Carter, R

Background Patients with atrial fibrillation (AF) often complain of dyspnea, raising the question that symptoms could be related to unrecognized heart failure (HF) with preserved ejection fraction (HFpEF). Methods We used the HFpEF‐age, body mass index and history of AF algorithm to estimate the probability of undiagnosed HFpEF in CABANA (Catheter Ablation vs. Antiarrhythmic Drug Therapy), and its interaction with catheter ablation on quality of life (QOL). Probable HFpEF was defined as patient‐reported dyspnea with HFpEF‐age, body mass index and history of AF probability≥75%. Absence of dyspnea or HFpEF‐age, body mass index and history of AF probability<75% was considered to reflect patients without HFpEF. The effect of randomization to catheter ablation on QOL (Mayo Atrial Fibrillation‐Specific Symptom Inventory questionnaire and EuroQol‐5 Dimension‐3 Level score) was assessed using mixed models. Results Of participants without known HF, 70% (n=1225) had probable HFpEF and the remaining 30% (n=522) did not. Those with probable HFpEF had worse New York Heart Association class, EuroQol‐5 Dimension‐3 Level score, Mayo Atrial Fibrillation‐Specific Symptom Inventory severity, and Mayo Atrial Fibrillation‐Specific Symptom Inventory frequency scores ( P <0.0001 for all), with higher risk of HF hospitalization (hazard ratio [HR], 2.19 [95% CI, 1.11–4.31], P =0.01). Ablation resulted in greater improvement in EuroQol‐5 Dimension‐3 Level and Mayo Atrial Fibrillation‐Specific Symptom Inventory severity/frequency scores in probable HFpEF (interactions P =0.005, P =0.05, and P =0.04 respectively). In probable HFpEF, catheter ablation was associated with lower risk of cardiovascular hospitalization (HR, 0.78 [95% CI, 0.66–0.92], P =0.003, interaction P =0.03) but not HF hospitalization (HR, 0.94 [95% CI, 0.55–1.64], P =0.84). Conclusions Nearly three quarters of CABANA participants had potentially undiagnosed HFpEF, with worse QOL and risk of HF hospitalization. Catheter ablation in probable HFpEF resulted in greater improvement in QOL, but residual QOL impairment and HF risk remained elevated despite ablation. These data reinforce the importance of diligent consideration of HFpEF among patients with symptomatic AF to ensure optimal use of foundational treatments for HF.

Abstract Background Artificial intelligence enhanced electrocardiograms (AI-ECG) have matured over the past 10 years. More recently, the AI-ECG estimate of biological sex has shown potential as a novel tool for cardiovascular risk stratification among women. Cardiovascular disease is a leading cause of pregnancy-related mortality; however, there have been limited approaches for estimating the physiological demands of pregnancy and its impact of maternal cardiovascular outcomes. Purpose To estimate pregnancy-related changes on AI-ECG-based sex and age estimation in a prospective cohort of otherwise healthy, complication free pregnant women. Methods This secondary analysis of the SPEC-AI Nigeria Clinical Trial identified pregnant women without any pregnancy-related complications (at baseline or during follow-up) who had at least two 12 lead ECGs recorded during the study. The ECGs were analyzed with previously-developed algorithms to predict age and biological sex. To address bias in the AI-ECG age estimate, a reference cohort of 25,144 healthy patients was analyzed using quantile regression to estimate the 75th percentile of the reference distribution as a function of age. The estimate of biological sex was used to determine a sex discordance score with a discordance of 0.2 considered clinically relevant. To simplify the trends in the data, the ECGs were categorized as during pregnancy (&lt;=28 days prior to birth), peripartum (+/- 28 days of birth), and postpartum (28+ days after birth), and mixed effects models with a random intercept were used for analysis. Results A total of 207 females with a median age of 31 years (IQR 28 to 36 years) and normal blood pressures (systolic: 110 (IQR 10 to 120); diastolic: 70 (IQR 60-77); mmHg) were included from 5 clinical sites in Nigeria. The majority (73%, 150/207) of participants were in their second trimester at time of enrollment and the median number of ECGs per participant was 3 (max 7). The sex discordance score was elevated in 38% of the women at baseline and this percentage fell to 23% postpartum; which resulted in a decrease in the mean discordance score of 0.03 (95% CI: -0.06 to 0.0; p=0.028). Conversely, the percentage of patients with elevated age prediction increased from 35% to 41% over the course of pregnancy. This resulted in an increase in the bias-corrected AI-ECG age of 3.0 years (95% CI: 1.5 to 4.6 years; p&lt;0.001). Conclusion This study demonstrates the preliminary utility of serial AI-ECG based assessment as a marker for the biological effects of pregnancy. The increase in sex discordance score during pregnancy may be related to changes in cardiac structure, such as increased ventricular mass and chamber dilation which often reverses postpartum. However, the impact on age estimation appeared to persist in the early postpartum period. Further study is warranted to associate any changes observed with long term maternal or fetal outcomes.Figure

