A Hybrid Decision Tree and Deep Learning Approach Combining Medical Imaging and Electronic Medical Records to Predict Intubation Among Hospitalized Patients With COVID-19: Algorithm Development and Validation.

BACKGROUND: Early prediction of the need for invasive mechanical ventilation (IMV) in patients hospitalized with COVID-19 symptoms can help in the allocation of resources appropriately and improve patient outcomes by appropriately monitoring and treating patients at the greatest risk of respiratory failure. To help with the complexity of deciding whether a patient needs IMV, machine learning algorithms may help bring more prognostic value in a timely and systematic manner. Chest radiographs (CXRs) and electronic medical records (EMRs), typically obtained early in patients admitted with COVID-19, are the keys to deciding whether they need IMV. OBJECTIVE: We aimed to evaluate the use of a machine learning model to predict the need for intubation within 24 hours by using a combination of CXR and EMR data in an end-to-end automated pipeline. We included historical data from 2481 hospitalizations at The Mount Sinai Hospital in New York City. METHODS: CXRs were first resized, rescaled, and normalized. Then lungs were segmented from the CXRs by using a U-Net algorithm. After splitting them into a training and a test set, the training set images were augmented. The augmented images were used to train an image classifier to predict the probability of intubation with a prediction window of 24 hours by retraining a pretrained DenseNet model by using transfer learning, 10-fold cross-validation, and grid search. Then, in the final fusion model, we trained a random forest algorithm via 10-fold cross-validation by combining the probability score from the image classifier with 41 longitudinal variables in the EMR. Variables in the EMR included clinical and laboratory data routinely collected in the inpatient setting. The final fusion model gave a prediction likelihood for the need of intubation within 24 hours as well. RESULTS: At a prediction probability threshold of 0.5, the fusion model provided 78.9% (95% CI 59%-96%) sensitivity, 83% (95% CI 76%-89%) specificity, 0.509 (95% CI 0.34-0.67) F1-score, 0.874 (95% CI 0.80-0.94) area under the receiver operating characteristic curve (AUROC), and 0.497 (95% CI 0.32-0.65) area under the precision recall curve (AUPRC) on the holdout set. Compared to the image classifier alone, which had an AUROC of 0.577 (95% CI 0.44-0.73) and an AUPRC of 0.206 (95% CI 0.08-0.38), the fusion model showed significant improvement (P<.001). The most important predictor variables were respiratory rate, C-reactive protein, oxygen saturation, and lactate dehydrogenase. The imaging probability score ranked 15th in overall feature importance. CONCLUSIONS: We show that, when linked with EMR data, an automated deep learning image classifier improved performance in identifying hospitalized patients with severe COVID-19 at risk for intubation. With additional prospective and external validation, such a model may assist risk assessment and optimize clinical decision-making in choosing the best care plan during the critical stages of COVID-19.

Implications of the Use of Artificial Intelligence Predictive Models in Health Care Settings : A Simulation Study.

BACKGROUND: Substantial effort has been directed toward demonstrating uses of predictive models in health care. However, implementation of these models into clinical practice may influence patient outcomes, which in turn are captured in electronic health record data. As a result, deployed models may affect the predictive ability of current and future models. OBJECTIVE: To estimate changes in predictive model performance with use through 3 common scenarios: model retraining, sequentially implementing 1 model after another, and intervening in response to a model when 2 are simultaneously implemented. DESIGN: Simulation of model implementation and use in critical care settings at various levels of intervention effectiveness and clinician adherence. Models were either trained or retrained after simulated implementation. SETTING: Admissions to the intensive care unit (ICU) at Mount Sinai Health System (New York, New York) and Beth Israel Deaconess Medical Center (Boston, Massachusetts). PATIENTS: 130 000 critical care admissions across both health systems. INTERVENTION: Across 3 scenarios, interventions were simulated at varying levels of clinician adherence and effectiveness. MEASUREMENTS: Statistical measures of performance, including threshold-independent (area under the curve) and threshold-dependent measures. RESULTS: At fixed 90% sensitivity, in scenario 1 a mortality prediction model lost 9% to 39% specificity after retraining once and in scenario 2 a mortality prediction model lost 8% to 15% specificity when created after the implementation of an acute kidney injury (AKI) prediction model; in scenario 3, models for AKI and mortality prediction implemented simultaneously, each led to reduced effective accuracy of the other by 1% to 28%. LIMITATIONS: In real-world practice, the effectiveness of and adherence to model-based recommendations are rarely known in advance. Only binary classifiers for tabular ICU admissions data were simulated. CONCLUSION: In simulated ICU settings, a universally effective model-updating approach for maintaining model performance does not seem to exist. Model use may have to be recorded to maintain viability of predictive modeling. PRIMARY FUNDING SOURCE: National Center for Advancing Translational Sciences.

