JMIR Publications

ABSTRACT

Background:

Opioids are often prescribed by doctors to relieve pain following surgery, accidents, or diseases such as cancer. However, taking pain medication is risky. According to the Centers for Disease Control and Prevention, 70,630 opioid overdose deaths were reported in the United States in 2019. Medication-assisted treatment (MAT) is an effective method for treating addiction that might lead to overdose. MAT combines the use of three FDA approved medications for opioid dependence: methadone, buprenorphine, and naloxone with counseling and behavioral therapies to reduce opioid use gradually and to reduce the occurrence of withdrawal symptoms and the desire to seek out unprescribed opioids.

Objective:

While MAT has been proven effective initially, very little is known about retention over the long-term leading to a desire for more information from the patients' perspective, especially in real-world settings. It would be challenging and time-consuming for health providers to assess the day-to-day experience of all MAT patients. However, a broad survey of patients' viewpoints can be obtained through social media and drug review forums and assessed using automated methods to discover experiences expressed at the moment over long periods of time. The primary aim of this study is to develop a predictive model to detect the patients' perceived effectiveness of two well-studied opioid dependence medications from text posted to health-related social media.

Methods:

To build the predictive model, we employ different analyses: 1) identifying non-clinical factors based on an unsupervised clustering technique, topic modeling, and 2) identifying clinical factors by applying MetaMap, a widely used tool created by researchers at the National Library of Medicine. We performed the topic modeling and MetaMap on 4,353 patient reviews related to methadone and buprenorphine/naloxone drugs using two healthcare forum websites from 2008 to 2021. We then built a dataset using the vectorized text, these topics, MetaMap categories, and duration of treatment as features, using the patients’ effectiveness rating as the class. We compared prediction models developed using Logistic Regression, Elastic Net, LASSO, Random Forest Classifier, Ridge Classifier, and XG Boost.

Results:

We developed a predictive model to recognize the effectiveness of opioid dependence treatments from health forums drug reviews. The F-measure of the predictive models across all methods ranged from 83.4% to 90.6%. The Elastic Net model, a regularized regression method, outperformed the other models.

Conclusions:

The patient's assessment of the effectiveness of opioid dependency treatment can be predicted by using automated text analysis. We found that adding clinical features such as symptoms, drug name, illness, and non-clinical factors like duration on treatment and topic models could improve predicting the effectiveness of treatment.