The American Journal of Managed Care November 2015
Using Sequence Discovery to Target Outreach for Diabetes Medication Adherence
Objectives: To investigate exposure sequences correlated with gaps in diabetes medication refills and to identify opportunities for targeted outreach for improved adherence.
Study Design: Sequence discovery was used to identify exposures from various data sources that preceded a gap in diabetes medication refills.
Methods: Patients who refilled a diabetes medication and had 6 months of continuous refill history were included. Patients with a therapy gap between February 1, 2012, and March 31, 2013, formed the gap group; those without formed the no-gap group. Gaps were defined as a prescription refill obtained 6 days or more after the days’ supply of the previous refill. Exposure sequences were explored in the 90 days before the gap, or before the date of last refill in the study period for the no-gap group. Exposures and sequences offering opportunity for health plan outreach were identified based on sequence length, confidence, number of intervention points, and higher gap group prevalence.
Results: Three exposure sequences with the greatest outreach opportunity to impact downstream adherence were identified within individuals taking diabetes medications who: 1) are prescribed a new medication—especially those with multiple out-of-network claims and/or visit a specialty physician after the new medication is prescribed; 2) have a prescription claim reversed by a pharmacist—particularly patients who are subsequently prescribed a new medication or visit a specialty physician; and 3) have multiple out-of-network claims and a hospitalization.
Conclusions: As medication adherence is a persisting challenge, novel application of sequence discovery techniques identified unique sequences of events with opportunities for outreach.
Am J Manag Care. 2015;21(11):e601-e608Take-Away Points
Sequence discovery techniques were used to identify sequences of events that preceded gaps in therapy (ie, prescription refills) with diabetes medications, using the diverse data sets uniquely available at the health plan level. This exploratory technique permitted identification of potentially previously unstudied variables. Given a health plan’s ability to act on identified exposures, the following exposures were identified to trigger outreach aimed at improving medication adherence:
- People taking diabetes medications who are prescribed a new medication, especially those who have multiple out of network claims and/or visit a specialty physician after the new medication is prescribed.
- People taking diabetes medications who reverse a prescription claim, especially if they subsequently are prescribed a new medication or visit a specialty physician.
- People taking diabetes medications who have multiple out of network claims, specifically those who also have a hospitalization.
Medication adherence can be measured in a variety of ways, including pill counts, patient reports, and pharmacy claims data. While researchers continue to search for a “gold standard” measurement of adherence, a multitude of predictors of medication nonadherence have been documented, including medication regimen complexity, multiple comorbidities, prescription cost, forgetfulness, depression, lack of patient understanding and engagement, and poor relationship with providers.8,9
Despite research efforts and numerous interventions, patients continue to struggle with adhering to prescribed medication regimens. In 2003, the World Health Organization reported medication adherence of only 50% in developed countries, with more recent data supporting this disappointing trend.5,8,10,11 Because nonadherence persists even after traditional epidemiological studies have identified multiple associated factors, it seems likely that other, yet unknown, factors may be related to poor medication nonadherence.
Data mining techniques have been used for decades in other industries to uncover correlations and understand patterns in large relational data sets.12-15 In recent years, these techniques have been applied to healthcare and biomedical data to identify unknown relationships between variables, generate new hypotheses, and support decision making.16-21 Different from traditional, hypothesis-driven approaches, data mining identifies correlations without consideration of prior knowledge and explores the effects of multiple combinations of exposures on outcomes.
Accordingly, we sought to investigate exposures and exposure sequences that are correlated with nonadherence, defined as a gap in prescription claims with diabetes medications—a therapeutic area with documented suboptimal adherence.22-27 Using the diverse data sets available at Humana Inc, a national healthcare company, we used association rule data mining and sequence discovery techniques to identify exposures from administrative, customer service, and consumer data. By understanding common exposure sequences indicative of nonadherence with diabetes medications, Humana and other healthcare companies can develop appropriately timed, patient-specific outreach aimed at improving adherence and patient outcomes.
The administrative claims, enrollment, customer service, and consumer data used for this study were collected from Humana Inc, which insures over 2.3 million Medicare Advantage members and 1.6 million commercial members (at the time of this analysis).28 Administrative claims data contained adjudication information for prescription medications, including drug name, dosage, quantity, days’ supply, and date of fill; International Classification of Diseases Ninth Revision, Clinical Modification (ICD-9-CM) codes for all inpatient and outpatient encounters; demographics; and coverage start and end dates. Enrollment data included address, plan enrollment, and premium information. Customer service data included detailed communications to and from health plan members, including inbound and outbound calls, e-mails, faxes, in-person contacts, and Web-based interactions. Consumer data included third party–compiled census data, buyer behavior, demographic information, proprietary models, and segmentation data. Data were analyzed from individuals with medical coverage and at least 1 prescription for a diabetes medication (ie, biguanides, dipeptidyl-peptidase-4 inhibitors, glucagon-like-peptide 1 agonist, meglitinides, sulfonylureas, and thiazolidinediones, alone or in combination) with continuous refill history between August 1, 2011, and February 1, 2012.
