A Machine Learning Approach to Identify Predictors of Severe COVID-19 Outcome in Patients With Rheumatoid Arthritis.

BACKGROUND: Rheumatoid arthritis (RA) patients have a lowered immune response to infection, potentially due to the use of corticosteroids and immunosuppressive drugs. Predictors of severe COVID-19 outcomes within the RA population have not yet been explored in a real-world setting. OBJECTIVES: To identify the most influential predictors of severe COVID-19 within the RA population. STUDY DESIGN: Retrospective cohort study. SETTING: Research was conducted using Optum's de-identified Clinformatics® Data Mart Database (2000-2021Q1), a US commercial claims database. METHODS: We identified adult patients with index COVID-19 (ICD-10-CM diagnosis code U07.1) between March 1, 2020, and December 31, 2020. Patients were required to have continuous enrollment and have evidence of one inpatient or 2 outpatient diagnoses of RA in the 365 days prior to index. RA patients with COVID-19 were stratified by outcome (mild vs severe), with severe cases defined as having one of the following within 60 days of COVID-19 diagnosis: death, treatment in the intensive care unit (ICU), or mechanical ventilation. Baseline demographics and clinical characteristics were extracted during the 365 days prior to index COVID-19 diagnosis. To control for improving treatment options, the month of index date was included as a potential independent variable in all models. Data were partitioned (80% train and 20% test), and a variety of machine learning algorithms (logistic regression, random forest, support vector machine [SVM], and XGBoost) were constructed to predict severe COVID-19, with model covariates ranked according to importance. RESULTS: Of 4,295 RA patients with COVID-19 included in the study, 990 (23.1%) were classified as severe. RA patients with severe COVID-19 had a higher mean age (mean [SD] = 71.6 [10.3] vs 63.4 [13.7] years, P < 0.001) and Charlson Comorbidity Index (CCI) (3.8 [2.4] vs 2.4 [1.8], P < 0.001) than those with mild cases. Males were more likely to be a severe case than mild (29.1% vs 18.5%, P < 0.001). The top 15 predictors from the best performing model (XGBoost, AUC = 75.64) were identified. While female gender, commercial insurance, and physical therapy were inversely associated with severe COVID-19 outcomes, top predictors included a March index date, older age, more inpatient visits at baseline, corticosteroid or gamma-aminobutyric acid analog (GABA) use at baseline or the need for durable medical equipment (i.e., wheelchairs), as well as comorbidities such as congestive heart failure, hypertension, fluid and electrolyte disorders, lower respiratory disease, chronic pulmonary disease, and diabetes with complication. LIMITATIONS: The cohort meeting our eligibility criteria is a relatively small sample in the context of machine learning. Additionally, diagnoses definitions rely solely on ICD-10-CM codes, and there may be unmeasured variables (such as labs and vitals) due to the nature of the data. These limitations were carefully considered when interpreting the results. CONCLUSIONS: Predictive baseline comorbidities and risk factors can be leveraged for early detection of RA patients at risk of severe COVID-19 outcomes. Further research should be conducted on modifiable factors in the RA population, such as physical therapy.

Concomitant Filled Prescriptions of Oxymorphone or Oxycodone with CYP3A Inhibitors and Inducers.

