Utilizing Twitter Data Analysis and Deep Learning to Identify Drug Use (Preprint)

BACKGROUND The collection and examination of social media has become a useful mechanism for studying the mental activity and behavior tendencies of users. OBJECTIVE Through the analysis of a collected set of Twitter data, a model will be developed for predicting positively referenced, drug-related tweets. From this, trends and correlations can be determined. METHODS Twitter social media tweets and attribute data were collected and processed using topic pertaining keywords, such as drug slang and use-conditions (methods of drug consumption). Potential candidates were preprocessed resulting in a dataset 3,696,150 rows. The predictive classification power of multiple methods was compared including regression, decision trees, and CNN-based classifiers. For the latter, a deep learning approach was implemented to screen and analyze the semantic meaning of the tweets. RESULTS The logistic regression and decision tree models utilized 12,142 data points for training and 1041 data points for testing. The results calculated from the logistic regression models respectively displayed an accuracy of 54.56% and 57.44%, and an AUC of 0.58. While an improvement, the decision tree concluded with an accuracy of 63.40% and an AUC of 0.68. All these values implied a low predictive capability with little to no discrimination. Conversely, the CNN-based classifiers presented a heavy improvement, between the two models tested. The first was trained with 2,661 manually labeled samples, while the other included synthetically generated tweets culminating in 12,142 samples. The accuracy scores were 76.35% and 82.31%, with an AUC of 0.90 and 0.91. Using association rule mining in conjunction with the CNN-based classifier showed a high likelihood for keywords such as “smoke”, “cocaine”, and “marijuana” triggering a drug-positive classification. CONCLUSIONS Predictive analysis without a CNN is limited and possibly fruitless. Attribute-based models presented little predictive capability and were not suitable for analyzing this type of data. The semantic meaning of the tweets needed to be utilized, giving the CNN-based classifier an advantage over other solutions. Additionally, commonly mentioned drugs had a level of correspondence with frequently used illicit substances, proving the practical usefulness of this system. Lastly, the synthetically generated set provided increased scores, improving the predictive capability. CLINICALTRIAL None