Regulation

Building an algorithmovigilance program

A build sequence for watching clinical algorithms after they go live — the operating model, the four signals to instrument, the people to name, the cadence to run, and the escalation path — each step tied to a primary source. As of July 2026.

By Jonas WeirReviewed by Jonas Weir · editorial reviewUpdated

The short version

  • Algorithmovigilance is the post-deployment surveillance of clinical algorithms — coined by Peter Embi in 2021 and modeled on pharmacovigilance in drug safety.
  • The reason a program is needed is written in two published failures: a deployed sepsis model that scored an AUC of 0.63 on external validation and missed 67% of cases at its alert threshold, and a care-management algorithm that under-served Black patients because it optimized on cost.
  • A working program instruments four signals continuously — discrimination and calibration, subgroup performance, alert burden, and drift — and gives clinicians a route to report suspected algorithm-related harm.
  • Vanderbilt's VAMOS is the first published health-system platform for this, organized around a performance-monitoring dashboard, model-specific pages, and snapshot views for governance review.
  • The mandate is now external: the HTI-1 rule imposes an ongoing maintenance duty on certified Predictive decision-support tools since January 2025, and the Joint Commission and CHAI make ongoing quality monitoring one of seven elements of responsible AI use.

A deployed clinical algorithm becomes a standing liability the moment it goes live, and most health systems still watch it with nothing. Validation happens once, before go-live; the model then runs for years while the patients, coding conventions, and upstream data pipelines beneath it shift. Algorithmovigilance is the discipline of watching for that decay. This page is a build sequence — the operating model to adopt, the signals to instrument, the people to name, the cadence to run, and the escalation path to wire — with each step tied to a primary source. It is orientation and general guidance; confirm any monitoring obligation that carries regulatory or contractual weight with your compliance function before you rely on it. As of July 2026.

What algorithmovigilance is, in one paragraph

Algorithmovigilance is the set of scientific methods and activities for evaluating, monitoring, understanding, and preventing adverse effects of algorithms used in healthcare. Peter Embi proposed the term in 2021, framing it deliberately on the model of pharmacovigilance and aiming it at two things at once — effectiveness and equity 1. A 2024 analysis carried the analogy into working machinery, showing how the pharmacovigilance toolkit adapts to deployed AI: "Drawing inspiration from pharmacovigilance methods, we discuss concepts that can be adapted for monitoring AI systems in healthcare" 2. That is the operating model this guide builds on: drug safety spent decades assembling infrastructure for spontaneous adverse-event reporting, cross-institution signal detection, and graded responses from label changes to withdrawal — and each has a plausible analogue for a deployed model.

Why a program: the two failures that define the field

A build guide earns its cadence from the failures it is designed to catch. Two published cases anchor the discipline, and between them they name the signals a program must instrument.

The first is a monitoring failure. When a widely implemented proprietary sepsis model was externally validated at one academic health system, it "predicted the onset of sepsis with an area under the curve of 0.63" and, at its live alerting threshold, "did not identify 1709 patients with sepsis (67%)" while generating alerts on roughly 18% of all hospitalized patients 3. The model had passed its developer's internal checks. Only surveillance at the deploying site surfaced the gap between the summary figure and the bedside miss rate.

The second is an equity failure. A 2019 study of a care-management algorithm applied to millions of patients found that, at a given risk score, "Black patients are considerably sicker than White patients, as evidenced by signs of uncontrolled illnesses" — because the model used healthcare cost as its proxy for need 4. Overall accuracy looked fine. The harm lived entirely in a subgroup, invisible to any aggregate metric.

Read together, these are the brief: watch the operating point, not the summary number, and watch every subgroup, rather than the mean alone.

The build, step by step

Six steps take a health system from nothing to a running program. Each is tied to the source that grounds it.

#StepWhat it means in practiceSource
1Adopt the operating modelTreat every deployed model as something to be surveilled for its whole life, on the pharmacovigilance pattern12
2Baseline before go-liveRecord external-validation performance as the reference every later measurement is compared against3
3Instrument the four signalsDiscrimination and calibration, subgroup performance, alert burden, and drift — on a dashboard345
4Set a risk-tiered cadenceCheck tools close to clinical decisions more often; check administrative tools periodically6
5Name owners and a reporting routeA designated owner per model, plus a channel for clinicians to report suspected harm26
6Wire escalation to actionPre-agree the thresholds that trigger retraining, restriction, or retirement78

1. Adopt the operating model

Start by settling what the program is: a standing process, never a one-time event. Embi's framing gives it purpose — effectiveness and equity — and the pharmacovigilance mapping gives it structure: signal detection, causality assessment, and incident reporting, each transposed from drugs to algorithms 12. Adopting this model up front decides everything downstream, because it reframes a deployed model as a supervised intervention rather than a finished purchase.

