Late-breaking science: AI at the bedside
AI-powered stethoscope, speech analysis for HF, interpreting ECGs with AI and prognostic tool for predicting MACE.
Monday’s Late-Breaking Science session, “Artificial Intelligence at the Bedside” also found that:
- AI-powered stethoscope doubles detection of peripartum cardiomyopathy.
- Speech analysis beats usual HF predictive model.
- AI interpretation of ECG speeds time to treatment.
- ORFAN: Novel prognostic tool improves MACE prediction, preventative treatment, by 40% versus conventional predictive models.
AI-powered stethoscope doubles detection of peripartum cardiomyopathy
A novel stethoscope that can record and interpret a single-lead electrocardiogram (ECG) has doubled the detection of peripartum cardiomyopathy in a multicenter study in Nigeria. Nigeria has the highest reported rate of peripartum cardiomyopathies in the world, 1:100 deliveries. The U.S. rate is about 1:2,000 deliveries and up to 1:700 deliveries in African Americans.
“The diagnosis of cardiomyopathies is challenging to make during pregnancy because there is a 
Demilade Adedinsewo
Mayo researchers developed an algorithm to detect the presence of reduced left ventricular ejection fraction from a 12-lead ECG. The artificial intelligence-powered algorithm was adapted for use with a single-lead ECG built into a digital stethoscope and tested across six teaching hospitals in Nigeria that have both cardiovascular and obstetric care. Testing involves little more than placing the stethoscope on the patient’s chest.
“We have become increasingly aware of the connection between heart diseases and pregnancy. However, current guidelines do not recommend routine screening for cardiac dysfunction during pregnancy. In addition, effective screening tests are not available,” Dr. Adedinsewo said. “This artificial intelligence algorithm gives us a simple, inexpensive and effective way to screen women for cardiomyopathy at the point of care, and treatment decisions can be made on the spot. This will allow providers to initiate appropriate medical therapy and refer women for cardiovascular care earlier, because this is a treatable condition, but life-threatening if we don’t treat it.”
The median age of women in the study was 31 years and 39% were in their third trimester. Four percent of women in the intervention arm had cardiomyopathy detected with AI-guided screening while 1.8% were detected in the control arm, suggesting that half of cases likely go undetected with usual care. AI-guided screening using the digital stethoscope doubled the diagnosis of cardiomyopathy, which was confirmed by echocardiography (OR=2.31, p=0.019). Among women in the intervention arm, the digital stethoscope had an area under the curve of 0.95 for detection of LVEF<50% and 0.98 for LVEF<40%.
“The implications are not just improving diagnosis and plugging women into appropriate care,” Dr. Adedinsewo said. “This can also reduce disparities in cardiovascular outcomes during pregnancy because Black women have a higher risk of cardiomyopathy compared to white women. Ultimately, this could help reduce mortality from cardiovascular causes, which is the leading causes of maternal death.”
Speech analysis algorithm can accurately predict heart failure earlier than clinical signs
An early-stage speech analysis algorithm outperforms traditional clinical signs and symptoms of heart failure. The initial smartphone-based system could detect heart failure three weeks prior to a heart failure event with 80% sensitivity when patients speak a set of standard sentences in their native language. A single-arm confirmation study showed similar sensitivity and a larger validation trial is currently under way. 
William T. Abraham
“Our current standard of care approach to monitoring heart failure patients has not helped us keep very many people well and out of the hospital,” said William T. Abraham, MD, FACC, FAHA, professor of medicine, physiology and cell biology and College of Medicine Distinguished Professor at The Ohio State University Wexner Medical Center. “Many studies of monitoring signs and symptoms, changes in daily weights, vital signs, all the usual measures, have failed. There is a huge need that speech analysis appears to meet. Just like your automated banking system can recognize and respond to your voice, so can this heart failure detection app.”
The Cordio HearO Community Study enrolled 416 outpatients: 263 in the algorithm development group and 153 in the testing group, with New York Heart Association (NYHA) Class II and III HF without regard to left ventricular ejection fraction (LFEF). Patients spoke five sentences into their smartphone daily, which were analyzed by a cloud-based system to assess changes in speech measures indicative of worsening heart failure.
Patients were a mean of 67 years, females were well represented, 60% had LVEF <40%, and mean NT-proBNP was about 3,000 pg/mL. Patients spoke in their native languages, a mixture of Arabic, English, Hebrew and Russian, and were followed for up to 44 months between March 2018 and April 2023. Patients recorded 82% of the expected occasions.
The first 263 patients and their 58 heart failure events in 43 patients were used to train the algorithm. Of the 43 first heart failure events, 35 (81%) were detected by the system 24 days prior to positive clinical signs of a heart failure event requiring hospitalization and/or intravenous therapies. The remaining 153 patients were used to evaluate the algorithm, which detected 77% of subsequent heart failure events a similar three weeks prior to the detection of conventional clinical signs. The algorithm is currently being validated in a larger U.S.-based population speaking English, Spanish and a variety of other languages represented in the United States.
