Purpose
This study introduces information value chain analysis by identifying essential information for use in gout care management. Part I reviews the essential concepts of information value chain analysis first introduced by Porter. Part II applies the analysis to determine the information values of patient health information and explores ways in which health information technologies can be best utilized to provide that information to patients with gout.
Methods We combined value chain analysis with natural language processing and machine learning techniques to develop algorithms that can identify patients with gout flares using clinical notes. As one of the first signs that the disease was not being controlled, variables found to be associated with gout flares were considered valuable information for patients with gout.
Results The best performing model, in terms of both gout flare prediction and association identification, was the comprehensive model that not only included concepts from all stages of the value chain but also designated natural language processing concepts from every care stage as surrogate variables. Additionally, all administrative codes traditionally associated with gout and its treatment were included as surrogate outcome variables.
Conclusion This study introduced information value chain analysis and applied it to develop a computer-based method with theoretical underpinnings to identify the concepts associated with gout flares. The findings can be used as a starting point for filtering the vast amounts of information patients must go through and identifying the most valuable information for patient with gout to adequately manage their symptoms.
,1
and Sujin Kim2
This study introduces information value chain analysis by identifying essential information for use in gout care management. Part I reviews the essential concepts of information value chain analysis first introduced by Porter. Part II applies the analysis to determine the information values of patient health information and explores ways in which health information technologies can be best utilized to provide that information to patients with gout.
We combined value chain analysis with natural language processing and machine learning techniques to develop algorithms that can identify patients with gout flares using clinical notes. As one of the first signs that the disease was not being controlled, variables found to be associated with gout flares were considered valuable information for patients with gout.
The best performing model, in terms of both gout flare prediction and association identification, was the comprehensive model that not only included concepts from all stages of the value chain but also designated natural language processing concepts from every care stage as surrogate variables. Additionally, all administrative codes traditionally associated with gout and its treatment were included as surrogate outcome variables.
This study introduced information value chain analysis and applied it to develop a computer-based method with theoretical underpinnings to identify the concepts associated with gout flares. The findings can be used as a starting point for filtering the vast amounts of information patients must go through and identifying the most valuable information for patient with gout to adequately manage their symptoms.
Michael Porter first introduced the concept of a value chain as the set of activities that an industry-specific firm uses to create value in the market. Value was originally defined as the amount that consumers are willing to pay for a product or service. Thus, it was measured by the revenue generated. To encompass the total value, the framework employs a process view of organizations in which firms are conceptualized as systems composed of subsystems, each with their own sets of inputs, transformation processes, and outputs [1]. Adopted from Porter's value chain theory, we developed an information value chain framework for use in chronic care management in the context of patient education.
This study is divided into two main parts. Part I reviews Porter's theory within the context of chronic care management applied to gout as an example. Section II introduces the application of the information value chain. In our newly published paper [2], we assessed the appropriate content and time of delivery in a research framework by examining information values at different stages of the care continuum within the context of systemic lupus erythematosus management. The previous study identified essential information to be recommended for different stages of systemic lupus erythematosus management. In this study, we expanded the previous approach by utilizing machine learning and clinical natural language processing methods to identify essential gout information in individual value chains.
Originally applied to the manufacturing context, Porter theorized the myriad activities that organizations must perform to transform inputs into outputs as primary or secondary activities. Primary activities are considered essential for creating value and competitive advantage. Secondary activities are undertaken to support and improve the primary activities. Value activities are identified and classified according to their technological and strategic distinctions. Thus, Porter proposed that primary activities consist of five generic components (inbound logistics, operations, outbound logistics, marketing and sales, and service), with the idea that every discrete activity that a firm engages in can be categorized to identify areas of competitive advantage. Support activities can be similarly classified into four broad categories (procurement, infrastructure, technological development, and human resource management), with the proposition that increased efficiency in any of the four will lead to beneficial results in at least one of the primary components. All activities contribute to a firm's profit margin, which is determined by the extent to which the value exceeds the costs [1].
Porter's [1] original assertion that industry-level and sector-wide value chains would be too broad to parse outsources of competitive advantage did not deter researchers. Buttigieg, Schuetz, and Bezzina [3] employed qualitative research methods to create and compare value chains for public and private healthcare sectors in Malta. This analysis was used to evaluate the feasibility of recommendations to maximize the provision of healthcare services. In addition, the Wharton School of Business at Harvard University developed a value chain for healthcare as a business featuring a conventional approach to value chain analysis, which involves mapping the activities involved in delivering a good or service [4].