Abstract Background Artificial intelligence (AI) augmented laboratory tests can improve quality, efficiency, cost-effectiveness, and staff satisfaction. However, the clinical success of these tests can introduce unforeseen challenges for model retraining. This study describes the discovery of an “AI label feedback loop” in a clinically implemented AI-augmented kidney stone composition test. Methods An AI-augmented kidney stone composition test (V1) has been previously deployed for clinical kidney stone characterization. After several years of clinical use, a retrained model (V2) was developed using 6 times more data. Model performance of both V1 and V2 were evaluated across 3 datasets: a recent production validation (hold-out) set (mostly AI-influenced labels), the original V1 validation set (pre-AI, entirely human-labeled), and a subset of recent cases with exclusively human-generated or human-corrected labels. Results V2 demonstrated a 10% lower concordance rate than V1 when evaluated on the recent production hold-out set, despite a much larger training dataset. Performance between V1 and V2 was similar when applied to the pre-AI validation set. Notably, V2 outperformed V1 on the recent subset of cases with human-only or human-corrected labels, particularly for less-common stone types. These findings revealed an AI label feedback loop, confounding retraining and evaluation. Conclusions The integration of AI into clinical practice can potentially influence reported test results, complicating the development and evaluation of future models. To mitigate AI label feedback loops, ongoing human annotation and careful validation set construction are essential. These strategies can ensure reliable performance assessment and support the safe evolution of clinical AI systems.

Abstract Disclosure: M.D. Hurtado: None. E. Tama: None. D. Hodge: None. A. Weston: None. S. Fahad: None. K. Ruddy: None. J. Olson: None. F. Coach: None. R. Carter: None. S. D'Andre: None. A. Acosta: Consulting Fee; Self; Amgen Inc. Grant Recipient; Self; Vivus USA, Novo Nordisk. C. Shufelt: None. Background: Breast cancer (BC) is the most common cancer in women. Weight gain after treatment for early-stage BC is associated with a higher risk of recurrence. Excess adiposity is also a major risk factor for cardiovascular disease, the leading cause of non-BC related death in BC survivors. This study aimed to describe weight trends in a large cohort of BC survivors and to identify predictors of weight gain after BC diagnosis. Methods: This was a retrospective study of BC survivors from the prospectively consented Mayo Clinic Breast Cancer Registry, which has enrolled &gt;10,000 patients seen at least once at Mayo Clinic Rochester for a BC diagnosed within the prior year. We extracted data on weights at baseline, and years 1, 4, and 6 after BC diagnosis to generate longitudinal weight trajectory curves on the change from baseline. Weights were either reported by patients in the registry questionnaires or extracted from the electronic health records. We identified BC survivors with weight gain &gt;10% from baseline weight at years 1, 4, and 6. We then conducted univariable logistic regression analyses to identify predictors of &gt;10% weight gain. Variables considered included: demographics, anthropometrics, BC clinical and pathologic characteristics, BRCA 1 and 2 genetic mutation status, and treatment (surgery, neoadjuvant or adjuvant chemotherapy, and/or radiotherapy). Results: We included 4744 BC survivors, mostly white (95%) and non-Hispanic (91%), with a mean age and weight at BC diagnosis of 58 ± 13 years and 76 ± 18 kg. Most patients had clinical stage I (43%) and ER+ and PR+ BC (83% and 76%, respectively). Most patients underwent breast preserving surgery (62%), half received radiation (54%), one third (36%) received systemic chemotherapy, and two thirds (61%) received endocrine therapy. At one year after BC diagnosis the average weight change was negligible compared to baseline (mean difference -0.2 ± 8 kg), but there were small but significant average increases by years 4 and 6 (0.5 ± 9 kg, p=0.002; and 0.8 ± 10 kg, p&lt;0.001; respectively). At years 1, 4, and 6, the percentage of BC survivors gaining &gt;10% of their baseline weight was 10%, 15%, and 18%, respectively. Baseline characteristics associated with &gt;10% weight gain at 6 years included lower baseline weight (p&lt;0.001), younger age at BC diagnosis (p&lt;0.001), more advanced clinical and pathologic BC stage (p=0.008), ER+ and PR+ tumor biology (p=0.02 for both), BRCA2 mutation (p&lt;0.001), mastectomy (compared to breast preserving surgeries, p=0.008), the use of systemic chemotherapy treatment (p&lt;0.001), and the use of endocrine therapy (p=0.03). Conclusion: While most BC survivors do not experience weight gain after survivorship, 1 out 5 may experience excessive weight gain of more than 10%. Baseline anthropometric and BC-related variables are associated with excessive weight gain and could be used as predictors of weight gain in this population. Presentation: 6/3/2024

