ABSTRACT
BACKGROUND: Left Ventricular Hypertrophy (LVH) is closely linked to the cardiovascular disease prognosis, and thus, timely diagnosis improves outcomes. Diagnosis is challenging due to dependency on doctor's visits and a 12lead ECG. In addition, the interpretation of LVH from ECGs is challenging due to variability of ECG measurements, body habitus, electrode positioning, several LVH ECG criteria and EP mechanisms. The aims of this study are to evaluate different big data-driven machine learning models for ECG LVH interpretation based on limb leads only, and to compare the performance of an ECG parameter-based statistical model with a deep learning-based model. METHODS AND DATA: The first two models are binary class Random Forest (RF) models, an ensemble learning method which constructs many decision trees at training time and predicts the class chosen by the greatest number of trees at inference time. One random forest is trained using the following five features: lead aVL R-wave amplitude, lead I, II, aVL ST segment amplitude, and QRS duration. The second RF model uses 54 features across all limb leads, including the five features used by the smaller model. The second type of model is a multi-class deep neural network (DNN) which takes median beats of 6 limb leads arranged in Cabrera sequence as input. The signal preprocessing included forming median beats, filtering with a 40-Hz lowpass filter, and down-sampling to 125 Hz. The DNN models consist of 1 lead-formation convolutional layer, 5 downsampling convolutional resnet blocks with skip connections, and 3 fully connected layers. The training dataset consisted of 1 million 10-s 12lead ECGs, and an independent test dataset consisted of 250,000 10-s ECGs from the Mayo Clinic. RESULTS: The five-parameter RF model has the prediction performance of Area Under the Receiver-Operator Curve (AUC) 0.78, and the larger RF model had AUC of 0.83. The DNN model for ECG LVH detection achieves AUC 0.92 using only the limb leads, compared to an AUC of 0.98 for the full 12lead DNN. CONCLUSION: The study shows that machine learning models trained only on limb leads achieve promising results with potential to add clinical value to early detection mechanisms. We also observe that the RF model splits parameters by thresholds known to be characteristic of LVH, and that the DNN model can automatically detect morphology differences from 6 limb lead ECGs. This will be meaningful for expanding the capabilities of potential electrical LVH detection in mobile 6lead ECG devices.
Subject(s)
Electrocardiography , Hypertrophy, Left Ventricular , Humans , Electrocardiography/methods , Hypertrophy, Left Ventricular/diagnosis , Neural Networks, Computer , Random Forest , Machine LearningABSTRACT
Digital technology is now an integral part of medicine. Tools for detecting, screening, diagnosis, and monitoring health-related parameters have improved patient care and enabled individuals to identify issues leading to better management of their own health. Wearable technologies have integrated sensors and can measure physical activity, heart rate and rhythm, and glucose and electrolytes. For individuals at risk, wearables or other devices may be useful for early detection of atrial fibrillation or sub-clinical states of cardiovascular disease, disease management of cardiovascular diseases such as hypertension and heart failure, and lifestyle modification. Health data are available from a multitude of sources, namely clinical, laboratory and imaging data, genetic profiles, wearables, implantable devices, patient-generated measurements, and social and environmental data. Artificial intelligence is needed to efficiently extract value from this constantly increasing volume and variety of data and to help in its interpretation. Indeed, it is not the acquisition of digital information, but rather the smart handling and analysis that is challenging. There are multiple stakeholder groups involved in the development and effective implementation of digital tools. While the needs of these groups may vary, they also have many commonalities, including the following: a desire for data privacy and security; the need for understandable, trustworthy, and transparent systems; standardized processes for regulatory and reimbursement assessments; and better ways of rapidly assessing value.
