Tom Pollard’s research while affiliated with Massachusetts Institute of Technology and other places

Background Valvular heart disease is a highly prevalent condition that contributes significantly to cardiovascular morbidity and mortality. Echocardiography is the gold-standard for valvular disease evaluation; however, diagnosis may be delayed due to limited resources or expertise. Deep learning analysis of electrocardiography (ECG) has demonstrated the ability to identify subtle structural and functional abnormalities within the cardiovascular system. Purpose The aim of this study was to evaluate the accuracy of a deep learning algorithm to identify valvular heart disease from a standard 12-lead ECG. Methods From 2008 to 2019, a total of 27,689 patients who were admitted to the Medical Center with an echocardiogram and ECG performed within 14 days of each other were included in the study. We trained a convolutional neural network to identify the presence of moderate or severe aortic stenosis (AS), aortic regurgitation (AR), mitral regurgitation (MR), and/or tricuspid regurgitation (TR) using a 12-lead ECG. Model performance was assessed using area under the receiver-operating characteristic (AUROC). Patients were randomly divided into development (n = 22,151; 80%) and independent validation sets (n = 5,538; 20[AN1] %). Data used for the study was collected from two open access databases, MIMIC-IV-ECG and MIMIC-IV-Note. Results Mean age of the cohort was 65.9 ± 16.6 years old and 51% were female. On the independent validation dataset, AS was present in 126 (2.3%) of patients, AR in 51 (1%), MR in 337 (6.1%), and TR in 450 (8.1 %) patients. During validation, the model accurately identified AS, AR, MR, and TR with AUROCs of 0.80 (95% CI 0.77-0.83), 0.74 (95% CI 0.69-0.80), 0.86 (95% CI 0.84-0.88), and 0.85 (95% CI 0.83-0.87), respectively (Table 1). Test characteristics were dependent on underlying prevalence and selected sensitivity levels. For a composite of any of AS, AR, MR, or TR, the model achieved a positive predictive value of 28% with 79.4% sensitivity. Conclusion A deep learning-enabled ECG demonstrates robust performance to identify patients living with valvular heart disease. This model may serve as a simple and promising tool for improving the early detection of valvular heart disease, showcasing the potential for an early development valvular disease screening program.

Without careful dissection of the ways in which biases can be encoded into artificial intelligence (AI) health technologies, there is a risk of perpetuating existing health inequalities at scale. One major source of bias is the data that underpins such technologies. The STANDING Together recommendations aim to encourage transparency regarding limitations of health datasets and proactive evaluation of their effect across population groups. Draft recommendation items were informed by a systematic review and stakeholder survey. The recommendations were developed using a Delphi approach, supplemented by a public consultation and international interview study. Overall, more than 350 representatives from 58 countries provided input into this initiative. 194 Delphi participants from 25 countries voted and provided comments on 32 candidate items across three electronic survey rounds and one in-person consensus meeting. The 29 STANDING Together consensus recommendations are presented here in two parts. Recommendations for Documentation of Health Datasets provide guidance for dataset curators to enable transparency around data composition and limitations. Recommendations for Use of Health Datasets aim to enable identification and mitigation of algorithmic biases that might exacerbate health inequalities. These recommendations are intended to prompt proactive inquiry rather than acting as a checklist. We hope to raise awareness that no dataset is free of limitations, so transparent communication of data limitations should be perceived as valuable, and absence of this information as a limitation. We hope that adoption of the STANDING Together recommendations by stakeholders across the AI health technology lifecycle will enable everyone in society to benefit from technologies which are safe and effective.

Objective: To challenge clinicians and informaticians to learn about potential sources of bias in medical machine learning models through investigation of data and predictions from an open-source severity of illness score. Methods: Over a two-day period (total elapsed time approximately 28 hours), we conducted a datathon that challenged interdisciplinary teams to investigate potential sources of bias in the Global Open Source Severity of Illness Score. Teams were invited to develop hypotheses, to use tools of their choosing to identify potential sources of bias, and to provide a final report. Results: Five teams participated, three of which included both informaticians and clinicians. Most (4/5) used Python for analyses, the remaining team used R. Common analysis themes included relationship of the GOSSIS-1 prediction score with demographics and care related variables; relationships between demographics and outcomes; calibration and factors related to the context of care; and the impact of missingness. Representativeness of the population, differences in calibration and model performance among groups, and differences in performance across hospital settings were identified as possible sources of bias. Discussion: Datathons are a promising approach for challenging developers and users to explore questions relating to unrecognized biases in medical machine learning algorithms.

We present the INSPIRE dataset, a publicly available research dataset in perioperative medicine, which includes approximately 130,000 surgical operations at an academic institution in South Korea over a ten-year period between 2011 and 2020. This comprehensive dataset includes patient characteristics such as age, sex, American Society of Anesthesiologists physical status classification, diagnosis, surgical procedure code, department, and type of anaesthesia. The dataset also includes vital signs in the operating theatre, general wards, and intensive care units (ICUs), laboratory results from six months before admission to six months after discharge, and medication during hospitalisation. Complications include total hospital and ICU length of stay and in-hospital death. We hope this dataset will inspire collaborative research and development in perioperative medicine and serve as a reproducible external validation dataset to improve surgical outcomes.

Background Identifying regional wall motion abnormalities (RWMAs) is critical for diagnosing and risk stratifying patients with cardiovascular disease, particularly ischemic heart disease. We hypothesized that a deep neural network could accurately identify patients with regional wall motion abnormalities from a readily available standard 12-lead electrocardiogram (ECG), reducing the need for echocardiography. Methods This observational, retrospective study included patients who were treated at Beth Israel Deaconess Medical Center and had an ECG and echocardiogram performed within 14 days of each other between 2008 and 2019. We trained a convolutional neural network to detect the presence of RWMAs, qualitative global right ventricular (RV) hypokinesis, and varying degrees of left ventricular dysfunction (left ventricular ejection fraction [LVEF] ≤50%, LVEF ≤40%, and LVEF ≤35%) identified by echocardiography, using ECG data alone. Patients were randomly split into development (80%) and test sets (20%). Model performance was assessed using area under the receiver operating characteristic curve (AUC). Cox proportional hazard models adjusted for age and sex were performed to estimate the risk of future acute coronary events. Results The development set consisted of 19,837 patients (mean age 66.7±16.4, 46.7% female) and the test set comprised of 4,953 patients (mean age 67.5±15.8 years; 46.5% female). On the test dataset, the model accurately identified the presence of RWMA, RV hypokinesis, LVEF ≤50%, LVEF ≤40%, and LVEF ≤35% with AUCs of 0.87 (95% CI 0.858-0.882), 0.888 (95% CI 0.878-0.899), 0.923 (95% CI 0.914-0.933), 0.93 (95% CI 0.921-0.939), and 0.876 (95% CI 0.858-0.896), respectively. Among patients with normal biventricular function at the time of the index ECG, those classified as having RMWA by the model were 3 times the risk (age- and sex-adjusted hazard ratio, 2.8; 95% CI 1.9-3.9) for future acute coronary events compared to those classified as negative. Conclusions We demonstrate that a deep neural network can help identify regional wall motion abnormalities and reduced LV function from a 12-lead ECG and could potentially be used as a screening tool for triaging patients who need either initial or repeat echocardiographic imaging.