Study setting and opioid misuse definition

This study utilized data from the EHR of an urban tertiary academic center between January 1, 2007 and September 30, 2017. An operational definition for opioid misuse was developed following the National Survey on Drug Use and Health criteria for opioid misuse with criterion input from a board-certified addiction specialist (ESA) and psychiatrist (NSK). The analysis cohort included consecutive adult (≥18 years of age) emergency department and inpatient encounters meeting criteria for opioid misuse during the study period. The criteria for opioid misuse were any of the following: (1) positive urine drug screen for an opiate with polysubstance use with any of the following: an illicit drug (phencyclidine or cocaine), a benzodiazepine that is not on the patient’s medication administration record, or an amphetamine that is not on the patient’s medication administration record; (2) positive urine drug screen for an opiate but without a prescription for an opioid on the patient’s admission administration record. Urine drug screens were eligible only if no opioid or benzodiazepine drug was dispensed by the hospital pharmacy before the urine drug screen was ordered.; (3) International Classification of Diseases (ICD)-9 and –10 codes for opioid-related hospitalizations were adopted from the Healthcare Cost and Utilization Project (HCUP) [14]. ICD codes reflect final billing diagnostic codes used for claims with payers. The codes include a variety of opioid-related events and opioid misuse codes and are detailed in S1 Appendix. Many of the ICD codes do not allow for heroin-related cases to be explicitly identified. In addition, the codes do not distinguish between illegal use of prescription drugs and their use as prescribed.

To validate the operational definition, a random sample of hospital encounters was extracted from the EHR during the study period for chart review. A sample of 1,000 patient encounters including age-sex matched controls were reviewed. The annotations included an oversampling of patients who met case criteria and non-cases who had ICD codes for chronic pain, naloxone administration, or a physician order for a urine drug screen. An annotator (KS) who is an MD, MPH candidate received substance use training through Loyola’s Institute for Transformative Interprofessional Education and completed Screening, Brief Intervention, and Referral to Treatment (SBIRT) through online training. Additional training was provided to screen for likelihood of opioid misuse on a Likert Scale (1–5), and the annotator met an inter-rater reliability of a Cohen’s kappa coefficient greater than 0.80 with a critical care physician and addiction specialist (MA, ESA) before independent review was performed.

The operational definition had a sensitivity of 88.6% (95% CI 85.2%-91.9%) and specificity of 78.5% (95% CI 75.4% - 81.7%). Chart review identified many false positives that occurred in outside hospital transfers that administered an opioid during care; therefore, hospital transfers were excluded from this analysis. Cases of overdose could not be reliably discriminated using billing codes or naloxone administration with many false positives occurring as well.

Multiple encounters by the same patient were included as independent observations during analysis. As patients’ severity and subtype of misuse may change over time, our primary unit of analysis is the patient encounter in order to provide actionable insight into the subtype of misuse at hospitalization which could inform timely custom interventions. To address the potentially high correlation of intra-patient encounters, sensitivity analysis was performed to remove multiple encounters by analyzing the most recent inpatient encounter by each patient in our cohort.

Identifying subtypes with latent class analysis (LCA)

Latent class analysis (LCA) is a statistical technique that uses mixture modelling to identify mutually exclusive and qualitatively different subgroups from multivariate categorical data [15,16]. LCA takes observed data as inputs to define a number of unobservable, distinct subtypes or classes from the population of interest. Model fit statistics are utilized to identify the appropriate number of latent classes, and the distributions of the observed class-defining indicator variables, called item response probabilities, are used to characterize the classes. Posterior probabilities for encounters indicate the likelihood of membership into each of the latent classes. The following were class-defining variables in the LCA model: (1) urine drug screen results; (2) ICD codes for opioid-related hospitalizations; (3) ICD codes for chronic pain; (4) age; (5) ICD codes for alcohol use disorders; (6) ICD codes for psychoses; (7) ICD codes for depression; and (8) ICD codes for liver disease. The eight class-defining variables for the LCA model were chosen a priori based on existing evidence for identifying cases of misuse and risk factors for misuse [17–21]. LCA models were considered for one to eight classes for these class-defining variables. The optimal number of latent classes was selected using fit statistics including the Bayesian information criterion (BIC), adjusted Bayesian information criterion (aBIC), consistent Akaike information criterion (cAIC), class prevalence, class separation, and model interpretability [22]. Each patient encounter was assigned a class according to the highest latent class posterior probability.

The face validity and clinical utility are examined by comparisons between latent classes using the following approaches: (1) descriptive statistics on patient characteristics and health utilization (structured EHR and census-level data) for each subtype; (2) topic models from natural language processing (clinical notes) and their probability assignment to each subtype; (3) association of subtypes with clinical outcomes (described below).