Machine learning using institution-specific multi-modal electronic health records improves mortality risk prediction for cardiac surgery patients.

Background: The Society of Thoracic Surgeons risk scores are widely used to assess risk of morbidity and mortality in specific cardiac surgeries but may not perform optimally in all patients. In a cohort of patients undergoing cardiac surgery, we developed a data-driven, institution-specific machine learning-based model inferred from multi-modal electronic health records and compared the performance with the Society of Thoracic Surgeons models. Methods: All adult patients undergoing cardiac surgery between 2011 and 2016 were included. Routine electronic health record administrative, demographic, clinical, hemodynamic, laboratory, pharmacological, and procedural data features were extracted. The outcome was postoperative mortality. The database was randomly split into training (development) and test (evaluation) cohorts. Models developed using 4 classification algorithms were compared using 6 evaluation metrics. The performance of the final model was compared with the Society of Thoracic Surgeons models for 7 index surgical procedures. Results: A total of 6392 patients were included and described by 4016 features. Overall mortality was 3.0% (n = 193). The XGBoost algorithm using only features with no missing data (336 features) yielded the best-performing predictor. When applied to the test set, the predictor performed well (F-measure = 0.775; precision = 0.756; recall = 0.795; accuracy = 0.986; area under the receiver operating characteristic curve = 0.978; area under the precision-recall curve = 0.804). eXtreme Gradient Boosting consistently demonstrated improved performance over the Society of Thoracic Surgeons models when evaluated on index procedures within the test set. Conclusions: Machine learning models using institution-specific multi-modal electronic health records may improve performance in predicting mortality for individual patients undergoing cardiac surgery compared with the standard-of-care, population-derived Society of Thoracic Surgeons models. Institution-specific models may provide insights complementary to population-derived risk predictions to aid patient-level decision making.

Proteomic characterization of acute kidney injury in patients hospitalized with SARS-CoV2 infection.

BACKGROUND: Acute kidney injury (AKI) is a known complication of COVID-19 and is associated with an increased risk of in-hospital mortality. Unbiased proteomics using biological specimens can lead to improved risk stratification and discover pathophysiological mechanisms. METHODS: Using measurements of ~4000 plasma proteins in two cohorts of patients hospitalized with COVID-19, we discovered and validated markers of COVID-associated AKI (stage 2 or 3) and long-term kidney dysfunction. In the discovery cohort (N = 437), we identified 413 higher plasma abundances of protein targets and 30 lower plasma abundances of protein targets associated with COVID-AKI (adjusted p < 0.05). Of these, 62 proteins were validated in an external cohort (p < 0.05, N = 261). RESULTS: We demonstrate that COVID-AKI is associated with increased markers of tubular injury (NGAL) and myocardial injury. Using estimated glomerular filtration (eGFR) measurements taken after discharge, we also find that 25 of the 62 AKI-associated proteins are significantly associated with decreased post-discharge eGFR (adjusted p < 0.05). Proteins most strongly associated with decreased post-discharge eGFR included desmocollin-2, trefoil factor 3, transmembrane emp24 domain-containing protein 10, and cystatin-C indicating tubular dysfunction and injury. CONCLUSIONS: Using clinical and proteomic data, our results suggest that while both acute and long-term COVID-associated kidney dysfunction are associated with markers of tubular dysfunction, AKI is driven by a largely multifactorial process involving hemodynamic instability and myocardial damage.

Acute kidney injury (AKI) is a sudden, sometimes fatal, episode of kidney failure or damage. It is a known complication of COVID-19, albeit through unclear mechanisms. COVID-19 is also associated with kidney dysfunction in the long term, or chronic kidney disease (CKD). There is a need to better understand which patients with COVID-19 are at risk of AKI or CKD. We measure levels of several thousand proteins in the blood of hospitalized COVID-19 patients. We discover and validate sets of proteins associated with severe AKI and CKD in these patients. The markers identified suggest that kidney injury in COVID-19 patients involves damage to kidney cells that reabsorb fluid from urine and reduced blood flow to the heart, causing damage to heart muscles. Our findings might help clinicians to predict kidney injury in patients with COVID-19, and to understand its mechanisms.

A study of gene expression in the living human brain.

