Electronic Health Record Dataset Research Articles

Complement-mediated hemolysis persists year round in patients with cold agglutinin disease.

Cold agglutinin disease (CAD) is a rare autoimmune hemolytic anemia mediated by immunoglobulin M autoantibodies that bind to the "I" antigen on erythrocytes. IgM binding results in either agglutination at ≤37°C, activation of the classical complement pathway, or both. Patients with CAD can have transient agglutination-mediated circulatory symptoms triggered by exposure to cold conditions. Separately, patients with CAD can experience complement-mediated symptoms such as anemia, hemolysis, and fatigue, but the effect of the season on these complement-mediated manifestations of CAD and clinical outcomes is not well understood. Using data from the Optum® de-identified Electronic Health Record dataset, we compared hemoglobin, markers of hemolysis (bilirubin and lactate dehydrogenase [LDH]), and healthcare resource utilization (HRU) between seasons for 594 patients (62% female; 66% aged ≥65 years) with CAD (defined as having CAD-related terms in their clinical notes on ≥3 separate occasions between December 2008 and May 2016). Laboratory parameters and HRU were compared between seasons using multivariate regression models. Estimated median hemoglobin (9.87 g/dL in summer and 9.86 g/dL in winter; P=0.944) and bilirubin (1.04 mg/dL in summer and 1.09 mg/dL in winter; P=0.257) were similar in winter versus summer. While LDH was statistically significantly higher in winter compared with summer (P < 0.001), the estimated median value was above normal for both seasons (309 U/L in summer and 367 U/L in winter). HRU measures and transfusion and thromboembolism rates were similar across seasons. Patients with CAD had evidence of persistent chronic hemolysis, HRU, and thromboembolism risk year round.

Abstract 13957: Predicting In-Hospital Mortality Among Patients Hospitalized With Heart Failure Using a Convolutional Neural Network and Transfer Learning on a Single Admission Chest X-Ray

Background: Early identification of mortality risk among patients hospitalized with heart failure (HF) is critical for therapeutic decisions. Early warning systems have relied on large datasets with frequent outcomes. We evaluate the use of transfer learning to use a single chest x-ray obtained at presentation in a small sample of HF hospitalizations to predict in-hospital mortality. Methods: In MIMIC-IV - a publicly available electronic health record dataset of consecutive hospitalizations from the ED or those requiring ≥1 day of ICU stay during 2008-2019 - we identified all hospitalizations with a primary diagnosis of HF. We combined this with chest x-rays (CXRs) drawn from the MIMIC-CXR dataset with available radiographic labels. We evaluated the risk of mortality based on CXR using a convolutional neural network (CNN) based on the Efficient Net B3 architecture. We compared 2 strategies - (1) training on primary HF hospitalizations directly for mortality risk (direct learning), or (2) pre-training the model on a non-mortality radiographic label, before training for mortality (transfer learning) (Fig A). Results: There were 3604 hospitalizations with a primary diagnosis of HF (mean age 70.8 years, 48% women) who had a corresponding CXR, of whom 87 (2.4%) died in the hospital. There were 102,551 hospitalizations without a diagnosis of HF with a CXR. The CNN model developed on HF data (60% training, 20% validation, 20% test) led to modest predictive performance (AUROC 0.67) for mortality (Fig B). In the second model that was pre-trained on identifying pleural effusion on CXRs in the absence of a HF diagnosis, and was tuned on the HF data for mortality, was associated with a large predictive gain (AUROC 0.80) (Fig B). Conclusion: A single CXR can accurately classify the mortality risk of patients with HF with high model discrimination. Transfer learning allows the use of unstructured data inputs in predictive models for small datasets and few inputs.

Abstract 11519: CHA 2 DS 2 VASc Score as a Predictor of Ablation Success Defined by Continuous Long-Term Monitoring

Introduction: There are few readily available and effective risk stratification tools for successful catheter ablation (CA) of atrial fibrillation (AF). CHA 2 DS 2 VASc score is a marker for vascular disease and has been shown to be mildly predictive of CA success, although prior studies have used short-term external monitors to define success. Hypothesis: Lower CHA 2 DS 2 VASc score correlates with CA success as measured by continuous monitoring via cardiac implantable electronic devices (CIED). Methods: Using the Optum® de-identified Electronic Health Record dataset (01/2007 to 06/2019) linked with the Medtronic CareLink device database, we identified patients who underwent a 1 st CA procedure for AF following CIED implantation. CA success was defined as absence of ≥1 hour of AF on any day following a 3-month blanking period. A proportional hazard model was used to estimate hazard ratios (HR) between CHA 2 DS 2 VASc subgroups (0-2, 3-4, ≥5). Results: A total of 665 patients (age 66.9±9.1 years, 73.4% male, CHA 2 DS 2 VASc 3.5±1.8, 37.3% paroxysmal AF, 62.7% persistent AF) were analyzed and included 36.2% insertable cardiac monitor, 28.0% PPM, 21.4% ICD, 13.2% CRT-D, and 1.2% CRT-P. CA success at 24 months post blanking period was 39.8% (95% CI 32.4%-48.9%), 36.9% (95% CI 29.7%-45.8%), and 20.9% (95% CI 13.9%-31.4%) for CHA 2 DS 2 VASc subgroups of 0-2, 3-4, and ≥5, respectively. Kaplan-Meier curves demonstrate a difference in survival free from AF between the three CHA 2 DS 2 VASc subgroups (p=0.007, Figure), with CHA 2 DS 2 VASc score being a significant predictor of CA ablation success (c-statistic 0.533). After controlling for AF type, patients with a score ≥5 remained at higher risk for AF recurrence compared to patients with a score of 0-2 (HR 1.28, 95% CI 1.001-1.63). Conclusions: In this real-world dataset of AF patients with continuous monitoring undergoing CA, CHA 2 DS 2 VASc score is a modest predictor of CA success.