The previous examples highlighted the complexities inherent in healthcare and demonstrated the potential applications of the value chain to identify ways to successfully utilize information technologies to address this complexity. Much work remains to be done, not the least of which is validating the previously proposed models. From a patient's perspective, lab tests by themselves have no value, although they still must pay for them. A test's value is in being able to connect the patient to needed treatment. Even treatment is not the goal but patient well-being. Thus, to analyze the value of healthcare delivery, the focus must shift from the products and services provided to the outcomes achieved. In this manner, the value for health outcomes cannot be attributed to a specific intervention at one point in time but must be redefined as the total outcomes achieved per cost over the total cycle of care. According to Kim, Farmer, and Porter [5], primary value is created by delivering care for specific medical conditions, and they created the care delivery value chain. This emphasizes the interrelated nature of healthcare and highlights the need for integrated care.
Porter [1] posited that his generic value chain could be used to compare and differentiate competing businesses. While the value added by each category may vary by industry, each component is considered integral in distinguishing competitive advantage to some extent. Therefore, although the model can be applied to any industry, including services, it is only relevant at the business unit level. In other words, value chain analysis should be performed on competing firms in the same industry. However, much work has been done to expand the model in the intervening decades, most of which has been achieved by Porter himself.
With the advent of the Internet, Rayport and Svlioka [6] argued that every business must compete in both physical and virtual worlds and that the processes for creating value in each are not the same. Porter's value chain views information in a supporting role, and managers often use data obtained from primary activities to improve those processes. However, this does not account for cases in which information explicitly adds value. In the virtual world, it is common for the information to be distributed. Rayport and Svlioka [6] posit five information-related activities required for value creation: gathering, organizing, selecting, synthesizing, and distributing. By analyzing the incorporation of these five activities into an organization's primary activities, businesses can identify where and how information and communication technologies can best be utilized for value creation. Each value-adding process for information can be combined with a generic value chain to create a value chain matrix to identify linkages and aid effective management.
Gout is an inflammatory disease caused by high levels of uric acid in blood [7]. It is the most prevalent form of inflammatory arthritis and continues to increase [8]. The most common symptom is gout flares, which occur when uric acid crystallizes into needle-like formations within joints [9]. Diagnostic codes have proven to be a poor source of flare identification in patients [10]. Furthermore, there is no standard definition for symptoms, leading researchers to search for better identification methods using clinical natural language processing [11].
Early research demonstrated that patients with electronic health record access to their health information are more involved in their health care, and this involvement can lead to long-term benefits, especially for low-income patients [12]. However, patients with gout have thus far been unable to consistently leverage available information to adequately improve health outcomes. Although much research has been conducted on related topics, particularly regarding patient education for effective treatment options, the information that is valuable to patients with gout throughout the continuum of care remains largely unanswered. The specific problem is that patients have access to an overwhelming amount of information without the knowledge or guidance to sort through, manage, or otherwise utilize that information to improve their health [13]. A knowledge gap exists regarding what personal health information is required to optimize care.
The Health Information Technology for Economic and Clinical Health Act of 2009 caused an upsurge in the adoption of electronic health records, and researchers rapidly moved to take advantage of this data source [14]. Early studies largely focused on usability [15] or structured data for risk prediction modelling [16]. However, data that are most useful for documentation and communication are usually available in an unstructured, free-text format [17]. Nevertheless, clinical natural language processing techniques for information extraction remain underutilized in clinical and translational research [18].
Part II in this current study sought to understand what information helps patients with gouts engage in long-term self-management using value chain analysis. As most previous work on patient information has focused on creating education and medication interventions for gout treatment, there is a knowledge gap regarding other types of information that might aid in wellness goals, including the information that is already available in electronic health records that can be used for this purpose. Specifically, this study aimed to use clinical natural language processing techniques to determine the prioritization of patient health information and explore ways in which health information technologies can be best utilized to provide that information for patients with gout. Utilizing a modified version of the surrogate-assisted feature extraction procedure outlined by Zhang et al. [19] and Porter's care delivery value chain, emphasis was placed on the stages at which information was most useful.