Abstract Background Analytical and post-analytical errors are a persistent problem, which can be benchmarked using sigma metrics. Manual processes are associated with error rates of 3–4 sigma. Secondary review of manual reporting is meant to reduce/eliminate errors and is widely used, yet the true efficacy of manual secondary review is difficult to assess. Kidney stone analysis is a manually intensive process, requiring personnel to identify stone composition from spectra and to transcribe stone percent composition into an LIS system. Recently, we developed an AI program that allows for quantitative determination of kidney stone composition from spectra. This allows for a unique opportunity to estimate error rates associated with manual reporting and those following secondary review by means of subsequent AI quality assurance analysis. Methods Experienced technologists assessed stone composition and estimated the percent of each component (± 10%), based on Fourier transform infrared (FTIR) spectroscopy. These manual spectral composition assessments are secondarily reviewed by other experienced technologists for identity and percentage composition accuracy. The recently described Kidney stone AI spectral analysis (RokkStar) was implemented to assess reported stone composition identity and percent composition for agreement following secondary review in 159 334 stone reports. AI flagged discrepant reports were reviewed by personnel and reports were revised when appropriate. Secondary reviewer error detection rate was obtained by monitoring detection in a subset of 7440 kidney stones, and was that rate was extrapolated to establish secondary review error detection rate over the entire data set. The sum (total error) of estimated secondary reviewer error detection, and well as AI-mediated error detection, and all client reported errors were used to determine error rates in sigma following manual reporting, secondary review, and AI quality assurance. Results Manual entry led to an estimated 2060 errors in 159 334 stone reports, yielding an initial sigma value of 3.73. Secondary manual review resulted in an estimated 1909 (error detection rate of 92.5%) of these errors being detected over the entire report set, improving the sigma value to 4.61. Post-verification AI caught an additional 138 errors, leading to an improved post-AI sigma value of 5.27. Post-AI quality assurance, there were 13 remaining errors, 9 of which were manual reporting process deviations subsequently caught by laboratory staff, and 4 were reporting errors observed by client physicians. All 13 of these errors involved usual stone constituents or typographical errors that were not assessed by the current AI quality assurance version. Conclusion This study represents a unique large-scale estimate of efficacy of secondary result review by means of AI quality assurance review. While the presence of undetected errors cannot be completely ruled out, in this case the relative paucity of client reported errors, as well the alignment of the estimated rate of manual errors with previous reports may indicate that most errors were detected in this system. Importantly, manual secondary review improved errors rates by ∼0.9 sigma, and the use of AI further improved estimated overall sigma metric by ∼0.7, leading to a “world class” 5+ sigma level for this complex reporting process.

