Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images.

BACKGROUND: Left ventricular (LV) systolic dysfunction is associated with a >8-fold increased risk of heart failure and a 2-fold risk of premature death. The use of ECG signals in screening for LV systolic dysfunction is limited by their availability to clinicians. We developed a novel deep learning-based approach that can use ECG images for the screening of LV systolic dysfunction. METHODS: Using 12-lead ECGs plotted in multiple different formats, and corresponding echocardiographic data recorded within 15 days from the Yale New Haven Hospital between 2015 and 2021, we developed a convolutional neural network algorithm to detect an LV ejection fraction <40%. The model was validated within clinical settings at Yale New Haven Hospital and externally on ECG images from Cedars Sinai Medical Center in Los Angeles, CA; Lake Regional Hospital in Osage Beach, MO; Memorial Hermann Southeast Hospital in Houston, TX; and Methodist Cardiology Clinic of San Antonio, TX. In addition, it was validated in the prospective Brazilian Longitudinal Study of Adult Health. Gradient-weighted class activation mapping was used to localize class-discriminating signals on ECG images. RESULTS: Overall, 385 601 ECGs with paired echocardiograms were used for model development. The model demonstrated high discrimination across various ECG image formats and calibrations in internal validation (area under receiving operation characteristics [AUROCs], 0.91; area under precision-recall curve [AUPRC], 0.55); and external sets of ECG images from Cedars Sinai (AUROC, 0.90 and AUPRC, 0.53), outpatient Yale New Haven Hospital clinics (AUROC, 0.94 and AUPRC, 0.77), Lake Regional Hospital (AUROC, 0.90 and AUPRC, 0.88), Memorial Hermann Southeast Hospital (AUROC, 0.91 and AUPRC 0.88), Methodist Cardiology Clinic (AUROC, 0.90 and AUPRC, 0.74), and Brazilian Longitudinal Study of Adult Health cohort (AUROC, 0.95 and AUPRC, 0.45). An ECG suggestive of LV systolic dysfunction portended >27-fold higher odds of LV systolic dysfunction on transthoracic echocardiogram (odds ratio, 27.5 [95% CI, 22.3-33.9] in the held-out set). Class-discriminative patterns localized to the anterior and anteroseptal leads (V2 and V3), corresponding to the left ventricle regardless of the ECG layout. A positive ECG screen in individuals with an LV ejection fraction ≥40% at the time of initial assessment was associated with a 3.9-fold increased risk of developing incident LV systolic dysfunction in the future (hazard ratio, 3.9 [95% CI, 3.3-4.7]; median follow-up, 3.2 years). CONCLUSIONS: We developed and externally validated a deep learning model that identifies LV systolic dysfunction from ECG images. This approach represents an automated and accessible screening strategy for LV systolic dysfunction, particularly in low-resource settings.

Severe aortic stenosis detection by deep learning applied to echocardiography.

BACKGROUND AND AIMS: Early diagnosis of aortic stenosis (AS) is critical to prevent morbidity and mortality but requires skilled examination with Doppler imaging. This study reports the development and validation of a novel deep learning model that relies on two-dimensional (2D) parasternal long axis videos from transthoracic echocardiography without Doppler imaging to identify severe AS, suitable for point-of-care ultrasonography. METHODS AND RESULTS: In a training set of 5257 studies (17 570 videos) from 2016 to 2020 [Yale-New Haven Hospital (YNHH), Connecticut], an ensemble of three-dimensional convolutional neural networks was developed to detect severe AS, leveraging self-supervised contrastive pretraining for label-efficient model development. This deep learning model was validated in a temporally distinct set of 2040 consecutive studies from 2021 from YNHH as well as two geographically distinct cohorts of 4226 and 3072 studies, from California and other hospitals in New England, respectively. The deep learning model achieved an area under the receiver operating characteristic curve (AUROC) of 0.978 (95% CI: 0.966, 0.988) for detecting severe AS in the temporally distinct test set, maintaining its diagnostic performance in geographically distinct cohorts [0.952 AUROC (95% CI: 0.941, 0.963) in California and 0.942 AUROC (95% CI: 0.909, 0.966) in New England]. The model was interpretable with saliency maps identifying the aortic valve, mitral annulus, and left atrium as the predictive regions. Among non-severe AS cases, predicted probabilities were associated with worse quantitative metrics of AS suggesting an association with various stages of AS severity. CONCLUSION: This study developed and externally validated an automated approach for severe AS detection using single-view 2D echocardiography, with potential utility for point-of-care screening.

