A significant challenge existing in extracting ADR terms from social media data is the diverse expressions of adverse reactions. Health consumers describe their health issues using different vocabularies than medical professionals due to their lack of medical training. As a result, standard medical lexicons used by professionals like the US National Library of Medicine’s Unified Medical Language System are not applicable in analyzing health consumer–contributed content. To address this problem, we used the Consumer Health Vocabulary (CHV) Wiki [13] to build our ADR lexicon. The CHV Wiki links everyday health-related words to professional terms and jargon, and the goal is to bridge the communication gap between consumers and health care professionals [5]. It provides a list of preferred names of ADRs and the corresponding consumer-contributed expressions to each of them. For example, “anorexia” is a professional expression of an ADR, and the CHV Wiki extends it to “appetite lost,” “appetite loss,” “appetite lack,” “no appetite,” and several other common expressions. In our study, we used all of the expressions suggested by the CHV Wiki to extract ADR entities in user-generated information [14].

After extracting ADR entities, we computed the association strength between ADR-disease, ADR-drug, and drug-disease pairs. In social media data, if 2 entities are mentioned together frequently, they are considered to be strongly associated [15]. We conducted co-occurrence analysis on those entity pairs to evaluate their association strength. In co-occurrence analysis, choosing proper analysis granularity is important. That means the co-occurrence analysis should be calculated within an appropriate scale. A post or comment might be set as an analysis unit, but they are usually very short and the users may jump into their problems without describing the context. Thus, a post or comment is too small an analysis unit to calculate the co-occurrence between entities. Instead, a thread, which contains a post and all the following comments, is more appropriate to be an analysis unit because it embodies all the discussions on a particular issue and contains all the keywords. Therefore, we calculated the co-occurrence frequency of entity pairs within each thread.

We applied association rule mining to determine the co-occurrence frequency between ADR-PD, ADR-drug, and drug-PD pairs. Mathematically, let I={I₁, I₂,..., I_m } be a set of items and let T={T₁, T₂,…, T_n} be a set of transactions, where each transaction is a subset of items such that T_i ⊆I. The association rule is an implication of the form A ⇒B, where A ⊂I, B ⊂I, and A ∩B=∅, which is referred to as an itemset. An itemset that contains k items is a k-itemset. Specifically, there are both 1-itemsets ({ADR}, {drug(R)}, {PD}) and 2-itemsets ({ADR, R}, {ADR, PD}, {R, PD}) in our experiments.

Lift is a commonly used measure based on probability theory in association rule mining. For instance, lift measures the strength of an association by considering not only the co-occurrence of R ∪ADR but also the correlation between itemsets {R } and {ADR }. To be specific, it computes the ratio of the proportion of threads containing both R and ADR above the expectation that R and ADR are independent of each other. The following shows the equations for computing Lift (R ⇒ADR):

A heterogeneous network is defined as a graph consisting of nodes connected by links, with at least 2 types of nodes and at least 2 types of links [16]. Let N={n₁,n₂,...,n_k} be a set of nodes and L={l₁,l₂,...,l_m } be a set of links, and G=(N, L) denotes the graph. In the graph G, each node n_i ∈N belongs to a particular type from γ and each link l_i ∈L belongs to a particular type from τ, where |γ |>1 or |τ |>1.

In our heterogeneous health care network for PD, there are 3 types of nodes (drug, PD, and ADR) and 3 types of links (drug-ADR, PD-ADR, and drug-PD), that is, γ={R, PD, ADR } and τ={L_R-ADR, L_PD-ADR, L_PD-R}. Figure 2 presents a nondirectional graph model of the described heterogeneous health care network that is weighted, and the association of the 2 end nodes of a link determines the weight of the link. For instance, the weight of the link between a drug and an ADR is represented and computed by Lift (R ⇒ADR).

In addition, we predicted the strength of association between disease-ADR pairs through all possible paths suggested by the network in this work and proposed a path mining method: Path (D-R-ADR,W_D-R*W_R-ADR). ADR represents the harmful and unpleasant reactions of medicine use, and hence, the ADRs that are associated with a disease are highly influenced by the drugs that are used for the disease. We use drug as the bridge between disease and ADR. The weight of a path is computed by the products of Lift (D ⇒R) and Lift (R ⇒ADR).

The association between disease and ADR is then computed by the equation seen in below, in which P denotes all of the possible paths between D and R:

The basic hypothesis of drug repositioning is that if ADR X is significantly associated with disease D, then drugs that list X as a side effect should be evaluated as candidates for treating disease D [5]. For all the disease-ADR associations we detected, we first used a sample t test to determine whether they were significant and then applied drug repositioning to those significant associations.

The repositioning process is completed in 3 steps: (1) for each ADR contained in those significant ADR-disease associations, we consulted the Side Effect Resource (SIDER) database [17] to identify the drugs that list the identified ADR as a frequent side effect, (2) we directed those drugs to diseases via the drug-ADR-disease paths and obtained the final drug-disease associations, and (3) we removed the drug-disease pairs that have already been published in the Pharmacogenomics Knowledgebase (PharmGKB) [18]. Thus, the remaining drug-disease pairs could be considered as repositioning drug–disease associations and reported as the drug repositioning results.

