Discovering major opioid-related research themes over time: A text mining technique

Abstract

The recent health crisis concerning opioid overdose has prompted watershed levels of publications. This study explores how opioid-related research themes have changed since 2000, using a text mining technique. The textual data were obtained from PubMed, and the research periods were divided into three periods. While a few topics appear throughout each period, many new health problems emerged as opioid abuse problems magnified. Topics such as HIV, methadone maintenance treatment, and world health organization appear consistently but diminish over time, while topics like injecting drugs, neonatal abstinence syndrome, and public health concerns are rapidly increasing. Recent widespread opioid abuse problems led to new research topics, including prescription drug monitoring programs, veteran’s health issue, posttraumatic stress disorder, HCV, opioid-related deaths, and emergency department visits. Examination of scholarly publications reveals that the expanded use of opioids worsened opioid abuse problems and accelerated the emergence of new health problems.

Introduction

The mortality rate of opioid overdose has increased by 402% from 2000 to 2016¹. In 2016, 11.5 million Americans misused prescription opioids and 42,249 died from overdosing on opioids, totaling to over 115 deaths a day^2,3. Additionally, opioid overdose has resulted in considerable health complications, prompting the Department of Health and Human Services to declare opioid overdose as a public health emergency². In accordance with the heightened health crisis, academia responded to the epidemic through publications. Because academic research addresses the major health problems of the time, investigating the trends of opioid-related research themes will reveal the critical health concerns of those time periods. As an example, in the early 2000s opioid-related research focused mainly on opioid treatment experimentation, and opioids at the time were primarily prescribed for cancer pain. Recently, the opioid prescription has become much more inclusive, including non-medical use in addition to cancer and non-cancer pain. As a result, deaths from opioid overdose and opioid abuse problems have resulted in a public health emergency. A few examples of the consequential health problems from opioid abuse, in recent years, include human immunodeficiency virus (HIV), hepatitis C virus (HCV), veterans’ health problems, deaths, injecting drugs, and neonatal abstinence syndrome.

Opioid overdose-related health problems have expanded over time. Despite the rapidly increasing opioid induced health problems, little to no studies have systematically investigated such changes. As such, there is very little knowledge on how the rapid increase of opioid abuse problems pose health threats over time. The difficulty of analyzing massive textual data with existing analytical tools may be to blame; however, with the advancement of data mining techniques, it is now possible to investigate such changes. This will be a small, first step towards systematically discovering how opioid abuse and related health problems have evolved, through the lens of scholarly publications.

The main purpose of this research is to discover how themes of opioid-related research have changed between 2000 and 2017. Research abstracts from PubMed were used for the analysis and were grouped into three periods: Period I (2000-2005), Period II (2006-2011), and Period III (2012-2017). This categorization has not only evenly spaced out the publication years but also coincides with changes in opioid-related deaths and the increasing number of publications as a result. The specific aim of this research is to investigate unique opioid-related research themes to reveal the health problems of each time period.

In order to answer the research question, a text mining (TM) strategy was employed. TM is a specialized data mining technique used to systematically extract major words or phrases from a collection of massive textual data^4,5. While this method is not new, very few studies have applied this method to healthcare research. One study in healthcare employed TM, but only counted frequent words⁶. A major problem with simple counts, such as counting frequent words, is that longer documents will naturally have higher frequencies of certain words. In this specific research, a certain document contains the term “pain” 25 times in the abstract, while some other documents have only 2 or 3 occurrences. Some studies in healthcare extracted single words (i.e., unigram)^7,8. While the unigram method is predominantly used in TM, regardless of the field, it creates some confusion when researchers attempt to interpret findings. For example, the term “pain” can occur in many different contexts (e.g., cancer pain, non-cancer pain, low back pain, chronic pain).

As such, this manuscript adopted three strategies: (1) the document normalization method was employed. This strategy will put longer and shorter documents on the same scale, so different lengths of documents can be directly comparable; (2) if a specific word appears across entire documents (or corpus in the TM terminology), the word will be given a lower value, which enables the study to exact distinct themes. The multiplication of (1) and (2) will finds distinct yet frequent themes from each period; (3) this research employed the three-word method (i.e.,trigram in the TM terminology). The trigram method has been proposed to capture the exact meaning from the corpus, but an application of this strategy to research has not been utilized; however, it is imperative to use trigrams to accurately interpret findings. All of these methods will be explicated later in the research methods section.

