Lopez-Jimenez, Francisco

Abstract Aims The increasing need for diagnostic echocardiography tests presents challenges in preserving the quality and promptness of reports. While Large Language Models (LLMs) have proven effective in summarizing clinical texts, their application in echo remains underexplored. Methods and results Adult echocardiography studies, conducted at the Mayo Clinic from 1 January 2017 to 31 December 2017, were categorized into two groups: development (all Mayo locations except Arizona) and Arizona validation sets. We adapted open-source LLMs (Llama-2, MedAlpaca, Zephyr, and Flan-T5) using In-Context Learning and Quantized Low-Rank Adaptation fine-tuning (FT) for echo report summarization from ‘Findings’ to ‘Impressions.’ Against cardiologist-generated Impressions, the models’ performance was assessed both quantitatively with automatic metrics and qualitatively by cardiologists. The development dataset included 97 506 reports from 71 717 unique patients, predominantly male (55.4%), with an average age of 64.3 ± 15.8 years. EchoGPT, a fine-tuned Llama-2 model, outperformed other models with win rates ranging from 87% to 99% in various automatic metrics, and produced reports comparable to cardiologists in qualitative review (significantly preferred in conciseness (P < 0.001), with no significant preference in completeness, correctness, and clinical utility). Correlations between automatic and human metrics were fair to modest, with the best being RadGraph F1 scores vs. clinical utility (r = 0.42) and automatic metrics showed insensitivity (0–5% drop) to changes in measurement numbers. Conclusion EchoGPT can generate draft reports for human review and approval, helping to streamline the workflow. However, scalable evaluation approaches dedicated to echo reports remains necessary.

Background A subset of individuals with major depressive disorder (MDD) have a high burden of cardiovascular risk factors and cerebral small‐vessel disease, implicating vascular disease in the development of depression. Cross‐sectional studies demonstrate a link between endothelial dysfunction and MDD, but the prospective association between peripheral endothelial dysfunction (PED) and an incident diagnosis of MDD is unknown. Methods and Results Patients undergoing a baseline assessment of cardiovascular risk were evaluated for PED using reactive hyperemia–peripheral arterial tonometry (≤1.8 consistent with PED). Patient medical records were reviewed to identify those who underwent a formal clinical evaluation of MDD after the index PED evaluation. The frequency of PED was compared in those with and without MDD. Logistic regression analyses were performed to assess the association between baseline PED and incident MDD. Between January 2006 and December 2020, 1614 patients underwent testing for PED. Four hundred eighty‐four (30.1%) patients underwent a formal evaluation for MDD after (0–15 years) the index procedure (mean±SD age, 52.8±13.8 years; 65.2% women). Of these, 157 (32.4%) had PED and 108 (31.0%) were diagnosed with MDD. Individuals with MDD had a higher frequency of PED (40.2% versus 30.2%; P =0.034) compared with those without MDD. In multivariable analyses, PED was significantly associated with MDD (odds ratio, 2.3 [95% CI, 1.4–3.8]; P <0.001). Conclusions PED is significantly associated with incident MDD. Thus, PED may be a useful marker to identify individuals at increased risk of depression who may benefit from more frequent and earlier management strategies.

Background Early detection of left and right ventricular systolic dysfunction (LVSD and RVSD respectively) in children can lead to intervention to reduce morbidity and death. Existing artificial intelligence algorithms can identify LVSD and RVSD in adults using a 12‐lead ECG; however, its efficacy in children is uncertain. We aimed to develop novel artificial intelligence–enabled ECG algorithms for LVSD and RVSD detection in pediatric patients. Methods and Results We identified 10 142 unique pediatric patients (age≤18) with a 10‐second, 12‐lead surface ECG within 14 days of a transthoracic echocardiogram, performed between 2002 and 2022. LVSD was defined quantitatively by left ventricular ejection fraction (LVEF). RVSD was defined semiquantitatively. Novel pediatric models for LVEF ≤35% and LVEF <50% achieved excellent test areas under the curve of 0.93 (95% CI, 0.89–0.98) and 0.88 (95% CI, 0.83–0.94) respectively. The model to detect LVEF <50% had a sensitivity of 0.85, specificity of 0.80, positive predictive value of 0.095, and negative predictive value of 0.995. In comparison, the previously validated adult data‐derived model for LVEF <35% achieved an area under the curve of 0.87 (95% CI, 0.84–0.90) for LVEF ≤35% in children. A novel pediatric model for any RVSD detection reached a test area under the curve of 0.90 (0.87–0.94). Conclusions An artificial intelligence–enabled ECG demonstrates accurate detection of both LVSD and RVSD in pediatric patients. While adult‐trained models offer good performance, improvements are seen when training pediatric‐specific models.

