TL;DR
This paper introduces a novel approach to predicting cardiovascular disease (CVD) risk by analyzing sentiment in social media posts. By developing a specialized dictionary of CVD-related keywords and employing sentiment analysis, the researchers demonstrate that social media data can be a more effective predictor of CVD risk than traditional demographic information alone.
What Problem Does This Paper Address?
Cardiovascular disease (CVD) remains a major global health concern. While demographic data has traditionally been used to assess CVD risk, it has limitations. The authors propose that social media, specifically Twitter, offers a rich, real-time source of information about individuals' psychological states and behaviors, which are known to influence CVD risk. The challenge lies in effectively extracting and interpreting this information to identify individuals at potential risk, especially when considering the nuances of language and the vastness of social media data.
Methods
The core of this research lies in a two-pronged approach:
- Tweet-Based Approach:
- Keyword Curation: A new dictionary of CVD-related keywords was developed by analyzing clinical and psychological risk factors from existing literature. This dictionary goes beyond formal medical terms to include everyday language and lifestyle-related phrases found on Twitter.
- Data Collection: Tweets were collected from eighteen US states (including the Appalachian region) using the curated keywords and the Tweepy library.
- Sentiment Analysis: The VADER (Valence Aware Dictionary and sEntiment Reasoner) model was employed to analyze the sentiment of the collected tweets. A threshold of -0.30 was used to classify users as potentially at CVD risk ('1') or not ('0').
- Machine Learning (ML) and Deep Learning (DL) Models: Various ML models (Bernoulli Naive Bayes, Support Vector Machine, Logistic Regression, CatBoost) and a hybrid CNN-LSTM DL model were trained and evaluated on the sentiment-labeled Twitter data.
- CDC Dataset-Based Approach:
- Data Source: A state-wise CDC dataset containing demographic information (gender, race, ethnicity, diagnosis year, disease type) for 2019-2020 was used.
- Preprocessing: Demographic features were numerically encoded, and the dataset was balanced using the SMOTE technique.
- ML Models: The same four ML models and the CNN-LSTM DL model were applied to this dataset for comparison.
Datasets
- Twitter Dataset: 269,969 tweets collected over three years (2019-2021) from eighteen US states, using a custom-curated dictionary of CVD-related keywords.
- CDC Dataset: A state-wise dataset from the US Department of Health & Human Services for 2019-2020, containing demographic information related to chronic diseases.
Key Insights and Ideas
- Social Media as a Health Indicator: The study strongly suggests that social media, particularly Twitter, can serve as a valuable and timely resource for public health monitoring, offering insights into population-level health behaviors and risks.
- Beyond Demographics: Analyzing the emotional tone and specific language used in tweets can provide a more nuanced and potentially more accurate prediction of CVD risk than demographic data alone. This highlights the power of understanding user sentiment and language patterns.
- The Importance of a Tailored Dictionary: The creation of a specialized CVD-related keyword dictionary, which includes colloquialisms and lifestyle terms, was crucial for effectively capturing relevant discussions on Twitter. This demonstrates that generic keyword lists might miss important signals.
- Sentiment as a Predictor: The VADER sentiment analysis, when applied to CVD-related tweets, proved to be a significant factor in identifying individuals at potential risk.
Key Results
- Twitter vs. CDC Performance: The ML and DL models trained on the Twitter dataset generally outperformed those trained on the CDC demographic dataset.
- Top Performers (Twitter): The Support Vector Machine (SVM) achieved the highest test accuracy of 88.75% on the Twitter dataset, followed closely by Logistic Regression (LR) at 87.82%. The CNN-LSTM model achieved 77.51% accuracy.
- Top Performers (CDC): On the CDC dataset, LR achieved the highest test accuracy of 58.03%, with other models performing similarly (around 57-58%).
- State-wise Comparison: Comparative analysis showed that the Twitter-based models (specifically SVM and LR) had better state-wise performance in predicting CVD prevalence compared to the CDC demographic data.
Limitations and Future Work
- Limitations:
- The study acknowledges that the VADER model's threshold was chosen based on empirical observation and preliminary work, and it's not asserted as universally optimal.
- The CDC dataset had limited demographic features, which posed a challenge despite the large number of patients.
- The paper mentions that while these findings suggest the utility of social media data, further analysis and validation are required to ensure reliability and generalizability.
- Future Work:
- The authors suggest incorporating more performance evaluation parameters to further assess the models.
- The study implies that further research into the nuances of social media language and its correlation with health outcomes is warranted.
Summary updated