(Figure 1) shows opioid overdose deaths between 2000 and 2016¹, and (Figure 2) provides opioid-related academic journal publications within the same time frame. The rate of opioid related deaths increased by 402.55% from 2000 to 2016. This number sharply rose from 2012, with a steep increase in 2016: 33,091 to 42,249 deaths from 2015 to 2016, around a 28% increase in deaths. (Figure 2) shows that the number of publications in Period I (2000-2005) were relatively stable, steadily increased in Period II (2006-2011), and sharply increased in Period III (2012-2017). The corresponding numbers are 532, 1340, and 3313 during Period I, II, and III respectively. The number of publications increased by about 522% over these periods. Interestingly, the number of deaths and publications follow similar trends, as shown by (Figure 1) & (Figure 2).

Figure 1.

Age-adjusted opioid overdose deaths between 2000 and 2016¹

Figure 2.

Opioid-related publications between 2000 and 2017⁹

Research Methods

Data acquisition and methods

Opioid-related articles were retrieved from PubMed⁹. Because the corpus of this research consists of opioid-related journal articles the search criteria were set to journal articles, abstract availability, human subject, English, and publication dates between 2000 and 2017. These criteria yielded the following results: 7,044 articles were retrieved, and 5,815 articles were retained after removal of duplicates. The search terms were the combinations opioid + health (3,230), opioid + medical (2,589), opium + health (83), opium + medical (87), opiate + health (606), and opiate + medical (449). All of the data cleaning and text mining analysis were performed using the tm package, version 0.7-5, in R.

Preprocessing corpus

All documents were preprocessed together in order to keep consistency across the three groups of the documents. More specifically, three terms were unified under the term “opioid”: opioid, opium and opiate because these terms were used to search for the same opioid-related research. TM algorithm treats the full name and the acronym differently, because of this, acronyms were converted to the full name while dropping the first appearing acronym in parenthesis in order to prevent them from counting twice. After completion of the preprocessing, the abstracts were categorized into three groups

Data cleaning

Like any other text mining, this project went through the standardized data cleaning process: removal of punctuation, numbers, stop words, white space, and the conversion of capital letters to lower case. Additionally, this project removed the words that commonly appear in research abstracts such as objective, purpose, aim, research methods, finding, background, and conclusion, however, the term “opioid” was not removed to capture its meaning in the corpus. To elaborate, this research paper chose trigrams, and the term, “opioid,” will be used in different contexts (e.g., opioid overdose death, opioid substitute program, non-medical prescription opioid, etc.). Lastly, stemming was applied, stemming in natural language process (NLP) is the process of reducing words to a common base form (i.e., stem). For example, “processing” becomes “process” after being stemmed, so the word “process” can then incorporate all the words stemming from process (e.g., processing, processes, processed).

N-gram

An N-gram is a consecutive word sequence of N terms from a given sentence. If it extracts a single word, it is referred to as unigram; if it retrieves a two-contiguous sequential word, it is bigram; and if it is a three-contiguous sequential word, it is called trigram. While unigrams take less memory and are easier to calculate than bigrams or trigrams, it can be difficult to interpret the findings. For this research, trigrams were employed because many of the important words (e.g., opioid overdose death, prescription opioid abuse, mental health service, etc.) may lose the exact meaning if they are unigrams or bigrams. Below are examples of unigrams, bigrams, and trigrams derived from the phrase, “due to rising public health concerns over escalating prescription opioid abuse”:

Unigrams

“due” “to” “rising” “public” “health” “concerns” “over” “escalating” “prescription” “opioid” “abuse”

Bigrams

“due to” “rising public” “health concerns” “over escalating” “prescription opioid”

“to rising” “public health” “concerns over” “escalating prescription” “opioid abuse”

Trigrams

“due to rising” “public health concerns” “over escalating prescription”

“to rising public” “health concerns over” “escalating prescription opioid”

“rising public health” “concerns over escalating” “prescription opioid abuse”

