Clinical Electronic Health Record Data Research Articles

Preeclampsia Prediction Using Machine Learning and Polygenic Risk Scores From Clinical and Genetic Risk Factors in Early and Late Pregnancies.

Preeclampsia, a pregnancy-specific condition associated with new-onset hypertension after 20-weeks gestation, is a leading cause of maternal and neonatal morbidity and mortality. Predictive tools to understand which individuals are most at risk are needed. We identified a cohort of N=1125 pregnant individuals who delivered between May 2015 and May 2022 at Mass General Brigham Hospitals with available electronic health record data and linked genetic data. Using clinical electronic health record data and systolic blood pressure polygenic risk scores derived from a large genome-wide association study, we developed machine learning (XGBoost) and logistic regression models to predict preeclampsia risk. Pregnant individuals with a systolic blood pressure polygenic risk score in the top quartile had higher blood pressures throughout pregnancy compared with patients within the lowest quartile systolic blood pressure polygenic risk score. In the first trimester, the most predictive model was XGBoost, with an area under the curve of 0.74. In late pregnancy, with data obtained up to the delivery admission, the best-performing model was XGBoost using clinical variables, which achieved an area under the curve of 0.91. Adding the systolic blood pressure polygenic risk score to the models did not improve the performance significantly based on De Long test comparing the area under the curve of models with and without the polygenic score. Integrating clinical factors into predictive models can inform personalized preeclampsia risk and achieve higher predictive power than the current practice. In the future, personalized tools can be implemented to identify high-risk patients for preventative therapies and timely intervention to improve adverse maternal and neonatal outcomes.

Clinical outcome prediction using observational supervision with electronic health records and audit logs

Objective:Audit logs in electronic health record (EHR) systems capture interactions of providers with clinical data. We determine if machine learning (ML) models trained using audit logs in conjunction with clinical data (“observational supervision”) outperform ML models trained using clinical data alone in clinical outcome prediction tasks, and whether they are more robust to temporal distribution shifts in the data. Materials and Methods:Using clinical and audit log data from Stanford Healthcare, we trained and evaluated various ML models including logistic regression, support vector machine (SVM) classifiers, neural networks, random forests, and gradient boosted machines (GBMs) on clinical EHR data, with and without audit logs for two clinical outcome prediction tasks: major adverse kidney events within 120 days of ICU admission (MAKE-120) in acute kidney injury (AKI) patients and 30-day readmission in acute stroke patients. We further tested the best performing models using patient data acquired during different time-intervals to evaluate the impact of temporal distribution shifts on model performance. Results:Performance generally improved for all models when trained with clinical EHR data and audit log data compared with those trained with only clinical EHR data, with GBMs tending to have the overall best performance. GBMs trained with clinical EHR data and audit logs outperformed GBMs trained without audit logs in both clinical outcome prediction tasks: AUROC 0.88 (95% CI: 0.85–0.91) vs. 0.79 (95% CI: 0.77–0.81), respectively, for MAKE-120 prediction in AKI patients, and AUROC 0.74 (95% CI: 0.71–0.77) vs. 0.63 (95% CI: 0.62–0.64), respectively, for 30-day readmission prediction in acute stroke patients. The performance of GBM models trained using audit log and clinical data degraded less in later time-intervals than models trained using only clinical data. Conclusion:Observational supervision with audit logs improved the performance of ML models trained to predict important clinical outcomes in patients with AKI and acute stroke, and improved robustness to temporal distribution shifts.

Retrospective study of propionic acidemia using natural language processing in Mayo Clinic electronic health record data

