Cardiovascular Disease Risk Prediction via Social Media

Technical Summary: Cardiovascular Disease Risk Prediction via Social Media

Problem Statement
Cardiovascular Disease (CVD) remains a leading cause of morbidity and mortality in the United States. Traditional risk assessment relies heavily on demographic data and clinical indicators, which may not fully capture the psychological and behavioral factors contributing to CVD risk. While social media platforms like Twitter offer a wealth of unstructured data regarding public discourse on health, psychological states, and lifestyle behaviors, the application of Natural Language Processing (NLP) and Machine Learning (ML) to predict CVD risk at a state level using this data remains an under-explored area. The core challenge addressed is the development of a robust methodology to extract, process, and classify tweets to identify individuals potentially at risk of developing CVD, and to evaluate whether this approach outperforms traditional demographic-based prediction models.

Methodology
The study employs a comparative framework utilizing two distinct datasets: a Twitter-based dataset and a Centers for Disease Control and Prevention (CDC) demographic dataset.

1. Data Collection and Curation:

   • Twitter Dataset: Tweets were collected from eighteen US states (including thirteen Appalachian states and five nearby states) over a three-year period (2019–2021). A custom dictionary of twelve CVD-related keywords was developed, encompassing clinical terms (e.g., "heart attack," "hypertension"), risk factors (e.g., "smoking," "cholesterol"), and psychological/lifestyle factors (e.g., "stress," "alcohol"). Approximately 269,969 tweets were extracted using the Tweepy library.
   • CDC Dataset: A state-wise dataset from the US Department of Health & Human Services (2019–2020) containing 124 chronic disease indicators and demographic information (gender, race, ethnicity, location) was utilized as a baseline for comparison.

2. Preprocessing and Labeling:

   • Twitter Data: Raw tweets underwent cleaning (removing URLs, emojis, repeating characters) and NLP preprocessing (tokenization, stemming, stop-word removal). The VADER (Valence Aware Dictionary and sEntiment Reasoner) model was employed for sentiment analysis. A threshold of -0.30 was established to classify sentiment polarity: users with scores above this threshold were labeled as '1' (potential CVD risk/positive sentiment regarding risk factors), and those below were labeled as '0' (no risk/negative sentiment).
   • CDC Data: Demographic features were numerically encoded (one-hot encoding for years, numeric encoding for gender/ethnicity). The dataset was balanced using the SMOTE technique to address class imbalance.

3. Model Architecture and Training:

   • Machine Learning (ML): Four conventional algorithms were implemented: Bernoulli Naive Bayes (BNB), Support Vector Machine (SVM), Logistic Regression (LR), and CatBoost.
   • Deep Learning (DL): A hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) architecture was constructed. The CNN layer extracts features from the input, while the LSTM layers capture temporal dependencies, aiming to classify individuals into high or low-risk categories.
   • Evaluation: Both datasets were split into 70% training and 30% testing sets. Models were evaluated using Test Accuracy, Precision, Recall, F1-score, Mathew's Correlation Coefficient (MCC), and Cohen's Kappa (CK) score.

Key Results
The study presents a direct performance comparison between the Twitter-based approach and the CDC demographic-based approach:

• Twitter Dataset Performance: The ML models significantly outperformed the DL model. The SVM achieved the highest test accuracy of 88.75%, followed closely by Logistic Regression (LR) at 87.82%. The hybrid CNN-LSTM model achieved 77.51% accuracy. BNB and CatBoost recorded accuracies of 74.55% and 76.67%, respectively.
• CDC Dataset Performance: All models performed considerably lower on the demographic data. LR ranked highest with 58.03% accuracy, followed by BNB (57.93%), CNN-LSTM (57.64%), SVM (57.55%), and CatBoost (57.42%).
• Comparative Analysis: The Twitter-based dataset consistently demonstrated superior predictive power across all metrics compared to the CDC dataset. State-wise analysis confirmed that the top-performing classifiers (SVM and LR) maintained higher accuracy when using tweet data versus demographic data.

Significance and Claims
The paper claims that analyzing emotions and language patterns expressed in social media tweets offers a more effective method for identifying individuals at potential risk of CVD than relying solely on traditional demographic information. By developing a specialized dictionary of CVD-related keywords and applying sentiment analysis, the study demonstrates that NLP and ML techniques can serve as a valuable supplementary tool for public health monitoring.

The authors emphasize that this research highlights the potential of using social media as an alternative or complementary data source to traditional demographic surveys for predicting health outcomes. The study does not claim to replace clinical diagnosis but suggests that social media analytics can provide early insights into population-level risks, potentially informing targeted public health interventions. The work underscores the utility of hybrid architectures and standard ML algorithms in processing unstructured social media data for health risk prediction.