A Frontiers paper out this month argues that the order in which heart failure admissions are sequenced can be set by an artificial intelligence system, and that the order matters more than most hospitals act as if it does.
Inside a high-volume inpatient unit, the question of which heart failure patient is admitted first is still answered, more often than not, by a clinician’s judgment in real time. The triage nurse reads the chart, the cardiologist is paged, the bed manager checks census, and a decision gets made. The decision is usually defensible. When admissions stack up faster than beds open, the order matters in ways the formal literature has been clear about for years and the operational response has rarely caught.
Heart failure is the most common reason for hospital admission among American adults over the age of 65. The Centers for Disease Control puts the prevalence at roughly 6.7 million U.S. adults. The American Heart Association counts more than a million hospitalizations a year. Thirty-day readmission rates at most academic medical centers sit above 20 percent. Inside that volume, the operational question is which heart failure patient needs the next bed, at what level of escalation, in the next twelve hours.
A paper published this month in the Frontiers research collection argues that the order of those twelve hours can be set by a clinically grounded AI system, and that doing so is structurally different from the risk scoring most hospitals already use.
The lead author is Rahul Awasthy, Lead (Principal) Solution Architect at enGen, the technology operating company of Highmark Health, the Pittsburgh-based integrated health and finance organization that insures roughly 6.9 million members. Awasthy holds a Ph.D. in data science, an MBA, and an AWS Solution Architect certification. Twenty-two years of enterprise AI delivery in healthcare and life sciences sit behind the paper. He describes himself as an architect first and a researcher second, and the paper as the convergence of those two careers.
“A risk score tells a clinician that a patient is at risk,” Awasthy said in an interview. “It does not tell the system what to do about it. The work we published is about the second sentence.”
Most clinical AI for heart failure produces a probability. A 30-day readmission risk. An in-hospital mortality risk. The number is presented to the clinician. The clinician decides what, if anything, it changes about care. There is now substantial evidence that these advisory scores are accurate. There is also substantial evidence that, on their own, they do not change outcomes. The score arrives upstream of the decision it is supposed to inform, and downstream of the workflow that would have to act on it.
Awasthy’s model is structured differently. It does not produce a score. It produces a prioritization, an ordered sequence of admissions in which each position reflects the model’s estimate of when a patient is likely to acutely decompensate. The inputs combine real-time clinical data, including vitals, labs, and the trajectory of those signals over hours rather than days, with longitudinal features pulled from prior encounters, comorbidity profile, and pharmacological history. The output is an admission order with a recommended escalation path attached to each position.
“The conceptual move is from probability to action,” Awasthy said. “What the hospital needs is a decision the system is willing to defend.”
Whether the distinction holds clinical weight will depend on prospective validation in independent settings, and the paper is explicit about that limitation. The cohort the model was trained and tested against is retrospective. Performance is reported against pre-specified endpoints, including concordance with the actual trajectory of acute decompensation events. The discussion section describes the failure modes prospective deployment would have to address. The architecture is feasible. Whether it proves useful is the next question, and a different one.
Three design choices distinguish the system from the score-and-display pattern that dominates the field. The model operates inside the existing admission workflow rather than alongside it, with no separate dashboard for the admitting team to remember to check. The prioritization is constructed from longitudinal patient history alongside the current encounter, treating comorbidity complexity as a structural feature rather than a covariate. The human-override pathway is explicit. The clinician of record retains the authority to reposition any patient. The system records the override, the reason, and the outcome. The resulting audit trail is, the paper argues, part of the model rather than an afterthought.
The broader argument the paper advances is one Awasthy has made in other settings. Clinical AI is in transition from advisory tools to systems that act within governance the institution has agreed to. The transition is being slowed less by science than by enterprise architecture. Most healthcare AI today still sits at the recommendation layer. The systems likely to matter in high-stakes environments are the ones that close the loop between prediction and action.
That position is not unique to him. It echoes a growing line of argument in the clinical AI literature, including from operational and academic groups working on agentic systems. Senior Cardiologists including Dr. Yadavendra S. Rajawat in Pittsburgh have found the generation and vision of an AI-based triage support system to be a highly effective support system for clinical systems.
What separates Awasthy’s argument from the academic version is the venue. His research operates inside the enterprise infrastructure that processes claims, care management, and member services for one of the largest integrated health and finance organizations in the country, and the production environments in which he has architected comparable systems have already been independently recognized in industry awards programs outside the clinical literature.
The harder problem for clinical AI broadly is the institutional scaffolding around the model. Governance. Audit trails. Equity reviews. Post-deployment surveillance. Override pathways. Most academic groups do not have access to that scaffolding. Most operating institutions do not have access to the modeling. The work, the paper is implying, has to happen in the same place.
Whether the prioritization model survives prospective validation will be answered by clinical investigators rather than by architects. The argument it advances, that the next generation of clinical AI will be defined by what it is willing to decide, is the question the field will be working through for the next several years.