Temporal relationship of computed and structured diagnoses in electronic health record data.

BACKGROUND: The electronic health record (EHR) holds the prospect of providing more complete and timely access to clinical information for biomedical research, quality assessments, and quality improvement compared to other data sources, such as administrative claims. In this study, we sought to assess the completeness and timeliness of structured diagnoses in the EHR compared to computed diagnoses for hypertension (HTN), hyperlipidemia (HLD), and diabetes mellitus (DM). METHODS: We determined the amount of time for a structured diagnosis to be recorded in the EHR from when an equivalent diagnosis could be computed from other structured data elements, such as vital signs and laboratory results. We used EHR data for encounters from January 1, 2012 through February 10, 2019 from an academic health system. Diagnoses for HTN, HLD, and DM were computed for patients with at least two observations above threshold separated by at least 30 days, where the thresholds were outpatient blood pressure of ≥ 140/90 mmHg, any low-density lipoprotein ≥ 130 mg/dl, or any hemoglobin A1c ≥ 6.5%, respectively. The primary measure was the length of time between the computed diagnosis and the time at which a structured diagnosis could be identified within the EHR history or problem list. RESULTS: We found that 39.8% of those with HTN, 21.6% with HLD, and 5.2% with DM did not receive a corresponding structured diagnosis recorded in the EHR. For those who received a structured diagnosis, a mean of 389, 198, and 166 days elapsed before the patient had the corresponding diagnosis of HTN, HLD, or DM, respectively, recorded in the EHR. CONCLUSIONS: We found a marked temporal delay between when a diagnosis can be computed or inferred and when an equivalent structured diagnosis is recorded within the EHR. These findings demonstrate the continued need for additional study of the EHR to avoid bias when using observational data and reinforce the need for computational approaches to identify clinical phenotypes.

Association of Use of an Intravascular Microaxial Left Ventricular Assist Device vs Intra-aortic Balloon Pump With In-Hospital Mortality and Major Bleeding Among Patients With Acute Myocardial Infarction Complicated by Cardiogenic Shock.

Importance: Acute myocardial infarction (AMI) complicated by cardiogenic shock is associated with substantial morbidity and mortality. Although intravascular microaxial left ventricular assist devices (LVADs) provide greater hemodynamic support as compared with intra-aortic balloon pumps (IABPs), little is known about clinical outcomes associated with intravascular microaxial LVAD use in clinical practice. Objective: To examine outcomes among patients undergoing percutaneous coronary intervention (PCI) for AMI complicated by cardiogenic shock treated with mechanical circulatory support (MCS) devices. Design, Setting, and Participants: A propensity-matched registry-based retrospective cohort study of patients with AMI complicated by cardiogenic shock undergoing PCI between October 1, 2015, and December 31, 2017, who were included in data from hospitals participating in the CathPCI and the Chest Pain-MI registries, both part of the American College of Cardiology's National Cardiovascular Data Registry. Patients receiving an intravascular microaxial LVAD were matched with those receiving IABP on demographics, clinical history, presentation, infarct location, coronary anatomy, and clinical laboratory data, with final follow-up through December 31, 2017. Exposures: Hemodynamic support, categorized as intravascular microaxial LVAD use only, IABP only, other (such as use of a percutaneous extracorporeal ventricular assist system, extracorporeal membrane oxygenation, or a combination of MCS device use), or medical therapy only. Main Outcomes and Measures: The primary outcomes were in-hospital mortality and in-hospital major bleeding. Results: Among 28 304 patients undergoing PCI for AMI complicated by cardiogenic shock, the mean (SD) age was 65.0 (12.6) years, 67.0% were men, 81.3% had an ST-elevation myocardial infarction, and 43.3% had cardiac arrest. Over the study period among patients with AMI, an intravascular microaxial LVAD was used in 6.2% of patients, and IABP was used in 29.9%. Among 1680 propensity-matched pairs, there was a significantly higher risk of in-hospital death associated with use of an intravascular microaxial LVAD (45.0%) vs with an IABP (34.1% [absolute risk difference, 10.9 percentage points {95% CI, 7.6-14.2}; P < .001) and also higher risk of in-hospital major bleeding (intravascular microaxial LVAD [31.3%] vs IABP [16.0%]; absolute risk difference, 15.4 percentage points [95% CI, 12.5-18.2]; P < .001). These associations were consistent regardless of whether patients received a device before or after initiation of PCI. Conclusions and Relevance: Among patients undergoing PCI for AMI complicated by cardiogenic shock from 2015 to 2017, use of an intravascular microaxial LVAD compared with IABP was associated with higher adjusted risk of in-hospital death and major bleeding complications, although study interpretation is limited by the observational design. Further research may be needed to understand optimal device choice for these patients.

