Using Natural Language Processing and Network Analysis to Develop a Conceptual Framework for Medication Therapy Management Research

Abstract

This paper describes a theory derivation process used to develop a conceptual framework for medication therapy management (MTM) research. The MTM service model and chronic care model were selected as parent theories. Review article abstracts targeting medication therapy management in chronic disease care were retrieved from Ovid Medline (2000-2016). Unique concepts in each abstract were extracted using MetaMap and their pairwise cooccurrence determined. The information was used to construct a network graph of concept co-occurrence that was analyzed to identify content for the new conceptual model. 142 abstracts were analyzed. Medication adherence is the most studied drug therapy problem and co-occurred with concepts related to patient-centered interventions targeting self-management. The enhanced model consists of 65 concepts clustered into 14 constructs. The framework requires additional refinement and evaluation to determine its relevance and applicability across a broad audience including underserved settings.

Introduction

Drug therapy problems are undesirable events or circumstances that involve drugs and interfere with the achievement of desired goals of therapy¹ They can be categorized according to medication indication (e.g., unnecessary treatment, need for additional treatment), effectiveness (e.g., ineffective drug, dosage too low), safety (e.g., adverse drug reactions, dosage too high), and adherence (non-adherence). Drug therapy problems are often deleterious and costly. In the United States, it is estimated that adverse drug events are responsible for an estimated 3.5 million physician office visits, 1 million emergency department visits and 125,000 hospital admissions per year ².

Medication therapy management (MTM) is a standard practice for assessing patient drug-related needs, and identifying and resolving drug therapy problems. MTM services are pharmacist-provided, non-dispensing services that aim to optimize drug therapy and improve clinical outcomes of patients ³. MTM services are the embodiment of the philosophy of pharmaceutical care. This philosophy asserts that it is the responsibility of the pharmacist to meet all drug-related needs of the patient, to be held accountable for those needs and to assist the patient in achieving his or her therapeutic goals through collaboration with other health professionals^1,3,4. MTM promotes medication adherence, safety, and effectiveness and is useful in improving the patient’s medication experience especially in chronic disease.

A conceptual framework is a meaningfully integrated collection of concepts, concept relationships, and assumptions that broadly explains phenomena of interest, expresses assumptions and reflects a philosophical stance ⁵. Conceptual frameworks are constructed to predict or explain the relationships, events or behaviors of phenomena. They are essential for strengthening theory-driven research whose aim is to investigate phenomena and support generalizability of findings ⁶. There are few examples of substantiated conceptual frameworks for guiding theoretical research in MTM. Consequently, studies within the MTM domain lack shared definitions and are highly heterogeneous with respect to the methods, predictors, and outcomes of interest ⁷. Furthermore, currently available conceptual frameworks such as the MTM Service Model in pharmacy practice³ describe specific dimensions of care delivery rather than specify relationships among concepts that could be scientifically investigated in hypothesis-driven studies. Therefore, it is important to develop new explanatory conceptual frameworks that could be used to strengthen correlational research within the MTM domain.

A common conceptual framework development strategy is theory derivation. In this process, a parent theory is selected and used to guide the development of a new model supported by evidence from current literature^8,9. Theory derivation is useful when related concepts could benefit from a structural representation of their inter-relationships, and when this insight can inform future research^8,10. Conventional theory derivation requires the application of thematic analysis where semantic themes are manually derived by coding literature excerpts and collating the codes into themes. While this is arguably the most logical approach for synthesizing information from text, it is often labor- and time-intensive and is subject to investigator bias ¹¹. Natural language processing is essential for automating knowledge discovery from literature corpora for decision support, guideline development, and medical literature indexing ^12,13. Text mining using natural language processing has the advantage of enabling fast and efficient analysis of large corpora of documents and may be less prone to investigator bias. Combining this approach with network analysis may provide a systematic way quantitatively analyzing and visualizing salient aspects in biomedical literature.

This article describes the first iteration of a theory derivation process for the development of a conceptual framework to guide medication therapy management research. The model, entitled Enhanced Medication Therapy Management Evaluation Framework, is based on the MTM Service Model in pharmacy practice³ and the Chronic Care Model¹⁴ that were selected as parent theories. Concepts incorporated into the enhanced model were extracted from abstracts of systematic review and meta-analysis articles using natural language processing and characterized using network theory. This study reports our current findings.

Methods

The theory derivation process applied in this study is a four-step process: a) selection of parent theory, b) literature search and review, c) semantic concept analysis, and d) content development. These are described further below.