The outcome of interest was a gap in diabetes medication therapy, defined as a prescription refill obtained 6 or more days after exhaustion of the days’ supply of the previous refill. Gaps were identified between February 1, 2012, and March 31, 2013. Patients were divided into 2 groups: those with and those without a gap in prescription refill history of diabetes medications. In each group, exposures were identified during the 90 days before the index date, which was defined as the actual next refill date in the no-gap group or the expected refill date in the gap group (eAppendix, available at www.ajmc.com). For the 90-day pre-index period, association rule mining and sequence discovery techniques were used to identify exposure sequences associated with a refill gap.
A sub-analysis within the gap group evaluated exposures during both a gap and no-gap period in the same patient. Exposure evaluation began on the actual (no-gap period) or expected (gap period) refill date. Refills with a 30-day supply had a 29-day look back, while refills with a 90-day supply had an 89-day look back to eliminate duplication of data between the gap and no-gap periods.
Exposure variables were identified based on availability in the databases, as well as the ability of Humana to provide outreach given the presence of a certain variable. Exposures related to an individual’s medical care included annual physical exam, hospitalization, emergency department (ED) visit, specialty physician visit, newly diagnosed condition, bariatric surgery, enrollment in a bariatric surgery program, and enrollment in a smoking cessation program. Prescription medication–related exposures included new medications, reversal of a claim for a prescription drug, adverse drug events, comprehensive medication reviews, prescription for smoking cessation product, change in mail order versus retail delivery channel, and mail order educational contacts. Personal exposures included address and religious affiliation changes, death of a family member in the same household, and natural disasters. Several communication-related exposures were also assessed, including inbound and outbound faxes, calls, e-mails, Web communication, and walk-in contacts. Insurance-related exposures included out-of-network claims, disruptions to plan coverage, entrance into the Part D coverage gap, and entrance into catastrophic prescription coverage.
Association Rule Mining and Sequence Discovery
Association rule mining is an efficient way to identify associations between variables in large data sets and determine the likelihood of variables occurring together. These rules count the number of times items occur in the data set, either alone or in combination, and differ from regression modeling, which assesses the independent strength of 2 variables, holding all others constant. An association rule is depicted as E O, where E represents exposure variables and O is the outcome variable. Association rule mining evaluates each association according to a minimum support, confidence, and lift level. "Support" for the rule represents the probability that both variables occur together, while "confidence" represents the conditional probability of the outcome occurring, given the exposure. "Lift" is defined as measured confidence divided by expected confidence level.
Support (E⇒O ) = P(E∩O)
Confidence (E⇒O) = P(O/E) = P(E∩O) / P(E)
Lift (E⇒O) = Confidence (E⇒O) / Expected Confidence (E⇒O)
Credible associations have a high support and confidence (both expressed as a percentage), with a lift level greater than 1. Sequence discovery utilizes association rule mining results and accounts for the timing of the relationship among items; for example, rule A B implies that event B occurred after event A occurred.29
We used sequence discovery to identify exposure sequences in the gap and no-gap groups and those common to both groups. The sequence discovery analysis built upon association rule mining analyses (using a minimum support of 5%) and was set to a minimum support threshold of 2%. The maximum length of sequences was limited to 7, and only patients with more than 1 exposure contributed to sequence discovery analysis. Sequences with the highest support and confidence for each sequence length were determined for each of the 3 groups (gap, no-gap, both). In addition to exposures sequences, singular exposures with the highest frequencies in the gap group and no-gap group were reported. Exposure sequences and singular exposure frequencies were also described for the gap group sub-analysis.
Using exposure sequences generated by sequence discovery analysis, we sought to identify exposures and exposure sequences that would be most practical for a healthcare company to target outreach aimed at improving adherence and patient outcomes. First, we identified the most frequently occurring initial or final exposures in gap group sequences. Initial exposures in sequences are important since they can be used as early events to trigger an intervention; final events provide a potential last opportunity for intervention. From those, we identified exposure sequences offering the greatest opportunity for intervention, using the following criteria: 1) sufficient sequence length allowing time for intervention; 2) higher support, combined with a high confidence level, compared with other sequences within the same sequence length group; 3) a variety of exposures in a sequence, supporting multiple intervention points; and 4) combinations of the same exposures, albeit in different order, that were more prevalent in the gap group.
These analyses were conducted as a part of Humana’s ongoing administrative activities aimed at improving medication adherence—not to generate scientific knowledge. Such quality improvement activities, which do not meet the regulatory definition of research under 45 CFR 46.102(d), do not require review by an institutional review board.30 Humana’s privacy and ethics board did review and approve this work, however. All analyses were run with SAS Enterprise Guide version 5.1 (SAS Institute, Cary, North Carolina).29