Part II of this study included patients with gout with available data from the University of Kentucky Healthcare electronic health record systems. The University of Kentucky Healthcare is a large academic tertiary referral center that contains detailed records of more than 1 million patients since 2004 and is stored in an electronic data warehouse with a relational database structure readily available for research purposes. This study was approved by the University of Kentucky Internal Review Board (IRB). First, we identified the study population and retrieved notes. We then used value chain analysis to develop annotation criteria for creating a beta dictionary. An algorithm was applied to counts extracted from structured electronic health record data to identify gout flares and to refine and develop a machine learning algorithm.
Our research first identified potential gout cases using two distinct types of electronic health record data. First, we searched for structured data from the electronic health records at the University of Kentucky Healthcare Systems. We screened the structured electronic health record data to create a highly sensitive dataset containing all potential patients with gout. The gout data mart consisted of all patients with ≥1 International Classification of Diseases (ICD) codes for gout (274.9, M10.0, M10.2, M10.3, M10.4, M10.9). Patients aged <18 years at the time of their first ICD code were excluded. Our preliminary search of the electronic health records of the University of Kentucky Healthcare Systems yielded 5,590 potential patients. For unstructured electronic health records, we obtained narrative data from all types of commonly available clinical notes. These included outpatient notes, rheumatology notes, discharge summaries, radiology, and pathology reports. For our purposes, a note was defined as any type of healthcare narrative containing more than 500 characters. Pre-processed narrative data were analyzed using a clinical natural language processing software called Clinical Language Annotation, Modeling, and Processing Toolkit to extract clinical variables at the patient level for patients with more than two notes (N=3,964) to ensure they have enough documentation for classification.
This study utilized a previously published gout information value chain developed using a systematic review of the relevant literature [20]. Based on this value chain, we created a comprehensive list of variables for the gout flare phenotyping algorithm classified by the six divisions of the care delivery value chain to create a gout value chain. The primary care cycle represents information believed to provide value. For each stage of the value chain, one category of information values was designated as a priority to indicate the minimum information a patient would need to navigate their treatment. Thus, if a patient could receive only one type of information from each care stage, information related to the categories identified as priority would be the most valuable.
Whenever possible, the list of terms was converted to structured data that were readily available from our electronic health records. Along with diagnostic codes as forms of ICD-9 or ICD 10 claims codes, the codes include current procedural terminology codes for procedural claims, national drug codes for electronic medication prescriptions, and logical observation identifier names and codes for laboratory tests. In addition, encounter records further identified the dates of outpatient visits and inpatient stay. Our entity extraction pipeline included a sentence splitter, context-sensitive tokenizer, part-of-speech tagger, section identifier (patient history, family history, etc.), a named entity recognition module created using machine learning algorithms (brown clustering, word embedding, etc.), and terminology/ontology mapping from the Unified Medical Language System (UMLS). Thus, this pipeline provided us with all the information required to further classify the extracted concepts according to our value chain. Figure 1 shows the flow of the study adopted from Surrogate-Assisted Feature selection (SAFE) papers [14, 16].
Figure 1
Study flow chart modified from SAFE.
Text articles describing gout were identified from publicly available knowledge sources, including Wikipedia, Medscape (eMedicine), Mayo Clinic, MedlinePlus, the American College of Rheumatology, WebMD, and UpToDate. The researchers then used the gout value chain above as criteria for annotating the information deemed important for gout patients contained in each article according to the stages of care delivery. The small number of articles included in this study (N=7) did not provide enough data to produce a model capable of reliably extracting the desired information from future documents. Therefore, after applying the poorly performing model to the total set of articles, the extracted terms were used to develop the dictionary. The output from this model provided a list of 494 terms that could be used as a beta dictionary. The list was reviewed and revised to exclude overly generic (e.g., symptoms and effects) and nonuseful terms (e.g., and, no). The resulting 450 terms constitute the final dictionary. This dictionary was then applied to our text articles and the resulting annotations were used to create a much better model. Terms were not case sensitive, and stemming was used to ensure that variations in terms (such as different tenses or plurals) were included as much as possible. All models were created using 5-fold cross-validation.