Abstract Aims Gender-affirming hormone therapy (GAHT) is used by some transgender individuals (TG), who comprise 1.4% of US population. However, the effects of GAHT on electrocardiogram (ECG) remain unknown. The objective is to assess the effects of GAHT on ECG changes in TG. Methods and results Twelve-lead ECGs of TG on GAHT at the Mayo Clinic were inspected using a validated artificial intelligence (AI) algorithm. The algorithm assigns a patient’s ECG male pattern probability on a scale of 0 (female) to 1 (male). In the primary analysis, done separately for transgender women (TGW) and transgender men (TGM), 12-lead ECGs were used to estimate the male pattern probability before and after GAHT. In a subanalysis, only patients with both pre- and post-GAHT EGCs were included. Further, the autopopulated PR, QRS, and QTc intervals were compared before and after GAHT. Among TGW (n = 86), the probability (mean ± SD) of an ECG male pattern was 0.84 ± 0.25 in the pre-GAHT group, and it was lowered to 0.59 ± 0.36 in the post-GAHT group (n = 173, P &lt; 7.8 × 10−10). Conversely, among TGM, male pattern probability was 0.16 ± 0.28 (n = 47) in the pre-GAHT group, and it was higher at 0.41 ± 0.38 in the post-GAHT group (n = 53, P &lt; 2.4×10−4). The trend persisted in the subanalysis. Furthermore, both the PR (P = 5.68 × 10−4) and QTc intervals (P = 6.65×10−6) prolonged among TGW. Among TGM, the QTc interval shortened (P = 4.8 × 10−2). Conclusion Among TG, GAHT is associated with ECG changes trending towards gender congruence, as determined by the AI algorithm and ECG intervals. Prospective studies are warranted to understand GAHT effects on cardiac structure and function.

Abstract Aims To test whether an index based on the combination of demographics and body volumes obtained with a multisensor 3D body volume (3D-BV) scanner and biplane imaging using a mobile application (myBVI®) will reliably predict the severity and presence of metabolic syndrome (MS). Methods and results We enrolled 1280 consecutive subjects who completed study protocol measurements, including 3D-BV and myBVI®. Body volumes and demographics were screened using the least absolute shrinkage and selection operator to select features associated with an MS severity score and prevalence. We randomly selected 80% of the subjects to train the models, and performance was assessed in 20% of the remaining observations and externally validated on 133 volunteers who prospectively underwent myBVI® measurements. The mean ± SD age was 43.7 ± 12.2 years, 63.7% were women, body mass index (BMI) was 28.2 ± 6.2 kg/m2, and 30.2% had MS and an MS severity z-score of −0.2 ± 0.9. Features β coefficients equal to zero were removed from the model, and 14 were included in the final model and used to calculate the body volume index (BVI), demonstrating an area under the receiving operating curve (AUC) of 0.83 in the validation set. The myBVI® cohort had a mean age of 33 ± 10.3 years, 61% of whom were women, 10.5% MS, an average MS severity z-score of −0.8, and an AUC of 0.88. Conclusion The described BVI model was associated with an increased severity and prevalence of MS compared with BMI and waist-to-hip ratio. Validation of the BVI had excellent performance when using myBVI®. This model could serve as a powerful screening tool for identifying MS.

Abstract Aims Mobile devices such as smartphones and watches can now record single-lead electrocardiograms (ECGs), making wearables a potential screening tool for cardiac and wellness monitoring outside of healthcare settings. Because friends and family often share their smart phones and devices, confirmation that a sample is from a given patient is important before it is added to the electronic health record. Methods and results We sought to determine whether the application of Siamese neural network would permit the diagnostic ECG sample to serve as both a medical test and biometric identifier. When using similarity scores to discriminate whether a pair of ECGs came from the same patient or different patients, inputs of single-lead and 12-lead medians produced an area under the curve of 0.94 and 0.97, respectively. Conclusion The similar performance of the single-lead and 12-lead configurations underscores the potential use of mobile devices to monitor cardiac health.