Abstract Background Screening for left ventricular (LV) systolic dysfunction (defined as ejection fraction ≤35%) based on data from a standard 12-lead electrocardiogram (ECG) has become well established when standard digital ECGs are available–8 independent leads sampled at least 250 hertz for 10 seconds. As the algorithm has been incorporated into various clinical scenarios and ancillary research projects, a limitation of the binary classification at 35% has become apparent. Purpose The objective of this study was to develop and validate a deep learning-based algorithm that would classify LVEF into three categories based on only the digital ECG input. Methods After IRB approval, native digital resting ECGs acquired between 1/1/2010 and 12/31/2021 on patients seen in Mayo Clinic in Jacksonville were extracted from the institutional electronic ECG database management system (MUSE, GE Healthcare). These ECGs were matched with transthoracic echocardiograms obtained up to four days prior or 30 days after the ECGs acquisition. A convolutional neural network consisting of 8 layers of convolutions, batch normalization and pooling was trained using Keras and Tensorflow with hyper-parameter optimization for L1 and L2 regularization, learning rate adjustments, and class weights to predict three classes of LVEF: ≤35%, 36–51%, and ≥52% based on clinical relevance. The primary measure of overall performance was the detection of LVEF ≤35%; however, the triad of model predictions was also considered in translating the model output to human interpretable findings. Results A total of 30,153 patients (60,169 ECG pairings; mean age 63 years; 48% male) were randomly split at the patient level into training (24,172 patients), validation (2,973 patients) and testing (3,008 patients). The trained model provided robust discrimination in the withheld testing data – AUROC of 0.941 (95% CI: 0.931 to 0.950). Using the optimal model threshold based on Youden's index from the validation data (0.186), sensitivity and specificity were estimated to be 87.9% (95% CI: 83.8% to 91.2%) and 86.3% (95% CI: 85.4% to 87.2%) in the testing data. In instances where discordant predictions were observed, the posterior distribution of model probabilities provide additional insights into the possible underlying value of LVEF (Figure 1). Conclusions The utilization of a multi-category deep learning classification model for the detection of reduced ejection fraction adds new dimensions to the use of AI technologies on digital ECGs. This work shows high discrimination can still be obtained when using three classes of LVEF. Funding Acknowledgement Type of funding sources: None.

Abstract Background Pregnancy related cardiomyopathy is a significant cause of maternal morbidity and mortality globally. A presumed overlap between normal pregnancy-associated symptoms and clinical symptoms of cardiomyopathy contributes to delays in diagnosis and increased risk of maternal mortality. Purpose We sought to evaluate the association between patient-reported cardiovascular symptoms and the presence of cardiomyopathy among pregnant and postpartum patients. We hypothesize that individual cardiovascular symptoms are unrelated to the presence of cardiomyopathy. We also evaluated the use of a novel adaptation of a validated health status questionnaire in relation to cardiomyopathy. Methods We enrolled 48 pregnant (&gt;13 weeks) and postpartum (up to 12 months) participants in a prospective study between October 2021 and February 2022. All study participants completed a baseline questionnaire, which included current cardiovascular symptoms, an assessment of health status using an adapted version of the Kansas City Cardiomyopathy Questionnaire (KCCQ-12), followed by a resting transthoracic echocardiogram on the same day. We defined cardiomyopathy as a left ventricular ejection fraction (LVEF) &lt;50% based on 2-D echocardiography. Fisher's exact and Wilcoxon rank-sum tests were employed to evaluate the association between reported cardiovascular symptoms, the adapted KCCQ-12 (KCC-A) score, and cardiomyopathy. Results At the time of enrollment, 67% were pregnant and 33% postpartum. Forty-eight percent identified as White, 31% as Black, 10% as Asian, and 10% as other race. The median age was 31 years (Q1: 28, Q3: 35) and 6% had an LVEF &lt;50%. We found no statistically significant association between four reported cardiovascular symptoms (shortness of breath, orthopnea, fast breathing, and episodes of “asthma” that did not improve with inhalers or other treatment) and cardiomyopathy or medial E/e' ratio. KCC-A scores were low in the study population overall (median 52; Q1:40, Q3: 61). We demonstrated a significantly lower KCC-A score among women with LVEF &lt;50% (median 24; Q1: 15, Q3: 44) compared to women with LVEF ≥50% (median 54; Q1: 44, Q3: 61) p=0.02. Conclusions We showed no significant association between individual cardiovascular symptoms and cardiomyopathy in an obstetric population. However, we demonstrate for the first time that an adapted KCCQ-12 questionnaire for health status assessment could potentially identify women with a high-likelihood of cardiomyopathy during the peripartum period who may benefit from additional evaluation including echocardiography. Larger studies are needed to validate this finding. Funding Acknowledgement Type of funding sources: Foundation. Main funding source(s): This study was funded by a research grant from the Miami Heart Research Institute, Florida Heart Foundation. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health