2. Baseline every model before go-live

You cannot detect decay without a reference. Before a model touches care, record its performance under external validation — tested on data from a different site, scanner, or era than it was built on — and store that as the baseline. The sepsis case is the warning: a model can arrive with a strong in-house number and post an AUC of 0.63 the first time it meets a new population 3. The baseline is the line every later measurement is read against. For the wider set of questions to ask of that pre-deployment evidence, see our guide on how to read an AI validation study.

3. Instrument the four signals

This is the core of the build. A dashboard should carry four families of metric, refreshed on a schedule:

  • Discrimination and calibration. Discrimination (whether the model still ranks sicker patients above healthier ones) degrades quietly; calibration — whether a predicted 20% risk still matches a 20% observed rate — degrades faster and is the property that most directly drives over- or under-treatment.
  • Subgroup performance. Break every headline metric down by the groups the model touches. The 2019 bias case is the reason: a model can hold its overall performance while failing a subgroup outright 4.
  • Alert burden. Track how often the model fires and what share of true cases it catches at its live threshold. The sepsis model's 18% firing rate against a 67% miss rate is exactly the pairing a dashboard makes visible 3.
  • Drift. Watch for model drift — shifts in the input data, the case mix, or upstream systems that move performance away from baseline over time.

Vanderbilt's published platform confirms the shape: it is organized around a performance-monitoring dashboard, model-specific pages, and snapshot views built for governance review 5.

4. Set a risk-tiered cadence

Not every model deserves the same watch. The Joint Commission and CHAI put the sorting rule plainly: scale monitoring to how close a tool sits to patient care. Their guidance tells organizations that "tools that inform or drive clinical decisions should be checked more often; tools that simply help with administrative or documentation tasks can be checked periodically," and it names the artifact — "developing an AI dashboard" — that carries the checks 6. Tier your models on that axis and set review frequency per tier.

5. Name owners and a reporting route

Surveillance with no owner is theater. Assign a designated individual accountable for each deployed model's performance, and stand up a channel for clinicians to report suspected algorithm-related harm — the incident-reporting limb of the pharmacovigilance analogy 2. The Joint Commission and CHAI reinforce both halves, calling for feedback channels to vendors on model changes and a route for reporting adverse events to leadership 6. This is where the program stops being a dashboard and becomes an accountable function.

6. Wire escalation to action

A signal that triggers nothing is noise. Pre-agree the thresholds that move a model from "watch" to "act" — retraining, restricting to a narrower population, or retirement. When the monitored model is a regulated device with a predetermined change control plan, those monitoring results feed the pre-authorized criteria that govern retraining, so escalation and the device's update pathway are the same wire. The HTI-1 rule adds a standing duty here for certified predictive tools: developers must "adhere to certain ongoing maintenance requirements to keep their DSI source attribute information complete and up-to-date as well as implement risk management practices for Predictive DSIs they supply" 7. And the Health AI Partnership's decision framework makes the end of the path explicit, treating "monitor the AI solution" and "update or decommission the AI solution" as distinct lifecycle stages with their own guides 8. Retirement is a planned move rather than a failure.

A reference blueprint: VAMOS

Health systems no longer have to design from scratch. Vanderbilt University Medical Center built the Vanderbilt Algorithmovigilance Monitoring and Operations System (VAMOS) through participatory design sessions and published the work in 2025, noting that few systems yet support systematic monitoring and governance of AI across a health system 5. The value for a team standing up its own program is the distilled interface set — a performance-monitoring dashboard, model-specific pages, and governance snapshots — which maps almost one-to-one onto steps 3 through 6 above. Treat it as a blueprint for what "instrumented" looks like in production.

How to read this build

Four cautions travel with the sequence.

First, a program is a standing process, never a project with an end date. Its whole premise is that validation is an event and vigilance is the process that outlasts it 1. Budget it as an ongoing function.

Second, responsibility is shared, and locally inescapable. Device manufacturers carry post-market obligations, but local performance depends on local data, workflows, and populations — which is why deploying systems build their own capacity rather than waiting on the vendor 5. Contract for vendor cooperation; own the local watch regardless.