As noted, trigrams can preserve the contextual meaning, but they require a lot of computing power and cleaning. In the trigram example above, only two are useable: “public health concerns” and “prescription opioid abuse.” All other trigrams should be filtered out. In order to preserve accuracy of the meaning, researchers need to double check some of the phrases in the document to avoid inadvertently removing meaningful trigrams. For example, in this manuscript, the terms, “human immunodefici virus” and “immunodefici virus infect” appear with similar frequencies, but they came from the same phrase “human immunodeficiency virus infection.” Another example would be the terms, “prescript drug monitor” and “drug monitor program,” both coming from the phrase “prescription drug monitoring programs.” Despite the labor-intensive efforts, in order to detect unique research themes over time, trigrams should be employed.

Tokenization (document-term frequency matrix)

Tokenization is the process of converting textual data into a data frame. Using the cleaned data, all words are decomposed into a consecutive three-word. For example, “people who inject drugs” becomes “peopl inject drug” after the application of cleaning and stemming, non-contextual words such as “who” are dropped in the stop word process, and the stemming process removed “s” after “drugs”, and “e” after “people.” (Table 1) shows a data frame, if a specific document has a specific phrase it is marked as 1; it is 2 if the phrase occurs twice. The examples of trigrams in (Table 1) are after the cleaning and stemming processes.

Table 1.

Data frame example of trigrams

	human immunodeficiviru	methadon mainten treatment	peopl inject drug	neonat abstin syndrom	…
Doc 1	2	2	1
Doc 2		1		1
Doc 3	1		1	1
…

Due to longer documents containing more terms, it is important to calculate the proportion of a specific word in a document, which is referred to as Term Frequency (TF) (i.e., number of times phrase t appears in a document / total number of trigrams in the document). For example, “methadon mainten treatment” (i.e., t) appears three times in a document (i.e., d ) with 150 trigrams . The proportion of the phrase is 3/150, and it is marked in data frame as “0.02.” It is inverse document frequency (IDF) that finds distinct terms from the corpus. For example, if “methadon mainten treatment” appears 600 times among 1,000 documents, its log of 1000/600=0.221849, and thus the IDF value is “0.221849.” TF-IDF is the multiplication of TF and IDF. In this example, the value of TF-IDF is “0.00443698” (0.02* 0.221849 =0.00443698). It is one value from one corpus. If “methadone mainten treatment” appears 1000 times in period I, for example, the trigram for the period is 4.43698 (0.00443698* 1000). TF-IDF allows researchers to find distinct yet frequent phrases in the corpus. TF-IDF is the best-known weighting scheme in text retrieval¹⁰. TF-IDF is valid not only for unigrams but also for bigrams, trigrams, and others⁴, the mathematical calculations are as follows:

$$\begin{array}{ccc}TF(t,d)=\frac{freq(t,d)}{{\displaystyle {\sum }_i^{n}freq(t_i,d)}}& IDF(t)=\log \left(\frac{N}{count(t)}\right)& TF-IDF(t,d)=TF(t,d)*IDF(t)\end{array}$$

Finding

As shown in the calculation of TF-IDF in the previous section, the value will go up as the number of publications increases, which means that the value of TF-IDF will be higher for Period III. In order to directly compare on the same scale, this research used Period III as the baseline and assigned weights of 5.23 for Period I and 2.08 for Period II. (Figure 3) is based on the weighted values of the three periods. Comparisons can contain a limited number of topics, therefore a wordcloud was used to provide a full picture of research topics within each period. Because wordclouds utilize different font sizes to convey frequency within the period, assigning weight is unnecessary. The comparison is shown in (Figure 3), and the wordclouds appear in (Figure 4).

Figure 3.

Weighted TF-IDF across groups

Figure 4.

Trigram using TF-IDF across groups

Interpretation

Common research themes

Concerning common research themes, HIV and methadone maintenance treatment have been major research themes across all periods, although these themes weaken over time. In regards to HIV, this may be the result of HIV-related pain being commonly treated with opioids¹¹. Methadone maintenance treatment (MMT) also emerged as a strong theme throughout each research period. MMT is not only effective for treating opioid addiction but can also reduce the risk of HIV transmission if patients completely adhere to the treatment¹². Period I experimented with MMT, this method may have been proven to be effective for treating opioid addiction, which could be the reason for its consistent appearance. For example, a meta-analysis based on 37 studies involving 3,029 patients discovered methadone is preferable to buprenorphine for treating opioid addiction, cocaine dependence, and HCV^13,14.

