Clinical decision support
A clinical model pays off only when it catches what clinicians would otherwise miss, at an alert volume they will act on. You can only prove that on your own patients and workflow, not on the vendor's validation set.
Part of AI for Life Sciences & Healthcare. 7 sources, checked .
- of sepsis cases missed by a widely implemented proprietary sepsis model in external validation at one academic health system (2018 to 2019 data)
- 67%
- of sepsis cases missed by a widely implemented proprietary sepsis model in external validation at one academic health system (2018 to 2019 data)
- JAMA Internal Medicine, 2021(opens in a new tab)
- of US non-federal acute care hospitals used predictive AI integrated with their EHR in 2024
- 71%
- of US non-federal acute care hospitals used predictive AI integrated with their EHR in 2024
- Assistant Secretary for Technology Policy / Office of the National Coordinator for Health IT, HHS, 2025(opens in a new tab)
- of hospitals answered "don't know" when asked whether their predictive AI was evaluated for bias (2024)
- 21%
- of hospitals answered "don't know" when asked whether their predictive AI was evaluated for bias (2024)
- Assistant Secretary for Technology Policy / Office of the National Coordinator for Health IT, HHS, 2025(opens in a new tab)
The problem
Most US hospitals already run predictive models. In 2024, 71% of hospitals reported using predictive AI integrated with their EHR, and the most common use was predicting health risks for inpatients.1 Most of it arrives with the EHR: 80% of hospitals used predictive AI from their EHR developer.1 Oversight has not kept pace. In the same survey, 21% of hospitals answered "don't know" when asked whether their models were evaluated for bias, and 15% did not know about accuracy.1
Skipping local validation has a documented cost. When Michigan Medicine tested a widely implemented proprietary sepsis model on 38,455 hospitalizations, the model reached an area under the curve of only 0.63. It missed 67% of patients who developed sepsis while generating alerts for 18% of all hospitalized patients.2 It identified just 183 of 2,552 sepsis patients (7%) who had not already received timely antibiotics.2 Lift over current practice, not headline accuracy, is the number that decides value.
The rules moved in 2026. FDA's revised Clinical Decision Support Software guidance, issued in January 2026, treats software that detects a life-threatening condition such as sepsis and raises an alarm as a device function, and notes that automation bias increases when decisions are urgent.3 ONC's HTI-5 rule removes the "model card" source attribute and risk management requirements from certified EHR decision support, effective December 28, 2026.4 Certification will no longer make vendors supply that transparency, so the burden of proof shifts to you.
Generative assistants add a second class of point-of-care tool. In a randomized trial of 238 outpatient physicians, one ambient AI scribe cut time-in-note by 9.5% against usual care, while a second showed no significant change. Clinicians reported clinically significant inaccuracies "occasionally" with both.7 Results differ by product and setting, so each tool needs its own evaluation before scale.
What we build
- 01
Local validation and silent-mode trials
Before go-live we run the model silently on your data, compare its flags with what clinicians actually did, and report lift over current practice, alert volume per unit and performance by subgroup.
- 02
Deterioration and sepsis alerting with tuned thresholds
Thresholds set against the alert volume each unit can absorb, alerts routed to a named role, and every response logged, so the alert becomes a workflow rather than a pop-up.
- 03
Evidence-linked point-of-care assistants
Assistants that answer from your approved guidelines, order sets and formulary, show the source passage for every answer, and decline when the evidence is not there.
- 04
Ambient documentation evaluation
A structured pilot of competing ambient scribe products on your specialties, measuring time-in-note, edit rates and error types before you sign an enterprise contract.
- 05
Decision support inventory and bias review
An inventory of every decision support tool in use, flagging inputs that measure race, color, national origin, sex, age or disability, with a documented mitigation for each.
Where it goes wrong
The failure modes we design against from the first week, and the control for each.
Trusting the vendor's validation
Performance measured on someone else's patients does not transfer by default: the sepsis model above reached an AUC of only 0.63 on one health system's data.2 Control: retrospective and silent-mode validation on your own patients, with go-live criteria agreed by clinical governance before results are seen.
Alert fatigue that buries the real signal
A model that alerts on 18% of inpatients teaches staff to dismiss it.2 Control: set thresholds by an alert budget per unit, review override rates weekly, and retire alerts that do not change what clinicians do.
