Diagnosing suicidal behavior is inherently complex, and recognizing it during initial clinical encounters is even more challenging, especially for general practitioners who may lack specialized training in mental health. Natural Language Processing (NLP) offers a promising avenue for developing tools to support early identification of suicide risk, such as during first contact in emergency departments, while leaving the final diagnosis to qualified mental health professionals. This study investigates the use of NLP to develop such a model by applying machine learning algorithms, including Support Vector Machines (SVMs) and BETO, to a real-world dataset comprising 374 anonymized patient interviews. These interviews represent initial consultations between patients and clinicians in mental health settings. The goal was to assess whether suicidal behavior could be detected solely from the brief textual exchanges recorded during these sessions. The best-performing SVM model achieved a macro \(F_1\) -score of 53.26% in binary classification, while the best BETO model reached 38% in multiclass classification. These results underscore the inherent difficulty of diagnosing mental health conditions and highlight the ongoing challenges in accurately identifying suicidal behavior from limited clinical text.