Structured EHR and census-level data

Individual patient measures from the EHR included the following: (1) demographics and insurance status; (2) Elixhauser mortality score; (3) ICD codes for chronic pain and other disease categories developed by the Agency for Healthcare Research and Quality [23,24]; (4) hospital utilization patterns; and (5) admission service (medicine, surgery, trauma); and (6) naloxone administration (only within first three hours of first recorded vital sign). Census tract measures were used as a proxy for individual level socioeconomic status (SES). An application program interface was built to match the housing addresses to corresponding geocodes for all patients in our health system’s clinical data warehouse and provide data at the census-tract level which is equivalent to a neighborhood established by the Bureau of Census for analyzing populations. The data were collected from the 2015 American Community Survey [25] and linked to corresponding geocodes at the patient level. The census-tract measures reported for this study were the following: (1) education level (more than high school vs. high school/less than high school); (2) employment status (employed vs. unemployed); (3) median household earnings; (4) homeowner status (any homeownership vs. none); and (5) poverty level. The census-tract variable for poverty level was shown to represent an important indicator of census-level SES that correlates well with other SES measures [26,27]; therefore, we categorized patients into high- poverty census-tract (20.0+ percent of households below federal poverty level) vs. low-(≤9.9 percent of households below federal poverty level) or middle-(10.0–19.9 percent of households below federal poverty level) [28].

For identifying conditions of chronic pain [23], ICD codes for chronic primary pain, psychogenic pain, chronic postsurgical and posttraumatic pain, and chronic neuropathic pain were included. Additional codes were included for chronic secondary musculoskeletal pain and chronic secondary visceral pain. Codes were excluded for acute pain, chronic cancer-related pain, and chronic secondary headache or orofacial pain. The final list of codes is provided in S2 Appendix.

Unstructured EHR data (clinical notes): Natural language processing and topic modelling

Pre-processing of all clinical notes was performed in the Apache clinical Text Analysis Knowledge Extraction System (cTAKES) [29]. cTAKES is a widely used software library for clinical Natural Language Processing (NLP) [30,31]. The pre-processing steps include text tokenization (splitting into words), sentence segmentation (splitting into sentences), part-of-speech tagging, and clinical concept lookup. The clinical concept lookup identified spans of clinically-relevant text, such as 'chronic pain', filtering out non-essential/non-clinical vocabulary. cTAKES maps each concept to a common standardized medical vocabulary in the Unified Medical Language System identified as a Concept Unique Identifier (CUI). This method is an approach to provide structure to otherwise unstructured text data, and the CUIs served as inputs for topic modelling.

Latent Dirichlet Allocation (LDA) is an unsupervised text mining approach used for topic modeling [32]. The LDA model was trained on the full cohort of patients with opioid misuse and produced topics expressed as a probability distribution over CUIs. LDA discovers latent topic structure by finding the mixture of CUIs that is associated with each topic and determining the probability of each topic for the individual patient encounter. We provide topic modelling for two reasons: (1) to demonstrate if the notes reflect what the administrative data and ICD codes describe for the subtypes, a representation of face validity; and (2) unstructured data from notes and reports comprise nearly 80% of the EHR data [33,34] and is a future direction in analyzing patient cohorts for clinical decision support. The optimal number of latent topics was identified by examining a range of topics with model fit statistics for topic coherence [35]. In LDA model development, 250 passes on the clinical notes were performed to learn the topics. To improve the efficiency of topic modelling, we restricted concepts, CUIs, to those observed in less than 70% and more than 10% of the EHR notes to eliminate concepts that were too commonly reported or too rare to be informative. To evaluate the contribution of a topic to an LCA-derived class, we average the probability of that topic across all the encounters in that class.

Clinical outcomes: 30-day unplanned hospital readmission and discharge dispositions

All-cause unplanned readmission was identified using the Center for Medicare and Medicaid Services (CMS) rules for index (eligible) admission and unplanned 30-day readmission [36]. Pre-specified billing codes for planned readmission were used from CMS rules and include obstetrical delivery, scheduled procedures, maintenance chemotherapy, and rehabilitation. To analyze data on diagnoses and procedures that met qualifying criteria for readmission, the Clinical Classification Software from the Agency for Healthcare Research and Quality was used to crosswalk with diagnoses from billing codes in the EHR [37]. Readmissions during the 30-day period that follow a planned readmission are not counted in the outcome. In the case of multiple readmissions during a 30-day period, we measured only one outcome. Readmissions on the same day were also not counted in the outcome. Ultimately, index admissions include any inpatient hospitalization during the study period and are excluded for the reasons described above. Additional outcomes examined include discharge status from the hospital (in-hospital death, psychiatric admission, against medical advice (AMA), and home).

Analysis was performed using Python Version 3.6.5 (Python Software Foundation) and RStudio Version 1.1.463 (RStudio Team, Boston, MA). Latent class analysis was performed with the poLCA package in R (http://dlinzer.github.com/poLCA) and followed the analysis plan by Zhang et al. [22]. Our open-source code to perform LCA may be viewed in S3 Appendix as well as at: https://bitbucket.org/afsharjoycelab/opioid-misuse-lca/. The GenSim package was used in Python to infer topic model structure [38]. The Institutional Review Board of Loyola University Chicago approved this study.