BackgroundPropionic acidemia (PA) is a rare autosomal recessive organic acidemia that classically presents within the first days of life with a metabolic crisis or via newborn screening and is confirmed with laboratory tests. Limited data exist on the natural history of patients with PA describing presentation, treatments, and clinical outcomes. ObjectiveTo retrospectively describe the natural history of patients with PA in a clinical setting from a real-world database using both structured and unstructured electronic health record (EHR) data using novel data extraction techniques in a unique care setting. Design/MethodsThis retrospective study used EHR data to identify patients with PA seen at the Mayo Clinic. Unstructured clinical text (medical notes, pathology reports) were analyzed using augmented curation natural language processing models to enhance analysis of data extracted by structured data fields (International Classification of Diseases 9th or 10th revision [ICD-9/−10] codes, Current Procedural Terminology [CPT] codes, and medication orders). De-identified health records were also manually reviewed by clinical scientists to ensure data accuracy and completeness. The index date was defined as the patient's date of PA diagnosis at the Mayo Clinic. Results were reported as aggregate descriptive statistics relative to patients' index dates. Complications, therapeutic interventions, laboratory tests, procedures, and hospitalization encounters related to PA were described at and within 6 months of the patient's index date, and from medical history available before the index date. ResultsIn total, 13 patients with PA were identified, with visits occurring from 1998 to 2022. Age at diagnosis ranged from birth to 3 years; age at initial evaluation at the Mayo Clinic ranged from 3 days to 28 years. The mean number of Mayo Clinic outpatient visits was 31 (median duration of care, 2 years). PA-related complications were documented in 85% of patients and included nutritional difficulties (46%), metabolic decompensation events (MDEs; 38%), neurologic abnormalities (38%), and cardiomyopathy (7%). One pair of affected siblings had mild symptoms and no complications or MDEs. All 5 patients with a history of MDEs presented with developmental delays. Among patients with MDEs, the mean frequency of outpatient clinical care visits was 10 per year, and 3 patients required inpatient hospitalization (mean duration, 16 days). The incidence of severe complications was higher among patients with MDEs than those without MDEs. Of the patients with MDEs, 2 experienced crises while receiving treatment at the Mayo Clinic, with 9 total MDEs occurring between the 2 patients. Symptoms at presentation included hyperammonemia (78%), fever and/or decreased nutritional intake (67%), hyperglycemia/hypoglycemia (56%), intercurrent upper respiratory infection and/or lethargy (44%), constipation (33%), altered mental status (33%), and cough (33%). ConclusionsThis study highlights the range and frequency of clinical outcomes experienced by patients with PA and demonstrates the clinical burden of MDEs.

Machine learning functional impairment classification with electronic health record data.

Poor functional status is a key marker of morbidity, yet is not routinely captured in clinical encounters. We developed and evaluated the accuracy of a machine learning algorithm that leveraged electronic health record (EHR) data to provide a scalable process for identification of functional impairment. We identified a cohort of patients with an electronically captured screening measure of functional status (Older Americans Resources and Services ADL/IADL) between 2018 and 2020 (N = 6484). Patients were classified using unsupervised learning K means and t-distributed Stochastic Neighbor Embedding into normal function (NF), mild to moderate functional impairment (MFI), and severe functional impairment (SFI) states. Using 11 EHR clinical variable domains (832 variable input features), we trained an Extreme Gradient Boosting supervised machine learning algorithm to distinguish functional status states, and measured prediction accuracies. Data were randomly split into training (80%) and test (20%) sets. The SHapley Additive Explanations (SHAP) feature importance analysis was used to list the EHR features in rank order of their contribution to the outcome. Median age was 75.3 years, 62% female, 60% White. Patients were classified as 53% NF (n = 3453), 30% MFI (n = 1947), and 17% SFI (n = 1084). Summary of model performance for identifying functional status state (NF, MFI, SFI) was AUROC (area under the receiving operating characteristic curve) 0.92, 0.89, and 0.87, respectively. Age, falls, hospitalization, home health use, labs (e.g., albumin), comorbidities (e.g., dementia, heart failure, chronic kidney disease, chronic pain), and social determinants of health (e.g., alcohol use) were highly ranked features in predicting functional status states. A machine learning algorithm run on EHR clinical data has potential utility for differentiating functional status in the clinical setting. Through further validation and refinement, such algorithms can complement traditional screening methods and result in a population-based strategy for identifying patients with poor functional status who need additional health resources.

434. Healthcare Personnel Interactive Pathogen Exposure Response System

Abstract Background Infectious disease exposure investigations in the hospital are labor-intensive for the infection prevention and control (IPC) team and vulnerable to healthcare personnel (HCP) recall bias. We developed an electronic health record (EHR) algorithm to identify and rank patient-HCP interactions based on the likelihood and intensity of exposure. Methods We compared conventional and EHR-based findings from seven exposure investigations conducted between November 1, 2020 and February 1, 2022 at The Johns Hopkins Hospital (JHH), a 1095 bed academic tertiary center in Baltimore, MD. Conventional exposure investigations were conducted for hospitalized patients who tested positive for SARS-CoV-2 while not in COVID-19 isolation precautions. IPC contacted department managers to identify and report potentially exposed HCPs to occupational health. The EHR-based method identified HCP-patient interactions based on clinical data such as documentation in the flowsheet, medication administration, etc. A score was calculated for each HCP based on the estimated duration and intensity of the contact. Genomic sequencing of available samples was performed to investigate transmission events. Results Overall, the EHR-based system identified 75% (59/79) of the HCPs identified by conventional exposure investigations and 100% of those who document in the EHR. In contrast, it was unable to identify any potentially exposed individuals who do not document in the EHR (Table 1). All patient-HCP COVID-19 transmissions identified by conventional investigation and confirmed through genomic sequencing were identified by the EHR-based system, and all had high-intensity scores (i.e., top quartile of the list of exposed individuals). Table 1.Comparison of Conventional and EHR Based COVID-19 Exposure Investigations Conclusion We found clinical EHR data was highly sensitive and specific in identifying potentially exposed HCPs compared with conventional exposure investigations. The inability to detect interactions with support staff or others who do not document in the EHR was a limitation and suggests that EHR data can augment but not replace conventional exposure investigations. The system’s speed, ease, and lower resource requirements make it a promising tool to more efficiently complete exposure investigations in healthcare settings. Disclosures Sara E. Cosgrove, MD, Basilea: Member of Infection Adjudication Committee.

