Can an AI-read ECG help flag acute aortic dissection? External sensitivity reached 77.4%
In an external cohort of 844 Chinese chest-pain patients, including 31 dissection events, an ECG-plus-age-and-sex model reached an AUC of 0.857 and flagged 25.1% of patients. Sensitivity was 77.4%, leaving missed cases; the authors say the model is not suitable for stand-alone diagnosis.
The short version
- What happened
- The retrospective two-centre study used 27,160 chest-pain patients at Zhongshan Hospital for model development and internal evaluation, then tested the fixed approach on 844 patients at a separate Qingpu Branch hospital.
- Who it affects
- Patients could benefit if a validated signal prompts earlier reassessment, but a negative score cannot safely exclude dissection at the reported sensitivity.
- What remains unproven
- The study was retrospective; clinicians did not receive alerts and no patient-care outcome was measured.
Keep in mind
The study was retrospective; clinicians did not receive alerts and no patient-care outcome was measured.
Only 30 internal-test and 31 external events were available, limiting precision and subgroup assessment.
The two hospitals are related institutions in Shanghai, so transport to independent regions, devices and health systems remains unproven.
External disease prevalence differed substantially from the internal test cohort, which affects predictive value and workflow burden.
Sensitivity of 77.4% externally and 63.3% internally is insufficient for stand-alone exclusion, while a 25.1% external alert rate could create substantial false-alert workload.
The model was not compared prospectively with established clinical risk pathways on time, imaging, complications, mortality or cost.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: it may provide a second look, but it cannot safely rule the condition in or out
An AI model reading a standard 12-lead electrocardiogram alongside age and sex separated acute aortic dissection cases from other chest-pain presentations better than age and sex alone in this retrospective Chinese study. In the external hospital cohort, its area under the receiver-operating-characteristic curve was 0.857, with a 95% confidence interval from 0.778 to 0.925. At a threshold fixed before external testing, sensitivity was 77.4% and specificity 76.9%.
Those figures do not make it a diagnostic test. With 31 events among 844 external patients, 77.4% sensitivity implies that some dissections were not flagged, while a 25.1% alert rate means roughly one quarter of all patients would trigger the signal. Clinicians would still need symptoms, examination, imaging decisions and established emergency pathways. The authors explicitly conclude that the observed sensitivity and alert burden do not support stand-alone diagnosis.
The value of the study is the external test at a second hospital and the decision to carry over the threshold unchanged. Many clinical AI papers report only random splits from one dataset, which can overstate transportability. Here, performance remained useful at a separate site, although both hospitals are linked to Fudan University in Shanghai and the event counts are small. This is peer-reviewed research published on 5 November, not a new clinical approval or emergency practice change, so it is Latest rather than Breaking.[1]
How the two-centre study was designed
The researchers retrospectively assembled 27,160 patients presenting with chest pain at Zhongshan Hospital. Patients, rather than individual ECG records, were split into development, tuning and internal-test sets to reduce leakage between stages. The separate external cohort comprised 844 patients from Zhongshan Hospital's Qingpu Branch. The internal test set contained 5,433 patients and 30 acute aortic dissection events; the external set contained 31 events.
Three approaches were compared: ECG plus age and sex, ECG alone, and age plus sex alone. The threshold for turning the continuous score into an alert was chosen in the tuning set using Youden's index, then applied without adjustment to the internal and external evaluation cohorts. That separation is important because choosing a threshold after seeing external results would make the apparent operating point less credible.
The paper also examined calibration, alert burden and a subgroup whose diagnosis was verified through aortic computed-tomography angiography. In the internal CTA-verified subset, which contained 94 patients, the full model's AUC was 0.781, with a 95% confidence interval from 0.671 to 0.877. This subset reduces uncertainty about how cases were established but is far too small to settle clinical utility. The study did not randomise care, deliver model alerts to clinicians or measure time to imaging, treatment, complications or survival.[1]
The distinction that matters
What the fixed threshold produced in two evaluation cohorts
The event count, sensitivity and alert burden all matter; AUC alone is not a clinical operating plan.