Opioid misuse cohort

The health system had 228,884 inpatient encounters during the study period and 6,224 (2.7%) met inclusion criteria for opioid misuse. In topic modelling, the final data corpus in the 6,224 patient encounters was comprised of 25,801 unique CUIs across 422,147 clinical notes. Twenty topics were identified to have the best model fit from the cohort (S4 Appendix). The top ten CUIs for each topic and a summarized topic theme for each topic are listed in Table 1. The topics spanned themes from chronic medical conditions to behavioral conditions and healthcare services.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Topic modelling across all clinical documents in patients with opioid misuse.

https://doi.org/10.1371/journal.pone.0219717.t001

Identification of four subtypes of opioid misuse

Fit statistics and model interpretability suggest a 4-class model was optimal. Improvements to model fit begin to diminish around 5 classes (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Goodness-of-fit statistics for models with varying number of classes.

Caption: The optimal number of latent classes was selected using fit statistics including the Bayesian information criterion (BIC), adjusted Bayesian information criterion (aBIC), consistent Akaike information criterion.

https://doi.org/10.1371/journal.pone.0219717.g001

Although lower model fit statistics could be achieved with more classes, the 4-class model represents meaningful and distinct clinical representation with better distribution of class prevalence and less complexity (Table 2). Details for the 5-class model are shown in S5 Appendix. In the 4-class model, the average latent class probabilities were 0.72 (sd 0.10) for Class 1, 0.87 (sd 0.12) for Class 2, and 0.88 (sd 0.11) for Class 3, and 0.95 (sd 0.12) for Class 4 indicating acceptable class separation. In addition, the 4-class model had appropriate face validity from topic modelling (Table 3). Polysubstance use as a topic had the highest probability for all classes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Patient characteristics by latent class.

https://doi.org/10.1371/journal.pone.0219717.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Distribution of probabilities for each topic across the 4-class latent model.

https://doi.org/10.1371/journal.pone.0219717.t003

In sensitivity analysis examining patient-level data using the most recent hospital encounter, model fit statistics and interpretability continued to show a 4-class model to be optimal (S6 Appendix). Good class separation was found with average latent class probabilities of 0.92 (sd 0.12) for Class 1, 0.91 (sd 0.12) for Class 2, 0.77 (sd 0.11) for Class 3, and 0.94 (sd 0.14) for Class 4. The patient-level 4-class model represented nearly identical distributions of patient characteristics across demographics, comorbidities, substance use, and SES (S6 Appendix).

Clinical distinctions between the four subtypes

The 6,224 encounters were categorized within one of the four latent classes. Class 1 represents encounters that carried a higher probability for topics on pain procedures and medical conditions associated with chronic pain (pancreatitis and metastatic cancer) (Table 3). In Table 2, approximately half were females and carry the greatest proportion with Medicare insurance among the classes. Nearly one-third had a diagnosis category for chronic pain and Class 1 had the fewest with urine drug screen testing and lowest rates for positive tests. All encounters in this class were for opioid-related hospitalizations with a greater proportion of prior 1-year inpatient and outpatient encounters compared to other classes. Class 1 comprised 36.5% of the cohort and received the label, “High hospital utilization with known opioid-related conditions”.

Class 2 are patient encounters with topics that had the highest probability for polysubstance use and mental health conditions and the greatest proportions with Elixhauser ICD codes for psychoses and drug use (Table 3). In Table 2, these patient encounters were mainly for patients 36–55 years of age and the majority were non-Hispanic black with Medicaid and uninsured status. All patient encounters in this class had positive cocaine urine drug screens. This class had the highest proportion of patients living in low socioeconomic census tracts and with the lowest median household income of all classes. Class 2 comprised 12.8% of the cohort and received the label, “Illicit use, low SES, and psychoses”.

Class 3 are patient encounters with topics that had the highest probability for alcohol use disorders and liver disease (Table 3). In Table 2, these encounters were for older and largely non-Hispanic white male patients compared to other classes, and represented the highest risk for mortality by Elixhauser score. The proportion with ICD codes for alcohol use disorders and liver disease were greatest in this class. Class 3 comprised 39.2% of the cohort and received the label, “Alcohol use disorders with complications”.

Class 4 patient encounters contained topics that had the highest probability for trauma and neurological diseases (seizures) (Table 3). In Table 2, these patients had more encounters in trauma centers with the highest proportion of urine drug testing and positive cases for opioids and benzodiazepines. Over one-third had chronic pain, similar to Class 1 but with more naloxone administration. Class 4 comprised 11.5% of the cohort and received the label, “Low hospital utilization and incidental opioid misuse”.

Clinical outcomes between the four classes/subtypes

The class labelled as “High hospital utilization with known opioid-related conditions” (Class 1) had the greatest proportion with 30-day unplanned hospital readmission at 13.9% (Table 4). This was followed by Class 3 with the label “Alcohol use disorders with complications”. Class 2 labelled as “Illicit use, low SES, and psychoses” had the greatest proportions for being discharged to inpatient psychiatry services and leaving against medical advice. Class 4 labelled as the “Low hospital utilization with incidental opioid misuse” had the greatest proportion with naloxone administration in the hospital and in-hospital death but the lowest proportion with readmission.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Latent class and clinical outcomes.

https://doi.org/10.1371/journal.pone.0219717.t004