Dr. Mónica Gracia, international director of admissions at the Iberoamerican University Foundation (FUNIBER), and Dr. Eduardo Silva, executive director of the Foundation in Guatemala, are participating in an international study to develop a new machine learning model capable of detecting suicidal ideation in social media posts with greater accuracy than traditional deep learning methods.
The study, entitled «Suicide Ideation Detection Using Social Media Data and Ensemble Machine Learning Model,» was recently published in Springer’s International Journal of Computational Intelligence Systems.
The research addresses a critical public health issue: the difficulty of early identification of at-risk individuals through conventional clinical assessments, which often require significant time and direct intervention by experts. Given the growing use of digital platforms to express emotional distress, this study proposes an automated solution that analyzes the language used in forums to identify warning signs.
The authors designed the Soft Voting Ensemble Model (SVEM). This system is not based on a single algorithm, but integrates three powerful classifiers: Random Forest, Logistic Regression, and Stochastic Gradient Descent. Instead of relying on a single decision, the model uses a «soft voting» mechanism. This means that the system calculates the average of the probabilities predicted by each of the three algorithms to arrive at a more robust and balanced final conclusion, drastically reducing the margin of error that an individual algorithm would have.
Training was performed using data from the Reddit platform, chosen for the anonymity it offers, which encourages more honest and in-depth discussions about mental health in subforums such as «Suicide Watch» and «Depression.» The model processes text using hybrid feature engineering: it analyzes keyword frequency (TF-IDF) and uses the «bag of words» model to detect linguistic patterns, phrases of despair, and semantic structures associated with imminent risk.
Detection model results
The study results indicate that the SVEM model achieved 94 % accuracy, outperforming more complex deep learning approaches such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, which achieved 91 % and 92 % accuracy, respectively. In addition to its high effectiveness, the proposed model stands out for its lower computational complexity, which facilitates its implementation in real-time monitoring systems without requiring supercomputers, facilitating its integration into digital health applications.
A key aspect of the research is the incorporation of LIME (Local Interpretable Model-agnostic Explanations), a technique that gives the model «explainability.» This tool is essential for clinical practice, as it opens up the «black box» of AI. LIME highlights exactly which words or fragments of the message (such as «useless,» «end it all,» or «burden») triggered the alert. This allows psychiatrists and psychologists to validate the machine’s diagnosis, understand the context, and make informed, human decisions.
If you want to learn more about this study, click here.
To read more research, check out the UNEATLANTICO repository.
The Iberoamerican University Foundation (FUNIBER) offers scholarships to study academic programs such as the Master’s Degree in Strategic Management in Software Engineering. Among the benefits of studying at FUNIBER’s International University Network are its financial aid scholarship program, a flexible distance learning methodology that adapts to professional schedules, and the possibility of obtaining a double university degree.