Enhancing the prediction of acute kidney injury risk after percutaneous coronary intervention using machine learning techniques: A retrospective cohort study.

BACKGROUND: The current acute kidney injury (AKI) risk prediction model for patients undergoing percutaneous coronary intervention (PCI) from the American College of Cardiology (ACC) National Cardiovascular Data Registry (NCDR) employed regression techniques. This study aimed to evaluate whether models using machine learning techniques could significantly improve AKI risk prediction after PCI. METHODS AND FINDINGS: We used the same cohort and candidate variables used to develop the current NCDR CathPCI Registry AKI model, including 947,091 patients who underwent PCI procedures between June 1, 2009, and June 30, 2011. The mean age of these patients was 64.8 years, and 32.8% were women, with a total of 69,826 (7.4%) AKI events. We replicated the current AKI model as the baseline model and compared it with a series of new models. Temporal validation was performed using data from 970,869 patients undergoing PCIs between July 1, 2016, and March 31, 2017, with a mean age of 65.7 years; 31.9% were women, and 72,954 (7.5%) had AKI events. Each model was derived by implementing one of two strategies for preprocessing candidate variables (preselecting and transforming candidate variables or using all candidate variables in their original forms), one of three variable-selection methods (stepwise backward selection, lasso regularization, or permutation-based selection), and one of two methods to model the relationship between variables and outcome (logistic regression or gradient descent boosting). The cohort was divided into different training (70%) and test (30%) sets using 100 different random splits, and the performance of the models was evaluated internally in the test sets. The best model, according to the internal evaluation, was derived by using all available candidate variables in their original form, permutation-based variable selection, and gradient descent boosting. Compared with the baseline model that uses 11 variables, the best model used 13 variables and achieved a significantly better area under the receiver operating characteristic curve (AUC) of 0.752 (95% confidence interval [CI] 0.749-0.754) versus 0.711 (95% CI 0.708-0.714), a significantly better Brier score of 0.0617 (95% CI 0.0615-0.0618) versus 0.0636 (95% CI 0.0634-0.0638), and a better calibration slope of observed versus predicted rate of 1.008 (95% CI 0.988-1.028) versus 1.036 (95% CI 1.015-1.056). The best model also had a significantly wider predictive range (25.3% versus 21.6%, p < 0.001) and was more accurate in stratifying AKI risk for patients. Evaluated on a more contemporary CathPCI cohort (July 1, 2015-March 31, 2017), the best model consistently achieved significantly better performance than the baseline model in AUC (0.785 versus 0.753), Brier score (0.0610 versus 0.0627), calibration slope (1.003 versus 1.062), and predictive range (29.4% versus 26.2%). The current study does not address implementation for risk calculation at the point of care, and potential challenges include the availability and accessibility of the predictors. CONCLUSIONS: Machine learning techniques and data-driven approaches resulted in improved prediction of AKI risk after PCI. The results support the potential of these techniques for improving risk prediction models and identification of patients who may benefit from risk-mitigation strategies.

Timing of Blood Draws Among Patients Hospitalized in a Large Academic Medical Center.

This study describes the degree to which blood draws occurred among hospitalized patients during traditional sleep hours and investigates trends over time.

Pulse2AI: An Adaptive Framework to Standardize and Process Pulsatile Wearable Sensor Data for Clinical Applications.