Abstract 11773: Impact of Race on Incident Atrial Fibrillation in a Large Cohort With Cardiac Implantable Electronic Devices

Introduction: Prior studies showed lower rates of atrial fibrillation (AF) in African Americans (AA) compared to Caucasians (CC) despite a higher prevalence of AF risk factors. These studies often relied on a clinical diagnosis of AF or limited analyses to Medicare beneficiaries which may bias the results. Our objective is to compare AF-free survival in AA vs CC patients with dual chamber cardiac implantable electronic devices (CIEDs) capable of continuous AF monitoring. Hypothesis: There is no significant difference in incident AF between AA and CC patients with dual chamber CIEDs. Methods: Using the Optum® de-identified Electronic Health Record dataset (01/2007 to 06/2019) linked with the Medtronic CareLink® device database, we identified self-described AA and CC patients with dual chamber CIEDs and no prior AF history. The primary endpoint was time to first AF ≥ 6 minutes. Cox proportional hazards models were used to examine the association between race and AF while adjusting for age, sex, and vascular risk factors in the pacemaker and ICD cohorts. Results: Of the 110,323 patients analyzed, mean age was 69±13 years, 64% were male, 8.9% were AA, and 39% had a pacemaker. CC were older (69.5 vs 63.4 years) and more were male (65% vs 54%). Heart failure, hypertension and diabetes were more common in AA (36% vs 22%; 60% vs 47%; 32% vs 22%, respectively). Over a mean follow-up of 2.3±1.8 years, unadjusted AF-free survival was higher in AA compared to CC in both CIED cohorts (p&lt;0.0001). After adjusting for age, sex and vascular risk factors, CC with pacemakers had a higher risk of incident AF (HR 1.10; 95% CI 1.03-1.17) but a lower risk was observed in CC with ICDs (HR 0.91; 95% CI 0.87-0.94) (Figure 1). Conclusions: In contrast to previous studies, this analysis of 110,323 CIED patients showed that adjusted AF-free survival was higher in AAs with pacemakers but not ICDs. These results challenge our understanding of the interaction between race and incident AF.

Characteristics, Comorbidities, and Outcomes of SARS-CoV-2 Infection in Patients With Autoimmune Conditions Treated With Systemic Therapies: A Population-based Study.

To describe characteristics and coronavirus disease 2019 (COVID-19) clinical outcomes of patients with rheumatoid arthritis (RA), psoriatic arthritis (PsA), or ulcerative colitis (UC) receiving systemic therapies vs the general population. This descriptive retrospective cohort study used data from the United States Optum deidentified COVID-19 electronic health record dataset (2007-2020). Adults with COVID-19 were stratified into 3 disease cohorts (patients with RA, PsA, or UC who had received systemic therapy) and a comparator cohort not meeting these criteria. Incidence proportions of hospitalization and clinical manifestations of interest were calculated. Using logistic regression analyses, risk of endpoints was estimated, adjusting for demographics and demographics plus comorbidities. This analysis (February 1 to December 9, 2020) included 315,101 patients with COVID-19. Adjusting for demographics, COVID-19 patients with RA (n = 2306) had an increased risk of hospitalization (OR 1.54, 95% CI 1.39-1.70) and in-hospital death (OR 1.61, 95% CI 1.30-2.00) compared with the comparator cohort (n = 311,563). The increased risk was also observed when adjusted for demographics plus comorbidities (hospitalization OR 1.25, 95% CI 1.13-1.39 and in-hospital death OR 1.35, 95% CI 1.09-1.68]). The risk of hospitalization was lower in COVID-19 patients with RA receiving tumor necrosis factor inhibitors (TNFi) vs non-TNFi biologics (OR 0.32, 95% CI 0.20-0.53) and the comparator cohort (OR 0.77, 95% CI 0.51-1.17). The risk of hospitalization due to COVID-19 was similar between patients receiving tofacitinib and the comparator cohort. Compared with the comparator cohort, patients with RA were at a higher risk of more severe or critical COVID-19 and, except for non-TNFi biologics, systemic therapies did not further increase the risk. (ENCePP; registration no. EU PAS 35384).