A goal of medical research is to determine the molecular basis of human brain health and illness. One way to achieve this goal is through observational studies of gene expression in human brain tissue. Due to the unavailability of brain tissue from living people, most such studies are performed using tissue from postmortem brain donors. An assumption underlying this practice is that gene expression in the postmortem human brain is an accurate representation of gene expression in the living human brain. Here, this assumption - which, until now, had not been adequately tested - is tested by comparing human prefrontal cortex gene expression between 275 living samples and 243 postmortem samples. Expression levels differed significantly for nearly 80% of genes, and a systematic examination of alternative explanations for this observation determined that these differences are not a consequence of cell type composition, RNA quality, postmortem interval, age, medication, morbidity, symptom severity, tissue pathology, sample handling, batch effects, or computational methods utilized. Analyses integrating the data generated for this study with data from earlier landmark studies that used tissue from postmortem brain donors showed that postmortem brain gene expression signatures of neurological and mental illnesses, as well as of normal traits such as aging, may not be accurate representations of these gene expression signatures in the living brain. By using tissue from large cohorts living people, future observational studies of human brain biology have the potential to (1) determine the medical research questions that can be addressed using postmortem tissue as a proxy for living tissue and (2) expand the scope of medical research to include questions about the molecular basis of human brain health and illness that can only be addressed in living people (e.g., "What happens at the molecular level in the brain as a person experiences an emotion?").

Proteomic Characterization of Acute Kidney Injury in Patients Hospitalized with SARS-CoV2 Infection.

Background Acute kidney injury (AKI) is a known complication of COVID-19 and is associated with an increased risk of in-hospital mortality. Unbiased proteomics using biological specimens can lead to improved risk stratification and discover pathophysiological mechanisms. Methods Using measurements of ~4000 plasma proteins in two cohorts of patients hospitalized with COVID-19, we discovered and validated markers of COVID-associated AKI (stage 2 or 3) and long-term kidney dysfunction. In the discovery cohort (N= 437), we identified 413 higher plasma abundances of protein targets and 40 lower plasma abundances of protein targets associated with COVID-AKI (adjusted p <0.05). Of these, 62 proteins were validated in an external cohort (p <0.05, N =261). Results We demonstrate that COVID-AKI is associated with increased markers of tubular injury ( NGAL ) and myocardial injury. Using estimated glomerular filtration (eGFR) measurements taken after discharge, we also find that 25 of the 62 AKI-associated proteins are significantly associated with decreased post-discharge eGFR (adjusted p <0.05). Proteins most strongly associated with decreased post-discharge eGFR included desmocollin-2 , trefoil factor 3 , transmembrane emp24 domain-containing protein 10 , and cystatin-C indicating tubular dysfunction and injury. Conclusions Using clinical and proteomic data, our results suggest that while both acute and long-term COVID-associated kidney dysfunction are associated with markers of tubular dysfunction, AKI is driven by a largely multifactorial process involving hemodynamic instability and myocardial damage.

A Web-Based Reporting System for Reviewing Local Practice Patterns of Anesthesiologists Derived from the Electronic Medical Record.

After completion of training, anesthesiologists may have fewer opportunities to see how colleagues practice, and their breadth of case experiences may also diminish due to specialization. We created a web-based reporting system based on data extracted from electronic anesthesia records that allows practitioners to see how other clinicians practice in similar cases. One year after implementation, the system continues to be utilized by clinicians.

Informative Censoring-A Cause of Bias in Estimating COVID-19 Mortality Using Hospital Data.

(1) Background: Several retrospective observational analyzed treatment outcomes for COVID-19; (2) Methods: Inverse probability of censoring weighting (IPCW) was applied to correct for bias due to informative censoring in database of hospitalized patients who did and did not receive convalescent plasma; (3) Results: When compared with an IPCW analysis, overall mortality was overestimated using an unadjusted Kaplan-Meier curve, and hazard ratios for the older age group compared to the youngest were underestimated using the Cox proportional hazard models and 30-day mortality; (4) Conclusions: An IPCW analysis provided stabilizing weights by hospital admission.

Multiplex Immunochromogenic Tissue Staining Employing Primary Antibodies from the Same Species.

Chromogenic immunohistochemistry (IHC) serves as an essential assay for the diagnoses of many diseases including cancer. Single-marker IHC detection is the standard used for clinical diagnostic assays. A technology to stain multiple biomarkers chromogenically on a single tissue will also yield contextual biomarker information. Methods to chromogenically stain multiple biomarkers simultaneously employing antibodies from the same species are limited and require complex protocols. Here we describe both manual and automated protocols using the UltraPlex™ mxIHC technology that allows simultaneous detection of up to three biomarkers on a single tissue using a single heat-induced antigen retrieval step in formaldehyde-fixed paraffin-embedded (FFPE) tissue and using primary antibodies from any species.