Goal: To establish Pulse2AI as a reproducible data preprocessing framework for pulsatile signals that generate high-quality machine-learning-ready datasets from raw wearable recordings. Methods: We proposed an end-to-end data preprocessing framework that adapts multiple pulsatile signal modalities and generates machine-learning-ready datasets agnostic to downstream medical tasks. Results: a dataset preprocessed by Pulse2AI improved systolic blood pressure estimation by 29.58%, from 11.41 to 8.03 mmHg in root-mean-square-error (RMSE) and its diastolic counterpart by 26.01%, from 7.93 to 5.87 mmHg in RMSE. For respiration rate (RR) estimation, Pulse2AI boosted performance by 19.69%, from 1.47 to 1.18 breaths per minute (BrPM) in mean-absolute-error (MAE). Conclusion: Pulse2AI turns pulsatile signals into machine learning (ML) ready datasets for arbitrary remote health monitoring tasks. We tested Pulse2AI on multiple pulsatile modalities and demonstrated its efficacy in two medical applications. This work bridges valuable assets in remote sensing and internet of medical things to ML-ready datasets for medical modeling.

Efficient deep learning-based automated diagnosis from echocardiography with contrastive self-supervised learning.

BACKGROUND: Advances in self-supervised learning (SSL) have enabled state-of-the-art automated medical image diagnosis from small, labeled datasets. This label efficiency is often desirable, given the difficulty of obtaining expert labels for medical image recognition tasks. However, most efforts toward SSL in medical imaging are not adapted to video-based modalities, such as echocardiography. METHODS: We developed a self-supervised contrastive learning approach, EchoCLR, for echocardiogram videos with the goal of learning strong representations for efficient fine-tuning on downstream cardiac disease diagnosis. EchoCLR pretraining involves (i) contrastive learning, where the model is trained to identify distinct videos of the same patient, and (ii) frame reordering, where the model is trained to predict the correct of video frames after being randomly shuffled. RESULTS: When fine-tuned on small portions of labeled data, EchoCLR pretraining significantly improves classification performance for left ventricular hypertrophy (LVH) and aortic stenosis (AS) over other transfer learning and SSL approaches across internal and external test sets. When fine-tuning on 10% of available training data (519 studies), an EchoCLR-pretrained model achieves 0.72 AUROC (95% CI: [0.69, 0.75]) on LVH classification, compared to 0.61 AUROC (95% CI: [0.57, 0.64]) with a standard transfer learning approach. Similarly, using 1% of available training data (53 studies), EchoCLR pretraining achieves 0.82 AUROC (95% CI: [0.79, 0.84]) on severe AS classification, compared to 0.61 AUROC (95% CI: [0.58, 0.65]) with transfer learning. CONCLUSIONS: EchoCLR is unique in its ability to learn representations of echocardiogram videos and demonstrates that SSL can enable label-efficient disease classification from small amounts of labeled data.

Artificial intelligence (AI) has been used to develop software that can automatically diagnose diseases from medical images. However, these AI models require thousands or millions of examples to properly learn from, which can be very expensive, as diagnosis is often time-consuming and requires clinical expertise. Using a technique called self-supervised learning (SSL), we develop an AI method to effectively diagnose heart disease from as few as 50 instances. Our method, EchoCLR, is designed for echocardiography, a key imaging technique to monitor heart health, and outperforms other methods on disease diagnosis from small amounts of data. This method can advance AI for echocardiography and enable researchers with limited resources to create disease diagnosis models from small medical imaging datasets.

Biometric contrastive learning for data-efficient deep learning from electrocardiographic images.