Selection of Parent Theory

We focused on delineating concepts and relationships at the intersection of medication therapy management and chronic disease care. As such, the MTM Service Model in pharmacy practice³ and the Chronic Care Model¹⁴ were selected as parent frameworks. The MTM service model includes five core elements: medication therapy review, personal medication record, medication-related action plan, intervention and/or referral for drug therapy problems, and documentation and follow-up. These core elements were established through consensus by a group of seven national pharmacy organizations in the United States in 2008. They form a framework for the delivery of MTM services in pharmacy practice ³. The MTM service model was used to provide an understanding of the goals, core features, process and outcomes associated with the delivery of MTM services. The chronic care model is an organizing framework for improving the management of chronic diseases. It asserts that comprehensive care improvement requires an approach that incorporates six interdependent components: community resources, health system support, self-management support, delivery system design, decision support and clinical information systems ¹⁴. Researchers have revised this model to demonstrate how eHealth tools can assist patients in managing their own chronic illnesses ¹⁵. The chronic care model was used in this study to provide insight into the role of self-management support, community involvement, healthcare delivery system design, decision support and clinical information systems in chronic disease management.

Literature search and Review

One author (WO) conducted a literature search to identify systematic review and meta-analysis articles that covered the topics related to medication therapy management in chronic care. For inclusion, articles must have reported results about antecedent factors associated with drug therapy problems or intervention strategies for drug therapy problems in chronic disease care. Studies were identified by electronically searching the Ovid Medline® database. The search strategy involved using medical subject headings (MeSH) and search strings associated with constructs of the MTM service model and the chronic care model frameworks (Table 1). The search was limited to abstracts of original or review articles published in English between 2000 and 2016. A total of 285 abstracts were identified out of which 12 duplicates were removed. An additional 131 articles were rejected through title and abstract review leaving 142 articles that were included in the analysis corpus.

Table 1.

Terms and strategy for literature search in Ovid Medline.

Searches		Results
1	(“Medication Therapy Management” or “Medication Reconciliation”).sh.	1374
2	(“MTM” or “DTM” or “drug therapy management” or “drug therapy problems” or “medication action plan” or “medication reconciliation” or “medication review” or “medication therapy management” or “medication therapy review” or “medication-related problem” or “medication- related problems” or “personal medication record” or “pharmaceutical care” or ((drug or medication) and adherence) or “adverse drug reaction” or “adverse reaction” or “adverse drug event” or “adverse reaction” or “adverse event”).ti,ab.	39199
3	(“Self Efficacy” or “Patient Participation” or “Self Care” or “Self-Help Groups” or “Social Support” or “Patient-Centered Care” or “Clinical Decision Support” or “Medical Informatics” or “Electronic Health Records” or “Mobile Applications” or “Cell Phones” or “Telemedicine” or “Text Messaging”).sh.	160439
4	(“Chronic Care Model” or “CCM” or “community health worker” or “frontline health worker” or “frontline worker” or “lay health worker” or “lay worker” or “outreach worker” or “allied health worker” or “patient activation” or “patient empowerment” or “patient participation” or “peer coach” or “peer counselling” or “peer counselor” or “peer education” or “peer educator” or “peer group” or “peer health worker” or “peer mentor” or “peer support worker” or “peer support” or “peer supporter” or “peer worker” or “peer-led” or “self-efficacy” or “selfmanagement” or “village health worker” or “volunteer health worker” or peer?to?peer or self?care or task?redistribution or task?shifting or “android app” or “android phone” or “cell phone” or “cellular phone” or “information communication technology” or “mobile app” or “mobile device” or “mobile technology” or “short message service” or text?messag$ or app or apps or e?health or electronic?health or “handheld computer” or handheld device or “interactive voice response” or IVR or mHealth or MMS or mobile?health or “reminder system” or “reminder” or smartphone or SMS or telecare or telehealth or telemedicine or “wireless technology”).ti,ab.	66962
5	(drug or medication or medicament or treatment or prescription or pharmaceutical or dose or remedy or cure or therapy). af	6711256
6	animal/ not (human/ and animal/)	4159388
7	((1 or 2) and (3 or 4) and 5) not 6	2414
8	limit 7 to (abstracts and english language and “review articles” and yr=“2000 - 2016”)	285

Semantic Concept Analysis

The goal of the semantic concept analysis step was to identify and select the key concepts that could be incorporated into the enhanced conceptual framework. This step involved using natural language processing to extract the concepts reported in the abstracts selected for analysis, computing the frequencies of the identified concepts and concept- concept pairs, and using graph analysis to describe and visualize the resulting information. This analysis was based on the assumption that each biomedical literature abstract consists of a set of concepts that reflect the core contents of what was studied and reported. Consequently, the frequencies of concepts and co-occurrent concept pairs in a corpus of abstracts about a given topic provide a reasonable quantitative description of the relevant themes related to that topic. Concepts in the abstracts selected for the study were extracted using MetaMap and analyzed using graph theory. MetaMap is a natural language processing program that maps biomedical text to the UMLS Metathesaurus¹⁶ Graph theory is the study of graphs - data structures that represent pairwise relationships between discrete objects ¹⁷.