Period I: Burgeoning opioid-related problems

The World Health Organization (WHO) is frequently used among researchers in Period I to cite the criteria provided by the organization, but the citing of WHO weakened as time progresses. Articles frequently refer to WHO to demonstrate their conformance to the treatment guidelines given by WHO, which might have been important at the time because knowledge of opioid addiction treatments was not yet established. This stage experimented with various treatments, which is validated by the research methods shown in (Figure 3). It is interesting though that treatments for cocaine and opioid dependence receive similar academic attention despite the fact that the corpus of this research was extracted based on opioid. It is clear that cocaine was a substantial public health threat in this time period, as it spreads HIV¹⁵. Since MMT is in a nascent stage, failure of treatment drove users to cocaine¹⁶. In this period nonsteroidal anti-inflammatory drugs (NSAIDs) are majorly used for pain medication but is not a major research theme beyond this period. The finding clearly shows that while opioid-related problems existed, they were not a critical health issue at the time.

Period II: Busting opioid-related problems

Unlike Period I, where opioids were mainly used for cancer pain and HIV patients, the opioid prescription was much more likely to include other pain types such as chronic noncancer and lower back pain, as shown in (Figure 4). As such, the portraits of opioid-related problems from this period and the previous period differ greatly. In other words, because of the expanded opioid prescription, many of the opioid-related problems had infiltrated to society and created critical public health concerns. Since primary care providers (PCPs) treat noncancer-related pain, and they account for approximately 50% of the prescription opioids dispensed^17,18, they are integral to the themes of Period II. Research shows that PCPs can play an important role in combating opioid addiction problems through methadone or buprenorphine treatments¹⁹, and they can collect critical information about opioid overdose, which is useful for controlling opioid use management^14,20. In fact, PCPs are often called upon to differentiate between appropriate medically indicated opioid use in pain management vs inappropriate abuse or addiction¹⁴. As such, the role of PCPs emerged as an important research theme in this period.

As opioid-related health problems escalated, scholars investigated target populations that were vulnerable to opioid abuse, which were identified as women and veterans. Women have an earlier age of initiation into substance use and a more rapid progression into drug involvement and dependence than men²¹. Pregnant women, especially, emerged as an important research topic because of neonatal abstinence syndrome (NAS). NAS occurs to a newborn who was exposed to addictive opioid drugs while in the mother’s womb. A study shows that 91% of the infants needed treatment in a neonatal care unit, 76% had NAS, and 57% needed morphine replacement therapy²². Also, many veterans enrolled in the veterans’ healthcare system suffer from high rates of chronic pain and mental health issues²³, which resulted in veterans becoming a vulnerable population. Therefore, causing veterans’ health administration to become a frequent research theme. Along with veterans’ mental health issues, the general public’s mental health issues exacerbated the opioid abuse problem. More specifically, 16% of mental disorders receive over half of all opioids prescribed in the United States²⁴, and thus mental health services frequently appear as a major topic in this period.

Drug injection as a problem intensifies in this period. While oral abuse decreased from 73.1% to 58.9%, injection abuse increased from 11.7% to 18.1% from 2004 to 2013²⁵. Injecting drugs poses serious health problems because it is associated with the most dangerous route of abuse, resulting in substantial health consequences including HIV, HCV, increased risk for overdose, emergency department visits, and mortality^26–28; however, some of the consequential problems such as HCV, emergency department (ED) visits, and deaths were not major research themes in this period.