BACKGROUND: Oxymorphone's metabolism does not involve the hepatic cytochrome P450 (CYP) system. The effect of this pharmacokinetic feature of oxymorphone on opioid prescribing is unknown. OBJECTIVE: To assess the relative frequency with which oxymorphone and oxycodone (a CYP3A-metabolized opioid analgesic) were each prescribed to patients concomitantly receiving CYP3A-modifying drugs (i.e., inducers and inhibitors) to characterize opioid-prescribing patterns in patients at risk for CYP3A-related drug interactions. METHODS: We analyzed the Sentinel Distributed Database from January 1, 2013, to December 31, 2016, to identify the proportion of patients with concomitant dispensing of selected CYP3A modifiers among initiators of oxymorphone. We then repeated the analysis using oxycodone instead of oxymorphone. We conducted sensitivity analyses that varied the washout periods for each opioid to account for potential opioid switching. RESULTS: In the primary analysis, the proportion of patients with concomitant incident dispensings of oxymorphone and selected CYP3A modifiers was 3.26% (95% CI = 3.09%-3.43%), and the proportion of patients with incident dispensings of oxycodone and selected CYP3A modifiers was 2.82% (95% CI = 2.79%-2.85%). The difference between proportions was 0.43% (95% CI = 0.26%-0.60%). Sensitivity analyses that varied the washout periods for each opioid with respect to the other opioid to account for switching yielded similar results. CONCLUSIONS: We observed similar proportions of patients using selected CYP3A modifiers concomitantly with both oxymorphone and oxycodone. While the CIs of the point estimates did not overlap, the absolute differences between the proportions were small. DISCLOSURES: This project was supported by Task Order HHSF22301001T under Master Agreement HHSF223201400030I from the U.S. Food and Drug Administration (FDA). The FDA approved the study protocol, including the statistical analysis plan, and reviewed and approved the manuscript. Coauthors from the FDA participated in the results interpretation and in the preparation and decision to submit the manuscript for publication. Coyle, Money, Staffa, Meyer, and Woods are employed by the FDA. The other authors have no financial conflicts of interest to report. The views expressed are those of the authors and not necessarily those of the U.S. Department of Health and Human Services, U.S. Food and Drug Administration.

Development of an algorithm to detect methotrexate wrong frequency error using computerized health care data.

PURPOSE: We validated an algorithm to detect frequency errors in computerized healthcare data and estimated the incidence of these errors in an integrated healthcare system. METHODS: We applied Sentinel System analytic tools on the electronic health records of Kaiser Permanente, Northern California, January 1, 2010, through May 30, 2015,to identify rheumatoid arthritis (RA) patients with new use of methotrexate (365-day baseline period). We identified potential methotrexate frequency errors using ICD-9 code 995.20 (adverse drug event), Current Procedural Terminology (CPT) code 96409 for injection of leucovorin and prescription refill patterns. We performed chart review to confirm the frequency errors, assessed performance for detecting frequency errors, and estimated the incidence of chart-confirmed errors. RESULTS: The study included 24,529 methotrexate dispensings among 3,668 RA patients. Among these, 722 (3%) had one dispensing and 23,807 (97.1%) had ≥2 dispensings during 1-year follow-up period. We flagged 653 (2.7%) with a potential medication error (46 with one dispensing and 607 with ≥2 dispensings). We sampled 94 for chart review, and confirmed three methotrexate errors. All three confirmed frequency errors involved a first methotrexate dispensing followed by injected rescue therapy, leucovorin, (positive predictive value, 60%; 95% confidence interval [CI], 15-95%). No potential errors were found among patients with ≥2 dispensings. We estimated the frequency error incidence among one methotrexate dispensing to be 0.4% (95%CI, 0.1% to 1.2%). CONCLUSION: Rescue therapy is a specific indicator of methotrexate overdose among first methotrexate dispensings. This method is generalizable to other medications with serious adverse events treated with antidotes.

Overall and cause-specific mortality in the Sentinel system: A power analysis.

PURPOSE: Mortality data within the Sentinel Death Tables remain generally uncharacterized. Assessment of mortality data within Sentinel will help inform its utility for medical product safety studies. METHODS: To determine if Sentinel contains sufficient all-cause and cause-specific mortality events to power postmarketing safety studies. We calculated crude rates of all-cause mortality and suicide and proportional mortality from suicide from 2004 to 2012 in seven Sentinel data partners. Results were stratified by data partner, sex, age group, and calendar year and compared with national estimates from Centers for Disease Control and Prevention Wide-ranging ONline Data for Epidemiologic Research. We performed sample size estimations for all-cause mortality and 10 leading causes of death. RESULTS: We observed 479 694 deaths, including 5811 suicides, during 68 million person-years of follow-up. Pooled mean death and suicide rates in the data partners were 710 and 8.6 per 100 000 person-years, respectively (vs 810 and 11.8 nationally). The mean proportional mortality from suicide among the data partners was 1.2%, compared with 1.5% nationally. National trends of decreasing overall mortality and increasing proportional mortality for suicide were reflected within Sentinel. We estimated that detecting hazard ratios of 1.25 and 3 would require 16 442 and 460 exposed patients, respectively, for overall mortality, and 1.3 million and 37 411, respectively, for suicide. CONCLUSIONS: This was the first study to investigate mortality data in the Sentinel death tables. We found that all-cause mortality appeared well powered for use as a safety outcome and cause-specific mortality outcomes may be adequately powered in certain circumstances. Further investigation into the quality of the Sentinel death data is needed.