Background Extended sedentary behavior is a risk factor for chronic disease and mortality, even among those who exercise regularly. Given the time constraints of incorporating physical activity into daily schedules, and the high likelihood of sitting during office work, this environment may serve as a potentially feasible setting for interventions to reduce sedentary behavior. Methods and Results A randomized cross‐over clinical trial was conducted at an employee wellness center. Four office settings were evaluated on 4 consecutive days: stationary or sitting station on day 1 (referent), and 3 subsequent active workstations (standing, walking, or stepper) in randomized order. Neurocognitive function (Selective Attention, Grammatical Reasoning, Odd One Out, Object Reasoning, Visuospatial Intelligence, Limited‐Hold Memory, Paired Associates Learning, and Digit Span) and fine motor skills (typing speed and accuracy) were tested using validated tools. Average scores were compared among stations using linear regression with generalized estimating equations to adjust standard errors. Bonferroni method adjusted for multiple comparisons. Healthy subjects were enrolled (n=44), 28 (64%) women, mean±SD age 35±11 years, weight 75.5±17.1 kg, height 168.5±10.0 cm, and body mass index 26.5±5.2 kg/m 2 . When comparing active stations to sitting, neurocognitive test either improved or remained unchanged, while typing speed decreased without affecting typing errors. Overall results improved after day 1, suggesting habituation. We observed no major differences across active stations, except decrease in average typing speed 42.5 versus 39.7 words per minute with standing versus stepping ( P =0.003). Conclusions Active workstations improved cognitive performance, suggesting that these workstations can help decrease sedentary time without work performance impairment. Registration URL: https://www.clinicaltrials.gov ; Unique identifier: NCT06240286.

Abstract Aims Cardiac amyloidosis (CA) is common in patients with severe aortic stenosis (AS) undergoing transcatheter aortic valve replacement (TAVR). Cardiac amyloidosis has poor outcomes, and its assessment in all TAVR patients is costly and challenging. Electrocardiogram (ECG) artificial intelligence (AI) algorithms that screen for CA may be useful to identify at-risk patients. Methods and results In this retrospective analysis of our institutional National Cardiovascular Disease Registry (NCDR)-TAVR database, patients undergoing TAVR between January 2012 and December 2018 were included. Pre-TAVR CA probability was analysed by an ECG AI predictive model, with &gt;50% risk defined as high probability for CA. Univariable and propensity score covariate adjustment analyses using Cox regression were performed to compare clinical outcomes between patients with high CA probability vs. those with low probability at 1-year follow-up after TAVR. Of 1426 patients who underwent TAVR (mean age 81.0 ± 8.5 years, 57.6% male), 349 (24.4%) had high CA probability on pre-procedure ECG. Only 17 (1.2%) had a clinical diagnosis of CA. After multivariable adjustment, high probability of CA by ECG AI algorithm was significantly associated with increased all-cause mortality [hazard ratio (HR) 1.40, 95% confidence interval (CI) 1.01–1.96, P = 0.046] and higher rates of major adverse cardiovascular events (transient ischaemic attack (TIA)/stroke, myocardial infarction, and heart failure hospitalizations] (HR 1.36, 95% CI 1.01–1.82, P = 0.041), driven primarily by heart failure hospitalizations (HR 1.58, 95% CI 1.13–2.20, P = 0.008) at 1-year follow-up. There were no significant differences in TIA/stroke or myocardial infarction. Conclusion Artificial intelligence applied to pre-TAVR ECGs identifies a subgroup at higher risk of clinical events. These targeted patients may benefit from further diagnostic evaluation for CA.

Abstract Aims Augmenting echocardiography with artificial intelligence would allow for automated assessment of routine parameters and identification of disease patterns not easily recognized otherwise. View classification is an essential first step before deep learning can be applied to the echocardiogram. Methods and results We trained two- and three-dimensional convolutional neural networks (CNNs) using transthoracic echocardiographic (TTE) studies obtained from 909 patients to classify nine view categories (10 269 videos). Transthoracic echocardiographic studies from 229 patients were used in internal validation (2582 videos). Convolutional neural networks were tested on 100 patients with comprehensive TTE studies (where the two examples chosen by CNNs as most likely to represent a view were evaluated) and 408 patients with five view categories obtained via point-of-care ultrasound (POCUS). The overall accuracy of the two-dimensional CNN was 96.8%, and the averaged area under the curve (AUC) was 0.997 on the comprehensive TTE testing set; these numbers were 98.4% and 0.998, respectively, on the POCUS set. For the three-dimensional CNN, the accuracy and AUC were 96.3% and 0.998 for full TTE studies and 95.0% and 0.996 on POCUS videos, respectively. The positive predictive value, which defined correctly identified predicted views, was higher with two-dimensional rather than three-dimensional networks, exceeding 93% in apical, short-axis aortic valve, and parasternal long-axis left ventricle views. Conclusion An automated view classifier utilizing CNNs was able to classify cardiac views obtained using TTE and POCUS with high accuracy. The view classifier will facilitate the application of deep learning to echocardiography.