Death and End-of-Life Care in Emergency Departments in the US

There are more than 140 million annual visits to emergency departments (EDs) in the US. The role of EDs in providing care at or near the end of life is not well characterized. To determine the frequency of death in the ED or within 1 month of an ED visit in an all-age, all-payer national database. The retrospective cohort study used patient-level data from the nationally representative Optum clinical electronic health record data set for 2010 to 2020. Data were analyzed from January to March 2022. Age, Charlson Comorbidity Index (CCI), and year of ED encounter. The primary outcome was death in the ED, overall and stratified by age, CCI, or year. A key secondary outcome was death within 1 month of an ED encounter. We extrapolated to make national estimates using US Census and Nationwide Emergency Department Sample data. Among a total of 104 113 518 individual patients with 96 239 939 ED encounters, 205 372 ED deaths were identified in Optum, for whom median (IQR) age was 72 (53 to >80) years, 114 582 (55.8%) were male, and 152 672 (74.3%) were White. ED death affected 0.20% of overall patients and accounted for 0.21% of ED encounters. An additional 603 273 patients died within 1 month of an ED encounter. Extrapolated nationally, ED deaths accounted for 11.3% of total deaths from 2010 to 2019, and 33.2% of all decedents nationally visited the ED within 1 month of their death. The proportion of total national deaths occurring in the ED decreased by 0.27% annually (P for trend = .003) but the proportion who died within 1 month of an ED visit increased by 1.2% annually (P for trend < .001). Compared with all ED encounters, patients with visits resulting in death were older, more likely to be White, male, and not Hispanic, and had higher CCI. Among ED encounters for patients aged older than 80 years, nearly 1 in 12 died within 1 month. This retrospective cohort study found deaths during or shortly after ED care were common, especially among patients who are older and with chronic comorbidities. EDs must identify patients for whom end-of-life care is necessary or preferred and be equipped to deliver this care excellently.

Comparing Machine Learning to Regression Methods for Mortality Prediction Using Veterans Affairs Electronic Health Record Clinical Data.

It is unclear whether machine learning methods yield more accurate electronic health record (EHR) prediction models compared with traditional regression methods. The objective of this study was to compare machine learning and traditional regression models for 10-year mortality prediction using EHR data. This was a cohort study. Veterans Affairs (VA) EHR data. Veterans age above 50 with a primary care visit in 2005, divided into separate training and testing cohorts (n= 124,360 each). The primary outcome was 10-year all-cause mortality. We considered 924 potential predictors across a wide range of EHR data elements including demographics (3), vital signs (9), medication classes (399), disease diagnoses (293), laboratory results (71), and health care utilization (149). We compared discrimination (c-statistics), calibration metrics, and diagnostic test characteristics (sensitivity, specificity, and positive and negative predictive values) of machine learning and regression models. Our cohort mean age (SD) was 68.2 (10.5), 93.9% were male; 39.4% died within 10 years. Models yielded testing cohort c-statistics between 0.827 and 0.837. Utilizing all 924 predictors, the Gradient Boosting model yielded the highest c-statistic [0.837, 95% confidence interval (CI): 0.835-0.839]. The full (unselected) logistic regression model had the highest c-statistic of regression models (0.833, 95% CI: 0.830-0.835) but showed evidence of overfitting. The discrimination of the stepwise selection logistic model (101 predictors) was similar (0.832, 95% CI: 0.830-0.834) with minimal overfitting. All models were well-calibrated and had similar diagnostic test characteristics. Our results should be confirmed in non-VA EHRs. The differences in c-statistic between the best machine learning model (924-predictor Gradient Boosting) and 101-predictor stepwise logistic models for 10-year mortality prediction were modest, suggesting stepwise regression methods continue to be a reasonable method for VA EHR mortality prediction model development.