In our experiment, user-generated data were collected from MedHelp [19], one of the pioneers in online health communities. Since its introduction in 1994, MedHelp has been the exemplar in online health communities. Currently, it empowers over 12 million people each month to describe and discuss their health and medical problems and find answers. We collected data from MedHelp by implementing an automatic Web crawler that went through MedHelp webpages in HTML format page by page and retrieved information including user name, post, comment, and timestamp. The returned data were saved in .txt files and organized thread by thread, where each thread contained the original post and all the following comments. We collected 12,571 threads (containing 504,097 comments) by using Parkinson and the corresponding 55 drugs suggested by PharmGKB as query words on October 10, 2016. After extracting disease, drug, and ADR entities from the consumer-contributed content, we constructed a heterogeneous network containing 3 types of nodes and 3 types of links. In total, the network contained 255 nodes: 55 drugs, 199 ADRs, and 1 disease.

When the drugs treating a disease share a common ADR, the ADR and the disease are considered to be connected via an underlying mechanism of action. By applying the proposed heterogeneous network mining method, we detected the ADRs associated with PD and filtered the significant associations based on statistical analysis. As a result, we found 9 ADRs significantly associated with PD: muscle cramp, gastrointestinal disorder, nervous system disorder, angiopathy, somnolence, orthostatic hypotension, carpal tunnel syndrome, hallucination, and influenza.

We retrieved the drugs listing the ADRs but not indicated for PD from the SIDER database with several filtering conditions: the drugs indicate 1 of the 9 ADRs as “frequent,” “common,” or “postmarketing” or with an occurring percentage higher than 10. As a result, we identified 44 repositioning drugs for PD. Table 1 lists all the repositioning drugs, as well as whether these drugs are associated with PD based on literature analysis. By using drug names and Parkinson as queries in PubMed, we then investigated all the findings to understand the implications behind the associations of disease and repositioning drugs. We found evidence in PubMed literature to support a positive effect of some drugs on PD, evidence to support negative or adverse effects of some drugs on PD, evidence that some drugs exert both positive and negative effects on PD, and no explicit association between some drugs and PD.

Multimedia Appendix 1 presents the repositioning drugs that are supported by evidence from PubMed literature, and each row illustrates the drug and how it is labeled with the ADRs in the SIDER database. The second-to-last row total shows how many drugs were detected via the ADR, and the last row total shows how many drugs were only detected via the ADR.

Multimedia Appendix 1 indicates that nervous system disorder and gastrointestinal disorder are the 2 ADRs that contributed the most number of repositioning drugs, while the ADR carpal tunnel syndrome suggested no repositioning drug. Among the repositioning drugs that have evidence support, 54% (15/28) were detected via only 1 ADR and 46% (13/28) were detected via more than 1 ADR. Moreover, within the 13 drugs, 6 of them were detected via the ADR combination: nervous system disorder + gastrointestinal disorder, which suggests that nervous system disorder and gastrointestinal disorder might be a potential ADR combination for identifying repositioning drugs for PD. In addition, to investigate whether there is relationship between the number of ADRs and the number of PubMed sources, we did correlation analysis. The calculated correlation coefficient (Pearson R) approximately equals 0, indicating little correlation between those 2 factors.

Table 2 summarizes the drugs analyzed in this section and their original labeled indications.

In drug repositioning research, results are only suggestions or predictions rather than confirmed truth until clinical trials are conducted. As a result, the evaluation of method performance becomes a difficult task, because there is no gold standard. In previous studies, there have been two common ways of evaluation: computational assessment, based on the co-occurrence of drug-disease terms in biomedical literature, and experimental assessment, based on in silico or in vitro experiments. Here we have adopted the computational assessment but beyond counting the co-occurrence frequency, we conducted a literature review to investigate whether the scientific studies in the published articles supported the repositioning of the drug to PD. In the future, experimental assessment could be exploited to reinforce the evaluation.

Among the discovered repositioning drugs, some have the potential to treat PD via their neuroprotection or neurotoxicity prevention function, while several others show the effects in clinical practice without the underlying mechanism being known; therefore, exhaustive analysis is needed of the drugs. In addition, we identified several psychological drugs that are proven to be effective without inducing adverse effects when used to treat PD. As psychological symptoms occur in a high ratio of PD patients, these findings are quite useful. Based on the results, the repositioning drugs that have positive functions on the neural system are the most promising findings and are the primary candidates for further experimental assessment. Meanwhile, the repositioning findings that are commonly used for other complicating diseases on PD patients should be tested more carefully for efficacy and risk evaluation.

In this work, we relied on the rational association between disease and ADR to find potential repositioning drugs and achieved an appreciable result. To investigate whether there is any relationship between the number of intermediary ADRs and the number of supporting literature sources, we did a correlation analysis, but there is little correlation between the 2 factors according to the calculated Pearson R coefficient (≈0). A possible reason could be the size of our dataset (eg, ADR and drug). In a future study, we can expand the dataset size to include more ADRs and involve more biomedical characteristics, such as chemical or biological characteristics or a combination of the two, hoping for better performance.

In this study, we proposed a computational drug repositioning method based on a heterogeneous network and used the rational association between disease and ADR to reveal repositioning drugs. Experiment results demonstrated that our approach is able to identify FDA-approved novel drugs for PD, and many of the predictions are supported by existing studies according to the literature-based investigation. Although further studies are needed to confirm the pharmaceutical effects of these repositioning drugs for PD, this study suggests the possibility and effectiveness of applying systematic drug repositioning methods to drug development.