Search

Supplemental Digital Content is available in the text.

Supplemental Digital Content is available in the text.

Introduction:Early diagnosis of aortic stenosis (AS) is critical for its timely management. However, despite emergence of hand-held echocardiography, the detection and grading of AS requires Doppler imaging, which is limited by both access and expertise. We developed a semi-supervised, contrastive learning approach to identify severe AS using limited labelled data of parasternal long axis (PLAX) videos from transthoracic echocardiography (TTE).Methods:We sampled TTE studies performed between 2015-2021 in a large health system. TTEs from 2015-2020 were used for training, with oversampling of AS for diagnostic enrichment (5311 studies, age 70±16 years, n=2601 [49%] women, 5029 unique patients). The testing set represented studies in 2021 without oversampling for AS (2040 studies, mean age 66±16 years, n=997 [49%] women, n=2034 unique patients). We performed self-supervised pretraining by selecting different PLAX videos from the same patient as positive samples for contrastive learning (multi-instance self-supervised learning) (A). The learned weights were used to initialize a 3D convolutional neural network to predict severe AS (B).Results:An ensemble model of three different weight initialization methods achieved an AUC of 0.97 (95% CI: 0.96-0.99) for severe AS detection, with 0.96 (95% CI: 0.83-0.97) specificity at 90% sensitivity. Among patients without severe AS, positive predictions were characterized by significantly higher peak aortic velocities compared to negative predictions, with no differences in LV function - a negative control (C). Saliency maps highlighted the aortic valve as most relevant to the final predictions (D, i-v:positive; vi:negative predictions).Conclusions:We have developed a novel method to detect severe AS using single-view TTE videos without requiring Doppler data. Our findings have significant implications for point-of-care ultrasound screening as part of routine clinic visits and in low-resource settings.

Introduction:Maximal oxygen uptake (VO2max), an indicator of cardiorespiratory fitness (CRF) requires exercise testing and, as a result, is rarely ascertained in large-scale population-based studies. There are many non-exercise algorithms that can estimate VO2max, but they are limited by their non-representative study population, lack of predictors, and the insufficient predictive power for the type of model chosen. In this study, we aim to improve the non-exercise algorithms using machine learning (ML) methods and data from U.S. national population surveys.Methods:We used the 1999-2004 data from the National Health and Nutrition Examination Survey (NHANES) because it included exercise testing to measure VO2max. Based on the literature review, predictors were identified from demographic, interview, examination, and laboratory data. The study population was split into a training set (80%) and a testing set (20%) for model development and validation. We built the improved model using Light Gradient Boosting Machine (LightGBM) for its outstanding performance and built-in functionality for handling missingness. The model was trained by optimizing Root-Mean Squared Error (RMSE) and using 5-fold cross-validation. Existing non-exercise algorithms for comparison were applied to the testing set.Results:Among the 5,668 participants included, the mean age was 32.5 and 49.9% were women. 40 predictors in total were selected for the final feature set. The fitted LightGBM model achieved an RMSE of 8.43 ml/kg/min (95% CI: 7.64 -9.19) on the testing set, significantly reducing the error by 14% (P <0.05) compared with the best existing non-exercise algorithms that could be cross-validated in NHANES (Table1).Conclusion:In conclusion, the extended non-exercise ML model in this study can provide a more accurate prediction of VO2max for NHANES participants, and those with similar data, than the existing non-exercise algorithms.