Patient Race, Ethnicity, Language, and Pain Severity in Primary Care: A Retrospective Electronic Health Record Study

BackgroundPatient race, ethnicity, and culture including language are intertwined and may influence patient reporting of pain severity. PurposeTo describe documentation of patient's self-reported pain presence and severity by race, ethnicity, and language, specifically, Spanish, Hmong, Lao, or Khmer requiring an interpreter or English. Design and SampleRetrospective, electronic health record clinical data mining study of 79,195 patient visits with documented pain scores from one primary care clinic. MethodsHurdle regression was used to explore the effect of race, ethnicity, and language on the chances of having any pain (vs. no pain) and pain severity for visits with pain scores ≥1, controlling for age, sex, and documentation of a pain diagnosis. Mann-Whitney tests were used to explore the influence of English vs. non-English language on pain severity within a race or ethnicity category. ResultsPain scores were higher for limited English proficiency, compared with English-speaking, patients within the Asian race or Hispanic/Latino ethnicity category. Older age, female sex, pain diagnosis, Black or African American race, and Spanish or Lao language increased the chance of having any pain. These same factors, plus American Indian or Alaska Native race, contributed to higher pain severity. Asian race, in contrast, decreased the chance of reporting any pain and contributed to lesser pain severity. ConclusionsRace, in addition to a new area of focus, language, impacted both the chances of reporting any pain and pain severity. Additional research is needed on the impact of language barriers on pain severity reporting, documentation, and differences in pain outcomes and disparities.

Studying patterns and predictors of HIV viral suppression using A Big Data approach: a research protocol

BackgroundGiven the importance of viral suppression in ending the HIV epidemic in the US and elsewhere, an optimal predictive model of viral status can help clinicians identify those at risk of poor viral control and inform clinical improvements in HIV treatment and care. With an increasing availability of electronic health record (EHR) data and social environmental information, there is a unique opportunity to improve our understanding of the dynamic pattern of viral suppression. Using a statewide cohort of people living with HIV (PLWH) in South Carolina (SC), the overall goal of the proposed research is to examine the dynamic patterns of viral suppression, develop optimal predictive models of various viral suppression indicators, and translate the models to a beta version of service-ready tools for clinical decision support.MethodsThe PLWH cohort will be identified through the SC Enhanced HIV/AIDS Reporting System (eHARS). The SC Office of Revenue and Fiscal Affairs (RFA) will extract longitudinal EHR clinical data of all PLWH in SC from multiple health systems, obtain data from other state agencies, and link the patient-level data with county-level data from multiple publicly available data sources. Using the deidentified data, the proposed study will consist of three operational phases: Phase 1: “Pattern Analysis” to identify the longitudinal dynamics of viral suppression using multiple viral load indicators; Phase 2: “Model Development” to determine the critical predictors of multiple viral load indicators through artificial intelligence (AI)-based modeling accounting for multilevel factors; and Phase 3: “Translational Research” to develop a multifactorial clinical decision system based on a risk prediction model to assist with the identification of the risk of viral failure or viral rebound when patients present at clinical visits.DiscussionWith both extensive data integration and data analytics, the proposed research will: (1) improve the understanding of the complex inter-related effects of longitudinal trajectories of HIV viral suppressions and HIV treatment history while taking into consideration multilevel factors; and (2) develop empirical public health approaches to achieve ending the HIV epidemic through translating the risk prediction model to a multifactorial decision system that enables the feasibility of AI-assisted clinical decisions.

Predicting Life Expectancy to Target Cancer Screening Using Electronic Health Record Clinical Data.

Guidelines recommend breast and colorectal cancer screening for older adults with a life expectancy >10 years. Most mortality indexes require clinician data entry, presenting a barrier for routine use in care. Electronic health records (EHR) are a rich clinical data source that could be used to create individualized life expectancy predictions to identify patients for cancer screening without data entry. To develop and internally validate a life expectancy calculator from structured EHR data. Retrospective cohort study using national Veteran's Affairs (VA) EHR databases. Veterans aged 50+ with a primary care visit during 2005. We assessed demographics, diseases, medications, laboratory results, healthcare utilization, and vital signs 1 year prior to the index visit. Mortality follow-up was complete through 2017. Using the development cohort (80% sample), we used LASSO Cox regression to select ~100 predictors from 913 EHR data elements. In the validation cohort (remaining 20% sample), we calculated the integrated area under the curve (iAUC) and evaluated calibration. In 3,705,122 patients, the mean age was 68 years and the majority were male (97%) and white (85%); nearly half (49%) died. The life expectancy calculator included 93 predictors; age and gender most strongly contributed to discrimination; diseases also contributed significantly while vital signs were negligible. The iAUC was 0.816 (95% confidence interval, 0.815, 0.817) with good calibration. We developed a life expectancy calculator using VA EHR data with excellent discrimination and calibration. Automated life expectancy prediction using EHR data may improve guideline-concordant breast and colorectal cancer screening by identifying patients with a life expectancy >10 years.