Internal test
5,433 patients; 30 events
AUC 0.803; sensitivity 63.3%
External test
844 patients; 31 events
AUC 0.857; sensitivity 77.4%
Fixed threshold
Apply tuning-set cutoff unchanged
External specificity 76.9%
Alert burden
Count all scores above threshold
25.1% of external patients flagged
What the comparisons reveal—and what they do not
The combined model's AUC was 0.803 internally and 0.857 externally. The ECG-only version reached 0.798 and 0.842, respectively, while age plus sex reached 0.670 internally and 0.553 externally. The small difference between the full and ECG-only models suggests most discrimination came from the ECG representation in these cohorts, while basic demographics alone travelled poorly to the external site.
AUC describes ranking across all possible thresholds; it does not tell a department how many people will be missed or sent for additional work at the chosen operating point. At the fixed threshold, internal sensitivity was 63.3%, specificity 81.0% and the alert rate 19.3%. Externally, sensitivity rose to 77.4%, specificity was 76.9% and alerts rose to 25.1%. The higher external AUC did not remove the trade-off: more than one in five events could still be missed, while many alerts would occur in patients without a dissection.
Because acute aortic dissection is uncommon and dangerous, the intended role matters more than a leaderboard number. A low-prevalence screening tool may generate many false alerts even with respectable specificity. The paper frames the score as a possible prompt for reassessment, not a replacement for imaging or specialist judgement. It does not report a prospective pathway showing whether an alert changes which patient receives CTA, whether it shortens delay or whether additional scanning creates harm and cost.[1]
Why the clinical denominator deserves close attention
The development dataset is large, but the outcome denominator is the number of confirmed dissection events, not the total ECG count. Thirty events in the 5,433-person internal test set and 31 in the 844-person external cohort yield wide uncertainty around sensitivity and subgroup behaviour. A few reclassified or missed cases can move an operating estimate materially. The external AUC confidence interval, 0.778 to 0.925, reflects some of that uncertainty.
The external prevalence is much higher than in the internal test set: 31 of 844 compared with 30 of 5,433. That can happen because hospitals, referral patterns or cohort definitions differ. AUC is theoretically less sensitive to prevalence than positive predictive value, but alert burden and the practical meaning of a positive result are not. A general emergency department should not assume that one quarter of its chest-pain patients would be flagged or that the same fraction would have disease.
Both sites are in Shanghai and share an institutional relationship, so external means a different hospital cohort, not a different health system or country. ECG devices, triage processes, population characteristics and diagnostic work-up may differ elsewhere. Before deployment, the model needs temporal validation on later patients, geographic validation across independent institutions and an analysis of performance by sex, age, comorbidity, ECG quality and clinical presentation with enough events in each group.[1]
What this could change for patients and emergency teams
Aortic dissection can resemble other causes of chest pain, and a routine ECG is already collected early for many patients. If a validated score could identify subtle patterns without delaying care, it might prompt a clinician to reconsider an initially low suspicion, review pulse or neurological findings, apply an established risk pathway or discuss imaging sooner. That is a decision-support hypothesis; this retrospective study did not test it.
The downside is alert fatigue and misplaced reassurance. A signal on one quarter of external patients could add cognitive and imaging pressure, while a negative result with sensitivity below 100% must not close the diagnostic process. The interface should display the intended role, operating point and uncertainty, and it should prevent staff from treating the score as clearance. Hospitals also need monitoring for drift when equipment, coding, referral patterns or patient populations change.