Background:Left ventricular systolic dysfunction (LVSD) is associated with an over 8-fold risk of heart failure and 2-fold risk of premature death. Echocardiographic screening is limited by access to the technology, and artificial intelligence-based tools have been limited to ECG signals, precluding access by end users.Methods:A total of 393,910 12-lead ECGs from patients undergoing echocardiography at Yale were used to develop a model identifying LV ejection fraction < 40% in printed ECG images, using a EfficientNet-B3 convolutional neural network with transfer learning. The algorithm was validated using real-world ECG images. To allow interpretability, gradient-weighted class activation mapping (Grad-CAM) was used to localize class-discriminating signals in the images.Results:A total of 60,293 ECGs (15.3%) were performed in patients with LVSD. The ECG-based image identified LVSD accurate across multiple image formats in the held-out test set (AUROC 0.91, AUPRC 0.69), and in two external sets of real-world ECG images from an outpatient academic center (AUROC 0.93, AUPRC 0.69) and a rural hospital system (AUROC 0.91, AUPRC 0.89). An ECG suggestive of LVSD portended over 24-fold higher odds of LVSD in the held-out set (OR 24.4, 95% CI, 22.4-26.6). Regardless of the ECG layout, class-discriminative patterns from Grad-CAM localized to leads V2 and V3, corresponding to the left ventricle. Moreover, a positive ECG screen in individuals with LV ejection fraction > 40% at the time of initial assessment was associated with 4-fold increased risk of developing incident LV systolic dysfunction in the future (HR 4.3, 95% CI 3.9-4.7, median follow up 2.3 years).Conclusions:We developed and validated a layout-independent deep learning model that identifies LVSD from images of ECGs. This approach represents an automated and accessible screening strategy for LVSD, particularly in low resource settings.

Background:The risk of acute kidney injury (AKI) for percutaneous coronary intervention (PCI) can inform decision-making and mitigation. However, risk assessment is hampered by the burden of manually abstracting and inputting data required by the established model developed on registry data. Accordingly, there is a need to develop AKI models that use automatically extracted electronic health records (EHR) data and evaluate their performance in comparison to the contemporary registry model.Methods:We used EHR data for PCIs performed at Yale New Haven Hospital (YNHH) and the linked local National Cardiovascular Data Registry (NCDR) CathPCI data to develop and compare the EHR models with the NCDR model. AKI was defined as an increase of ≥0.3 mg/dL or 50% in serum creatinine. To maximize interoperability across systems and usability at the point of care, we only considered structured EHR readily extractable before PCI, and implemented simple but universal data preprocessing. Longitudinal data was aggregated into 24h, 7d and 14d prior to PCI. We used Lasso regression and gradient descent boosting (GDB) to train the models and tested their capacity to approximate the NCDR model predictions and predict AKI independently. We evaluated performance via cross-validation by c-statistic, calibration slope, and predictive range defined by event rate difference in the lowest and highest risk deciles. We used paired t tests for model comparison.Results:The cohort included 9,202 PCIs from December 2012 to November 2019 at YNHH for 8,445 patients and had a mean age of 67.0±12.0 years with 27.6% females and 798 (8.7%) AKIs. Evaluated on the same test sets, the EHR models approximating the NCDR model predictions achieved similar performance to that model (all p>0.05). Further, the EHR model employing GDB to independently predict AKI achieved slightly better performance than the NCDR model in c-statistic (0.84 vs 0.83, p=0.02) and predictive range (41.2% vs 39.5%, p=0.13) with similar calibration (slope 0.98 vs 0.99).Conclusion:The automatically extracted EHR data produced models with similar performance to the registry model. These models can automate AKI risk assessment for PCI, alleviating provider burden and increase uptake of risk assessment.

Introduction:Community-level social determinants of health (SDOH) can facilitate the identification of vulnerable populations with poor health outcomes.Hypothesis:We assess whether machine learning methods utilizing area-level SDOH indicators can identify populations at higher risk for CV mortality.Methods:144 SDOHs were extracted from the SDOH database compiled, at a county level, by the Agency for Healthcare Research and Quality between 2009-2018. The Person’s correlation coefficient between SDOH with CV mortality was used to select 12 key SDOH (p < 0.05) and linear and nonlinear models were tested to build a predictive model for CVD mortality rates per 100,000 people for each county from SDOH. Datasets were randomly split in 80/20 for training and validation, and the accuracy of the models was compared in terms of mean relative error (MRE).Results:SDOH related to income (median household income), poverty (fraction population with a poverty ratio more than 2.00), and educational attainment (high school, BSc, MSc, doctorate) were the principal factors affecting CV mortality. Factors related to race/ethnicity, living conditions (children living with grandparents), and physical infrastructure (median household value and fraction of mobile homes) were also included in the model. The most performant regression method is the Bayesian optimized artificial neural network (Figure). The performance of linear models (linear elastic-networks, MRE = 17.6%; linear regression, MRE = 38.8%) was inferior to nonlinear models. Among nonlinear models, optimized Bayesian feedforward neural networks OBFNN presented the best performance (MRE = 9.7%).Conclusions:We developed a county-level risk score for cardiovascular mortality using 12 SDOH including demographics, income and poverty, education, and physical infrastructure. Future work should look at extending beyond CVD mortality and combining this SDOH-based score for patient-level cardiovascular risk.