Plasma Renin Activity Increases With Cardiopulmonary Bypass and is Associated With Vasoplegia After Cardiac Surgery.

OBJECTIVES: To describe the trend in plasma renin activity over time in patients undergoing cardiac surgery on cardiopulmonary bypass, and to investigate if increased plasma renin activity is associated with postcardiopulmonary bypass vasoplegia. DESIGN: A prospective cohort study. SETTING: Patients were enrolled from June 2020 to May 2021 at a tertiary cardiac surgical institution. PATIENTS: A cohort of 100 adult patients undergoing cardiac surgery on cardiopulmonary bypass. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Plasma renin activity was measured at 5 time points: baseline, postoperatively, and at midnight on postoperative days 1, 2, and 3. Plasma renin activity and delta plasma renin activity were correlated with the incidence of vasoplegia and clinical outcomes. The median plasma renin activity increased approximately 3 times from baseline immediately after cardiac surgery, remained elevated on postoperative days 0, 1, and 2, and began to downtrend on postoperative day 3. Plasma renin activity was approximately 3 times higher at all measured time points in patients who developed vasoplegia versus those who did not. CONCLUSIONS: In patients undergoing cardiac surgery on cardiopulmonary bypass, plasma renin activity increased postoperatively and remained elevated through postoperative day 2. Additionally, patients with vasoplegic syndrome after cardiac surgery on cardiopulmonary bypass had more robust elevations in plasma renin activity than nonvasoplegic patients. These findings support the need for randomized controlled trials to determine if patients undergoing cardiac surgery with high plasma renin activity may benefit from targeted treatment with therapies such as synthetic angiotensin II.

Automated electrocardiogram signal quality assessment based on Fourier analysis and template matching.

We developed and tested a novel template matching approach for signal quality assessment on electrocardiogram (ECG) data. A computational method was developed that uses a sinusoidal approximation to the QRS complex to generate a correlation value at every point of an ECG. The strength of this correlation can be numerically adapted into a 'score' for each segment of an ECG, which can be used to stratify signal quality. The algorithm was tested on lead II ECGs of intensive care unit (ICU) patients admitted to the Mount Sinai Hospital (MSH) from January to July 2020 and on records from the MIT BIH arrhythmia database. The algorithm was found to be 98.9% specific and 99% sensitive on test data from the MSH ICU patients. The routine performs in linear O(n) time and occupies O(1) heap space in runtime. This approach can be used to lower the burden of pre-processing in ECG signal analysis. Given its runtime (O(n)) and memory (O(1)) complexity, there are potential applications for signal quality stratification and arrhythmia detection in wearable devices or smartphones.

Predicting Adult Hospital Admission from Emergency Department Using Machine Learning: An Inclusive Gradient Boosting Model.

BACKGROUND AND AIM: We analyzed an inclusive gradient boosting model to predict hospital admission from the emergency department (ED) at different time points. We compared its results to multiple models built exclusively at each time point. METHODS: This retrospective multisite study utilized ED data from the Mount Sinai Health System, NY, during 2015-2019. Data included tabular clinical features and free-text triage notes represented using bag-of-words. A full gradient boosting model, trained on data available at different time points (30, 60, 90, 120, and 150 min), was compared to single models trained exclusively at data available at each time point. This was conducted by concatenating the rows of data available at each time point to one data matrix for the full model, where each row is considered a separate case. RESULTS: The cohort included 1,043,345 ED visits. The full model showed comparable results to the single models at all time points (AUCs 0.84-0.88 for different time points for both the full and single models). CONCLUSION: A full model trained on data concatenated from different time points showed similar results to single models trained at each time point. An ML-based prediction model can use used for identifying hospital admission.

The fourth wave: vaccination status and intensive care unit mortality at a large hospital system in New York City.

