. Saria and Johns Hopkins University are entitled to revenue distributions related to the technology referenced in this report. A conflict-of-interest management plan governs the work described, and the work was conducted in accordance with Johns Hopkins University’s conflict-of-interest policies. This report includes a description of investigational and research uses of AI-based clinical decision support that fall outside the scope of current U.S. Food and Drug Administration clearance. For full disclosures of the other study authors, visit aacr.org.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.