Skip to content
Physician Case Review Consult

When a Deterioration Model Stays Silent

Predictive models now watch for patients getting worse. When one fails to alert, or alerts and nothing happens, the model's own records become evidence.

Consider a hypothetical. A patient on a medical ward has a rising heart rate, a falling blood pressure, and new confusion over about eight hours. The hospital uses a predictive deterioration model that scores every inpatient and alerts the rapid response team when a score crosses a threshold. This patient’s score climbed but never crossed it. By the time a nurse called for help, the patient needed intensive care and had suffered lasting harm.

Many hospitals now run models like this, either built into the EHR or added by a vendor. They are designed to catch the patient who is slipping before anyone notices. When they work, they give clinicians earlier warning. When a patient deteriorates anyway, the model becomes part of the story, and its records can answer questions the chart cannot.

A new national reporting category. The 2025 update to the National Quality Forum’s Serious Reportable Events list added a new event: “Patient harm associated with unrecognized clinical deterioration.” The guidance describes it as harm from missed, delayed, or ineffective recognition of and response to deterioration, when the information was available, or should have been available, to the care team. It names electronic monitoring, rapid response teams, and warning systems as examples of the reliable processes hospitals should have. It excludes harm from rapid, unforeseeable changes despite timely monitoring, and the expected progression of illness. The Joint Commission will fold the updated list into its sentinel event list in January 2027.

That classification is a reporting and improvement framework, not a legal standard. But it tells counsel what a hospital is expected to look at after an event like this one, and it signals that deterioration is now treated as a recognized system failure, not only an individual one.

Why the model’s silence needs explaining. A model that did not alert can mean several different things, and each points to different evidence:

  • The threshold was set where it was. Hospitals usually choose the alert threshold locally, trading missed patients against false alarms. The choice, and the reasoning behind it, should be documented.
  • The inputs were late. A model can only score data that has been entered. If vital signs were taken at 2:00 and charted at 5:00, the model saw them at 5:00. This is where model evidence and EHR audit evidence meet: entry times show what the model could have known and when.
  • The model was not performing as expected. Models can drift as patient populations, practice patterns, or documentation habits change, and vendors release new versions. Was it validated locally before go-live, and was its performance monitored after?
  • The alert fired and went nowhere. If the score did cross the threshold, the questions shift to routing: who received the alert, whether it was acknowledged, and what happened next.

What the guidance expects. The Joint Commission and the Coalition for Health AI (CHAI) 2025 guidance on responsible AI use is voluntary, but it is specific about monitoring. It recommends asking vendors how a tool was validated, checking performance locally, monitoring on a schedule scaled to risk, and checking tools that inform clinical decisions more often than administrative ones. It also recommends treating AI-related safety events like other patient safety events, including capturing them in internal incident systems. A deterioration model sits squarely in the higher-risk group.

Records to request. These are rarely in a standard production and often need to be named specifically:

  • The patient’s model score history, with timestamps, for the whole admission.
  • Alert logs: what fired, to whom, and whether and when it was acknowledged.
  • The threshold and configuration in effect on the dates of care, and the model version.
  • Local validation and ongoing monitoring reports for the model.
  • Vital-sign and assessment entry times, compared with the times recorded.
  • Rapid response policies, including when staff were expected to call regardless of the score.

What to expect from the defense. A model is a safety net, not a replacement for clinical judgment, and that cuts in both directions. A silent model does not excuse a failure to recognize deterioration that was visible at the bedside; many rapid response policies expect staff to call on their own concern, whatever the score. But expect the defense to argue that a hospital’s decision to use a model, or where it set the threshold, is not negligent simply because one patient fell below it. Every threshold misses some patients, so a case built on the threshold alone is the easiest one to answer. The real questions are whether the system as a whole, the model, the alert routing, the staff, and the escalation policy, gave this patient a reasonable chance of being recognized in time.

The model’s records can help answer that. They show what the system knew, when it knew it, and what it did with that knowledge.

Sources

  • Joint Commission and National Quality Forum (NQF). Aligning Patient Safety Event Reporting: 2025 Updates to Sentinel Events and Serious Reportable Events. Washington, DC: NQF, January 2026 (Updated May 2026). jointcommission.org. Content reproduced with credit to Joint Commission and National Quality Forum.
  • Joint Commission and Coalition for Health AI. The Responsible Use of AI in Healthcare (RUAIH). 2025. jointcommission.org

The scenario above is hypothetical. Disclosure: Dr. Bansal is Chief Medical Officer of Diagnose Data, Inc., a healthcare technology company. This article does not describe or endorse any product.

This article is educational and does not constitute legal advice. Safety classifications and voluntary guidance describe reporting and improvement frameworks, not a legal standard of care or any outcome. Sources are listed at the end of the article.