“This app is language-independent,” Dr. Abraham said. “It develops an individual patient baseline during periods of stability, then continually refines and compares changes in voice and delivery indicating worsening heart failure. There are many implantable and wearable technologies being developed for the same purpose, but speech analysis is more straightforward and easier for patients to use. It is as simple as picking up your smartphone daily and speaking five sentences.”
AI beats standard care to diagnose STEMI
Adding AI to standard ECG shows promise in improving timely diagnosis and treatment of ST-elevation myocardial infarction (STEMI) in the emergency and inpatient departments. AI-enabled ECG (AI-ECG) reduced ECG to cath lab time to 43.3 minutes from 52.3 for usual care (p=0.003). AI-ECG also reduced non-imaging validated STEMI from 15.8% for usual care to 6.5% (OR=0.37). 
Chin-Sheng Lin
“Our study is the first randomized controlled clinical trial to evaluate the effects of AI-ECG on the reduction of ECG to lab time,” said Chin-Sheng Lin, MD, PhD, chief of cardiology at Tri-Service General Hospital, National Defense Medical Center in Taipei, Taiwan. “We showed that AI-ECG significantly shortens time to cath lab and had fewer non-validated STEMIs compared to usual care. Our AI-ECG exhibited high diagnostic accuracy with a positive predictive value of 88.0% and a negative predictive value of 99.9%.”
The Artificial Intelligence Enabled Rapid Identification of ST-Elevation Myocardial Infarction Using Electrocardiogram (ARISE) trial included 43,177 patients presenting at the emergency department and inpatient department with at least one ECG without coronary angiography within three days. Patients received either standard care, which was a conventional ECG with interpretation by on-duty cardiologists (21,622 patients) or AI-ECG (21,555 patients) for rapid identification and triage of STEMI. A subset of patients received coronary angiograms, 77 in the intervention group and 68 in the standard care group.
The mean age for the entire cohort was 60 years and half were female. The CAG-STEMI group was slightly older, 65 years, and predominately male (80%).
The primary outcome was time from ECG to admission to the catheterization lab. Secondary outcomes included event analysis, clinical outcomes and diagnostic accuracy.
“This study is a signal for benefit from AI-ECG, at a single institution, a small sample size and short follow-up time,” Dr. Lin said. “Large scale, multicenter studies are needed to confirm the effects of AI-ECG on STEMI diagnosis and treatment.”
AI-powered prognostic tool improves MACE prediction and changes the management in 40% of people having CT heart scans
A novel prognostic algorithm based on routine coronary CT angiograms (CCTA) can improve prediction of cardiovascular events by 47% versus the conventional QRISK3 clinical risk factor prediction model. ORFAN, a prospective, real-world trial of Fat Attenuation Index (FAI) Score in the United Kingdom, triggered initiation of statin treatment (25% of patients), statin dose-intensification (14%) or addition of colchicine or some other cardiovascular treatment (8%). The trial followed >40,000 consecutive patients undergoing CCTA for up to a decade. 
Charalambos Antoniades
“We developed a technology that analyzes the fat surrounding coronary arteries that can tell you how inflamed each artery is,” said Charalambos Antoniades, MD, PhD, British Heart Foundation chair of cardiovascular medicine and director of the Acute Multidisciplinary Imaging and Intervention Centre at Oxford University in Oxford, U.K. “These changes can be quantified using images from routine CCTA. FAI Score identifies patients at risk for a major adverse cardiovascular event: If you are in the highest 25% of coronary inflammation measured in this way, your risk of dying from a heart attack over the next 10 years is 20 times higher than people with low inflammation. This is a clinical game changer.”
FAI score was developed and trained on a U.S. population with an AI-assisted prognostic model (AI-Risk) using plaque burden and other data in routine CCTA images. The prospective ORFAN study validated the model in a British population of 40,091 patients undergoing CCTA at seven U.K. hospitals. A nested study followed 3,392 patients for a mean of 7.7 years to evaluate the prognostic value of FAI Score and the performance of AI-Risk. An evaluation of 744 consecutive patients undergoing CCTA for chest pain at four hospitals assessed the impact of AI-Risk on clinical management.
The primary endpoint was major adverse cardiac events (MACE) a composite of non-fatal myocardial infarction, new onset heart failure or cardiac death.
MACE occurred in 20.1% of patients with obstructive CAD and in 9.5% of patients without obstructive CAD. Patients with FAI Score above the 75th percentile who had obstructive CAD had 6.7 times higher risk of MACE, those without obstructive CAD had 4.8-fold higher risk. AI-Risk reclassified 45% of patients and triggered new or intensified preventive treatment versus the QRISK3 clinical risk factor prediction model.
“This is a highly cost-effective approach that allows you to identify individuals at risk early and treat them with cheap agents to prevent heart attacks and complex, expensive treatments later,” Dr. Antoniades said. “You can apply this kind of technology in health care systems anywhere in the world.”