Evaluation of COVID-19 Mortality and Adverse Outcomes in US Patients With or Without Cancer

As the COVID-19 pandemic continues, understanding the clinical outcomes of patients with cancer and COVID-19 has become critically important. To compare the outcomes of patients with or without cancer who were diagnosed with COVID-19 and to identify the factors associated with mortality, mechanical ventilation, intensive care unit (ICU) stay, and hospitalization. This cohort study obtained data from the Optum de-identified COVID-19 electronic health record data set. More than 500 000 US adults who were diagnosed with COVID-19 from January 1 to December 31, 2020, were analyzed. The patient groups were (1) patients without cancer, (2) patients with no recent cancer treatment, and (3) patients with recent cancer treatment (within 3 months before COVID-19 diagnosis) consisting of radiation therapy or systemic therapy. Mortality, mechanical ventilation, ICU stay, and hospitalization within 30 days of COVID-19 diagnosis were the main outcomes. Unadjusted rates and adjusted odds ratios (ORs) of adverse outcomes were presented according to exposure group. A total of 507 307 patients with COVID-19 were identified (mean [SD] age, 48.4 [18.4] years; 281 165 women [55.4%]), of whom 493 020 (97.2%) did not have cancer. Among the 14 287 (2.8%) patients with cancer, 9991 (69.9%) did not receive recent treatment and 4296 (30.1%) received recent treatment. In unadjusted analyses, patients with cancer, regardless of recent treatment received, were more likely to have adverse outcomes compared with patients without cancer (eg, mortality rate: 1.6% for patients without cancer, 5.0% for patients with no recent cancer treatment, and 7.8% for patients with recent cancer treatment). After adjustment, patients with no recent cancer treatment had similar or better outcomes than patients without cancer (eg, mortality OR, 0.93 [95% CI, 0.84-1.02]; mechanical ventilation OR, 0.61 [95% CI, 0.54-0.68]). In contrast, a higher risk of death (OR, 1.74; 95% CI, 1.54-1.96), ICU stay (OR, 1.69; 95% CI, 1.54-1.87), and hospitalization (OR, 1.19; 95% CI, 1.11-1.27) was observed in patients with recent cancer treatment. Compared with patients with nonmetastatic solid tumors, those with metastatic solid tumors and hematologic malignant neoplasms had worse outcomes (eg, mortality OR, 2.36 [95% CI, 1.96-2.84]; mechanical ventilation OR, 0.87 [95% CI, 0.70-1.08]). Recent chemotherapy and chemoimmunotherapy were also associated with worse outcomes (eg, chemotherapy mortality OR, 1.84 [95% CI, 1.51-2.26]). This cohort study found that patients with recent cancer treatment and COVID-19 had a significantly higher risk of adverse outcomes, and patients with no recent cancer treatment had similar outcomes to those without cancer. The findings have risk stratification and resource use implications for patients, clinicians, and health systems.

Community-acquired pneumonia in infants: Not simply an acute event with complete recovery

BackgroundPneumonia in infancy has been linked to long-term consequences for the rapidly developing lung. We examined the impact of hospitalized community-acquired pneumonia (CAP) on subsequent respiratory health. MethodsWe conducted a retrospective matched-cohort study using the Optum® de-identified Electronic Health Record Dataset (2009–2018). Study population comprised healthy infants hospitalized for CAP (“CAP patients”), and matched comparators without pneumonia (“comparison patients”), before age 2 years. Study outcomes included any chronic respiratory disorder, reactive airway disease (asthma, hyperactive airway disease, recurrent wheezing), and CAP hospitalization occurring between age 2–5 years, and were evaluated overall as well as by age and etiology at first CAP hospitalization. ResultsStudy population totaled 1,343 CAP patients and 6,715 comparison patients. Rates per 100 patient-years and relative rates (RR) of study outcomes from age 2–5 years for CAP patients versus comparison patients were: any chronic respiratory disorder, 11.6 vs. 4.9 (RR = 2.4 [95% CI: 2.1–2.6]); reactive airway disease, 6.1 vs 1.9 (RR = 3.2 [2.6–3.8]); and CAP hospitalization, 1.0 vs 0.2 (RR = 6.3 [3.6–10.9]). Rates of study outcomes were highest among CAP patients who had their initial hospitalization in the second year of life. ConclusionsInfant CAP foreshadows an increased risk of subsequent chronic respiratory disorders, which may be elevated when CAP occurs closer to pre-school age (i.e., age 2–5 years). These findings are most consistent with the hypothesis that inflammation persists beyond the acute stage of pneumonia and plays a role in the development of chronic respiratory sequelae.