Trends in the multiple myeloma treatment landscape and survival: a U.S. analysis using 2011–2019 oncology clinic electronic health record data

Multiple myeloma treatment has evolved with approvals of new immunomodulatory imide drugs (IMiDs), monoclonal antibodies (MoABs), and proteasome inhibitors (PIs). We characterized U.S. treatment trends and survival from 2011 to 2019 using Flatiron data from multiple myeloma patients followed from treatment index until death/end of data. Patients (n = 10,553) were primarily (88%) treated in community centers. Frontline PI-IMiD-dexamethasone use increased over time, while IMiD-dexamethasone and PI-dexamethasone use decreased. MoAB-IMiD-dexamethasone use increased in relapsed/refractory disease. In all lines, use of doublets decreased and triplets increased, with triplets becoming the most prescribed combination by 2018–2019, especially in first line (62%). Monotherapy use decreased in first line (19% to 10%) but remained steady in relapsed/refractory disease (∼20%). With each increasing line of therapy, median overall survival decreased (60, 48, 36, 29, 23 months). Survival increased with more recent diagnosis. Our results indicate that the multiple myeloma landscape has evolved significantly in the last decade.

Analysis of Employee Patient Portal Use and Electronic Health Record Access at an Academic Medical Center.

Patient portals provide patients and their caregivers online access to limited health results. Health care employees with electronic health record (EHR) access may be able to view their health information not available in the patient portal by looking in the EHR. In this study, we examine how employees use the patient portal when they also have access to the tethered EHR. We obtained patient portal and EHR usage logs corresponding to all employees who viewed their health data at our institution between January 1, 2013 and November 1, 2017. We formed three cohorts based on the systems that employees used to view their health data: employees who used the patient portal only, employees who viewed health data in the EHR only, and employees who used both systems. We compared system accesses and usage patterns for each employee cohort. During the study period, 35,172 employees accessed the EHR as part of patients' treatment and 28,631 employees accessed their health data: 25,193 of them used the patient portal and 13,318 accessed their clinical data in EHR. All employees who accessed their records in the EHR viewed their clinical notes at least once. Among EHR accesses, clinical note accesses comprised more than 42% of all EHR accesses. Provider messaging and appointment scheduling were the most commonly used functions in the patient portal. Employees who had access to their health data in both systems were more likely to engage with providers through portal messages. Employees at a large medical center accessed clinical notes in the EHR to obtain information about their health. Employees also viewed other health data not readily available in the patient portal.

Health information exchange between hospital and skilled nursing facilities not associated with lower readmissions.

To assess whether an electronic health record (EHR) portal to enable health information exchange (HIE) between a hospital and three skilled nursing facilities (SNFs) reduced likelihood of patient readmission. Secondary data; all discharges from a large academic medical center to SNFs between July 2013 and March 2017, combined with portal usage records from SNFs with HIE access. We use difference-in-differences to determine whether portal implementation reduced likelihood of readmission over time for patients discharged to HIE-enabled SNFs, relative to those discharged to nonenabled facilities. Additional descriptive analyses of audit log data characterize portal use within enabled facilities. Encounter-level clinical EHR data were merged with EHR audit log data that captured portal usage in the timeframe associated with a patient transition from hospital to SNF. Declines in likelihood of 30-day readmission were not significantly different for patients in HIE-enabled vs control SNFs (diff-in-diff=0.022; P=.431). We observe similar null effects with shorter readmission windows. The portal was used for 46 percent of discharges, with significant usage pattern variation within/across facilities. Implementation of a hospital-SNF EHR portal did not reduce readmissions from enabled SNFs. Emergent HIE use cases need to be better defined and leveraged for design and implementation that generates value in the context of postacute transitions.

