Did AI-guided ECG screening find more hidden rhythm problems? A 14,665-person trial found a fragile 0.26-point gain
In Taiwan's LATENT randomized trial, the 90-day composite occurred in 0.74% of the AI-guided group and 0.48% with usual care. The absolute difference was small, one event could erase conventional significance, and atrial fibrillation alone was not significantly higher.
The short version
- What happened
- LATENT randomized 14,665 adults aged 40–85 at one Taiwanese centre to an AI-guided pathway or usual care and followed the primary composite outcome for 90 days.
- Who it affects
- Patients may be offered extra rhythm monitoring after an apparently normal ECG, but the AI score is a risk flag rather than a diagnosis.
- What remains unproven
- Single-centre, open-label trial in Taiwan; transfer to other populations and pathways is unproven.
Keep in mind
Single-centre, open-label trial in Taiwan; transfer to other populations and pathways is unproven.
The 0.26-point primary difference was fragile to reclassification of one event.
Atrial fibrillation or flutter alone was not significantly higher in the AI group.
Only 48 of 84 high-risk intervention participants completed monitoring.
The short follow-up did not establish effects on stroke, hospitalization, quality of life or mortality.
Multi-source analysis · 3 linked sources. Primary analysed source date: 7 November 2026. Earliest linked supporting source: 1 September 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer: a small composite gain that needs cautious reading
The LATENT trial gives a more demanding answer than a retrospective accuracy study. Among 14,665 randomized adults, the study's 90-day composite outcome occurred in 0.74% of people assigned to AI-guided screening and 0.48% assigned to usual care. That is an absolute difference of 0.26 percentage points, or roughly three additional composite events per 1,000 people screened. The paper reports conventional statistical significance, but its own fragility check shows that reclassifying a single event would move the p value to 0.057. A positive result this close to the threshold should be treated as provisional rather than decisive.
The composite combined newly documented atrial fibrillation or flutter with related clinical actions. That can make a trial sensitive to changes in testing and care, even when the underlying arrhythmia difference is uncertain. Looking only at atrial fibrillation or flutter, detection was 0.3% in the AI group and 0.2% in usual care, a difference that was not statistically significant. At 180 days, antiarrhythmic therapy was almost identical: 2.3% versus 2.4%. The study supports a narrow claim about a short-term composite, not proof that it prevented strokes or improved survival.[1]
How the trial moved from an ECG score to monitoring
Participants were adults aged 40 to 85 receiving care at a single Taiwanese centre. The trial was open-label: clinicians and participants knew which pathway they were in. A previously developed AI-ECG model identified people considered at high risk of developing atrial fibrillation. In the intervention arm, those flagged as high risk were offered ambulatory rhythm monitoring and follow-up; usual-care participants continued through ordinary practice. Randomization is the design's major strength because it reduces baseline differences between the two large groups.
The operational denominator matters. Eighty-four people in the intervention arm were categorized as high risk, but only 48—57.1%—actually wore the offered monitor. Researchers also monitored 48 matched people from the control group for a subgroup comparison. The trial is therefore an evaluation of a risk model, an invitation process and real-world uptake. A patient who declines or cannot manage monitoring receives no diagnostic benefit from a good risk score. The headline enrollment should not be mistaken for 14,665 people receiving continuous monitoring.[1][2]
The distinction that matters
How to read the LATENT result
The study tested a care pathway, not whether an AI score alone diagnoses atrial fibrillation.
Randomize
Assign 14,665 adults to AI-guided screening or usual care
Comparable full groups
Flag
Identify high-risk ECGs
84 intervention participants classified high risk
Monitor
Offer ambulatory rhythm monitoring
48 high-risk intervention participants wore it
Compare
Count the 90-day composite
0.74% versus 0.48%; one-event fragility
The high-risk subgroup was striking but small
Within the monitored high-risk comparison, the primary composite occurred in 32.1% of 84 intervention participants and 2.6% of 76 control participants, with a reported interaction p value below 0.001. That pattern suggests the pathway concentrated testing where it was most likely to change care. It also comes from a small subgroup selected by the model and then affected by whether people accepted monitoring. Subgroup percentages can move sharply when only a few events change, and the matched comparison does not carry the same protection against bias as the trial's full randomized analysis.
Monitoring-related issues were reported in 22.9% of monitored intervention participants and 18.8% of monitored controls. The abstract does not establish a clear difference in harm, but screening is not free of burden. Patches can irritate skin, alarms can create anxiety, and additional tests can generate uncertain findings. A useful programme must show that earlier detection leads to better outcomes at acceptable cost and burden, not only that it produces more clinical activity among people a model labels high risk.[1]
What randomization settles—and what it does not
Randomization makes the full-group comparison substantially more credible than asking whether people with a high AI score later receive diagnoses. The large total sample also captures how rare the 90-day events were. Yet the open-label design can influence clinician investigation and coding, especially when the primary outcome includes care actions. The study was conducted at one health system in Taiwan, so referral pathways, ECG equipment, background risk and access to ambulatory monitors may differ elsewhere. External validation is needed before transferring the programme to another population.
