Natural Language Processing–Identified Problem Opioid Use and Its Associated Health Care Costs

Abstract

Use of prescription opioids and problems of abuse and addiction have increased over the past decade. Claims-based studies have documented substantial economic burden of opioid abuse. This study utilized electronic health record (EHR) data to identify chronic opioid therapy (COT) patients with problem opioid use (POU) and compared costs with those for COT patients without POU. This study utilized EHR and claims data from an integrated health care system. Patients received COT (≥70 days’ supply in ≥1 calendar quarter, 2006–2012). Natural language processing (NLP) identified notations of opioid addiction, abuse, misuse, or overuse, and manual validation was performed. Cases had evidence of POU (index = first POU notation), and controls, sampled 9:1, did not. Health care resource utilization was measured and costs estimated using Medicare reimbursement rates. A longitudinal analysis of costs was conducted using generalized estimating equations. Adjusted analyses controlled for baseline age, gender, region, specific comorbidities, and a comorbidity index. The analysis population included 1,125 cases and 10,128 controls. Unadjusted costs were higher for cases in all three years. After controlling for covariates, total costs remained higher in cases and were significantly higher in the first year of follow-up ($38,064 vs. $31,674, P = .0048). The largest cost difference was observed in the first month of follow-up. COT patients with POU experienced significantly higher costs compared with COT patients without POU in the first year of follow-up. The greatest difference in costs was observed around identification of POU.