A novel multi-model feature generation technique for suicide detection.

Area of Science:

• Computational psychiatry and digital phenotyping.
• Application of machine learning in mental health.
• Social media analytics for public health surveillance.

Background:

• Traditional depression assessment methods have limitations in patient comfort and data availability.
• Social media offers a rich, accessible data source for understanding mental states.
• Early prediction of depression and suicide risk is crucial for timely intervention.

Purpose of the Study:

• To propose an innovative approach for predicting suicide risk using social media content and machine learning.
• To enhance machine learning model performance through a novel multi-model feature generation technique.
• To develop a compact and highly correlated feature set for efficient depression prediction.

Main Methods:

• Utilized term frequency-inverse document frequency (TF-IDF) for feature extraction.
• Developed a probability-based feature set (ProBFS) by combining logistic regression and support vector machine models.
• Evaluated the performance of machine learning models using the ProBFS on social media data.

Main Results:

• The Support Vector Machine (SVM) model achieved a high accuracy of 0.96 using the ProBFS.
• The proposed method demonstrated a low computational time of 5.63 seconds for extensive datasets.
• The ProBFS feature set proved to be compact and highly correlated with depression risk indicators.

Conclusions:

• The novel ProBFS technique significantly enhances the accuracy and efficiency of suicide risk prediction from social media.
• Machine learning models, particularly SVM, show strong potential for early depression and suicide risk detection.
• This approach offers a scalable and effective tool for mental health monitoring using readily available social media data.