The follow-up window is short for the outcomes patients ultimately care about. Ninety days can show whether a pathway changes diagnosis or management, but not whether it reduces stroke, heart failure, hospitalization or death. More detection can include brief episodes whose treatment value is uncertain. The result is peer-reviewed and consequential because it tests an AI pathway prospectively, but it is not Breaking: there is no urgent clinical alert or immediate change to standard care.[1][2]
What this means for clinicians and patients
For clinicians, LATENT argues against treating an AI-ECG score as a diagnosis. The score selected people for additional monitoring, and many flagged participants did not complete that monitoring. A safe pathway needs an explanation of what the score does, a way to confirm rhythm findings, review of false positives and a plan for people who cannot use the device. Decisions about anticoagulation or antiarrhythmic treatment still require clinical evidence and shared decision-making, not a probability generated from a resting ECG.
For patients, the honest message is that the trial found a small increase in a combined short-term outcome, while the difference in atrial fibrillation itself remained uncertain. Someone offered monitoring should ask how long it will last, what counts as a meaningful episode and what treatment would follow. Health systems should track uptake, abandoned monitoring, downstream tests, treatment and patient-reported burden. Otherwise a programme can look productive because it orders more care while leaving health benefit unresolved.[1][2]
What evidence would change the assessment
Confidence would rise if multicentre trials in other countries reproduce the full randomized effect, prespecify a clinically meaningful endpoint and follow participants long enough to measure stroke, hospitalization, quality of life and bleeding from treatment. Reports should show calibration by age, sex, ethnicity, comorbidity and device type; disclose missing ECGs and monitoring adherence; and compare the pathway with simpler clinical risk rules. Cost-effectiveness should include staff time, patches, confirmatory tests and treatment, not only the price of running the model.
The assessment would weaken if the composite advantage disappears in replication, diagnostic activity rises without better outcomes, or false positives and unequal access concentrate harms. LATENT is valuable because it moves AI-ECG evidence into a randomized care pathway and reveals the friction between a model flag and completed monitoring. Its 14,665-person denominator is impressive, but the signal rests on uncommon events and a fragile primary result. That warrants continued study rather than immediate population-wide screening.[1][2]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring randomized trials: compare the finding and its evidence limits.
Can a respiratory chatbot improve lay triage?
A preregistered Chinese trial found that 2,176 compliant participants answered simulated respiratory questions more accurately with a task-constrained GPT-4o chatbot than with ordinary web search. The largest gain was diagnosis; triage accuracy remained modest, and the trial did not measure real care-seeking or patient outcomes.
Read the next reportExplore the full source and evidence record
What this means for people
- Patients may be offered extra rhythm monitoring after an apparently normal ECG, but the AI score is a risk flag rather than a diagnosis.
- Clinicians need confirmation, treatment thresholds and a way to explain uncertain or incidental findings.
- Health systems should measure monitoring uptake and downstream burden as well as diagnoses, because 42.9% of flagged intervention participants did not wear the offered monitor.
Global context
Atrial fibrillation screening is debated because earlier detection could prevent strokes, while broad monitoring can also find brief episodes of uncertain treatment value. Taiwan's integrated care setting and population risk profile may not match fragmented systems or under-served communities elsewhere. AI-guided selection could reduce the number who need monitoring, but only if the score remains calibrated and patients can complete the next step. The trial should inform—not settle—screening policy.
What the evidence does not yet show
- Single-centre, open-label trial in Taiwan; transfer to other populations and pathways is unproven.
- The 0.26-point primary difference was fragile to reclassification of one event.
- Atrial fibrillation or flutter alone was not significantly higher in the AI group.
- Only 48 of 84 high-risk intervention participants completed monitoring.
- The short follow-up did not establish effects on stroke, hospitalization, quality of life or mortality.
What to watch next
- Multicentre replication with patient-centred outcomes
- Stroke, hospitalization, bleeding and quality-of-life results over longer follow-up
- Calibration and outcomes across demographic and clinical groups
- Monitoring uptake, false positives and downstream testing costs
- Comparison with simpler clinical risk rules
Living evidence record
Impact record IAI-0LZDBIZ
Evidence stage
Studied
Confidence
Corroborated
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
8 November 2026
Source trail
3 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 8 November 2026
Research question, topics and regions
Whether an AI-ECG risk score followed by ambulatory monitoring increases detection and treatment of atrial fibrillation-related conditions
Topics: Cardiology · AI-ECG · Screening · Randomized trials · Atrial fibrillation
Regions: Taiwan · Asia · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.