Natural language processing for suicidal behavior detection in short Spanish clinical notes from first-contact...

Suicide is a multifactorial phenomenon influenced by social, cultural, biological, and psychological factors, whose early detection remains challenging. According to the World Health Organization, suicide is the third leading cause of death among individuals aged 15-29 years, representing a major public health concern for young populations worldwide. In first-contact clinical settings, screening opportunities are often constrained by limited time, sparse patient information, and operational workflows. This study evaluates the feasibility of detecting suicidal behavior from short Spanish-language texts collected during initial screening interviews.

The dataset comprises 374 anonymized records labeled as "no suicidal behavior," "ideation," "planning," or "attempt," with a strong imbalance toward "attempt," reflecting operational clinical usage rather than population-representative sampling. To address data scarcity, we applied data augmentation methods (Easy Data Augmentation for Spanish, backtranslation, and large language model-based synthesis), validated through statistical analysis and clinician-based plausibility assessments.

Exploratory analyses revealed highly constrained and overlapping lexical patterns across classes, indicating the challenging nature of the task. A Support Vector Machine achieved 0.53 macro F ₁ in binary classification ("suicidal behavior" vs. "no behavior"), while BETO, a Spanish BERT model, reached 0.36 macro F ₁ in multiclass classification. Under the same text-only conditions, human specialists achieved macro F ₁ scores of 0.29 and 0.15 respectively; however, these results were based on only 20 real texts evaluated with double annotation across five specialists and should therefore be interpreted cautiously.

Although performance is modest compared to studies using full clinical records, these findings suggest that short-text machine learning could support nonspecialists in first-contact settings. Future work should expand datasets and explore multimodal approaches, including voice and text, to improve clinical applicability and early detection.