Automation bias in urgent decisions
FDA describes automation bias as over-reliance on automated suggestions and notes it increases when there is no time to weigh other information.3 Control: show the inputs and basis behind each recommendation, train staff on known failure modes, and audit cases where clinicians followed a wrong output.
Silent drift after go-live
Case mix, coding and documentation change, and a model validated last year can degrade unnoticed. Control: monthly performance and calibration checks by subgroup, a named owner and a rollback plan. For a cleared device, only changes described in an FDA-reviewed change control plan avoid a new submission.6
Rules that apply
FDA Clinical Decision Support Software guidance (January 2026)
Primary text on fda.gov(opens in a new tab)Non-device CDS must let the clinician independently review the basis for its recommendations; software that detects conditions such as sepsis and generates an alarm is a device function.
HHS Section 1557 rule, 45 CFR 92.210
Primary text on ecfr.gov(opens in a new tab)Covered entities must make reasonable efforts to identify patient care decision support tools that use race, color, national origin, sex, age or disability as inputs, and to mitigate the risk of discrimination from each.
ONC HTI-5 final rule, decision support interventions
Primary text on federalregister.gov(opens in a new tab)Removes the source attribute and intervention risk management requirements for certified decision support from December 28, 2026, so model transparency must come from your contracts and governance.
FDA guidance on Predetermined Change Control Plans for AI-enabled devices (August 2025)
Primary text on fda.gov(opens in a new tab)Manufacturers can describe planned model modifications and how they will be validated in advance, and make those changes without a new marketing submission for each one.
How to measure it
- Lift over current practice: cases the model flagged before clinicians acted
- Alerts per 100 patient-days and the share acted on, by unit
- Sensitivity, positive predictive value and calibration by subgroup
- Time from alert to intervention for the target condition
- Clinician override rate and the documented reasons
Questions buyers ask
Is our sepsis or deterioration model regulated by the FDA?
It depends on what it does. Under FDA's January 2026 guidance, software that analyzes patient data to detect a life-threatening condition such as sepsis and generates an alarm is a device function. Tools that present recommendations a clinician can independently review, without time pressure, may fall outside device rules. If you bought the model, ask the vendor for its regulatory status in writing. If you built it, document your own determination.
How do we validate a vendor's clinical AI model on our own patients?
Run it retrospectively on at least a year of your data, then silently in production for several weeks with outputs hidden from clinicians. Compare its flags with what actually happened and with what clinicians already did. Report sensitivity, positive predictive value, calibration and alert volume, overall and by subgroup. Agree the go-live threshold with clinical governance before anyone sees the results.
Do ambient AI scribes save clinicians time?
Sometimes, and by less than vendors suggest. A 2025 randomized trial of 238 physicians found one scribe cut note-writing time by about a tenth while another showed no significant change, though both showed possible improvements in burnout. Clinicians also saw occasional clinically significant errors. Pilot more than one product on your own specialties and measure edit rates and errors, not just satisfaction.
Sources
Every figure on this page comes from the sources below. We checked each link on .
- 1.Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024 (ASTP Data Brief No. 80)(opens in a new tab)Assistant Secretary for Technology Policy / Office of the National Coordinator for Health IT, HHS, September 2025 · healthit.gov
- 2.External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients(opens in a new tab)JAMA Internal Medicine, June 21, 2021 · jamanetwork.com
- 3.Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff(opens in a new tab)U.S. Food and Drug Administration, January 29, 2026 · fda.gov
- 4.Health Data, Technology, and Interoperability: ONC Deregulatory Actions To Unleash Prosperity (Final Rule, 91 FR 67916)(opens in a new tab)Office of the National Coordinator for Health Information Technology, HHS (Federal Register), October 28, 2026 · federalregister.gov
- 5.45 CFR 92.210: Nondiscrimination in the use of patient care decision support tools(opens in a new tab)Electronic Code of Federal Regulations (HHS), July 5, 2024 · ecfr.gov
- 6.Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions(opens in a new tab)U.S. Food and Drug Administration, August 2025 · fda.gov
- 7.Ambient AI Scribes in Clinical Practice: A Randomized Trial(opens in a new tab)NEJM AI, November 26, 2025 · pmc.ncbi.nlm.nih.gov