Abstract Background Screening for left ventricular (LV) systolic dysfunction (defined as ejection fraction ≤35%) based on data from a standard 12-lead electrocardiogram (ECG) has become well established when standard digital ECGs are available–8 independent leads sampled at least 250 hertz for 10 seconds. As the algorithm has been incorporated into various clinical scenarios and ancillary research projects, a limitation of the binary classification at 35% has become apparent. Purpose The objective of this study was to develop and validate a deep learning-based algorithm that would classify LVEF into three categories based on only the digital ECG input. Methods After IRB approval, native digital resting ECGs acquired between 1/1/2010 and 12/31/2021 on patients seen in Mayo Clinic in Jacksonville were extracted from the institutional electronic ECG database management system (MUSE, GE Healthcare). These ECGs were matched with transthoracic echocardiograms obtained up to four days prior or 30 days after the ECGs acquisition. A convolutional neural network consisting of 8 layers of convolutions, batch normalization and pooling was trained using Keras and Tensorflow with hyper-parameter optimization for L1 and L2 regularization, learning rate adjustments, and class weights to predict three classes of LVEF: ≤35%, 36–51%, and ≥52% based on clinical relevance. The primary measure of overall performance was the detection of LVEF ≤35%; however, the triad of model predictions was also considered in translating the model output to human interpretable findings. Results A total of 30,153 patients (60,169 ECG pairings; mean age 63 years; 48% male) were randomly split at the patient level into training (24,172 patients), validation (2,973 patients) and testing (3,008 patients). The trained model provided robust discrimination in the withheld testing data – AUROC of 0.941 (95% CI: 0.931 to 0.950). Using the optimal model threshold based on Youden's index from the validation data (0.186), sensitivity and specificity were estimated to be 87.9% (95% CI: 83.8% to 91.2%) and 86.3% (95% CI: 85.4% to 87.2%) in the testing data. In instances where discordant predictions were observed, the posterior distribution of model probabilities provide additional insights into the possible underlying value of LVEF (Figure 1). Conclusions The utilization of a multi-category deep learning classification model for the detection of reduced ejection fraction adds new dimensions to the use of AI technologies on digital ECGs. This work shows high discrimination can still be obtained when using three classes of LVEF. Funding Acknowledgement Type of funding sources: None.

Abstract Background Current approaches utilizing non-invasive methods to screen for cardiac allograft rejection (gene expression profiling and cell free DNA) have yet to be broadly integrated into heart transplant management and have shown limited discrimination (AUCs of 0.68 and 0.77, respectively). Changes in the electrocardiogram (ECG) have been reported at the time of severe cardiac rejection, including low voltages and conduction abnormalities. It remains unknown if subtle ECG changes correlating with cardiac allograft rejection can be detected earlier using machine learning methods. Purpose We sought to develop an artificial intelligence (AI) model to detect cardiac allograft rejection based on the 12 lead ECG. Methods We identified all patients who underwent a heart transplant at 3 hospital sites within a single health system from Jan 1998 through Apr 2021 and extracted digital 12-lead ECG data as well as endomyocardial biopsy pathology results from the electronic medical record. We partitioned our data into a training (80%), validation (10%), and test set (10%) based on a group stratification sampling. Each patient was present in only one set and each set had a positivity rate of 2.6% with 6,074/758/758 ECGs belonging to 1,146/140/141 unique patients in each set respectively. Cardiac allograft rejection was defined as moderate or severe acute cellular rejection based on International Society for Heart and Lung Transplantation (ISHLT) guidelines. A convolutional neural network, using the 12-lead ECG data as input, was trained with hyperparameter optimization for regularization, learning rate adjustments, and class weights. Model performance metrics were based on the test data and estimated using the final model architecture. Results 1,587 heart transplant recipients who had at least one endomyocardial biopsy were evaluated for inclusion. We limited our sample to ECGs performed within 30 days of the biopsy date (7,590 ECGs, representing 1,425 unique patients). Our study population had a median age of 55.8 years and 28.7% were female. The median number of ECG-biopsy pairs per patient was 5. The majority of endomyocardial biopsy results were classified as none or mild rejection (97.1%), and 2.9% had moderate/severe rejection. The ECG-based AI model detected cardiac allograft rejection with an area under the receiver operative curve (AUC) of 0.84 in the test set. The sensitivity, specificity, positive and negative predictive values were 95%, 52.6%. 5.2% and 99.7% respectively. Conclusions An AI-ECG model appears to outperform novel non-invasive laboratory tests (gene expression profiling and cell free DNA) for detecting cardiac allograft rejection and does not require a blood draw or the additional complexities surrounding sample processing. This model relies on a readily available and relatively inexpensive test, the ECG. In addition, AI predictions can be made available within a few minutes following ECG acquisition. Funding Acknowledgement Type of funding sources: Private hospital(s). Main funding source(s): Mayo Clinic