BACKGROUND: We aim to describe the demographics and outcomes of patients with severe disease with the Omicron variant. The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) virus continues to mutate, and the availability of vaccines and boosters continue to rise, it is important to understand the health care burden of new variants. We analyze patients admitted to intensive care units (ICUs) in a large Academic Health System during New York City's fourth surge beginning on November 27, 2021. METHODS: All patients admitted to an ICU were included in the primary analysis. Key demographics and outcomes were retrospectively compared between patients stratified by vaccination status. Univariate and multivariate logistic regression was used to identify risk factors for in-hospital mortality. RESULTS: In-hospital mortality for all admitted patients during the fourth wave was significantly lower than in previous waves. However, among patients requiring intensive care, in-hospital mortality was high across all levels of vaccination status. In a multivariate model older age was associated with increased in-hospital mortality, vaccination status of overdue for booster was associated with decreased in hospital mortality, and vaccination status of up-to-date with vaccination showed a trend to reduced mortality. CONCLUSIONS: In-hospital mortality of patients with severe respiratory failure from coronavirus disease 2019 (COVID-19) remains high despite decreasing overall mortality. Vaccination against SARS-CoV-2 was protective against mortality. Vaccination remains the best and safest way to protect against serious illness and death from COVID-19. It remains unclear that any other treatment will have success in changing the natural history of the disease.

How businesses are working together to deliver NASA/JPL-designed ventilators to the world in the fight against COVID-19.

In response to the COVID-19 pandemic, NASA Jet Propulsion Laboratory (JPL) engineers had embarked on an ambitious project to design a reliable, easy-to-use, and low-cost ventilator that was made of readily available parts to address the unexpected global shortage of these lifesaving devices. After successfully designing and building the VITAL (Ventilator Intervention Technology Accessible Locally) ventilator in record time, FDA Emergency Use Authorization (EUA) was obtained and then the license to manufacture and sell these ventilators was made available to select companies through a competitive process. STARK Industries, LLC (STARK), located in Columbus, OH, USA, was one of only eight U.S. companies to be selected to receive this worldwide license. Motivated by its mission to improve human health and well-being through innovated medical technologies, STARK accepted the challenge of further developing the VITAL technology and manufacturing the ventilators in large quantities and making them available to those in need around the world. To this end, Spiritus Medical, Inc (Spiritus) was spun off from STARK to focus on the ventilator business. Through collaborative efforts with various corporate, academic, governmental, and non-profit partners, Spiritus was able to successfully begin manufacturing and selling its ventilators. Due to its low-cost nature and its straightforward design, this ventilator is ideal for use in developing countries where ventilators are in short supply and affordability is a major consideration. This is a story of how NASA's ingenuity, based on space-based know-how and experience, was used to rapidly design this innovative ventilator. And by forging partnerships with highly qualified and motivated partners such as STARK and Spiritus, NASA has succeeded in translating this work into technology that could potentially save thousands of lives in the fight against the COVID-19 pandemic.

Proteomic Characterization of Acute Kidney Injury in Patients Hospitalized with SARS-CoV2 Infection.

Acute kidney injury (AKI) is a known complication of COVID-19 and is associated with an increased risk of in-hospital mortality. Unbiased proteomics using biological specimens can lead to improved risk stratification and discover pathophysiological mechanisms. Using measurements of ∼4000 plasma proteins in two cohorts of patients hospitalized with COVID-19, we discovered and validated markers of COVID-associated AKI (stage 2 or 3) and long-term kidney dysfunction. In the discovery cohort (N= 437), we identified 413 higher plasma abundances of protein targets and 40 lower plasma abundances of protein targets associated with COVID-AKI (adjusted p <0.05). Of these, 62 proteins were validated in an external cohort (p <0.05, N =261). We demonstrate that COVID-AKI is associated with increased markers of tubular injury (NGAL) and myocardial injury. Using estimated glomerular filtration (eGFR) measurements taken after discharge, we also find that 25 of the 62 AKI-associated proteins are significantly associated with decreased post-discharge eGFR (adjusted p <0.05). Proteins most strongly associated with decreased post-discharge eGFR included desmocollin-2, trefoil factor 3, transmembrane emp24 domain-containing protein 10, and cystatin-C indicating tubular dysfunction and injury. Using clinical and proteomic data, our results suggest that while both acute and long-term COVID-associated kidney dysfunction are associated with markers of tubular dysfunction, AKI is driven by a largely multifactorial process involving hemodynamic instability and myocardial damage.

Machine learning to predict in-hospital mortality among patients with severe obesity: Proof of concept study.