The final "gout value chain" model was used to process all available clinical notes (N=310,519) for our gout cohort. Each mention of a concept per care phase is counted at the patient level. As terms and their associated concept unique identifiers could appear labeled as more than one phase of the value chain, the concept unique identifiers for each extracted term were first aggregated by the care delivery value chain phase before occurrences were counted for each patient using a Perl script. For example, a symptom might be labeled as diagnosing if it serves as a diagnostic criterion for the disease or intervening if it is a side effect of the medication used to treat the disease. It could also be included in any of the other phases, depending on the context in which it was used in the clinical note. Therefore, multiple instances of the concept unique identifiers in the screening phase are aggregated and counted as "symptom concept unique identifiers; screening".
The same concept appearing in other phases would be similarly aggregated and counted according to the number of times it appears for each patient with the associated phase label. In addition to natural language processing concepts, we included the frequency of codified data readily extracted from structured electronic health record data using MySQL as candidate features. These features include the phenotype for gout, competing diagnoses (such as rheumatoid arthritis), and medications (allopurinol, colchicine, and febuxostat) that were determined by our clinical domain expert to be relevant to the target phenotype. The identified terms and concepts were then mapped to their associated diagnoses (ICD-9/10) and national drug codes and counted for each patient. We also counted the total number of unique billing codes (ICD), healthcare visits, and clinical documents per patient as potential measures of patient healthcare utilization to round out our candidate feature set.
Gout flares were identified and counted using an adapted version of a previously published algorithm [21]. Using a Perl script, patients in our cohort were classified as having a flare if they had either a diagnosis code for gout in the system (ICD-9,274. xx; ICD-10: M10.0x, M10.1x, M10.2x, M10.3x, M10.4x, M10.9x) followed by at least one code for any medication (colchicine, nonsteroidal anti-inflammatory drugs, corticosteroids) or procedure (radiograph of a joint, magnetic resonance imaging for any part of an extremity, joint aspiration, microscopic examination of joint fluid, urate testing) commonly prescribed for gout flares within seven days or a medical claim with a diagnosis code for joint pain (ICD-9:719.4x; ICD-10: M25.5x) followed by a code for colchicine within seven days. Consistent with prior studies, care for gout flares was expected to last for 30 days. Therefore, any code indicating repeated flares within that time frame was counted as one flare, with the date associated with the first applicable diagnosis code considered the flare start date.
To train and evaluate algorithms to predict gout flares, this study utilized an available R package, a Common semi-supervised Approach for Phenotyping (PheCAP), that allows the use of various classification methods to identify patients with a specific disease or condition [19]. Using gout flares as the outcome variable, the PheCAP algorithm allowed us to identify the candidate features that were most closely associated with flares at each stage of care delivery. To this end, we first identified surrogate variables to act as "silver standard labels". These features are believed to be highly predictive of textbook cases of a condition. For this study, surrogate variables were chosen using the concepts identified as priorities in our gout value chain. For instance, any concepts identified as specifying or ordering labs or tests are used as surrogate variables for the diagnosis stage. This applies to both natural language processing concepts extracted from clinical documents and structured codes identified as candidate features. These features were then used as response variables to select potential features using penalized logistic regression for the final algorithm training. The selected features were then used to train models for each care stage using adaptive Least Absolute Shrinkage and Selection Operator (LASSO) penalized regression to identify features associated with screening, diagnosis, preparation, rehabilitation, and monitoring. A comprehensive model was developed in the same manner to identify features pertinent to patients with gout over the entire care continuum. The performance of each model was evaluated and compared using the Area Under the receiver operator characteristic Curve (AUC) along with the false positive rate, true positive rate, positive predictive value, negative predictive value, and weighted average of precision and recall, denoted as F1 scores.
Data from 3,964 patients were included in the study. Additionally, 71.3% of the population was male (N=2,827) and 28.7% were female (N=1,137). All but 15 patients had race-related information available in the database. Further, 82.6% of the sample were identified as white (N=3,276), 15.8% as black/African American (N=627), and 1.0% as Asian (N=40). Two patients were identified as Hawaiian/Pacific Islander and four as American Indian/Alaskan. Additionally, 98.0% of the patients were designated as non-Hispanic/Latino. The median age of the study cohort was 59 years, with most patients aged between 55 and 69 years. The model successfully retrieved 3,892 terms and phrases that could be mapped to concept-unique identifiers in unified medical language systems. Many terms, although potentially medically relevant, are not necessarily related to gout ("history of adopted child"). Therefore, we set a threshold wherein a concept needs to be extracted from the notes of at least four patients to be included in the final analysis. This was performed to minimize noise and maximize the sample size in the final analyses. In total, 2,020 terms appeared in only one patient. A further 521 terms appeared in only two patient records. An additional 252 terms were removed to appear in only three patient records. The remaining 1,094 terms were examined for association with gout flares in the next step of the project. In the screening stage, 329 potential variables were identified. The priority information for this stage was signs and symptoms.
