Mayo Clinic primary-care practices
Verified evidenceContinuous decisioningA routine ECG can trigger evaluation for asymptomatic low ejection fraction before specialist suspicion.
Customer service · Minnesota and Wisconsin, United States
Executive brief
A low-cost model created value because it changed which existing ECGs triggered confirmatory action, while the clinician retained the diagnosis.
AI value · New diagnosis of ejection fraction at or below 50% within 90 days
Verified2.1% with clinician access to AI-ECG results; odds ratio 1.32, P=0.007
Pragmatic cluster-randomized result supports causality for diagnosis, not downstream morbidity or mortality benefit.
Before
Order ECGs for routine clinical reasons and evaluate heart failure using usual-care signals. → Order echocardiography when symptoms or judgment indicate it.
After
Analyzes routine ECGs and reports high likelihood of low ejection fraction. → Interprets the AI result and decides whether to order echocardiography and diagnose low EF.
Human boundary
The model flags risk; clinicians decide whether to order an echocardiogram and make the diagnosis.
Why it matters
A routine ECG can trigger evaluation for asymptomatic low ejection fraction before specialist suspicion.
Before
How the work ran before the change.
Primary-care clinician
Order ECGs for routine clinical reasons and evaluate heart failure using usual-care signals.
ControlUsual care
Clinician
Order echocardiography when symptoms or judgment indicate it.
ControlHuman suspicion
What changed
A routine ECG can trigger evaluation for asymptomatic low ejection fraction before specialist suspicion.
Decision rightSelection moves from a fixed rule to the model
After
How the same work runs now.
AI-ECG model
Analyzes routine ECGs and reports high likelihood of low ejection fraction.
ControlPositive/negative decision support
Primary-care clinician
Interprets the AI result and decides whether to order echocardiography and diagnose low EF.
ControlClinician retains diagnostic authority
Exception path
A positive result does not diagnose low EF; clinicians can decline imaging, and usual care continues for negative results.
Work removed
Decision authority
The model flags risk; clinicians decide whether to order an echocardiogram and make the diagnosis.
Before
Primary-care clinician
Order ECGs for routine clinical reasons and evaluate heart failure using usual-care signals.
Control: Usual care
Clinician
Order echocardiography when symptoms or judgment indicate it.
Control: Human suspicion
After
AI-ECG model
Analyzes routine ECGs and reports high likelihood of low ejection fraction.
Control: Positive/negative decision support
Primary-care clinician
Interprets the AI result and decides whether to order echocardiography and diagnose low EF.
Control: Clinician retains diagnostic authority
Work that left the path
Human role before
Clinicians selected patients for echocardiography from symptoms, examination, and conventional ECG interpretation.
Human role after
Clinicians receive an additional AI risk signal and selectively order confirmatory imaging, retaining all diagnostic and treatment rights.
AI roleNeural network that predicts the likelihood of ejection fraction at or below 50% from a routine 12-lead ECG.
1.6% in usual-care control practices→2.1% with clinician access to AI-ECG results; odds ratio 1.32, P=0.007
90 days after routine ECG during the 2019-2020 trial · 22,641 adults, 120 primary-care teams, 45 practices
Pragmatic cluster-randomized result supports causality for diagnosis, not downstream morbidity or mortality benefit.
Anti-pattern
Calling the AI output a diagnosis or extrapolating increased case finding to mortality reduction.
Questions
Portability conditions
Reputation risk
low
Evidence and authority
Current · updated
1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 304701f2a0cd4f80
Google DeepMind and EMBL-EBI · Continuous decisioning
Researchers can use precomputed structure predictions, then decide what experimental validation is required.
Google Research Flood Hub · Continuous decisioning
A global model can transfer hydrology into ungauged basins.
Google · Continuous decisioning
Novel breakages need not wait for a comprehensive post-submit test cycle.
Google and Google DeepMind · Continuous decisioning
That variable wind generation cannot be committed to the grid on a useful day-ahead schedule.