A novel method for causal structure discovery from EHR data and its application to type-2 diabetes mellitus

Modern AI-based clinical decision support models owe their success in part to the very large number of predictors they use. Safe and robust decision support, especially for intervention planning, requires causal, not associative, relationships. Traditional methods of causal discovery, clinical trials and extracting biochemical pathways, are resource intensive and may not scale up to the number and complexity of relationships sufficient for precision treatment planning. Computational causal structure discovery (CSD) from electronic health records (EHR) data can represent a solution, however, current CSD methods fall short on EHR data. This paper presents a CSD method tailored to the EHR data. The application of the proposed methodology was demonstrated on type-2 diabetes mellitus. A large EHR dataset from Mayo Clinic was used as development cohort, and another large dataset from an independent health system, M Health Fairview, as external validation cohort. The proposed method achieved very high recall (.95) and substantially higher precision than the general-purpose methods (.84 versus .29, and .55). The causal relationships extracted from the development and external validation cohorts had a high (81%) overlap. Due to the adaptations to EHR data, the proposed method is more suitable for use in clinical decision support than the general-purpose methods.

Contrastive learning improves critical event prediction in COVID-19 patients

SummaryDeep learning (DL) models typically require large-scale, balanced training data to be robust, generalizable, and effective in the context of healthcare. This has been a major issue for developing DL models for the coronavirus disease 2019 (COVID-19) pandemic, where data are highly class imbalanced. Conventional approaches in DL use cross-entropy loss (CEL), which often suffers from poor margin classification. We show that contrastive loss (CL) improves the performance of CEL, especially in imbalanced electronic health records (EHR) data for COVID-19 analyses. We use a diverse EHR dataset to predict three outcomes: mortality, intubation, and intensive care unit (ICU) transfer in hospitalized COVID-19 patients over multiple time windows. To compare the performance of CEL and CL, models are tested on the full dataset and a restricted dataset. CL models consistently outperform CEL models, with differences ranging from 0.04 to 0.15 for area under the precision and recall curve (AUPRC) and 0.05 to 0.1 for area under the receiver-operating characteristic curve (AUROC).

Discovering disease\u2013disease associations using electronic health records in The Guideline Advantage (TGA) dataset

Certain diseases have strong comorbidity and co-occurrence with others. Understanding disease–disease associations can potentially increase awareness among healthcare providers of co-occurring conditions and facilitate earlier diagnosis, prevention and treatment of patients. In this study, we utilized the valuable and large The Guideline Advantage (TGA) longitudinal electronic health record dataset from 70 outpatient clinics across the United States to investigate potential disease–disease associations. Specifically, the most prevalent 50 disease diagnoses were manually identified from 165,732 unique patients. To investigate the co-occurrence or dependency associations among the 50 diseases, the categorical disease terms were first mapped into numerical vectors based on disease co-occurrence frequency in individual patients using the Word2Vec approach. Then the novel and interesting disease association clusters were identified using correlation and clustering analyses in the numerical space. Moreover, the distribution of time delay (Δt) between pair-wise strongly associated diseases (correlation coefficients ≥ 0.5) were calculated to show the dependency among the diseases. The results can indicate the risk of disease comorbidity and complications, and facilitate disease prevention and optimal treatment decision-making.

Risks of Multiple Skin Cancers in Organ Transplant Recipients

There are limited reports on the risks of multiple primary skin cancers in organ transplant recipients (OTRs). To determine the risks over time and risk factors for OTRs developing (1) any skin cancer posttransplant, (2) a subsequent skin cancer after the first posttransplant skin cancer in the data sets used in the study, and (3) 10 or more skin cancers. This retrospective cohort study used data from Optum deidentified electronic health record data set (7.7 million patients) and Truven Health MarketScan insurance claims data set (161 million patients) from 2007 to 2017. Skin cancers were identified using diagnosis plus treatment codes for basal cell carcinoma, squamous cell carcinoma, and melanoma; OTRs were identified using 4 or more diagnosis codes for organ transplant. Data analysis took place from January 1, 2007, to December 31, 2017. Cumulative risks of (1) any skin cancer treatment posttransplant, (2) a subsequent skin cancer treatment after the first posttransplant skin cancer treatment in our data, and (3) 10 or more skin cancer treatments in OTRs. A Wei-Lin-Weissfeld marginal model was used to evaluate risk factors for any skin cancer. A total of 7390 OTRs in Optum and 133 651 in MarketScan were identified, 4.5% and 13.3% of which had had at least 1 skin cancer treatment, respectively. At 2 years after the initial posttransplant skin cancer in the data sets, OTRs had a 44.0% to 57.0% risk of a subsequent skin cancer treatment and a 3.7% to 6.6% risk of having 10 or more skin cancer treatments. Statistically significant risk factors for any skin cancer included age, history of skin cancer, and history of actinic keratosis in both data sets, and male sex and thoracic transplant in MarketScan. In this retrospective cohort study, approximately half of the OTRs who developed at least 1 posttransplant skin cancer developed a subsequent skin cancer within 2 years, and approximately 1 in 20 developed 10 or more skin cancers. Identifying OTRs at highest risk for multiple primary skin cancers may help target strategies for prevention and early detection.