Each of the 142 abstracts selected for the study was preprocessing by removing all non-ASCII characters to enable processing via MetaMap. Next, the MetaMap 2016 Web API was used to process the abstracts. The MetaMap output files were further analyzed in R Statistical Programming Language ¹⁸. All identified concepts were reviewed to eliminate those that were not relevant based on their UMLS semantic types. Relevant concepts that had similar or related meanings such as therapy adherence, adherence to medication regime, and medication compliance were identified and merged after looking up concept-concept relationships in a local installation of the UMLS 2015AB Metathesaurus¹⁶ All unique concept-concept pairs in each abstract were identified and their frequencies across the entire corpus computed. The results of this process were a list of all relevant concepts (nodes/vertices), and a list of all unique pairs of co-occurrent concepts (edges) as well as their frequencies (edge weights). These data were used to construct and visualize a weighted graph of concepts and their relationships in Gephi 0.9.1 ¹⁹. Key concepts were identified by computing the weighted degree centrality and decomposing the graphs at various pairwise co-occurrence frequency cutoffs.

Content development

The final step of the conceptual framework creation was the manual refinement and organization of the concepts identified in the previous steps. This involved the process of concept reduction where abstract constructs were refined to concepts which can be reduced to measurable variables. The Donabedian’s Quality Framework²⁰ was used to organize the entities of the new conceptual framework meaningfully. The Donabedian model is a model for assessing the quality of health. It is based on the assertion that structures of care influence processes of care, and that the latter influence Health outcomes ²⁰. Structures of care describe the physical and organizational aspects such as infrastructure, equipment, financing, and personnel. Processes of care are the resources and mechanisms for carrying out healthcare activities. Health outcomes are the goals and effects of healthcare on patients such as recovery, survival, and patient satisfaction.

Results

Semantic Concept Analysis Results

Table 2 provides the summary statistics of the concepts, concept pairs and graph analyzed of the 142 abstracts included in the study. A total of 27,094 text tokens in the corpus were mapped to 4,090 unique UMLS concepts. Of these, only 50 concepts (1.2%) were considered relevant and included. A majority of the excluded concepts resulted from the one-to-many mapping of text tokens including stop words to multiple concepts. Medication adherence was the most prevalent concept in the corpus, occurring in 93% of the abstracts. Other high-frequency concepts include selfmanagement, tailored intervention and patient engagement (Table 3). These highly frequent concepts also formed the most frequent concept-concept pairs in the corpus (Table 4).

Table 2.

Descriptive statistics of the concepts, concept pairs, and graph analyzed in the study.

Description	Statistic
Concepts
Total number of text to concept mappings in corpus (n)	27,094
Number of text to concept mappings per abstract (Median [Range])	172 [56-626]
Total number of unique concepts in corpus(n)	4090
Number of unique concepts per abstract (Median [Range])	149 [56-501]
Total number of unique concepts in corpus included (n(%))	50(1.2%)
Number of unique concepts included per abstract (Median [Range])	4 [2-13]
Concept pairs
Total number of unique concept pairs in corpus included in graph analysis (n)	102
Number of unique concept pairs included per abstract (Median [Range])	3 [1-14]
Graph
Number of nodes (n)	50
Number of edges (n)	102
Weighted Degree Centrality (Mean [Range])	18.64 [1-388]
Graph Density (%)	8.3%

Table 3.

Top 10 concepts realized from graph analysis and ranked by weighted degree centrality (Concept ID is UMLS Unique Concept Identifier or CUI)

Concept ID	UMLS Preferred Name	% Occurrence in Abstracts	Weighted Degree Centrality	Degree Centrality
1.	C1171369	Adherence	93.7%	388	47
2.	C0086969	Self-Management	42.6%	24	13
3.	C2986593	Tailored Intervention	24.6%	22	11
4.	C0162648	Telemedicine	17.6%	22	12
5.	C3508152	Patient Engagement	21.8%	10	8
6.	C0030688	Patient education	19.0%	10	3
7.	C0546816	Persistence	4.2%	10	8
8.	C0600564	Self Efficacy	13.4%	5	4
9.	C0877248	Adverse event	9.9%	5	4
10.	C0517785	Medication Knowledge	4.2%	4	4

Table 4.

