A Twitter Discourse Analysis of Negative Feelings and Stigma Related to NAFLD, NASH and Obesity

Jeffrey V. Lazarus; Christine Kakalou; Adam Palayew; Christina Karamanidou; Christos Maramis; Pantelis Natsiavas; Camila A Picchio; Marcela Villota-Rivas; Shira Zelber-Sagi; Patrizia Carrieri


Liver International. 2021;41(10):2295-2307. 

In This Article


There are some limitations to this study, such as the sampling method, given that there are distinct temporal trends that could have impacted the study results. The interactivity aspect of online discussions (including the network analysis) was not thoroughly examined. Another limitation is the bias in labelling tweets, which is an inherently subjective task. However, we took several steps to reduce this bias, including screening in duplicate and having experts validate a subsection of tagged tweets. Another limitation is the representativeness of obesity discourse within the NAFLD/NASH dataset, which seems to be limited mainly to academic discussions with regards to the calculated negative sentiment of the obesity-related tweets collected from a wider audience. Although at this point any direct comparison of the findings from the two phases would be unreliable, we believe that in the future and as the general public becomes aware of the relationship between obesity and NAFLD and such discussions become more frequent, the current obesity discourse analysis will greatly enrich future NAFLD/obesity insights on stigmatizing and otherwise affective language. Lastly, the inherent limitations of ML methods apply to the classifications of the custom sentiment analysis pipeline. For instance, we cannot be sure that the annotated dataset that has been employed for training the ML parts of the pipeline is an adequate representation of the target tweet population. If this is not the case, our estimations for the distribution of sentiment polarity classes in the population cannot be trusted. In addition, despite the fact that imbalance-tolerant classifiers have been employed, the extreme imbalance of the sentiment polarity classification target could lead to an underestimation of the classes with a lower sample size (neutral and positive).

One of the strengths of the study is that we applied a NLP technique to classify tweets related to obesity by using a variety of terms, including slang words/phrases, which we were able to do successfully. This approach is more sensitive and helps to detect pertinent tweets when compared to previous approaches. It also allows for inclusion of further evidence compared to previous studies on the topic, based on tweet analysis.