Dynamic optimization of CRT pacing using a novel algorithm improves survival in a real-world cohort of patients

Abstract Background The AdaptivCRT (aCRT) algorithm continuously adjusts cardiac resynchronization therapy (CRT) according to intrinsic atrioventricular (AV) conduction status, providing synchronized left-ventricular pacing in patients (pts) with a normal PR interval and adaptive biventricular (BiV) pacing in pts with a prolonged PR interval. Prior secondary analyses of prospective trials demonstrated a reduction in mortality and atrial fibrillation incidence with aCRT. Objective To compare mortality in a large real-world population of pts with and without aCRT using electronic health record (EHR) data. Methods An Optum® de-identified EHR dataset and Medtronic CareLink device data were linked to identify pts (age&amp;gt;18 years) with a CRT implanted between 2014–2018 who survived ≥35 days post implant and with a documented baseline left ventricular ejection fraction (LVEF). Pts were grouped with aCRT set to either Adaptive BiV+LV (aCRT group) or standard CRT (control group). Two pt subgroups were identified: (i) pts with left bundle branch block (LBBB) and no AV block, (ii) pts with either no LBBB or with AV block. Kaplan-Meier curves and log-rank were used to estimate and compare survival between treatment and control groups in both the full cohort and subgroups. Multivariate Cox regression was used to estimate the aCRT treatment effect while controlling for baseline characteristics. Results Among 8624 patients implanted with a CRT device with EHR data available for a median of 23.5 months post implant [IQR: 12.6, 37.3], 3994 patients had the aCRT feature activated. Pts in the aCRT group were younger (median 69 [IQR:60, 76] vs 73 [IQR: 65, 79] years), more frequently female (35% vs 26%), and more frequently had LBBB (72% vs 51%); p's&amp;lt;0.001. Median QRS duration was 150ms (IQR: 133, 164) for the aCRT group and 152ms (IQR: 134, 172) for the control group (p&amp;lt;0.001). Median LVEF was 22% (IQR: 18, 30) for the aCRT group and 25% (IQR: 20, 30) for the control group (p&amp;lt;0.001). Through 60 months post implant, the overall mortality rate was 23.6% (95% CI: 20.7%-26.3%) for the aCRT group and 37.6% for the control group (95% CI: 34.9%-40.1%), p&amp;lt;0.001. Subgroup analysis demonstrated a significant reduction in mortality with aCRT in both pts with LBBB without AV block (Figure, left panel) and pts without LBBB and with AV block (Figure, right panel). After accounting for baseline differences, the reduction in mortality remained significant in the LBBB/no AV block group (HR: 0.54, 95% CI: 0.39–0.76, p&amp;lt;0.001) and trended toward significance in the no LBBB/with AV block group (HR: 0.89, 95% CI: 0.79–1.01; p=0.08). Conclusion In a large, real-world population of pts implanted with CRT devices, the AdaptivCRT algorithm was associated with significantly reduced mortality, with a greater reduction observed among pts with LBBB and without AV block. These results are being tested prospectively in the ongoing AdaptResponse multi-national, randomized clinical trial. Funding Acknowledgement Type of funding sources: Private company. Main funding source(s): Medtronic, Inc. Cumulative probability of death by group

Electronic health records features associated to coronary artery disease in patients with chest pain

Abstract Introduction Cardiovascular diseases remain the main cause of death in most European countries and coronary artery disease (CAD) follows an ascendent tendency. Current guidelines on chronic coronary syndromes recommend to estimate the CAD pre-test probability based on age, sex and symptoms in patients with suspected CAD. Besides, they underline the importance of adapting this likelihood with the patient information, although the risk the physician attributes to the medical history might not be concordant among colleagues. Electronic health records (EHR) are the main patient-level data collection tools in clinical care and data capture systems compile this clinical information into research-oriented datasets. We can analyze which variables in EHR have stronger associations with CAD in this context to facilitate a more concise and personalized adjustment of the pre-test CAD risk. Purpose To analyze the association between features in EHR with the development of CAD in patients who attended the local public health system for thoracic pain in the 2016–2018 period. Methods We queried the system for patients who required medical attendance for thoracic pain or equivalent symptoms from 2016 to 2018 in the public health system. We extracted data within this subset of the potentially predictive features according to a literature review. We investigated whether patients developed CAD in the year following their attendance. A database cleaning process removed subjects without further diagnostic tests. We used parametric tests to assess the distribution differences of these features between patients who developed CAD and those who did not and we applied logistic regression to document their prediction potential. Results The EHR dataset extracted from the system contained information about 2,199,711 individuals, out of which 43,835 required medical attention for thoracic pain in the selected period. 10,463 patients had no further diagnostic tests, so were excluded. Among the remaining 33,372 patients, 5,379 (16.1%) developed CAD in the follow-up period. Heart failure and traditional cardiovascular risk factors were predominant in the CAD group, and so were related blood test features and therapies. Any history of CAD diagnostic test was significantly associated to CAD presentation. Ischemic etiology of the chest pain was more unlikely in patients with anxiety history and benzodiazepine prescription. Proton pump inhibitors use and hearth rhythm disorders did not increase CAD risk in adjusted models. Conclusions Medical history, basal blood tests and treatment information available in EHR are related to CAD risk in patients with chest pain. Considering this information seems valuable for a personalized risk adjustment of the CAD pre-test probability in symptomatic patients. Further studies with multivariable models and machine learning techniques applied to this dataset might entail more specific predictive algorithms. Funding Acknowledgement Type of funding sources: None.