Third, the mandate is hardening. What reads today as good practice is becoming an expectation: the HTI-1 maintenance duty is live for certified predictive tools 7, and accreditor guidance now names ongoing monitoring as an element of responsible use 6. The moving global picture is tracked in our global AI in health regulation tracker.

Fourth, monitoring lives inside governance. A dashboard without a committee to read it and act changes nothing; the escalation path in step 6 assumes a body with the authority to restrict or retire a model. That structure is the subject of our companion hospital AI governance committee playbook. Because some monitoring obligations are contractual or regulatory, confirm the specific duties for any deployed tool with your compliance or regulatory function before relying on this outline.

Sources and method

This guide is built from primary material. The operating model comes from the two papers that define the field — Embi's 2021 framing of algorithmovigilance 1 and the 2024 pharmacovigilance-mapping analysis 2. The signals to instrument come from two documented failures used as worked examples: a sepsis-model external validation 3 and a care-management bias study 4. The platform blueprint is Vanderbilt's published VAMOS design 5. The cadence, ownership, and mandate come from three governance instruments: the Joint Commission and CHAI's 2025 responsible-use guidance 6, the HTI-1 Final Rule's ongoing-maintenance duty 7, and the Health AI Partnership's eight-decision-point framework 8. Every figure is drawn from the source cited beside it. We revisit this page every ninety days and whenever a tracked instrument changes. Statuses are current as of July 2026.

Questions & answers

  • What is an algorithmovigilance program?

    It is the standing capability a health system uses to watch its deployed clinical algorithms for degraded, unsafe, or inequitable performance — the post-deployment counterpart to the validation done before go-live. The term was coined by Peter Embi in 2021 and modeled on pharmacovigilance, the drug-safety discipline of monitoring medicines in real-world use.

  • What should an algorithmovigilance program actually measure?

    Four signals, continuously: discrimination and calibration (does the model still rank and does its predicted risk still match observed risk), subgroup performance (does it hold across the populations it touches), alert burden (how often it fires and what share of true cases it catches at its threshold), and drift (has the input data or the population shifted). A route for clinicians to report suspected algorithm-related harm sits alongside the metrics.

  • Is post-deployment AI monitoring required, or optional?

    Increasingly required. In the US, the HTI-1 rule places an ongoing maintenance duty on developers of certified Predictive decision-support tools, effective since January 2025, and the Joint Commission and CHAI's 2025 guidance names ongoing quality monitoring as one of seven elements of responsible AI use. Confirm the specific obligations for your setting with your compliance function.

Sources

  1. Embi PJ. Algorithmovigilance—Advancing Methods to Analyze and Monitor Artificial Intelligence–Driven Health Care for Effectiveness and Equity. JAMA Network Open. 2021;4(4):e214622. doi.org/10.1001/jamanetworkopen.2021.4622
  2. Balendran A, Benchoufi M, Evgeniou T, Ravaud P. Algorithmovigilance, lessons from pharmacovigilance. npj Digital Medicine. 2024;7:270. doi.org/10.1038/s41746-024-01237-y
  3. Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Internal Medicine. 2021;181(8):1065-1070. doi.org/10.1001/jamainternmed.2021.2626
  4. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453. doi.org/10.1126/science.aax2342
  5. Human-centered design of an artificial intelligence monitoring system: the Vanderbilt Algorithmovigilance Monitoring and Operations System. JAMIA Open. 2025;8(5):ooaf136. doi.org/10.1093/jamiaopen/ooaf136
  6. The Joint Commission and Coalition for Health AI (CHAI). Guidance on the Responsible Use of AI in Healthcare (RUAIH). 17 September 2025. digitalassets.jointcommission.org/api/public/content/dcfcf4f1a0cc45cdb526b3cb034c68c2
  7. Federal Register. Health Data, Technology, and Interoperability: Certification Program Updates, Algorithm Transparency, and Information Sharing (HTI-1 Final Rule). 9 January 2024. www.federalregister.gov/documents/2024/01/09/2023-28857/health-data-technology-and-interoperability-certification-program-updates-algorithm-transparency
  8. Health AI Partnership (Duke Institute for Health Innovation). Key decisions in adopting an AI solution — eight key decision points. Accessed July 2026. healthaipartnership.org/key-decisions-in-adopting-an-ai-solution