1498P - Real-world effectiveness of nivolumab monotherapy after prior systemic therapy in advanced non-small cell lung cancer (NSCLC) in the United States

BackgroundNivolumab demonstrated overall survival (OS) benefit over docetaxel in patients (pts) with pre-treated squamous (SQ) and non-SQ (NSQ) NSCLC in 3 phase III clinical trials. We investigated US real-world outcomes of pts treated with nivolumab after ≥1 prior systemic therapy in predominantly community-based clinical practice. MethodsThis retrospective analysis of clinical electronic health record data in the Flatiron Health database included adult pts with advanced NSCLC receiving nivolumab monotherapy in second or later line (2L+) treatment between Mar 1, 2015 and Dec 31, 2018 (SQ) or Oct 1, 2015 and Dec 31, 2018 (NSQ). The index date was the start date of nivolumab. Prior treatment with ≥1 platinum-based and, if indicated, targeted therapy was required for inclusion. Pts with primary cancers other than NSCLC or who had received prior immunotherapy were excluded. Median (95% confidence interval [CI]) OS and progression-free survival (PFS) from index were estimated using Kaplan–Meier methodology in pt subgroups by histology. ResultsOf 3023 pts included in the analysis, 65.2% and 34.8% had NSQ and SQ histology, respectively. In the 2 groups at index, median age (range) was 68 (21–84) yrs and 70 (36–84) yrs, 51% and 63% were male, 11.3% and 5.9% had brain metastasis, 19.7% and 22.1% had ECOG PS≥2 (∼26% had missing PS data in both groups). Most pts received nivolumab as 2L (NSQ 83.4% and SQ 85.3%); pts receiving nivolumab as 3L and 4L were included. Survival outcomes are reported in the Table. ConclusionsPts receiving 2L+ nivolumab monotherapy for advanced NSQ and SQ NSCLC in US real-world clinical practice had numerically lower OS than pts in registrational trials, likely due to inclusion of pts with more impaired function. PFS was similar to that in registrational trials, but may be measured differently in routine clinical practice. Analyses with additional follow-up are planned to confirm long-term benefit of nivolumab in 2L+ NSCLC.Table1498PTableSurvival outcomes of pts receiving 2L+ nivolumab monotherapyOSPFSnMedian (95% CI), months12-month rate, %18-month rate, %Median (95% CI), months12-month rate, %18-month rate, %NSQ19728.5 (7.9–9.3)40313.2 (3.0–3.5)1712SQ10517.4 (6.8–8.5)36243.0 (2.8–3.3)1510 Editorial acknowledgementLeah Bernstein of Evidence Scientific Solutions Inc, funded by Bristol-Myers Squibb. Legal entity responsible for the studyBristol-Myers Squibb. FundingBristol-Myers Squibb. DisclosureD.D. Stenehjem: Advisory / Consultancy, Research grant / Funding (institution): BMS; Research grant / Funding (institution): Novartis; Research grant / Funding (institution): Bioverativ; Research grant / Funding (institution): AstraZeneca; Advisory / Consultancy: Salarius Pharmaceuticals; Advisory / Consultancy: Iterion Therapeutics; Advisory / Consultancy: GlycosBio. S.J. Lubinga: Full / Part-time employment: BMS. K. Gupte-Singh: Full / Part-time employment: BMS. Y. Zhang: Full / Part-time employment: BMS. K. Le Trong: Full / Part-time employment: BMS. J.R. Penrod: Full / Part-time employment: BMS.

Discovery of Noncancer Drug Effects on Survival in Electronic Health Records of Patients With Cancer: A New Paradigm for Drug Repurposing.

PURPOSEDrug development is becoming increasingly expensive and time consuming. Drug repurposing is one potential solution to accelerate drug discovery. However, limited research exists on the use of electronic health record (EHR) data for drug repurposing, and most published studies have been conducted in a hypothesis-driven manner that requires a predefined hypothesis about drugs and new indications. Whether EHRs can be used to detect drug repurposing signals is not clear. We want to demonstrate the feasibility of mining large, longitudinal EHRs for drug repurposing by detecting candidate noncancer drugs that can potentially be used for the treatment of cancer.PATIENTS AND METHODSBy linking cancer registry data to EHRs, we identified 43,310 patients with cancer treated at Vanderbilt University Medical Center (VUMC) and 98,366 treated at the Mayo Clinic. We assessed the effect of 146 noncancer drugs on cancer survival using VUMC EHR data and sought to replicate significant associations (false discovery rate < .1) using the identical approach with Mayo Clinic EHR data. To evaluate replicated signals further, we reviewed the biomedical literature and clinical trials on cancers for corroborating evidence.RESULTSWe identified 22 drugs from six drug classes (statins, proton pump inhibitors, angiotensin-converting enzyme inhibitors, β-blockers, nonsteroidal anti-inflammatory drugs, and α-1 blockers) associated with improved overall cancer survival (false discovery rate < .1) from VUMC; nine of the 22 drug associations were replicated at the Mayo Clinic. Literature and cancer clinical trial evaluations also showed very strong evidence to support the repurposing signals from EHRs.CONCLUSIONMining of EHRs for drug exposure–mediated survival signals is feasible and identifies potential candidates for antineoplastic repurposing. This study sets up a new model of mining EHRs for drug repurposing signals.