An interpretable outcome prediction model based on electronic health records and hierarchical attention

Outcome prediction aims to predict the future health condition of patients from Electronic Health Record (EHR) data. Because of the sequential characteristic of EHR data, recurrent neural network (RNN)-based outcome prediction methods have achieved state-of-the-art results. However, the major drawback of RNN-based outcome prediction methods is lack of interpretability, which would lead to trust issues. Aiming at this problem, this paper proposes interpretable outcome prediction model with hierarchical attention (IoHAN), an interpretable outcome prediction model by leveraging attention mechanism. The main novelty of IoHAN is that it can pinpoint the fine-grained influence on the final prediction result of each medical component by decomposing the attention weights hierarchically into hospital visits, medical variables, and interactions between medical variables. We evaluated IoHAN on MIMIC-III, a large real-world EHR data set. The experiment results demonstrate that IoHAN can achieve higher prediction accuracy than state-of-the-art outcome prediction models. In addition, the hierarchical decomposed attention weights can interpret the prediction results in a more natural and understandable way.

A Real-Time Data Clustering Scheme Using K-Medoids Based Optimal Neural Network Approach for Integrating Demographics and Diagnosis Codes

Data clustering from the Electronic Health Record (EHR) system is an important system for diagnosing the association between the preprocessing phase and the clinical profiles of patients. According to prior studies, data heterogeneity, high costs associated with medical equipment, privacy and utility preservation, computational complexity, and other aspects make developing new clustering algorithms difficult. In this study, we introduce K-Medoids-based Optimal Neural Network (Km-ONN) to cluster the real-time (RT) dataset. Both diagnosis codes and demographic attributes of patient’s values are present in an RT dataset. Based on diagnosis codes and demographic codes, the identity disclosure is prevented from attackers by applying (K, Kn ) in which the information loss is minimized. The data owner specifies the utility constraint that restricts the number of generalizations. In this analysis, we used two EHR datasets, INFORMS and VERMONT, on which the proposed model performed better. With a number of 1000 medoids, the proposed method offers an accuracy, recall, Mathews correlation coefficient(MCC), and F-score of 0.99, 0.998,0.999, and 1 during training. The proposed model, on the other hand, is well suited for clustering RT datasets, resulting in more scalable and efficient outputs than other existing models.

Predicting hypoglycemia in critically Ill patients using machine learning and electronic health records.

Hypoglycemia is a common occurrence in critically ill patients and is associated with significant mortality and morbidity. We developed a machine learning model to predict hypoglycemia by using a multicenter intensive care unit (ICU) electronic health record dataset. Machine learning algorithms were trained and tested on patient data from the publicly available eICU Collaborative Research Database. Forty-four features including patient demographics, laboratory test results, medications, and vitals sign recordings were considered. The outcome of interest was the occurrence of a hypoglycemic event (blood glucose < 72mg/dL) during a patient's ICU stay. Machine learning models used data prior to the second hour of the ICU stay to predict hypoglycemic outcome. Data from 61,575 patients who underwent 82,479 admissions at 199 hospitals were considered in the study. The best-performing predictive model was the eXtreme gradient boosting model (XGBoost), which achieved an area under the received operating curve (AUROC) of 0.85, a sensitivity of 0.76, and a specificity of 0.76. The machine learning model developed has strong discrimination and calibration for the prediction of hypoglycemia in ICU patients. Prospective trials of these models are required to evaluate their clinical utility in averting hypoglycemia within critically ill patient populations.

A Poisson binomial-based statistical testing framework for comorbidity discovery across electronic health record datasets.