Incidence of Heart Failure and Cardiomyopathy Following Initiation of Medications for Attention-Deficit/Hyperactivity Disorder: A Descriptive Study.

PURPOSE/BACKGROUND: Stimulant abuse is associated with cardiomyopathy, but cardiomyopathy rates with therapeutic use of stimulants for attention-deficit/hyperactivity disorder (ADHD) are poorly characterized. Labels for methylphenidate, amphetamine, and atomoxetine caution against use in patients with cardiovascular disease. We sought to assess the incidence of new-onset heart failure or cardiomyopathy among initiators of these medications. METHODS/PROCEDURES: Using the Sentinel distributed database, we analyzed new-onset heart failure or cardiomyopathy among initiators of selected ADHD medications (amphetamine products including lisdexamfetamine, methylphenidate, and atomoxetine), by duration of use (0-90, 91-180, 181-270, 271-365, 366-730, and 731-1095 days) and age group (<22, 22-44, 45-64, and ≥65 years). FINDINGS/RESULTS: In our sample of 2,012,948 initiators of ADHD medications, 44.6% were female, and 54.1% were younger than 22 years. Heart failure/cardiomyopathy rates in the age groups younger than 22 and 22 to 44 years old were less than 50 per 10,000 person-years, without clear trends by duration of use. The highest rates occurred soon after treatment initiation in the age group 65 years or older, with 1 case per 10.5 person-years of follow-up, or 950 cases per 10,000 person-years, for days 0-90. IMPLICATIONS/CONCLUSIONS: Heart failure/cardiomyopathy rates were not higher over 3 years of ADHD medication use compared with shorter-term treatment. In older age groups, lower rates later in treatment could reflect depletion of patients predisposed to the outcome if they develop it soon after starting treatment.

Early impact of the ICD-10-CM transition on selected health outcomes in 13 electronic health care databases in the United States.

PURPOSE: To describe the consistency in the frequency of 5 health outcomes across the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) and Tenth Revision, Clinical Modification (ICD-10-CM) eras in the US. METHODS: We examined the incidence of 3 acute conditions (acute myocardial infarction [AMI], angioedema, ischemic stroke) and the prevalence of 2 chronic conditions (diabetes, hypertension) during the final 5 years of the ICD-9-CM era (January 2010-September 2015) and the first 15 months of the ICD-10-CM era (October 2015-December 2016) in 13 electronic health care databases in the Sentinel System. For each health outcome reviewed during the ICD-10-CM era, we evaluated 4 definitions, including published algorithms derived from other countries, as well as simple-forward, simple-backward, and forward-backward mapping using the General Equivalence Mappings. For acute conditions, we also compared the incidence between April to December 2014 and April to December 2016. RESULTS: The analyses included data from approximately 172 million health plan members. While the incidence or prevalence of AMI and hypertension performed similarly across the 2 eras, the other 3 outcomes did not demonstrate consistent trends for some or all the ICD-10-CM definitions assessed. CONCLUSIONS: When using data from both the ICD-9-CM and ICD-10-CM eras, or when using results from ICD-10-CM data to compare to results from ICD-9-CM data, researchers should test multiple ICD-10-CM outcome definitions as part of sensitivity analysis. Ongoing assessment of the impact of ICD-10-CM transition on identification of health outcomes in US electronic health care databases should occur as more data accrue.

Assessing the impact of the new ICD-10-CM coding system on pharmacoepidemiologic studies-An application to the known association between angiotensin-converting enzyme inhibitors and angioedema.