OBJECTIVE: Artificial intelligence (AI) detects heart disease from images of electrocardiograms (ECGs). However, traditional supervised learning is limited by the need for large amounts of labeled data. We report the development of Biometric Contrastive Learning (BCL), a self-supervised pretraining approach for label-efficient deep learning on ECG images. MATERIALS AND METHODS: Using pairs of ECGs from 78 288 individuals from Yale (2000-2015), we trained a convolutional neural network to identify temporally separated ECG pairs that varied in layouts from the same patient. We fine-tuned BCL-pretrained models to detect atrial fibrillation (AF), gender, and LVEF < 40%, using ECGs from 2015 to 2021. We externally tested the models in cohorts from Germany and the United States. We compared BCL with ImageNet initialization and general-purpose self-supervised contrastive learning for images (simCLR). RESULTS: While with 100% labeled training data, BCL performed similarly to other approaches for detecting AF/Gender/LVEF < 40% with an AUROC of 0.98/0.90/0.90 in the held-out test sets, it consistently outperformed other methods with smaller proportions of labeled data, reaching equivalent performance at 50% of data. With 0.1% data, BCL achieved AUROC of 0.88/0.79/0.75, compared with 0.51/0.52/0.60 (ImageNet) and 0.61/0.53/0.49 (simCLR). In external validation, BCL outperformed other methods even at 100% labeled training data, with an AUROC of 0.88/0.88 for Gender and LVEF < 40% compared with 0.83/0.83 (ImageNet) and 0.84/0.83 (simCLR). DISCUSSION AND CONCLUSION: A pretraining strategy that leverages biometric signatures of different ECGs from the same patient enhances the efficiency of developing AI models for ECG images. This represents a major advance in detecting disorders from ECG images with limited labeled data.

SOFA score performs worse than age for predicting mortality in patients with COVID-19.

The use of the Sequential Organ Failure Assessment (SOFA) score, originally developed to describe disease morbidity, is commonly used to predict in-hospital mortality. During the COVID-19 pandemic, many protocols for crisis standards of care used the SOFA score to select patients to be deprioritized due to a low likelihood of survival. A prior study found that age outperformed the SOFA score for mortality prediction in patients with COVID-19, but was limited to a small cohort of intensive care unit (ICU) patients and did not address whether their findings were unique to patients with COVID-19. Moreover, it is not known how well these measures perform across races. In this retrospective study, we compare the performance of age and SOFA score in predicting in-hospital mortality across two cohorts: a cohort of 2,648 consecutive adult patients diagnosed with COVID-19 who were admitted to a large academic health system in the northeastern United States over a 4-month period in 2020 and a cohort of 75,601 patients admitted to one of 335 ICUs in the eICU database between 2014 and 2015. We used age and the maximum SOFA score as predictor variables in separate univariate logistic regression models for in-hospital mortality and calculated area under the receiver operator characteristic curves (AU-ROCs) and area under precision-recall curves (AU-PRCs) for each predictor in both cohorts. Among the COVID-19 cohort, age (AU-ROC 0.795, 95% CI 0.762, 0.828) had a significantly better discrimination than SOFA score (AU-ROC 0.679, 95% CI 0.638, 0.721) for mortality prediction. Conversely, age (AU-ROC 0.628 95% CI 0.608, 0.628) underperformed compared to SOFA score (AU-ROC 0.735, 95% CI 0.726, 0.745) in non-COVID-19 ICU patients in the eICU database. There was no difference between Black and White COVID-19 patients in performance of either age or SOFA Score. Our findings bring into question the utility of SOFA score-based resource allocation in COVID-19 crisis standards of care.

Automated Diagnostic Reports from Images of Electrocardiograms at the Point-of-Care.

Timely and accurate assessment of electrocardiograms (ECGs) is crucial for diagnosing, triaging, and clinically managing patients. Current workflows rely on a computerized ECG interpretation using rule-based tools built into the ECG signal acquisition systems with limited accuracy and flexibility. In low-resource settings, specialists must review every single ECG for such decisions, as these computerized interpretations are not available. Additionally, high-quality interpretations are even more essential in such low-resource settings as there is a higher burden of accuracy for automated reads when access to experts is limited. Artificial Intelligence (AI)-based systems have the prospect of greater accuracy yet are frequently limited to a narrow range of conditions and do not replicate the full diagnostic range. Moreover, these models often require raw signal data, which are unavailable to physicians and necessitate costly technical integrations that are currently limited. To overcome these challenges, we developed and validated a format-independent vision encoder-decoder model - ECG-GPT - that can generate free-text, expert-level diagnosis statements directly from ECG images. The model shows robust performance, validated on 2.6 million ECGs across 6 geographically distinct health settings: (1) 2 large and diverse US health systems- Yale-New Haven and Mount Sinai Health Systems, (2) a consecutive ECG dataset from a central ECG repository from Minas Gerais, Brazil, (3) the prospective cohort study, UK Biobank, (4) a Germany-based, publicly available repository, PTB-XL, and (5) a community hospital in Missouri. The model demonstrated consistently high performance (AUROC≥0.81) across a wide range of rhythm and conduction disorders. This can be easily accessed via a web-based application capable of receiving ECG images and represents a scalable and accessible strategy for generating accurate, expert-level reports from images of ECGs, enabling accurate triage of patients globally, especially in low-resource settings.

