AI Reads Brief Clinical Notes to Flag Suicide Risk, Sometimes Outperforming Specialists

Suicide remains one of the most difficult public health problems to catch in time. The World Health Organization ranks it as the third leading cause of death among people aged 15 to 29, and research consistently shows that many individuals see a healthcare provider shortly before an attempt, yet only a fraction are recognized as being at risk during those visits. A new study from researchers at the Center for Scientific Research and Higher Education at Ensenada in Mexico now asks a deceptively simple question: if all a clinician has is a few short, free-text sentences typed during a first-contact interview, can artificial intelligence detect signs of suicidal behavior from those fragments alone?

The research, published in Frontiers in Neuroinformatics, is unusual because it deliberately strips away almost everything that makes suicide risk prediction feasible. Most previous work in this field has relied on rich longitudinal electronic health records, years of clinical notes, structured variables, or multimodal data. The Mexican team, by contrast, worked exclusively with minimal first-contact textual inputs in Spanish, assuming no prior clinical history, no follow-up information, and no standardized assessment scores. This mirrors the harsh reality of frontline screening scenarios, such as initial interviews, online intake forms, or early telehealth interactions, where a general practitioner may have only minutes and a handful of fragmented notes.

The dataset comprised 374 anonymized records collected over roughly five years by a deployed screening platform serving health services in Nayarit, Mexico. Each record had been labeled by mental health specialists into one of four categories: no suicidal behavior, suicidal ideation, suicide planning, or suicide attempt. The labels reflected expert judgment supported by validated instruments such as the Plutchik Suicide Risk Scale. The class distribution was starkly imbalanced, with 279 attempt cases, 48 planning cases, 36 ideation cases, and only 11 records of no suicidal behavior, a skew that reflects real clinical referral patterns rather than a representative sample of the population. Most participants were young: 63 percent were 30 years old or younger, and 74 percent were female.

Before building any models, the researchers probed whether the texts even contained separable signals. Statistical tests showed that note length did not meaningfully distinguish the classes, with only the ideation-versus-attempt comparison reaching significance after correction. Word-frequency analysis revealed that the binary classes shared just about 5 percent of their vocabulary, but the suicidal behavior subcategories overlapped heavily. When the team projected the texts into low-dimensional space using embeddings from Spanish GloVe word vectors and the BETO transformer, no discernible cluster boundaries emerged at all; data points from different diagnostic categories were thoroughly intermingled. In other words, the semantic signatures of ideation, planning, and attempt are not visually distinct, a finding that foreshadowed the difficulty of the task.

To overcome the scarcity of data, which would normally starve modern machine learning models that expect thousands of examples, the team turned to data augmentation. They adapted the Easy Data Augmentation framework for Spanish using two large language models, Llama 3 70B and GPT-4o, to perform synonym replacement, random insertion, swapping, and deletion while preserving clinical tone. They also used back-translation through nine intermediate languages to generate paraphrases, and prompted GPT-4o to synthesize roughly 7,000 new first-person, interview-style responses conditioned on each diagnostic category. Crucially, they validated these synthetic texts on two fronts: word-frequency distributions followed Zipf’s law just like natural language, and five licensed clinical psychologists, rating texts blind, judged most of them plausible as realistic patient responses and broadly consistent with their intended labels.

The classification pipeline was split into two tasks. The first, a binary identification problem, asked whether a text indicated any suicidal behavior at all. A Support Vector Machine operating on averaged 300-dimensional Spanish word embeddings proved most robust, achieving a macro F1 score of 0.53 and an overall accuracy of 0.86. Its recall for the suicidal behavior class reached 0.97, meaning it almost never missed a true positive case, though recall for the non-suicidal class was only 0.09, reflecting a deliberate bias toward minimizing false negatives, the clinically costlier error. The second task assigned positive cases to ideation, planning, or attempt. Here BETO, a Spanish BERT transformer fine-tuned with a classification head, reached a macro F1 of 0.36 and an accuracy of 0.73, but its confusion matrix showed misclassifications across all categories, consistent with the overlapping vocabulary revealed earlier.

The most striking comparison came when five mental health specialists, all experienced with the same screening platform, were given the identical text-only task. Reading 20 real transcripts without access to patient history, visual cues, or standardized instruments, the specialists achieved a macro F1 of just 0.29 on the binary task and 0.15 on the multiclass task. Confusion matrices suggested clinicians tended to underestimate suicidal behavior when relying solely on free text, particularly for the intermediate categories of ideation and planning. On these limited numbers, the automated models scored higher than the human experts under the same information constraints.

The authors are careful to stress what this comparison does not mean. It was based on only 20 real texts evaluated with double annotation, lacked formal statistical significance testing or inter-rater reliability analysis, and involved an artificial task that removed the contextual information clinicians normally use. The dataset itself was an operational census shaped by referral policies rather than a population-representative sample, which limits generalization to lower-risk settings such as primary care or community screening. An additional methodological caveat is that the binary classifier learned the negative class exclusively from synthetic examples, because all 11 authentic non-suicidal records were reserved for evaluation, so the learned decision boundary for that class depends on how representative the generated texts really are.

Even with these caveats, the findings carry a provocative message. Under an information bottleneck so severe that even experienced clinicians struggled, machine learning combined with carefully validated data augmentation managed to extract meaningful risk signals from a few dozen words of Spanish text. The performance levels, modest in absolute terms compared with studies that enjoy full clinical records, may represent practical upper bounds for what text-only, first-contact screening can achieve, and further gains are likely to depend more on richer or multimodal inputs than on incremental architectural tweaks. The researchers envision the system not as a diagnostic tool but as a decision-support aid embedded in a human-in-the-loop workflow, flagging patients for referral to trained mental health professionals in settings where specialized assessment is scarce. Future work, they suggest, should expand the corpus of authentic records, especially non-suicidal ones, and explore combining text with voice and sociodemographic data to sharpen early detection where it matters most: the very first conversation.

Subject of Research: Natural language processing and machine learning for detecting suicidal behavior from short Spanish-language first-contact clinical notes

Article Title: Natural language processing for suicidal behavior detection in short Spanish clinical notes from first-contact interviews

Article References: Natural language processing for suicidal behavior detection in short Spanish clinical notes from first-contact interviews. (n.d.). https://doi.org/10.3389/fninf.2026.1919962

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1919962

Keywords: suicidal behavior detection, natural language processing, machine learning, clinical notes, data augmentation, BETO, support vector machine, first-contact screening, mental health, Spanish language, synthetic data, clinical decision support