Therefore, "high fever", "warm skin", and "during the night" were chosen as natural language processing surrogates. ICD codes were associated with joint pain (719.4x, M25.5x). As one of the most common symptoms of gout, the general, unspecified versions of these codes were also used as surrogates. The model developed using these variables as surrogates did not perform well (training AUC=.63), although it performed slightly better on the validation dataset (AUC=.65). None of the natural language processing concepts' unique identifiers contributed to the model by themselves. As a group, the concept of unique identifiers and ICD codes showed a weak association (all beta coefficients <±.1). The model's .5 F1 score (.46) occurred at the lowest cutoff (.74). The positive predictive value (.78) was the lowest at this point. While all other metrics were the highest at this cutoff, all were <.50.
The model for the diagnosis stage was developed using the concept of unique identifiers and current procedural terminology codes used for laboratory tests involved in diagnosing gout. Specifically, "fluid analysis" and "urate crystals" were used as natural language processing surrogates, and the current procedural terminology code for serum urate testing was used as the surrogate for coded data. A total of 238 variables were considered for diagnosis. This model performed similarly to the screening model (training AUC=.64, validation AUC=.65). However, the code for serum urate testing demonstrated a strong association with the outcome variables, with a beta coefficient >.65. Furthermore, the natural language processing variables had a beta coefficient of >8.18.
The preparation stage contained only eight natural language processing variables. The priority information for this stage of the value chain includes all terms related to education concerning treatment options. Although this category includes information about side effects, compliance, and efficacy, it is not limited to medication or drug interventions. Thus, "red meat", "taking nonsteroidal anti-inflammatory drugs", and "kidney" were chosen as natural language processing surrogates. Because nonsteroidal anti-inflammatory drugs are likely to be the first-line treatment for gout flares that patients can take while deciding on long-term treatment options, the national drug codes for a common over-the-counter treatment (ibuprofen) were also designated as surrogate variables. In total, 121 possible variables were considered in this model. The model performed similarly to previous stages. Ibuprofen had the greatest effect in this model. "Red meat" was the only natural language processing variable to have an individual effect on the model.
For the intervention, there were 297 variables. Priority information for the intervening stage of the care cycle consists of ordering and administering drug therapy. The natural language processing concepts chosen as surrogates were all drugs commonly prescribed for gout treatment, including both acute attacks and long-term urate-lowering therapy. The national drug codes for these drugs were also included as surrogate outcome variables for 46 surrogates for consideration in surrogate-assisted feature extraction. This model performed better (training AUC=.73). Sulfinpyrazone showed a strong association with the model, with a beta coefficient of >.50.
The recovery stage provided only three natural language processing concepts. Priority information is related to fine-tuning of therapies. Two concepts were selected as related to this designation ("daily" and "renal impairment"). ICD codes for renal disease were included as surrogate representations of the structured data. The model was developed by using 111 variables. The performance metrics for this model were slightly lower than those previously presented (validation AU C=.64). The main variables all demonstrated small associations with gout flares; however, the natural language processing concept of "renal impairment" alone showed a beta coefficient of -96.27.
The monitoring stage presented 647 variables for developing the model. The priority for monitoring is to manage and avoid complications, particularly acute attacks. From a coded data perspective, avoidance is handled through the continued monitoring of serum urate levels. Management of an acute attack frequently consists of using colchicine to alleviate symptoms. Therefore, the current procedural terminology and national drug codes were used as surrogate variables. A total of 19 natural language processing variables related to priority information for this stage were selected to act as surrogates. This model performed slightly worse than the intervening model (training AUC=.72). The concept of unique identifiers for "fluid intake" and "fish oil" stands out as strongly associated with gout flares.
Finally, a model encompassing the entire value chain was developed. While the first model was meant to be a general model to look for associations between broadly defined surrogates, this model was designed to look specifically at each stage of the entire value chain. The natural language processing surrogates defined for use in all previous models were also assigned as surrogate outcome variables in this model. For medical codes, all codes used in the algorithm to count flares were included as surrogates. A total of 1,207 variables were considered for the development of this model, 112 of which were designated as surrogates. The model created improved performance metrics compared with the other models. AUC for training data was .81. At the lowest threshold value (.67), positive predictive value was at it s lowest value of .88. All the other statistics were at their highest, with an F1 score of .76.