Predicting Major Adverse Events in Patients Undergoing Transcatheter Left Atrial Appendage Occlusion.

BACKGROUND: The National Cardiovascular Data Registry Left Atrial Appendage Occlusion Registry (LAAO) includes the vast majority of transcatheter LAAO procedures performed in the United States. The objective of this study was to develop a model predicting adverse events among patients undergoing LAAO with Watchman FLX. METHODS: Data from 41 001 LAAO procedures with Watchman FLX from July 2020 to September 2021 were used to develop and validate a model predicting in-hospital major adverse events. Randomly selected development (70%, n=28 530) and validation (30%, n=12 471) cohorts were analyzed with 1000 bootstrapped samples, using forward stepwise logistic regression to create the final model. A simplified bedside risk score was also developed using this model. RESULTS: Increased age, female sex, low preprocedure hemoglobin, no prior attempt at atrial fibrillation termination, and increased fall risk most strongly predicted in-hospital major adverse events and were included in the final model along with other clinically relevant variables. The median in-hospital risk-standardized adverse event rate was 1.50% (range, 1.03%-2.84%; interquartile range, 1.42%-1.64%). The model demonstrated moderate discrimination (development C-index, 0.67 [95% CI, 0.65-0.70] and validation C-index, 0.66 [95% CI, 0.62-0.70]) with good calibration. The simplified risk score was well calibrated with risk of in-hospital major adverse events ranging from 0.26% to 3.90% for a score of 0 to 8, respectively. CONCLUSIONS: A transcatheter LAAO risk model using National Cardiovascular Data Registry and LAAO Registry data can predict in-hospital major adverse events, demonstrated consistency across hospitals and can be used for quality improvement efforts. A simple bedside risk score was similarly predictive and may inform shared decision-making.

A Review of Digital Innovations for Diet Monitoring and Precision Nutrition.

This article provides an up-to-date review of technological advances in 3 key areas related to diet monitoring and precision nutrition. First, we review developments in mobile applications, with a focus on food photography and artificial intelligence to facilitate the process of diet monitoring. Second, we review advances in 2 types of wearable and handheld sensors that can potentially be used to fully automate certain aspects of diet logging: physical sensors to detect moments of dietary intake, and chemical sensors to estimate the composition of diets and meals. Finally, we review new programs that can generate personalized/precision nutrition recommendations based on measurements of gut microbiota and continuous glucose monitors with artificial intelligence. The article concludes with a discussion of potential pitfalls of some of these technologies.

Automated Diet Capture Using Voice Alerts and Speech Recognition on Smartphones: Pilot Usability and Acceptability Study.

BACKGROUND: Effective monitoring of dietary habits is critical for promoting healthy lifestyles and preventing or delaying the onset and progression of diet-related diseases, such as type 2 diabetes. Recent advances in speech recognition technologies and natural language processing present new possibilities for automated diet capture; however, further exploration is necessary to assess the usability and acceptability of such technologies for diet logging. OBJECTIVE: This study explores the usability and acceptability of speech recognition technologies and natural language processing for automated diet logging. METHODS: We designed and developed base2Diet-an iOS smartphone application that prompts users to log their food intake using voice or text. To compare the effectiveness of the 2 diet logging modes, we conducted a 28-day pilot study with 2 arms and 2 phases. A total of 18 participants were included in the study, with 9 participants in each arm (text: n=9, voice: n=9). During phase I of the study, all 18 participants received reminders for breakfast, lunch, and dinner at preselected times. At the beginning of phase II, all participants were given the option to choose 3 times during the day to receive 3 times daily reminders to log their food intake for the remainder of the phase, with the ability to modify the selected times at any point before the end of the study. RESULTS: The total number of distinct diet logging events per participant was 1.7 times higher in the voice arm than in the text arm (P=.03, unpaired t test). Similarly, the total number of active days per participant was 1.5 times higher in the voice arm than in the text arm (P=.04, unpaired t test). Furthermore, the text arm had a higher attrition rate than the voice arm, with only 1 participant dropping out of the study in the voice arm, while 5 participants dropped out in the text arm. CONCLUSIONS: The results of this pilot study demonstrate the potential of voice technologies in automated diet capturing using smartphones. Our findings suggest that voice-based diet logging is more effective and better received by users compared to traditional text-based methods, underscoring the need for further research in this area. These insights carry significant implications for the development of more effective and accessible tools for monitoring dietary habits and promoting healthy lifestyle choices.