PURPOSE: To replicate the well-established association between angiotensin-converting enzyme inhibitors versus beta blockers and angioedema in the International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM) era. METHODS: We conducted a retrospective, inception cohort study in a large insurance database formatted to the Sentinel Common Data Model. We defined study periods spanning the ICD-9-CM era only, ICD-10-CM era only, and ICD-9-CM and ICD-10-CM era and conducted simple-forward mapping (SFM), simple-backward mapping (SBM), and forward-backward mapping (FBM) referencing the General Equivalence Mappings to translate the outcome (angioedema) and covariates from ICD-9-CM to ICD-10-CM. We performed propensity score (PS)-matched and PS-stratified Cox proportional hazards regression to estimate hazard ratios (HRs) and 95% confidence intervals (CIs). RESULTS: In the ICD-9-CM and ICD-10-CM eras spanning April 1 to September 30 of 2015 and 2016, there were 152 017 and 145 232 angiotensin-converting enzyme inhibitor initiators and 115 073 and 116 652 beta-blocker initiators, respectively. The PS-matched HR was 4.19 (95% CI, 2.82-6.23) in the ICD-9-CM era, 4.37 (2.92-6.52) in the ICD-10-CM era using SFM, and 4.64 (3.05-7.07) in the ICD-10-CM era using SBM and FBM. The PS-matched HRs from the mixed ICD-9-CM and ICD-10-CM eras ranged from 3.91 (2.69-5.68) to 4.35 (3.33-5.70). CONCLUSION: The adjusted HRs across different diagnostic coding eras and the use of SFM versus SBM and FBM produced numerically different but clinically similar results. Additional investigations as ICD-10-CM data accumulate are warranted.

Electronic clinical laboratory test results data tables: lessons from Mini-Sentinel.

PURPOSE: Developing electronic clinical data into a common data model posed substantial challenges unique from those encountered with administrative data. We present here the design, implementation, and use of the Mini-Sentinel Distributed Database laboratory results table (LRT). METHODS: We developed the LRT and guided Mini-Sentinel data partners (DPs) in populating it from their source data. Data sources included electronic health records and internal and contracted clinical laboratory systems databases. We employed the Logical Observation Identifiers, Names, and Codes (LOINC®) results reporting standards. We evaluated transformed results data using data checks and an iterative, ongoing characterization and harmonization process. RESULTS: Key LRT variables included test name, subcategory, specimen source, LOINC, patient location, specimen date and time, result unit, and unique person identifier. Selected blood and urine chemistry, hematology, coagulation, and influenza tests were included. Twelve DPs with outpatient test results participated; four also contributed inpatient test results. As of September 2013, the LRT included 385,516,239 laboratory test results; data are refreshed at least quarterly. LOINC availability and use varied across DP. Multiple data quality and content issues were identified and addressed. CONCLUSION: Developing the LRT brought together disparate data sources with no common coding structure. Clinical laboratory test results obtained during routine healthcare delivery are neither uniformly coded nor documented in a standardized manner. Applying a systematic approach with data harmonization efforts and ongoing oversight and management is necessary for a clinical laboratory results data table to remain valid and useful.

Design considerations, architecture, and use of the Mini-Sentinel distributed data system.

PURPOSE: We describe the design, implementation, and use of a large, multiorganizational distributed database developed to support the Mini-Sentinel Pilot Program of the US Food and Drug Administration (FDA). As envisioned by the US FDA, this implementation will inform and facilitate the development of an active surveillance system for monitoring the safety of medical products (drugs, biologics, and devices) in the USA. METHODS: A common data model was designed to address the priorities of the Mini-Sentinel Pilot and to leverage the experience and data of participating organizations and data partners. A review of existing common data models informed the process. Each participating organization designed a process to extract, transform, and load its source data, applying the common data model to create the Mini-Sentinel Distributed Database. Transformed data were characterized and evaluated using a series of programs developed centrally and executed locally by participating organizations. A secure communications portal was designed to facilitate queries of the Mini-Sentinel Distributed Database and transfer of confidential data, analytic tools were developed to facilitate rapid response to common questions, and distributed querying software was implemented to facilitate rapid querying of summary data. RESULTS: As of July 2011, information on 99,260,976 health plan members was included in the Mini-Sentinel Distributed Database. The database includes 316,009,067 person-years of observation time, with members contributing, on average, 27.0 months of observation time. All data partners have successfully executed distributed code and returned findings to the Mini-Sentinel Operations Center. CONCLUSION: This work demonstrates the feasibility of building a large, multiorganizational distributed data system in which organizations retain possession of their data that are used in an active surveillance system.