Friedman, Paul A.

Abstract Aims Cardiotoxicity is a serious side effect of anthracycline treatment, most commonly manifesting as a reduction in left ventricular ejection fraction (EF). Early recognition and treatment have been advocated, but robust, convenient, and cost-effective alternatives to cardiac imaging are missing. Recent developments in artificial intelligence (AI) techniques applied to electrocardiograms (ECGs) may fill this gap, but no study so far has demonstrated its merit for the detection of an abnormal EF after anthracycline therapy. Methods and results Single centre consecutive cohort study of all breast cancer patients with ECG and transthoracic echocardiography (TTE) evaluation before and after (neo)adjuvant anthracycline chemotherapy. Patients with HER2-directed therapy, metastatic disease, second primary malignancy, or pre-existing cardiovascular disease were excluded from the analyses as were patients with EF decline for reasons other than anthracycline-induced cardiotoxicity. Primary readout was the diagnostic performance of AI-ECG by area under the curve (AUC) for EFs < 50%. Of 989 consecutive female breast cancer patients, 22 developed a decline in EF attributed to anthracycline therapy over a follow-up time of 9.8 ± 4.2 years. After exclusion of patients who did not have ECGs within 90 days of a TTE, 20 cases and 683 controls remained. The AI-ECG model detected an EF < 50% and ≤ 35% after anthracycline therapy with an AUC of 0.93 and 0.94, respectively. Conclusion These data support the use of AI-ECG for cardiotoxicity screening after anthracycline-based chemotherapy. This technology could serve as a gatekeeper to more costly cardiac imaging and could enable patients to monitor themselves over long periods of time.

Background: The gap between electrocardiogram (ECG) artificial intelligence-derived age (ECG-Age) and chronological age (Age-Gap) predicts cardiovascular (CV) comorbidities and mortality reflecting physiologic age. Adverse childhood experiences (ACEs) are traumatic and harmful events that impact adulthood CV outcomes. We hypothesized that ACEs would be associated with physiologic age as determined by the Age-Gap. Methods: We applied a previously validated ECG-Age algorithm to the ECGs of 1,988 consecutive individuals aged ≥18, that within 2 years completed an ACEs questionnaire as part of psychological evaluations for weight management. The ACE questionnaire assesses 3 psychosocial domains (i) childhood abuse (psychological, physical or sexual), (ii) household dysfunction (substance abuse, mental illness, female figure treated violently, criminal behavior and parental separation/divorce) and (iii) neglect (emotional or physical) before age 18; results ranged 0 to 10 and were categorized as burden (0, 1-3, and ≥4) and presence (abuse, household dysfunction, neglect). Age-Gap was operationalized as (≤mean or > mean) positive values reflect physiological aging. Logistic regression analyses tested the association between ACE and Age-Gap. Results: Mean ± SD chronological age was 49±13 yr, ECG-Age 54±12 yr, Age-Gap 4.26±8.95 yr,75% women, average ACE 2±2, 32% experienced 0, 44% 1-3, and 24% ≥4 ACEs. Age-Gap was 1.4 years higher in those who experienced at least 1 ACE. Compared to those who didn’t experience ACE those who faced 1-3 ACE were more likely to have an Age-Gap >mean (OR 1.28 95%CI 1.57-1.04) independent of sex and the age-adjusted Charlson comorbidity index Fig A. Also, they were more prone to experiencing neglect and household dysfunction Fig B. Conclusions: Age-gap was associated with ACEs burden and domains, underscoring the need to identify ACEs during clinical evaluations to prevent ACEs at the individual, family, and community levels.

Background: An electrocardiogram (ECG)-based artificial intelligence (AI) algorithm has shown good performance in detecting hypertrophic cardiomyopathy (HCM AI-score). However, its clinical application may be challenging due to the low disease prevalence and potentially high false-positive rate when screening unenriched cohorts. Aims: To identify clinical characteristics associated with false-positive HCM AI-score results to improve screening tool application. Methods: From a cohort of 20,677 patients with a digital 12-lead ECG in routine clinical practice in 2021, we performed a detailed chart review of the 200 patients with highest HCM AI-score to determine their HCM status and collect clinical data. Stepwise forward logistic regression was used to create a clinical variable-based “candidacy for HCM screening” score (CHS score), differentiating true- and likely true-positive from likely false-positive HCM AI-score results. We then validated the CHS score in 200 patients with the highest HCM AI-scores from an independent cohort of 15,147 patients with an ECG in 2022. Results: Among the 200 patients in 2021 (median age 71 years, 48% female), 96 (48%) were likely false-positive HCM AI-score results, 71 (36%) definite or likely true-positive results, and 33 (16%) indeterminate. The CHS score included a history of surgical septal myectomy, coronary artery bypass graft surgery, cardiac valve replacement, and pacemaker, and had an AUC of 0.88 (95% CI 0.83-0.92) for the differentiation of true- versus false-positive AI results. In the 2022 cohort, the CHS score had an AUC of 0.82 (95% CI 0.76 - 0.88). After limiting the cohort to HCM candidates identified by the CHS score, the false-positive results rate was reduced to 27/114 (24%). Conclusions: By applying the CHS score to patients with positive HCM AI-screening results, the false-positive rate decreased from 49% to 24%. Utilizing a clinical model used in tandem with an AI model improved diagnostic yield.