Period III: Becoming an epidemic

The opioid prescription expands into non-medical use: a stark contrast between the use for cancer pain in Period I and medically deemed pain in Period II. As such, opioid overdoses increased by 30% from July 2016 to September 2017²⁹. About 21% to 29% of patients prescribed opioids for chronic pain misuse them³⁰. As a result of widespread opioid use in this period, some of the problems were further exacerbated from Period II. Drug injection is one example: injecting drugs increases the chances of getting HCV and HIV. In the US, about 9% of HIV infections diagnosed in 2015 were attributed to injection drug use³¹, and injecting drugs is responsible for about 25% of new HIV cases³². Injecting drugs also accounts for more than 40% of HCV cases annually³³, and the prevalence of HCV infection is far greater than that of HIV for injecting drug users^34,35. Furthermore, these two diseases accelerate the progression of each other. HIV speeds the course of HCV infection by accelerating the progression of hepatitis-associated liver disease³⁶. Consequently, it is not surprising to observe a high occurrence of HCV with HIV in this period. The continued appearance of MMT might be attributed to the treatment of HCV, as this method is proven to be more effective than buprenorphine¹⁴.

NAS, one of the consequential problems from opioid abuse, continues to pose as a major public health concern³⁷. While MMT is an effective method to treat pregnant women, recent investigations have shown a trend for a milder NAS under buprenorphine. Buprenorphine exposure was associated with lower NAS scores³⁸ and less severe NAS relative to methadone³⁹. This may be the reason buprenorphine treatment research accelerated in this period. One other reason for this acceleration could be that buprenorphine treatment was approved in 2002 by WHO, and scholars could then experiment with this method.

The dominant and yet distinct health problems from opioid overdose in this period are ED visits, opioid overdose deaths, post-traumatic stress disorder (PTSD), and opioid-induced constipation. Studies show that PTSD and mental health diagnoses are associated with 30% increased odds of high-dose opioid prescribing among veterans⁴⁰, and veterans tend to receive higher dose therapy^40,41. This problem is a major source of morbidity and mortality among US veterans⁴². As such, veterans’ health issues and veterans’ health administration are frequent subjects of scholarly investigation in this period. Opioid-related deaths are a serious problem in this period as well. In fact, opioid overdose deaths increased by about 28% from 2015 to 2016¹, which makes opioid overdose deaths account for two thirds of drug overdose deaths between 2015 and 2016⁴³. ED visits emerged as a health threat from July 2016 through September 2017, 142,557 opioid-related ED visits⁴⁴. This rate increased, on average, by 5.6% per quarter⁴⁴. Because of the sharp increases in ED visits and opioid-related deaths, it is not surprising to capture these themes emerging as new dominant research subjects.

Another set of unique research topic in this period is prescription drug monitoring programs (PDMPs), electronic health/medical records, and evidence-based treatments. These research topics became possible with the aid of recent advanced health technology. PDMPs are designed to combat the abuse of opioids and other prescription drugs by sharing statewide electronic databases that store controlled drug dispensing information⁴⁵. In this way, doctor shopping can be controlled. As a consequence of this, it became difficult to obtain opioids in a legitimate way, drug users then turn to illegal or street drugs, which might have increased the rates of HIV and HCV in Period III. Furthermore, street-drugs are responsible for 60% of overdose deaths in 2016⁴⁶, which might contribute to the increased number of “deaths” in this period.

Conclusion

This study investigated opioid-related research themes over the last 18 years by categorizing research abstracts into three groups. A text mining analysis shows distinct research themes over time with the rapid expansion of opioid prescriptions and opioid abuse problems. The main research theme of Period I was experimenting with MMT, and opioids were mainly used for HIV and cancer pain during this time. It is clear that opioid overdose problems were not at critical levels in Period I. The emerging research themes in Period II were PCPs, women’s opioid addiction problems, mental health issues, veterans’ pain and mental health, and prescription opioid abuse problems, and these problems were exacerbated in Period III. In Period III, some of the research themes from Period II were magnified, but the dominant health problems revealed were those of the current health crisis, such as opioid-related deaths, HCV, ED visits, and PTSD. With the advent of advanced health technology, PDMPs, electronic health/medical records, and evidence-based treatments emerged as major research topics.

This study is a small, first attempt to systematically uncover how opioid-related research has evolved over the last 18 years. This study’s corpus is only from PubMed: if article abstracts were not included in PubMed, they were not included in this analysis. As such, interpretation of this study’s findings requires caution. Future research may need to expand the scope of the search and compare how those studies published in PubMed differ from other sources. Another suggested future research topic is a comparison of unigrams, bigrams, and trigrams to investigate how these methods differ when capturing topics.