The results of this study in Part II show that the stages of the value chain are fluid, with patients potentially moving back and forth multiple times through the chain. In particular, it is difficult to determine which care stage a patient is in or which stage information is most applicable to the patient. This is reflected throughout the value chain, patient education materials, clinical note output, and models created to identify gout flares and associated concepts. The gout information value chain was developed using a collection of studies concerning the information requirements of patients with gout and best practices for treatment and self-management. For a variety of reasons, healthcare providers do not always implement best practices. The repetition of topics throughout the value chain indicates that information might best be given multiple times in the care cycle. It is likely that patients might need slightly different information about the same topic depending on the care stage. However, patients are usually provided with only one worksheet of basic gout information when they are first diagnosed with the condition. Much of the information identified in the gout value chain is not included in patient education materials or clinical notes, nor are they told about the information by any of their healthcare teams.
The British Society for Rheumatology Guideline for the Management of Gout specifically advocates the provision of written and verbal instructions for most of the information included in our value chain [22]. Research routinely shows that patients do not recall receiving educational information [23], and our study demonstrates that many concepts related to necessary information are not readily available in patient education materials or patient notes. When using concepts collected from patient notes to develop models for each stage of the value chain, there was a noticeable lack of available data for several care phases. The effects of this on algorithm creation for each stage of the care cycle revealed possible implications for clinical practice. There are plenty of available natural language processing concepts for screening; however, when searching for priority information to use as surrogate variables, none of the symptoms that are considered among the first signs of gout (painful joints, redness, swelling) were included in the screening data.
These concepts appeared in diagnosis and monitoring data. The concepts were annotated as screening when creating the keyword dictionary; however, the concepts were used more frequently in the context of diagnosing or monitoring within patient education materials. The named entity recognition model subsequently learned to label these signs and symptoms as one of those stages. Consequently, natural language screening was proliferated with an over-abundance of terms related to family history and risk factors. Similarly, priority information for diagnosis is related to the ordering and administration of laboratory tests used to diagnose gout. Synovial fluid analysis is considered the gold standard for diagnosing gout [22]. There was a concept similar to this (fluid analysis), which was used as a surrogate variable for this model; however, the diagnostic data included more information on possible signs and symptoms. This is in line with research that states that clinicians generally use signs and symptoms for gout diagnosis, and screening mostly takes place outside of the healthcare cycle [5].
The lack of concepts found in the preparation stage and abundance of terms labeled intervening support research highlights the focus on treatment in healthcare, with the preparation stage largely being skipped [22]. Recovering similarly had a low number of concepts, and research strongly indicates that this stage is mostly ignored in gout because patients do not realize that urate-lowering therapy can often set off flares as a side effect. This can be remedied by adjusting the medication; however, many patients stop the treatment [22]. The preponderance of the terms found for monitoring is likely similar to that of intervening.
Information is critical for self-management of patients with gout, especially for patient education by professional nurses. Many factors are involved in ensuring that patients with gout have the right information to make informed decisions about their healthcare. This is the first study to use value chain analysis to develop a named entity recognition model. Although the models created in this study demonstrated only marginal success, the results highlight several interesting findings. This study validates the gaps in the information perceived by patients and provides us with particular areas to start filling in. Nevertheless, it is also important to determining how to fill these gaps. The results of this study indicate that patient data do not necessarily contain the required information. The weak associations could be used to identify stronger connections to gout flares and more useful information to provide patients with access to important information tailored to their needs, thereby optimizing their healthcare. Future studies should focus on identifying patients' information needs throughout the care continuum.
CONFLICTS OF INTEREST:The authors declared no conflict of interest.
AUTHORSHIP:
Study conception and design acquisition - RM and KS.
Data collection - RM and KS.
Analysis and interpretation of the data - RM and KS.
Drafting and critical revision of the manuscript - RM and KS.
This article is an extended study based on the first author's (Dr. Maranda Russell) doctoral dissertation from University of Kentucky (Lexington, KY, USA).
University of Kentucky College of Medicine' internal faculty research fund was partially used in this research.
E-SUBMISSION