Introduction: We have previously shown that ECG-derived age (ECG-Age) predicts all-cause and cardiovascular (CV) mortality, reflecting physiologic age. We hypothesized that ECG-Age minus chronological age (Age-Gap) would be associated with coronary artery calcium (CAC) assessed by computer tomography. Methods: We developed a neural network model to predict ECG-Age using 12 lead ECGs from 399,750 training, 99,977 validation and 275,056 testing patients. We then applied the algorithm to 41,202 consecutive patients that from 1997-2020 underwent clinically indicated CAC and ECG within 1 year. Major modifiable CV risk factors were collected as part of preventive cardiology or general medical evaluations. Exclusion criteria were statin therapy, pacemakers, or history of coronary artery disease. Results: We included 41,202 subjects (68.5% males, 83.6% Caucasian), mean chronological age was 55.2±9.1 yrs and the ECG-Age was 56.01±9.3 yrs. The mean Age-Gap was -0.78±7 yrs, R 2 =0.78, p<0.0001. Mean CAC score was 140.2±532.3, and 21,483 (52%) had CAC >0. After adjustment for age and sex, those older by ECG (defined as having a positive Age-Gap) were more likely to have CAC>0 (OR 1.35, 95% CI 1.34-2.37 for those with >1 but <2 SD of a positive Age-Gap; and OR 1.80, 95% CI 1.34-2.37 for those with ≥2 SD of a positive Age-Gap), while those younger by ECG were less likely to have any CAC (OR 0.70, 95% CI 0.57-0.87 for those with >1 but <2 SD of a negative Age-Gap; and OR 0.75, 95% CI 0.67-0.84 for those with ≥2 SD of a negative Age-Gap), respectively. This association remained significant after multivariate adjustment, Fig. A and increased with CAC burden, Fig. B , all P for trend <0.01. Conclusions: The difference between physiologic AI-ECG age and chronologic age, something likely representing rate of biological aging, is associated with presence of CAC, further supporting the hypothesis that AI-ECG age can identify people more likely to have subclinical coronary atherosclerosis.

Evaluating the Electrocardiographic Characteristics of Transgender Patients Undergoing Gender-affirming Hormone Therapy with Artificial Intelligence Introduction: An artificial intelligence (AI) electrocardiogram (ECG) algorithm has been demonstrated to identify a patient’s birth designated sex with 90.4% accuracy with an AUC = 0.97. However, the algorithm has not been evaluated in transgender and gender diverse individuals. We aimed to determine how the algorithm output varies among transgender individuals seen at the Mayo Clinic Transgender and Intersex Specialty Care Clinic (TISCC) who are undergoing gender-affirming hormonal treatment (GAHT). Methods: We applied the algorithm (which assessed the probability of “male” between 0 and 1, with < 0.5 considered as “female” and > 0.5 as “male”) to ECGs obtained from patients in the TISCC before and after the initiation of GAHT. GAHT frequently involves the use of testosterone in transgender men, and estrogen and anti-androgen agents in transgender women. Patient characteristics, including the date of the initiation of GAHT, were collected through chart review. Results: Among transgender women, the AI model probability of “male” decreased from 0.84± 0.25 (12/86 classified as males) to 0.59±0.36 (68/173 classified as male) with GAHT (p < 7.8e-10). Among transgender men, the AI model probability of “male” increased from 0.16 ± 0.28 (7/47 classified as male) to 0.41 ±0.38 (22/53 classified as male) with GAHT (p < 2.4e-4). Conclusion: After GAHT initiation, the Mayo Clinic AI ECG algorithm assigns a higher probability of being male among transgender men and a lower probability of being male among transgender women. These findings warrant further investigation into the electrocardiographic differences across genders and the effects of GAHT on the human heart.

