Opinion Mining and Multi-Label Classification of Patient-Generated Long-Term Anti-Obesity Medication Reviews (Preprint)

Abstract

With the explosion of web 2.0 technology, patients have taken to the internet to share experiences about their health conditions and treatments. Online drug review portals currently allow patients to their experiences with drugs they used in managing their conditions. These data sources could be harnessed for patient-reported real-world evidence to understand the impact of drugs on the users.

To understand patients’ opinions about long-term AOMs (phentermine-topiramate, orlistat, naltrexone-bupropion, lorcaserin, and liraglutide) through online patient-posted user reviews. To determine the frequency of occurrence of key obesity treatment outcomes and build a multi-label classification model for detecting key obesity outcome topics.

We crawled drug.com, askaapatient.com, webmd.com, druglib.com, and extracted reviews posted by the users of long-term AOMs about their experience with the drugs. Next, we carried out a generic lexicon-based document-level sentiment analysis by matching the words in the reviews of each AOM with their polarity classes in the sentiment dictionary. We then calculated the scaled sentiment score to measure how averagely positive the patient’s opinion is towards the drugs. The frequencies of occurrence of weight, adverse effect, glycemic, blood pressure, lipidemic outcome topics in the posted reviews were analyzed. A Multi-label classification model for classifying obesity outcome related topics was built and tested.

Patients expressed the most positive opinion for lorcaserin with a scaled sentiment score of 0.139, followed by phentermine-topiramate with scaled sentiment score of 0.04. Orlistat and naltrexone-bupropion had scaled-sentiment scores of -0.008 and -0.02 respectively. Having a scaled sentiment score -0.036, liraglutide was the most negatively appraised long-term AOM by patients’ reviews. Comparing the frequency of occurrence of weight and cardiometabolic topic in the reviews, weight loss outcome was the dominant topic, occurring in 1585 reviews, adverse effect topic occurred in 1273 reviews, glycemic outcome topic occurred in 92 reviews, blood pressure outcome topic occurred in 72 reviews, lipidemic outcome topic occurred in 48 reviews and topic on pulse outcome occurred in 31 reviews. The Multi-label classification model trained with the patient-posted AOM reviews has F1 score of 0.98, 0.55, 0.67, 0.80, and 0.67 in predicting AOM-related weight loss, adverse effect, , glycemic, blood pressure, lipidemic and pulse topics respectively in free text form.

Sentiment analysis of patient-posted long-term AOM reviews could be useful in understanding patient‘s experience with long-term AOMs. Despite having being withdrawn for the market, lorcaserin was the most positively appraised long-term AOM followed by phentermine-topiramate, orlistat, naltrexone-bupropion, and liraglutide. The users of AOMs commented most on the weight and safety (adverse effects) outcomes of AOMs than cardio-metabolic outcomes of their treatments. Classification model trained with patient posted AOM reviews had a good performance in detecting efficacy and safety signals occurring in text documents. sentiments/opinions formed by obese and overweight patients from their experience with long-term AOMs could be used in demonstrating the values of the medications, as part of patient-reported real-world evidence.