Top 10 concept pairs ranked by percent proportion of occurrence in the analyzed abstracts (Concept ID is UMLS Unique Concept Identifier or CUI)

	Concept 1		Concept 2
	Concept ID	Preferred Name	Concept ID	Preferred Name	% Occurrence in Abstracts
1.	C0086969	Self-Management	C1171369	Adherence	36.6%
2.	C1171369	Adherence	C2986593	Tailored Intervention	18.3%
3.	C0030688	Patient education	C1171369	Adherence	16.9%
4.	C1171369	Adherence	C3508152	Patient Engagement	16.9%
5.	C0237125	Medication regimen	C1171369	Adherence	16.2%
6.	C0085519	Reminder Systems	C1171369	Adherence	14.8%
7.	C0162648	Telemedicine	C1171369	Adherence	13.4%
8.	C0814098	health-related beliefs	C1171369	Adherence	13.4%
9.	C0037438	Social support	C1171369	Adherence	12.7%
10.	C0600564	Self Efficacy	C1171369	Adherence	10.6%

The network graph identified and analyzed in the study had 50 nodes and 102 edges (Figure 1). It is a sparse undirected graph with a density of 8.3%. The most dominant nodes and edges in the graph are highlighted in blue and further described in Tables 3 and 4 respectively. Medication adherence, connected to 47 nodes, is the most central node in the network. 19 nodes (38% of the full graph) are one-degree vertices connected to the medication adherence node. Its removal from the network results in a much smaller graph with 32 nodes and 55 edges. The other dominant nodes are closely related to the medication adherence as contributors or interventions to medication non-adherence. Interestingly, no other drug-related categories were identified in the network. The adverse drug event concept identified in the network analysis was discussed in the reviewed abstracts as an antecedent contributor to medication non-adherence rather than a drug therapy problem per se.

Figure 1.

Network graph of concepts identified from literature abstracts on drug therapy problems. The graph has a total of 50 nodes and 102 edges. Nodes represent UMLS concepts. Larger nodes have higher weighted degree centrality. Edges represent the frequency of co-occurrence of UMLS concept pairs across the document corpus. Thicker edges correspond to co-occurrence frequency. The top 10 nodes and edges, based on weighted degree centrality, are highlighted in blue.

Conceptual Model Development

The enhanced model developed in this study meaningfully organizes the synthesized information in a manner that clearly shows the relationships and fluidity of the constructs and concepts about the application of medication therapy management for addressing drug therapy problems as illustrated in Figure 2. The model consists of 65 concepts clustered into 14 constructs that are further organized into structural, process and outcome constructs. Structural constructs are shown in blue, process constructs in orange and outcome constructs in green. The constructs are derived from the parent theories and represent clusters of concepts that may contribute to drug therapy problems that were identified by network analysis and abstract review. The process constructs represent actors, activities, drug therapy problems and critical measurements associated with MTM service provision. The outcome constructs represent the key process, intermediate and clinical outcomes related to MTM service delivery. Because the concepts in this domain are highly heterogeneous, specific concept to concept relationships are not shown, and concepts have not been reduced to measurable variables.

Figure 2.

The Enhanced Medication Therapy Management Evaluation Framework. Structure entities are shown in blue, process entities in orange and outcomes in green. Arrows indicate fluidity between structure, process and outcome constructs.

Discussion

This paper describes the first iteration of a process used to develop a conceptual framework to guide a program of research in medication therapy management (MTM). We selected the MTM Service Model in pharmacy practice³ and the Chronic Care Model¹⁴ as parent theories and enhanced these through theory derivation ⁸. Content included in the enhanced model were extracted from abstracts of review articles using natural language processing and analyzed using network theory. The derived model was organized according to the structure, process and outcome constructs of the Donabedian Model ²⁰. The framework requires additional elaboration as specific concept-concept relationships are currently not described.

Medication adherence was the most studied drug therapy problem. This may be attributed to the fact that medication adherence is often considered the single most significant predictor of treatment success for chronic diseases ²¹. Indeed, an estimated 70% of hospital admissions due to drug therapy problems are related to medication adherence ²². Our findings confirm the fact that adherence is a complex phenomenon influenced by a variety of heterogeneous factors, and that there appears to be no consensus on the most practical approaches for measuring and promoting adherence in chronic diseases ^23–25. However, providing patient-centered care and enhancing patient self-management through patient engagement, education, and psychosocial support are accepted strategies for improving adherence.