Introduction:Transcatheter septal puncture is an important procedure for the success of left-sided structural heart interventions, including left atrial appendage occlusion devices and transcatheter mitral valve repair therapy. Although a transcatheter septal puncture is generally safe under image guidance, serious complications can occur, including accidental puncture of the aorta or pericardial tamponade. Enhanced visualization technologies may improve procedural accuracy and reduce complications.Hypothesis:Augmented reality as a 3D visualization technology can better guide transseptal puncture procedures than views on conventional 2D screens.Methods:Here, we demonstrate an augmented reality guidance system for a transcatheter septal puncture procedure. Pre-procedural cardiac CT was used as a patient-specific 3D image data for rendering a patient?s heart. The 3D position of a catheter was automatically calculated from biplane X-ray fluoroscopy. The vertebrae were segmented using deep learning algorithm (convolutional neural network) and used as an intrinsic fiduciary marker for 3D-2D co-registration. From the patient cardiac CT images, a 3D-printed heart and spine model was generated and the accuracy of the augmented reality guidance system was measured using an electromagnetic sensor.Results:The augmented reality guidance system showed high success rate for registration, low registration error, and clinically acceptable computational cost. The measured catheter position by the sensor at the tip of the catheter showed good correlation with our algorithm-based catheter position calculations. Automatic 3D-2D overlay image was generated by machine learning-based image segmentation and registration. Final 3D images were reconstructed in the augmented reality environment within the Microsoft HoloLens headset.Conclusions:We developed a fully interactive method to guide cardiac interventions based on pre-procedural and intra-procedural imaging with an advanced holographic system and machine learning algorithms. Further validation for heart displacement during respiratory and cardiac cycle is mandatory.

Introduction:We use 4D-Flow MRI to investigate the effects of bicuspid aortic valve (BAV) disease on downstream pressure drop (PD), wall shear stress (WSS), and viscous energy loss (EL) in the ascending aorta (AAo).Hypothesis:BAV patients exhibit greater PD, WSS, and EL in the AAo when compared to healthy controls.Methods:Fifteen healthy controls (32?13 years, 7 female) and 48 BAV patients (45?16 years, 17 female) underwent cardiac MRI at 3T, inclusive of cine imaging and 4D flow. Four cross sections (Figure 1) were placed along aortas in distinct locations: left ventricular outflow tract (LVOT), sinuses of Valsalva (SOV), mid-ascending aorta (MAA), and proximal to first aortic branch (AA1). Cross sections were analyzed for (i) net flow, (ii) peak velocity, (iii) aortic diameter (normalized to body surface area), (iv) PD (LVOT used as reference point), (v) EL (entire LVOT-AA1 volume measured; normalized by LVOT net flow), and (vi) WSS. Two sub-vectors of WSS, axial (WSSax) and circumferential (WSScirc), were also analyzed (Figure 1).Results:Compared to healthy volunteers the BAV patients exhibited significantly greater PD (MAA: 3.24?2.7 vs. 9.10?7.8mmHg, respectively; p<0.01), EL (LVOT-AA1: 0.038?0.01 vs. 0.096?0.05mW/mL; p<0.01), and WSScirc (MAA: 0.17?0.06 vs. 0.27?0.1Pa; p<0.01), but less WSS (MAA: 0.23?0.08 vs 0.15?0.07; p<0.01). Ascending aorta diameter correlated positively with EL (MAA: R=0.52, p<0.01), but negatively with WSS (MAA: R=-0.607, p<0.01) and WSSax (MAA: R= -0.306, p=0.013).Conclusions:BAV disease significantly alters PD, EL, and WSS in the AAo. Viscous energy loss and WSS appear to associate with AAo diameter.