Objectives: Hospitalized patients with severe obesity require adapted hospital management. The aim of this study was to evaluate a machine learning model to predict in-hospital mortality among this population. Methods: Data of unselected consecutive emergency department admissions of hospitalized patients with severe obesity (BMI ≥ 40 kg/m2) was analyzed. Data was retrieved from five hospitals from the Mount Sinai health system, New York. The study time frame was between January 2011 and December 2019. Data was used to train a gradient-boosting machine learning model to identify in-hospital mortality. The model was trained and evaluated based on the data from four hospitals and externally validated on held-out data from the fifth hospital. Results: A total of 14,078 hospital admissions of inpatients with severe obesity were included. The in-hospital mortality rate was 297/14,078 (2.1%). In univariate analysis, albumin (area under the curve [AUC] = 0.77), blood urea nitrogen (AUC = 0.76), acuity level (AUC = 0.73), lactate (AUC = 0.72), and chief complaint (AUC = 0.72) were the best single predictors. For Youden's index, the model had a sensitivity of 0.77 (95% CI: 0.67-0.86) with a false positive rate of 1:9. Conclusion: A machine learning model trained on clinical measures provides proof of concept performance in predicting mortality in patients with severe obesity. This implies that such models may help to adopt specific decision support tools for this population.

Predictors of antiplatelet cessation in a real-world patient population undergoing non-cardiac surgery after PCI.

BACKGROUND: The optimal perioperative management of antiplatelet therapy (APT) therapy in patients undergoing noncardiac surgery (NCS) following percutaneous coronary intervention (PCI) is unclear. We sought to identify predictors of APT cessation in a real-world cohort of patients undergoing NCS within 1 year of PCI. METHODS: Consecutive patients undergoing PCI at a tertiary center between 2011 and 2018 were prospectively enrolled. Perioperative interruption of APT was defined as cessation of either aspirin or P2Y12 inhibitor between 1 and 14 days prior to NCS. Predictors of APT discontinuation were identified by multivariable Cox regression with stepwise selection of candidate variables. RESULTS: A total of 1092 surgeries corresponding to 747 patients were identified. Overall, there were 487 (44.6%) preoperative antiplatelet interruptions: discontinuation of either P2Y12 inhibitors only (47.4%), aspirin only (7.9%), or both agents (44.7%). Both patient-specific risk factors (prior stroke, lower BMI, anemia, MI) and procedure specific risk factors (chronic total occlusions, multivessel disease, drug-eluting stent use) affected decisions regarding APT cessation. Likelihood of APT cessation increased in higher-risk surgeries and in patients on more potent P2Y12 inhibitors (ticagrelor/prasugrel vs clopidogrel) whereas those undergoing NCS <90 days post PCI were less likely to have cessation of APT. CONCLUSION: In this contemporary cohort of post-PCI patients undergoing NCS, patient-, angiographic- and surgery-specific factors all affected decision-making regarding APT cessation. Our findings reflective of real-world practice, highlight the importance of a multidisciplinary team approach to individualize decision making in these patients.

Development of a machine learning model using electrocardiogram signals to improve acute pulmonary embolism screening.

Aims: Clinical scoring systems for pulmonary embolism (PE) screening have low specificity and contribute to computed tomography pulmonary angiogram (CTPA) overuse. We assessed whether deep learning models using an existing and routinely collected data modality, electrocardiogram (ECG) waveforms, can increase specificity for PE detection. Methods and results: We create a retrospective cohort of 21 183 patients at moderate- to high suspicion of PE and associate 23 793 CTPAs (10.0% PE-positive) with 320 746 ECGs and encounter-level clinical data (demographics, comorbidities, vital signs, and labs). We develop three machine learning models to predict PE likelihood: an ECG model using only ECG waveform data, an EHR model using tabular clinical data, and a Fusion model integrating clinical data and an embedded representation of the ECG waveform. We find that a Fusion model [area under the receiver-operating characteristic curve (AUROC) 0.81 ± 0.01] outperforms both the ECG model (AUROC 0.59 ± 0.01) and EHR model (AUROC 0.65 ± 0.01). On a sample of 100 patients from the test set, the Fusion model also achieves greater specificity (0.18) and performance (AUROC 0.84 ± 0.01) than four commonly evaluated clinical scores: Wells' Criteria, Revised Geneva Score, Pulmonary Embolism Rule-Out Criteria, and 4-Level Pulmonary Embolism Clinical Probability Score (AUROC 0.50-0.58, specificity 0.00-0.05). The model is superior to these scores on feature sensitivity analyses (AUROC 0.66-0.84) and achieves comparable performance across sex (AUROC 0.81) and racial/ethnic (AUROC 0.77-0.84) subgroups. Conclusion: Synergistic deep learning of ECG waveforms with traditional clinical variables can increase the specificity of PE detection in patients at least at moderate suspicion for PE.