Clinical Risk Prediction Models with Meta-Learning Prototypes of Patient Heterogeneity.

Hospitalized patients sometimes have complex health conditions, such as multiple diseases, underlying diseases, and complications. The heterogeneous patient conditions may have various representations. A generalized model ignores the differences among heterogeneous patients, and personalized models, even with transfer learning, are still limited to the small amount of training data and the repeated training process. Meta-learning provides a solution for training similar patients based on few-shot learning; however, cannot address common cross-domain patients. Inspired by prototypical networks [1], we proposed a meta-prototype for Electronic Health Records (EHR), a meta-learning-based model with flexible prototypes representing the heterogeneity in patients. We apply this technique to cardiovascular diseases in MIMIC-III and compare it against a set of benchmark models, and demonstrate its ability to address heterogeneous patient health conditions and improve the model performances from 1.2% to 11.9% on different metrics and prediction tasks.Clinical relevance- Developing an adaptive EHR risk prediction model for outcomes-driven phenotyping of heterogeneous patient health conditions.

Variational Autoencoders for Biomedical Signal Morphology Clustering and Noise Detection.

Accurate estimation of physiological biomarkers using raw waveform data from non-invasive wearable devices requires extensive data preprocessing. An automatic noise detection method in time-series data would offer significant utility for various domains. As data labeling is onerous, having a minimally supervised abnormality detection method for input data, as well as an estimation of the severity of the signal corruptness, is essential. We propose a model-free, time-series biomedical waveform noise detection framework using a Variational Autoencoder coupled with Gaussian Mixture Models, which can detect a range of waveform abnormalities without annotation, providing a confidence metric for each segment. Our technique operates on biomedical signals that exhibit periodicity of heart activities. This framework can be applied to any machine learning or deep learning model as an initial signal validator component. Moreover, the confidence score generated by the proposed framework can be incorporated into different models' optimization to construct confidence-aware modeling. We conduct experiments using dynamic time warping (DTW) distance of segments to validated cardiac cycle morphology. The result confirms that our approach removes noisy cardiac cycles and the remaining signals, classified as clean, exhibit a 59.92% reduction in the standard deviation of DTW distances. Using a dataset of bio-impedance data of 97885 cardiac cycles, we further demonstrate a significant improvement in the downstream task of cuffless blood pressure estimation, with an average reduction of 2.67 mmHg root mean square error (RMSE) of Diastolic Blood pressure and 2.13 mmHg RMSE of systolic blood pressure, with increases of average Pearson correlation of 0.28 and 0.08, with a statistically significant improvement of signal-to-noise ratio respectively in the presence of different synthetic noise sources. This enables burden-free validation of wearable sensor data for downstream biomedical applications.

Hypothesis Scoring for Confidence-Aware Blood Pressure Estimation With Particle Filters.

We propose our Confidence-Aware Particle Filter (CAPF) framework that analyzes a series of estimated changes in blood pressure (BP) to provide several true state hypotheses for a given instance. Particularly, our novel confidence-awareness mechanism assigns likelihood scores to each hypothesis in an effort to discard potentially erroneous measurements - based on the agreement amongst a series of estimated changes and the physiological plausibility when considering DBP/SBP pairs. The particle filter formulation (or sequential Monte Carlo method) can jointly consider the hypotheses and their probabilities over time to provide a stable trend of estimated BP measurements. In this study, we evaluate BP trend estimation from an emerging bio-impedance (Bio-Z) prototype wearable modality although it is applicable to all types of physiological modalities. Each subject in the evaluation cohort underwent a hand-gripper exercise, a cold pressor test, and a recovery state to increase the variation to the captured BP ranges. Experiments show that CAPF yields superior continuous pulse pressure (PP), diastolic blood pressure (DBP), and systolic blood pressure (SBP) estimation performance compared to ten baseline approaches. Furthermore, CAPF performs on track to comply with AAMI and BHS standards for achieving a performance classification of Grade A, with mean error accuracies of -0.16 ± 3.75 mmHg for PP (r = 0.81), 0.42 ± 4.39 mmHg for DBP (r = 0.92), and -0.09 ± 6.51 mmHg for SBP (r = 0.92) from more than test 3500 data points.

