Analyzing Social Network Data for National Self-Harm Trend Predictions

Abstract

Introduction: Self-harm means the conscious demonstration of self-harming or self-injury, regardless of the basic inspiration or the level of self-destructive aim, which might prompt injury or casualty. Self-mischief and self destruction are huge issues, especially in unfortunate countries. A new report shows that around 77% of self destruction occasions happen in low-and center pay countries. This pattern is connected to the reception of specialized advancements and the quick movement of urbanization here.

Objectives: The goal of this research is to create and test a framework called FAST (Forecasting self-harm using Aggregated Social media patterns), which uses extensive social media data to forecast and nowcast patterns in self-harm at the national level. The research assesses the predictive ability of machine learning regressors to improve the accuracy of self-harm trend predictions, particularly in areas with little historical statistical data, by extracting psychological signals using language-agnostic language models and turning them into time series data.

Methods: This exploration presents FAST, a framework expected to foresee self-harm drifts broadly by utilizing mental signs got from broad online entertainment information. These signs go about as a proxy for real populace psychological well-being, which may be used to work on the consistency of self-harm designs. Language-skeptic calculations are at first prepared to remove different mental signs from amassed virtual entertainment interchanges. Accordingly, these signs are united and handled into multivariate time series, whereupon the time-delay implanting strategy is utilized to change over them into worldly inserted examples. At last, different ML regressors are surveyed for their prescient ability. We have explored different avenues regarding two of the most conspicuous ML calculations: the Decision Tree method and the Voting Regressor. In contrast with elective techniques, it yields a decreased MAE error.

Results: This study divides as Mean Absolute Error (MAE): Showed better overall accuracy with a lower average prediction error as compared to baseline models. The model's ability to handle significant deviations was confirmed by the Root Mean Squared Error (RMSE), which displayed a decreased squared error size. Mean Absolute Percentage Error (MAPE): The model's average improvement over the ARIMA baseline was 43.56% in predicting deaths from self-harm and 36.48% in predicting injuries.

Conclusions: This study presents FAST, a unique approach that uses psychological signals from social media to forecast national trends in self-harm. FAST predicted self-harm deaths and injuries in Thailand with high accuracy using 12 mental health variables from tweets. Based on MAPE, FAST increased forecast accuracy by 43.56% for deaths and 36.48% for injuries when compared to the conventional ARIMA model. Although additional enhancements utilizing models like as Decision Tree and Voting Regressor are recommended for future research, the results demonstrate FAST's potential for prompt public health intervention.