Our review and model development highlights several knowledge gaps in the literature. First, clearer definitions of factors associated with drug therapy problems are needed. This would help researchers to formulate research questions and evaluate relationships in this domain more effectively. Second, there are knowledge gaps about factors associated with medication indication, effectiveness, and safety. More research on these drug therapy problems is needed to demonstrate their importance and the role of MTM in addressing them. Third, researchers have not adequately examined the process of task redistribution in the provision of MTM services. Task redistribution is the appropriate delegation of health service responsibilities to less specialized but more readily available health professionals, and in some cases non-professionals ^26–29 Studies have shown that task-redistribution enhances patient engagement and selfefficacy through education and psychosocial support across different populations and intervention modalities ^30–32. However, its role in MTM service provision remains unstudied. In high-resource settings where MTM services are primarily provided by pharmacists, the role of task-redistribution in MTM services is not adequately described.

Conversely, in resource-poor settings where task redistribution is popular, little is known of the standard practice of MTM. The application of mobile health (mHealth) technology to complement this approach by improving recording keeping and decision support has been shown to increase the quality and efficiency of this strategy ^33,34. Lastly, while acknowledging the importance of complementing drug therapy problem intervention strategies with reminder tools and decision to support adherence, ^35,36 further informatics research is needed on the role these technology solutions in operationalizing of MTM service activities such as care coordination and eligibility determination.

Two key challenges should be expected when using natural language processing and network analysis to extract automatically and characterize concepts during conceptual framework development. First, the effectiveness of this approach is limited by currently available natural language processing technologies. For example, while natural language processing tools such as MetaMap can effectively automate the extraction of concepts from biomedical texts, technologies that could effectively extract semantic relationships between concepts are currently available. It is, therefore, necessary to rely on surrogate approaches such as pairwise concept co-occurrence analysis to identifying and quantify the strength of plausible relationships between concepts reported in biomedical texts. Second, natural language processing tools such as MetaMap often map idiosyncratic texts including stop words, acronyms, and abbreviations to concepts ¹². Therefore, it may be necessary to review the extracted concepts to eliminate those that are not relevant. Because this review is manual, it may not be useful to scale the approach to larger numbers of biomedical texts.

There are several limitations in this study. First, due to pragmatic constraints associated with using MetaMap, we restricted our concept extraction to published abstracts of systematic reviews and meta-analyses indexed in Ovid Medline. This was done with the assumption that theoretical saturation would be achieved as long as a critical mass of literature texts were analyzed. Using full texts, including other types of published articles or incorporating articles from other bibliographic databases could have resulted in richer content, but would imply additional computation and manual review. Second, we relied on pairwise concept co-occurrence analysis as a surrogate measure of the relationships between concepts occurring in the analyzed abstracts. However, doing so does not give an accurate interpretation of the true strength of the relationship between co-occurring concepts in the analyzed abstracts. For example, a pair of concepts that occur in the same abstract may not be related. Conversely, just because two concepts do not co-occur together does not imply that they are not related directly or indirectly. We lost information about the directionality of plausible concept-concept relationships. Nonetheless, our goal at this stage was not to establish the true relationships between concepts, but to quantify what biomedical researchers have been interested in studying within the MTM domain. Our next steps are to refine the framework further by characterizing the interrelationships among the individual concepts and describing the measurable variables associated with these concepts. This will be followed by the use of domain experts to evaluate the framework extensively to determine its relevance across a variety of settings including in underserved settings.

Conclusions

This paper describes the use of natural language processing and network analysis for the development of a conceptual framework to guide a program of research in medication therapy management. The developed model consists of 65 concepts clustered into 14 constructs derived from the MTM service model and the chronic care model. It requires further refinement and evaluation to determine its applicability across a broad audience including underserved settings. Our analysis confirms that medication adherence is the most studied drug therapy problem. Patient-centered care approaches and enhancing patient self-management through engagement, education, regimen simplification and psychosocial support are important strategies for promoting adherence. Drug therapy problems associated with medication indication, safety and effectiveness require more studies. Further research is also needed to understand the roles of task redistribution and medical informatics interventions in operationalizing MTM activities. We recommend that researchers evaluating tools to address drug therapy problems should consider the fact that the efficacy of such tools could be influenced by a multitude of complex factors. In situations where controlled experiments cannot be carried out, these factors should be investigated as potential confounders and effect measure modifiers.