Subject(s)
Cardiology , Cardiovascular Diseases , Heart Failure , Telemedicine , Wearable Electronic Devices , Arrhythmias, Cardiac/diagnosis , Arrhythmias, Cardiac/therapy , Artificial Intelligence , Glucose , Heart Failure/diagnosis , Heart Failure/therapy , HumansABSTRACT
Despite the recent explosion of machine learning applied to medical data, very few studies have examined algorithmic bias in any meaningful manner, comparing across algorithms, databases, and assessment metrics. In this study, we compared the biases in sex, age, and race of 56 algorithms on over 130,000 electrocardiograms (ECGs) using several metrics and propose a machine learning model design to reduce bias. Participants of the 2021 PhysioNet Challenge designed and implemented working, open-source algorithms to identify clinical diagnosis from 2- lead ECG recordings. We grouped the data from the training, validation, and test datasets by sex (male vs female), age (binned by decade), and race (Asian, Black, White, and Other) whenever possible. We computed recording-wise accuracy, area under the receiver operating characteristic curve (AUROC), area under the precision recall curve (AUPRC), F-measure, and the Challenge Score for each of the 56 algorithms. The Mann-Whitney U and the Kruskal-Wallis tests assessed the performance differences of algorithms across these demographic groups. Group trends revealed similar values for the AUROC, AUPRC, and F-measure for both male and female groups across the training, validation, and test sets. However, recording-wise accuracies were 20% higher (p < 0.01) and the Challenge Score 12% lower (p = 0.02) for female subjects on the test set. AUPRC, F-measure, and the Challenge Score increased with age, while recording-wise accuracy and AUROC decreased with age. The results were similar for the training and test sets, but only recording-wise accuracy (12% decrease per decade, p < 0.01), Challenge Score (1% increase per decade, p < 0.01), and AUROC (1% decrease per decade, p < 0.01) were statistically different on the test set. We observed similar AUROC, AUPRC, Challenge Score, and F-measure values across the different race categories. But, recording-wise accuracies were significantly lower for Black subjects and higher for Asian subjects on the training (31% difference, p < 0.01) and test (39% difference, p < 0.01) sets. A top performing model was then retrained using an additional constraint which simultaneously minimized differences in performance across sex, race and age. This resulted in a modest reduction in performance, with a significant reduction in bias. This work provides a demonstration that biases manifest as a function of model architecture, population, cost function and optimization metric, all of which should be closely examined in any model.
Subject(s)
Arrhythmias, Cardiac , Electrocardiography , Female , Humans , Male , Sex Factors , Age FactorsABSTRACT
BACKGROUND: Atrial fibrillation (AFib) is the most common cardiac arrhythmia associated with stroke, blood clots, heart failure, coronary artery disease, and/or death. Multiple methods have been proposed for AFib detection, with varying performances, but no single approach appears to be optimal. We hypothesized that each state-of-the-art algorithm is appropriate for different subsets of patients and provides some independent information. Therefore, a set of suitably chosen algorithms, combined in a weighted voting framework, will provide a superior performance to any single algorithm. METHODS: We investigate and modify 38 state-of-the-art AFib classification algorithms for a single-lead ambulatory electrocardiogram (ECG) monitoring device. All algorithms are ranked using a random forest classifier and an expert-labeled training dataset of 2,532 recordings. The seven top-ranked algorithms are combined by using an optimized weighting approach. RESULTS: The proposed fusion algorithm, when validated on a separate test dataset consisting of 4,644 recordings, resulted in an area under the receiver operating characteristic (ROC) curve of 0.99. The sensitivity, specificity, positive-predictive-value (PPV), negative-predictive-value (NPV), and F1-score of the proposed algorithm were 0.93, 0.97, 0.87, 0.99, and 0.90, respectively, which were all superior to any single algorithm or any previously published. CONCLUSION: This study demonstrates how a set of well-chosen independent algorithms and a voting mechanism to fuse the outputs of the algorithms, outperforms any single state-of-the-art algorithm for AFib detection. The proposed framework is a case study for the general notion of crowdsourcing between open-source algorithms in healthcare applications. The extension of this framework to similar applications may significantly save time, effort, and resources, by combining readily existing algorithms. It is also a step toward the democratization of artificial intelligence and its application in healthcare.