The research was funded by China's Noncommunicable Chronic Diseases National Science and Technology Major Project and a Fudan University AI for Science grant. The authors state that funders had no role in the study and declare no competing interests. Ethics approvals and waivers for retrospective data were reported. Those disclosures are reassuring but do not substitute for independent prospective validation or evidence that real-world use improves patient outcomes.[1]
What would change the assessment
The next decisive study would prospectively run the model in several independent emergency departments while initially hiding scores to measure calibration and workflow burden. A subsequent controlled evaluation could compare standard care with model-supported reassessment, using outcomes such as time to definitive imaging, missed or delayed diagnosis, unnecessary CTA, contrast or radiation exposure, treatment delay, complications, mortality and clinician workload. The threshold should be preregistered rather than retuned at every site without explanation.
Independent replication should include health systems outside Shanghai and enough dissection events to estimate sensitivity precisely. Researchers should publish positive and negative predictive values at each site's prevalence, calibration curves, failure cases and subgroup results. Comparison with established risk scores and biomarkers would show whether the ECG model adds information beyond what clinicians already use, rather than merely restating age, sex or obvious ECG abnormalities.
Evidence of improved outcomes without disproportionate scanning or missed cases would strengthen the case for clinical decision support. Failure to calibrate across sites, unstable performance on newer ECGs or no benefit over established pathways would weaken it. For now, the model is a promising externally tested signal in two related Chinese centres. It is not a medical device approval, a stand-alone diagnostic system or proof that an AI-read ECG makes patients safer.[1]
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 emergency medicine: compare the finding and its evidence limits.
Does AI help doctors read post-arrest ECGs? Specificity improved, but wrong advice pulled decisions with it
Fifteen physicians made 2,070 interpretations of 69 ECGs recorded after cardiac arrest. Algorithm support raised AUC from 0.882 to 0.912 and specificity from 84.64% to 90.43%, while 97% of changed decisions followed the AI—including cases where it was wrong.
Read the next reportExplore the full source and evidence record
What this means for people
- Patients could benefit if a validated signal prompts earlier reassessment, but a negative score cannot safely exclude dissection at the reported sensitivity.
- Emergency clinicians would need the score integrated into an established pathway with clear escalation rules, not presented as an autonomous diagnosis.
- Hospitals must count additional imaging, alert fatigue, missed cases and subgroup performance alongside discrimination metrics before deployment.
Global context
Chest-pain pathways and access to CTA vary widely across health systems. A model built in Shanghai may behave differently where disease prevalence, referral thresholds, ECG hardware or emergency staffing differ. External validation at a second centre is an important step, but safe global use requires independent geographic and temporal testing, transparent calibration and evidence that decision support improves care rather than simply increasing alerts.
What the evidence does not yet show
- The study was retrospective; clinicians did not receive alerts and no patient-care outcome was measured.
- Only 30 internal-test and 31 external events were available, limiting precision and subgroup assessment.
- The two hospitals are related institutions in Shanghai, so transport to independent regions, devices and health systems remains unproven.
- External disease prevalence differed substantially from the internal test cohort, which affects predictive value and workflow burden.
- Sensitivity of 77.4% externally and 63.3% internally is insufficient for stand-alone exclusion, while a 25.1% external alert rate could create substantial false-alert workload.
- The model was not compared prospectively with established clinical risk pathways on time, imaging, complications, mortality or cost.
What to watch next
- Prospective silent validation followed by controlled clinical-impact trials at independent hospitals.
- Comparison with established aortic-dissection risk tools, biomarkers and clinician assessment.
- Positive and negative predictive values, calibration and alert burden reported at each site's disease prevalence.
- Adequately powered subgroup and drift monitoring across sex, age, comorbidity, presentation, ECG device and geography.
Living evidence record
Impact record IAI-0PFAZPX
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
5 November 2026
Source trail
1 direct source across 1 source type.
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.
Single-source reporting disclosure
This record analyses one direct source. It can establish what BMC Medical Informatics and Decision Making published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 5 November 2026
Research question, topics and regions
Whether a 12-lead ECG model combined with age and sex can support reassessment of acute aortic dissection risk among chest-pain patients
Topics: Emergency medicine · Electrocardiography · Aortic dissection · Clinical decision support · External validation
Regions: China · Asia · Global
This report is labelled 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.