Discovering the concomitant occurrence of distinct medical conditions in a patient, also known as comorbidities, is a prerequisite for creating patient outcome prediction tools. Current comorbidity discovery applications are designed for small datasets and use stratification to control for confounding variables such as age, sex or ancestry. Stratification lowers false positive rates, but reduces power, as the size of the study cohort is decreased. Here we describe a Poisson binomial-based approach to comorbidity discovery (PBC) designed for big-data applications that circumvents the need for stratification. PBC adjusts for confounding demographic variables on a per-patient basis and models temporal relationships. We benchmark PBC using two datasets to compute comorbidity statistics on 4,623,841 pairs of potentially comorbid medical terms. The results of this computation are provided as a searchable web resource. Compared with current methods, the PBC approach reduces false positive associations while retaining statistical power to discover true comorbidities.

S855 Characteristics, Comorbidities, and Outcomes of SARS-CoV-2 Infection in Patients With Conditions Treated With Systemic Immunosuppressive Therapies: A Population-Based Study

Introduction: We describe the characteristics and COVID-19 clinical outcomes in this pt population, compared with the general population. Methods: This descriptive retrospective cohort study used data from the US Optum® deidentified COVID-19 electronic health record dataset (2007–2020). Adults with COVID-19 were stratified into 3 disease cohorts (pts with rheumatoid arthritis [RA], psoriatic arthritis [PsA], or UC who had received systemic immunosuppressive therapy) and a comparator cohort that did not meet these criteria. Incidence proportions of hospitalization, intensive care unit admission, in-hospital death, all-cause mortality, and clinical manifestations of interest were calculated. Logistic regression was used in exploratory analyses to estimate the risk of endpoints, adjusting for demographics (age, sex, race, region), and demographics plus comorbidities. Results: This analysis (Feb 01–Dec 09, 2020) included 315,101 COVID-19 pts. The incidence of hospitalization due to COVID-19 was greatest in the RA cohort compared with the PsA, UC, and comparator cohorts (Table). Adjusting for demographics, and demographics plus comorbidities, COVID-19 pts with RA had an increased risk of hospitalization and in-hospital death, compared with the comparator cohort (Figure). Adjusting for demographics, pts with UC had an increased risk of hospitalization; after adjustment for demographics plus comorbidities, the risk was no longer statistically significant (Figure). There was no statistically significant increased risk of hospitalization among pts with PsA (Figure). The risk of hospitalization was lower in COVID-19 pts with RA receiving tumor necrosis factor inhibitors (TNFi) vs non-TNFi biologics (0.32 [0.20, 0.53]) and vs the comparator cohort (0.77 [0.51, 1.17]). Compared with the comparator cohort, the risk of hospitalization due to COVID-19 was similar between pts with RA receiving tofacitinib (1.06 [0.55, 2.05]), JAK inhibitors (1.26 [0.74, 2.16]) or csDMARDs (1.21 [1.02, 1.44]). The risk of hospitalization due to COVID-19 was greater in pts with RA receiving non-TNFi biologics vs the comparator cohort (2.33 [1.79, 3.04]). A similar analysis was not performed for the UC and PsA cohorts due to small sample sizes. Conclusion: Compared with the comparator cohort, pts with RA were at a higher risk of more severe COVID-19, and except for non-TNFi biologics, immunosuppressive therapies in general did not further increase the risk. COVID-19 outcomes in pts with UC were generally similar to the comparator cohort.Figure 1.: Adjusted odds ratios of each endpoint in the (A) RA cohort, (B) UC cohort, and (C) PsA cohort vs the comparator cohort. a.Covariates included in logistic regression models were age, sex, race, and region; b.Covariates included in logistic regression models were age, sex, race, region, lung disease (interstitial lung disease, asthma, and chronic obstructive pulmonary disease), VTE, hypertension, coronary artery disease, serious infections (hospitalized), cancer, diabetes, CKD or dialysis, and liver disease, except for UC, which did not include serious infections (hospitalized). Odds ratios using this model are only displayed for endpoints for which there were ≥ 30 events; c.Serious infections (hospitalized) was not included as a covariate; d.CKD/dialysis was not included as a covariate; e. VTE was not included as a covariate. ARDS, acute respiratory distress syndrome; CI, confidence interval; CKD, chronic kidney disease; ECMO, extracorporeal membrane oxygenation; ICU, intensive care unit; PsA, psoriatic arthritis; RA, rheumatoid arthritis; UC, ulcerative colitis.Table 1.: Baseline Characteristics and IPs of Clinical Manifestations/Outcomes in the RA, PsA, UC, and Comparator Cohorts. a.At least 1 diagnosis of RA, PsA, or UC, and treatment with IM therapy (JAK inhibitor, TNFi, non-TNFi biologic, or csDMARD) within 2 yrs prior to SARS-CoV-2 diagnosis date; b.Absence of a diagnosis of RA, PsA, or UC, and no treatment with IM therapy within 2 yrs prior to SARS-CoV-2 diagnosis date; c.Baseline was defined as within 6 months prior to SARS-CoV-2 diagnosis date; d.csDMARDs included auranofin, aurothioglucose, azathioprine, chloroquine hydrochloride, chloroquine phosphate, cyclophosphamide, cyclosporine, gold sodium thiomalate, hydroxychloroquine sulfate, leflunomide, mercaptopurine, mesalamine, methotrexate, minocycline hydrochloride, n-acetylpenicillamine, penicillamine, primaquine, sulfasalazine, tacrolimus, and thalidomide; e.Within 90 days prior to SARS-CoV-2 diagnosis date; f.Denominator is the number of pts hospitalized within 30 days of SARS-CoV-2 diagnosis; defined as death during hospitalization; g.Within 90 days of SARS-CoV-2 diagnosis date; h.Within 30 days of SARS-CoV-2 diagnosis date. ARDS, acute respiratory distress syndrome; CI, confidence interval; csDMARD, conventional synthetic disease-modifying antirheumatic drug; ECMO, extracorporeal membrane oxygenation; ICU, intensive care unit; IM, immunomodulatory; IP, incidence proportion; IV, intravenous; JAK, Janus kinase; N, number of pts in the cohort; n, number of pts in the specified category; PsA, psoriatic arthritis; pts, patients; RA, rheumatoid arthritis; TNFi, tumor necrosis factor inhibitor; UC, ulcerative colitis.