Phenome-Wide Association Studies Uncover a Novel Association of Increased Atrial Fibrillation in Male Patients With Systemic Lupus Erythematosus.

Phenome-wide association studies (PheWAS) scan across billing codes in the electronic health record (EHR) and re-purpose clinical EHR data for research. In this study, we examined whether PheWAS could function as an EHR-based discovery tool for systemic lupus erythematosus (SLE) and identified novel clinical associations in male versus female patients with SLE. We used a de-identified version of the Vanderbilt University Medical Center EHR, which includes more than 2.8 million subjects. We performed EHR-based PheWAS to compare SLE patients with age-, sex-, and race-matched control subjects and to compare male SLE patients with female SLE patients, controlling for multiple testing using a false discovery rate (FDR) P value of 0.05. We identified 1,097 patients with SLE and 5,735 matched control subjects. In a comparison of patients with SLE and matched controls, SLE patients were shown to be more likely to have International Classification of Diseases, Ninth Revision codes related to the SLE disease criteria. In the PheWAS of male versus female SLE patients, with adjustment for age and race, male patients were shown to be more likely to have atrial fibrillation (odds ratio 4.50, false discovery rate P = 3.23 × 10-3 ). Chart review confirmed atrial fibrillation, with the majority of patients developing atrial fibrillation after the SLE diagnosis and having multiple risk factors for atrial fibrillation. After adjustment for age, sex, race, and coronary artery disease, SLE disease status was shown to be significantly associated with atrial fibrillation (P = 0.002). Using PheWAS to compare male and female patients with SLE, we identified a novel association of an increased incidence of atrial fibrillation in male patients. SLE disease status was shown to be independently associated with atrial fibrillation, even after adjustment for age, sex, race, and coronary artery disease. These results demonstrate the utility of PheWAS as an EHR-based discovery tool for SLE.

Evaluation of Frailty as an Unmeasured Confounder in Observational Studies of Antidiabetic Medications.

It is unknown whether observational studies evaluating the association between antidiabetic medications and mortality adequately account for frailty. Our objectives were to evaluate if frailty was a potential confounder in the relationship between antidiabetic medication regimen and mortality and how well administrative and clinical electronic health record (EHR) data account for frailty. We conducted a retrospective cohort study in a single Veterans Health Administration (VHA) healthcare system of 500 hospitalizations-the majority due to heart failure-of Veterans who received regular VHA care and initiated type 2 diabetes treatment from 2001 to 2008. We measured frailty using a modified frailty index (FI, >0.21 frail). We obtained antidiabetic medication regimen and time-to-death from administrative sources. We compared FI among patients on different antidiabetic regimens. Stepwise Cox proportional hazards regression estimated time-to-death by demographic, administrative, clinical EHR, and FI data. Median FI was 0.22 (interquartile range 0.18, 0.27). Frailty differed across antidiabetic regimens (p < .001). An FI increase of 0.05 was associated with an increased risk of death (hazard ratio 1.45, 95% confidence interval 1.32, 1.60). Cox proportional hazards model for time-to-death including demographic, administrative, and clinical EHR data had a c-statistic of 0.70; adding FI showed marginal improvement (c-statistic 0.72). Frailty was associated with antidiabetic regimen and death, and may confound that relationship. Demographic, administrative, and clinical EHR data, commonly used to balance differences among exposure groups, performed moderately well in assessing risk of death, with minimal gain from adding frailty. Study design and analytic techniques can help minimize potential confounding by frailty in observational studies.

Using electronic health records to characterize prescription patterns: focus on antidepressants in nonpsychiatric outpatient settings.

