Smartphone AI spotted 84% of hidden heart attacks that standard ECG pathways missed

A smartphone-based AI algorithm correctly identified occlusive myocardial infarction in 84 percent of cases where the standard diagnostic pathway caught just 42 percent. The study, presented March 20 at ESC Acute CardioVascular Care 2026 in Lisbon, tested the algorithm on 1,490 patients who arrived with symptoms suggesting a heart attack but whose initial ECG showed no ST-segment elevation - the classic warning sign that triggers immediate intervention.

The problem with missing the ST elevation

When a coronary artery becomes completely blocked, part of the heart muscle starts dying. If a standard 12-lead ECG shows ST-segment elevation - a specific pattern in the electrical trace - the diagnosis is clear and the response is immediate: the patient goes to the catheterization lab for an emergency procedure to reopen the artery. This protocol saves lives every day.

But not every complete coronary occlusion produces ST elevation on the ECG. A substantial fraction of patients with a fully blocked artery present without that telltale sign, and in those cases, the diagnosis becomes uncertain. Clinicians must rely on serial troponin blood tests, clinical judgment, and sometimes coronary angiography - an invasive imaging procedure - to determine whether the artery is blocked. The process takes time. For a patient whose heart muscle is actively dying, time is the thing they have least of.

"Many patients without an ST elevation have an occlusive MI, but it can be difficult for clinicians to quickly and accurately recognise this, leading to delays in providing emergency treatment," said Dr. Federico Nani from Central Hospital Bolzano, Italy, who presented the study.

1,490 patients, one phone, one algorithm

The study was a single-center prospective trial. All 1,490 patients had symptoms suggestive of acute coronary syndrome but no ST elevation on their initial ECG. Their mean age was 63 years; 42 percent were female. Each patient went through the standard diagnostic pathway: clinician interpretation of the ECG, troponin testing, and coronary angiography when warranted, following European Society of Cardiology guidelines.

In parallel, the initial ECG was fed into a CE-certified AI algorithm running on a smartphone. The algorithm analyzed the same electrical data the clinicians saw but applied pattern recognition trained on large datasets of ECG recordings with known outcomes.

The AI flagged 108 patients - 7 percent of the total - as having an occlusive MI. It ruled out occlusion in the remaining 1,382. Against the final confirmed diagnoses, the AI achieved:

• Overall accuracy: 84 percent correct identification of occlusive MI
• Sensitivity: 77 percent (proportion of true occlusions detected)
• Specificity: 99 percent (proportion of non-occlusions correctly excluded)
• Negative predictive value: 98 percent (when the AI said no occlusion, it was right 98 percent of the time)
• False negatives: 27 patients (2 percent)
• False positives: 17 patients (1 percent)

Standard pathway: 42 percent

The comparison with the conventional approach is stark. Under the standard diagnostic pathway, human ECG interpretation correctly identified occlusive MI in 42 percent of cases. The remaining patients required additional testing - troponin levels ruled out occlusion in 1,207 patients, while 283 underwent coronary angiography to reach a definitive diagnosis.

The conventional pathway works. It eventually reaches the right answer for most patients. But "eventually" is the operative word. Each additional test takes time, and for patients with a complete coronary occlusion, the delay between symptom onset and artery reopening directly affects how much heart muscle survives. An AI tool that can flag likely occlusions from the initial ECG, before any blood tests return, could shorten that critical window.

What the 99 percent specificity means in practice

The high specificity deserves particular attention, because in emergency medicine, false alarms are expensive. Every patient incorrectly flagged for occlusive MI would potentially be rushed to the catheterization lab for an invasive procedure they did not need. With 99 percent specificity, the AI produced only 17 false positives out of nearly 1,500 patients. That is a rate most emergency departments could absorb without overwhelming their resources.

The 77 percent sensitivity is less reassuring. It means the AI missed roughly one in four true occlusions. Those 27 false negatives represent patients whose blocked arteries the algorithm did not catch. In clinical practice, the AI would function as a complement to existing tools, not a replacement - a flag that says "look harder at this patient" rather than a final diagnosis. Patients not flagged by the AI would still proceed through the standard troponin and angiography pathway.

A phone-based tool, not a hospital system

One practical detail stands out: the algorithm runs on a smartphone. It does not require hospital-grade computing infrastructure, specialized hardware, or integration with electronic health records. A clinician can photograph or digitize an ECG strip and get a result on the spot. For emergency departments, ambulances, and settings where rapid triage decisions must be made with limited resources, that portability matters.

The CE certification indicates the algorithm has met European regulatory standards for medical devices, though the designation alone does not validate clinical efficacy - that requires the kind of prospective testing this study provides.

Single center, single study

The limitations are important and the researchers stated them clearly. This was a single-center study at one hospital in Bolzano, Italy. The patient population, the ECG equipment, the clinical workflows, and the prevalence of occlusive MI in the study cohort may not be representative of other settings. A 7 percent prevalence of occlusive MI among non-ST-elevation patients is within the expected range, but higher or lower prevalence would shift the predictive values.

The study was prospective, which is a strength - the AI was tested on patients as they arrived, not retrospectively on curated datasets. But it compared the AI's initial ECG interpretation against a pathway that includes troponin testing and angiography, which are fundamentally different information sources. The AI had access to the ECG alone; the standard pathway used blood tests and imaging. That the AI still outperformed the conventional approach on overall accuracy, despite having less information, is notable - but the comparison is not quite apples to apples.

Multi-center validation across different populations, healthcare systems, and ECG equipment is needed before the tool could be recommended for routine clinical use. The researchers acknowledged this directly.

Where AI meets the catheterization lab

The study fits into a broader trend of AI-assisted cardiac diagnostics that will be explored as the spotlight theme at this year's ESC Congress in Munich. The appeal of these tools is clear: heart attacks are time-sensitive emergencies where faster, more accurate triage can preserve heart muscle and save lives. AI algorithms can analyze ECGs in seconds, do not get fatigued during overnight shifts, and can be deployed wherever there is a phone.

But clinical adoption requires more than accuracy numbers. Emergency physicians need to trust the tool, understand its failure modes, and integrate it into workflows without creating alert fatigue or decision paralysis. Whether this particular algorithm - or one like it - becomes standard equipment in emergency departments depends on validation work that is still ahead.