Introduction: Over 100 FDA-approved medications, electrolyte perturbations, and many disease states can prolong the QTc beyond its 99 th percentile value resulting in acquired QT prolongation. In contrast, approximately 1 in 2000 people have congenital long QT syndrome (LQTS) hallmarked by pathological QT prolongation secondary to genetic defects in the heart. Hypothesis: An artificial intelligence (AI) deep neural network (DNN) can distinguish patients with LQTS from those with acquired QT prolongation. Methods: The study cohort included all patients with LQTS evaluated in the Windland Smith Rice Genetic Heart Rhythm Clinic and controls from Mayo Clinic’s ECG data vault comprising over 2.5 million patients. For the AI-DNN model, every patient/control with ≥ 1 ECG above age- and sex- specific 99 th percentile values for QTc [> 460 ms for all patients (male/female) < 13 years of age, or > 470 ms for men and > 480 ms for women above this age] was included. An AI-DNN involving a multi-layer c onvolutional n eural n etwork (CNN) was developed to classify patients. LQTS patients were age- and sex- matched to controls at 1:5 ratio. Results: Among the 1,599 patients with genetically confirmed LQTS, 808 had ≥ 1 ECG with QTc above the threshold (2,987 ECGs) compared to 361,069/2.5M controls (14% of Mayo Clinic patients getting an ECG, ‘presumed negative’; 989,313 ECGs). Following age- and sex- matching and splitting, 3,309 (training), 411 (validation) and 887 (testing) control ECGs were used. This model distinguished LQTS from those with acquired QT prolongation with an AUC of 0.896 (accuracy 85%, sensitivity 77%, specificity 88%, PPV 0.58, NPV 0.94). After exclusion of patients with a wide QRS (>150 ms) or pacemaker, the model remained successful in distinguishing the two groups (AUC 0.85, accuracy 78%, sensitivity 78%, specificity 87%, PPV 59%, NPV 94%). Conclusions: For patients with a QTc exceeding its 99 th percentile values, this novel AI-DNN functions as a LQTS mutation detector being able to identify patients with abnormal QT prolongation secondary to a LQTS-causative mutation rather than acquired QT prolongation with a >50% positive predictive value. This algorithm may facilitate screening for this potentially lethal, yet highly treatable, genetic heart disease.

Introduction: Deep-learning artificial intelligence (AI) algorithms can identify hypertrophic cardiomyopathy (HCM) from the electrocardiogram (ECG). Preliminary data indicate that AI-ECG may also track treatment response in HCM patients (pts) receiving mavacamten (Mava). In this placebo-controlled analysis, we investigated the performance of AI-ECG in HCM detection and its changes during Mava treatment compared to placebo in the phase 3, EXPLORER-HCM trial. Methods: Pts with non-paced ECGs enrolled in EXPLORER-HCM, evaluating Mava over 30 weeks of treatment, were included. All pts had a 12-lead ECG pre-treatment, at each study visit, and 8 weeks after treatment end. Two independently developed AI-ECG convolutional neural network algorithms from the Mayo Clinic and the University of California, San Francisco (UCSF) were applied to all ECGs suitable for analysis. Serial AI-ECG-predicted scores, which correlate with HCM likelihood (range 0-1), were examined between Mava and placebo groups. Results: A total of 101 Mava and 110 placebo pts were included (mean age 58.6, female 40%). At baseline, the mean Mayo and UCSF AI-ECG scores for HCM were 0.82 and 0.65, with 94% and 88% of pts exceeding the respective diagnostic thresholds of 0.11 (Mayo) and 0.4 (UCSF). During 30 weeks of follow-up in the Mava group, the algorithms showed a mean AI-ECG score reduction of 0.32 (Mayo) and 0.17 (UCSF) in all pts between start and end of treatment. In comparison, AI-ECG scores decreased by 0.05 (Mayo) and increased by 0.02 (UCSF) in the placebo group (Mava vs placebo, t-test p-values <0.001 for both algorithms; Figure ) Conclusion: In the EXPLORER-HCM trial population, two independent deep-learning AI-ECG algorithms correctly identified most HCM pts and demonstrated significant reductions in predicted HCM scores after treatment with Mava when compared to placebo. This relative normalization of the ECG suggests disease-modifying properties of Mava, deserving further investigation.

Background: We developed a deep learning algorithm to predict elevated coronary artery calcium (CAC) score from 12-lead electrocardiograms (CAC- ECG). We tested the hypothesis that this CAC-ECG algorithm will independently predict long-term survival. Methods: We leveraged a historical cohort of 43,210 consecutive patients who from the years 1997-2020 underwent clinically indicated ECG-gated unenhanced chest computed tomography (CT) to identify and quantify CAC and had an ECG within 1 year of the CT. Data on cardiovascular risk (CV) factors, and to calculate the Pooled Cohort Equation (PCE) for ASCVD was collected as part of preventive cardiology or general medical evaluations. We used the oldest CAC in record and excluded those taking statins at the time of the CAC. The algorithm was trained in 60% of the cohort, and the association between CAC-ECG and survival was evaluated with multivariate cox proportional hazard models in 40% of the remaining observations. Results: Of the 17,284 evaluated patients, mean ± SD age 55.9 ± 9.9, 33% female, 3,714 (21%) had elevated CAC ≥ 300. During an average of 15±5.9 years follow-up, 848 (5%) patients died. The algorithm’s area under the receiving operating characteristics curve (ROC), sensitivity, specificity, and accuracy to detect a CAC ≥ 300 were 0.83, 0.90%, 0.56%, 0.60%. Those with elevated CAC had a nearly two-fold risk of death, a value similar to those deemed positive by the ECG-CAC algorithm ( Figure A ). Risk of death increased with CAC-ECG probability quartiles ( Figure A ). The CAC-ECG algorithm enhanced predicted capabilities of the PCE across all ASCVD risk subgroups ( Figure B ), as well as in those with no CAC and elevated CAC ( Figure C ), all p for trend <0.001. Conclusions: A deep learning-enabled CAC-ECG algorithm was independently associated with long-term survival and enhanced current risk prediction paradigms. The CAC-ECG algorithm could help identify individuals at risk in primary prevention of CV disease.