Real World Long-term Assessment of The Efficacy of Tocilizumab in Patients with COVID-19: Results From A Large De-identified Multicenter Electronic Health Record Dataset in the United States

BackgroundStudies have shown conflicting results on the efficacy of tocilizumab (TCZ) for patients with COVID-19, with many confounders of clinical status and limited duration of the observation. Here, we evaluate the real-world long-term efficacy of TCZ in COVID-19 patients. MethodsWe conducted a retrospective study of hospitalized adult patients with COVID-19 using a large US-based multicenter COVID-19 database (Cerner Real-World Data; updated in September, 2020). The TCZ group was defined as patients who received at least one dose of the drug. Matching weight (MW) and a propensity score weighting method were used to balance confounding factors. ResultsA total of 20,399 patients were identified. 1,510 and 18,899 were in the TCZ and control groups, respectively. After MW adjustment, no statistically significant differences in all-cause mortality were found for the TCZ vs. control group (Hazard Ratio [HR]:0.76, p=0.06). Survival curves suggested a better trend in short-term observation, driven from a subgroup of patients requiring oxygen masks, BIPAP or CPAP. ConclusionWe observed a temporal (early) benefit of TCZ, especially in patients on non-invasive high-flow supplemental oxygen. However, the benefit effects faded with longer observation. The long-term benefits and risks of TCZ should be carefully evaluated with follow-up studies.

COVID-19 Mortality Prediction From Deep Learning in a Large Multistate Electronic Health Record and Laboratory Information System Data Set: Algorithm Development and Validation.

BackgroundCOVID-19 is caused by the SARS-CoV-2 virus and has strikingly heterogeneous clinical manifestations, with most individuals contracting mild disease but a substantial minority experiencing fulminant cardiopulmonary symptoms or death. The clinical covariates and the laboratory tests performed on a patient provide robust statistics to guide clinical treatment. Deep learning approaches on a data set of this nature enable patient stratification and provide methods to guide clinical treatment.ObjectiveHere, we report on the development and prospective validation of a state-of-the-art machine learning model to provide mortality prediction shortly after confirmation of SARS-CoV-2 infection in the Mayo Clinic patient population.MethodsWe retrospectively constructed one of the largest reported and most geographically diverse laboratory information system and electronic health record of COVID-19 data sets in the published literature, which included 11,807 patients residing in 41 states of the United States of America and treated at medical sites across 5 states in 3 time zones. Traditional machine learning models were evaluated independently as well as in a stacked learner approach by using AutoGluon, and various recurrent neural network architectures were considered. The traditional machine learning models were implemented using the AutoGluon-Tabular framework, whereas the recurrent neural networks utilized the TensorFlow Keras framework. We trained these models to operate solely using routine laboratory measurements and clinical covariates available within 72 hours of a patient’s first positive COVID-19 nucleic acid test result.ResultsThe GRU-D recurrent neural network achieved peak cross-validation performance with 0.938 (SE 0.004) as the area under the receiver operating characteristic (AUROC) curve. This model retained strong performance by reducing the follow-up time to 12 hours (0.916 [SE 0.005] AUROC), and the leave-one-out feature importance analysis indicated that the most independently valuable features were age, Charlson comorbidity index, minimum oxygen saturation, fibrinogen level, and serum iron level. In the prospective testing cohort, this model provided an AUROC of 0.901 and a statistically significant difference in survival (P<.001, hazard ratio for those predicted to survive, 95% CI 0.043-0.106).ConclusionsOur deep learning approach using GRU-D provides an alert system to flag mortality for COVID-19–positive patients by using clinical covariates and laboratory values within a 72-hour window after the first positive nucleic acid test result.