Biometric Contrastive Learning for Data-Efficient Deep Learning from Electrocardiographic Images.

Objective: Artificial intelligence (AI) detects heart disease from images of electrocardiograms (ECGs), however traditional supervised learning is limited by the need for large amounts of labeled data. We report the development of Biometric Contrastive Learning (BCL), a self-supervised pretraining approach for label-efficient deep learning on ECG images. Materials and Methods: Using pairs of ECGs from 78,288 individuals from Yale (2000-2015), we trained a convolutional neural network to identify temporally-separated ECG pairs that varied in layouts from the same patient. We fine-tuned BCL-pretrained models to detect atrial fibrillation (AF), gender, and LVEF<40%, using ECGs from 2015-2021. We externally tested the models in cohorts from Germany and the US. We compared BCL with random initialization and general-purpose self-supervised contrastive learning for images (simCLR). Results: While with 100% labeled training data, BCL performed similarly to other approaches for detecting AF/Gender/LVEF<40% with AUROC of 0.98/0.90/0.90 in the held-out test sets, it consistently outperformed other methods with smaller proportions of labeled data, reaching equivalent performance at 50% of data. With 0.1% data, BCL achieved AUROC of 0.88/0.79/0.75, compared with 0.51/0.52/0.60 (random) and 0.61/0.53/0.49 (simCLR). In external validation, BCL outperformed other methods even at 100% labeled training data, with AUROC of 0.88/0.88 for Gender and LVEF<40% compared with 0.83/0.83 (random) and 0.84/0.83 (simCLR). Discussion and Conclusion: A pretraining strategy that leverages biometric signatures of different ECGs from the same patient enhances the efficiency of developing AI models for ECG images. This represents a major advance in detecting disorders from ECG images with limited labeled data.

Quantifying Blood Pressure Visit-to-Visit Variability in the Real-World Setting: A Retrospective Cohort Study.

BACKGROUND: Visit-to-visit variability (VVV) in blood pressure values has been reported in clinical studies. However, little is known about VVV in clinical practice and whether it is associated with patient characteristics in real-world setting. METHODS: We conducted a retrospective cohort study to quantify VVV in systolic blood pressure (SBP) values in a real-world setting. We included adults (age ≥18 years) with at least 2 outpatient visits between January 1, 2014 and October 31, 2018 from Yale New Haven Health System. Patient-level measures of VVV included SD and coefficient of variation of a given patient's SBP across visits. We calculated patient-level VVV overall and by patient subgroups. We further developed a multilevel regression model to assess the extent to which VVV in SBP was explained by patient characteristics. RESULTS: The study population included 537 218 adults, with a total of 7 721 864 SBP measurements. The mean age was 53.4 (SD 19.0) years, 60.4% were women, 69.4% were non-Hispanic White, and 18.1% were on antihypertensive medications. Patients had a mean body mass index of 28.4 (5.9) kg/m2 and 22.6%, 8.0%, 9.7%, and 5.6% had a history of hypertension, diabetes, hyperlipidemia, and coronary artery disease, respectively. The mean number of visits per patient was 13.3, over an average period of 2.4 years. The mean (SD) intraindividual SD and coefficient of variation of SBP across visits were 10.6 (5.1) mm Hg and 0.08 (0.04). These measures of blood pressure variation were consistent across patient subgroups defined by demographic characteristics and medical history. In the multivariable linear regression model, only 4% of the variance in absolute standardized difference was attributable to patient characteristics. CONCLUSIONS: The VVV in real-world practice poses challenges for management of patients with hypertension based on blood pressure readings in outpatient settings and suggest the need to go beyond episodic clinic evaluation.