ObjectiveTo characterize nonpsychiatric prescription patterns of antidepressants according to drug labels and evidence assessments (on-label, evidence-based, and off-label) using structured outpatient electronic health record (EHR) data.MethodsA retrospective analysis was conducted using deidentified EHR data from an outpatient practice at a New York City-based academic medical center. Structured “medication–diagnosis” pairs for antidepressants from 35 325 patients between January 2010 and December 2015 were compared to the latest drug product labels and evidence assessments.ResultsOf 140 929 antidepressant prescriptions prescribed by primary care providers (PCPs) and nonpsychiatry specialists, 69% were characterized as “on-label/evidence-based uses.” Depression diagnoses were associated with 67 233 (48%) prescriptions in this study, while pain diagnoses were slightly less common (35%). Manual chart review of “off-label use” prescriptions revealed that on-label/evidence-based diagnoses of depression (39%), anxiety (25%), insomnia (13%), mood disorders (7%), and neuropathic pain (5%) were frequently cited as prescription indication despite lacking ICD-9/10 documentation.ConclusionsThe results indicate that antidepressants may be prescribed for off-label uses, by PCPs and nonpsychiatry specialists, less frequently than believed. This study also points to the fact that there are a number of off-label uses that are efficacious and widely accepted by expert clinical opinion but have not been included in drug compendia. Despite the fact that diagnosis codes in the outpatient setting are notoriously inaccurate, our approach demonstrates that the correct codes are often documented in a patient’s recent diagnosis history. Examining both structured and unstructured data will help to further validate findings. Routinely collected clinical data in EHRs can serve as an important resource for future studies in investigating prescribing behaviors in outpatient clinics.

Predicting Survival From Large Echocardiography and Electronic Health Record Datasets: Optimization With Machine Learning

ObjectivesThe goal of this study was to use machine learning to more accurately predict survival after echocardiography. BackgroundPredicting patient outcomes (e.g., survival) following echocardiography is primarily based on ejection fraction (EF) and comorbidities. However, there may be significant predictive information within additional echocardiography-derived measurements combined with clinical electronic health record data. MethodsMortality was studied in 171,510 unselected patients who underwent 331,317 echocardiograms in a large regional health system. The authors investigated the predictive performance of nonlinear machine learning models compared with that of linear logistic regression models using 3 different inputs: 1) clinical variables, including 90 cardiovascular-relevant International Classification of Diseases, Tenth Revision, codes, and age, sex, height, weight, heart rate, blood pressures, low-density lipoprotein, high-density lipoprotein, and smoking; 2) clinical variables plus physician-reported EF; and 3) clinical variables and EF, plus 57 additional echocardiographic measurements. Missing data were imputed with a multivariate imputation by using a chained equations algorithm (MICE). The authors compared models versus each other and baseline clinical scoring systems by using a mean area under the curve (AUC) over 10 cross-validation folds and across 10 survival durations (6 to 60 months). ResultsMachine learning models achieved significantly higher prediction accuracy (all AUC >0.82) over common clinical risk scores (AUC = 0.61 to 0.79), with the nonlinear random forest models outperforming logistic regression (p < 0.01). The random forest model including all echocardiographic measurements yielded the highest prediction accuracy (p < 0.01 across all models and survival durations). Only 10 variables were needed to achieve 96% of the maximum prediction accuracy, with 6 of these variables being derived from echocardiography. Tricuspid regurgitation velocity was more predictive of survival than LVEF. In a subset of studies with complete data for the top 10 variables, multivariate imputation by chained equations yielded slightly reduced predictive accuracies (difference in AUC of 0.003) compared with the original data. ConclusionsMachine learning can fully utilize large combinations of disparate input variables to predict survival after echocardiography with superior accuracy.

Effect of an Evidence-based Inpatient Tobacco Dependence Treatment Service on 30-, 90-, and 180-Day Hospital Readmission Rates.

Smoking is a risk factor for hospitalization and interferes with patient care due to its effects on pulmonary function, wound healing, and interference with treatments and medications. Although benefits of stopping smoking are well-established, few hospitals provide tobacco dependence treatment services (TDTS) due to cost, lack of mandatory tobacco cessation standards and lack of evidence demonstrating clinical and financial benefits to hospitals and insurers for providing services. This study explored the effect of an inpatient TDTS on 30-, 90-, and 180-day hospital readmissions. To carry out this work, 3 secondary datasets were linked, which included clinical electronic health record data, tobacco cessation program data, and statewide health care utilization data. Odds ratios (ORs) were calculated using inverse propensity score-weighted logistic regression models, with program exposure as the primary independent variable and 30 (90 and 180)-day readmission rates as the dependent variable, and adjustment for putative covariates. Odds of readmission were compared for patients who did and did not receive TDTS. At 30 days postdischarge, smokers exposed to the TDTS had a lower odds of readmission (OR=0.77, P=0.031). At 90 and 180 days, odds of readmission remained lower in the TDTS group (ORs=0.87 and 0.86, respectively), but were not statistically significant. Findings from the current study, which are supported by prior studies, provide evidence that delivery of TDTS is a strategy that may help to reduce hospital readmissions.