2026· Workshop on Computational Linguistics and Clinical Psychology· pp. 547-553· 1 citation· 18 references
TL;DR
This paper presents a system for the CLPsych 2026 Shared Task on longitudinal mental health modeling from social media timelines, grounded in the MIND framework, achieving the top Fit and Specificity scores in Task 3.2, demonstrating the benefits of explicit clinical grounding for conceptual accuracy.
Abstract
This paper presents a system for the CLPsych 2026 Shared Task on longitudinal mental health modeling from social media timelines, grounded in the MIND framework (Atzil-Slonim, 2025). MIND conceptualizes mental health as evolving self-states defined by A ffect, B ehavior, C ognition, and D esire (ABCD), providing a structured lens on mental health trajectories. The system centers on a theory-explicit prompting framework for structured sequence summarization (Task 3.1) and recurrent dynamic signature extraction (Task 3.2), encoding the full ABCD taxonomy directly into the LLM prompt to ensure clinically grounded, inter-pretable outputs. A three-stage pipeline infers a direction-of-change label per sequence, produces structured ABCD summaries with few-shot exemplar augmentation, and aggregates these summaries to derive cross-individual recurrent patterns. The system ranks first on deterioration-related recurrent signatures and second overall, achieving the top Fit and Specificity scores in Task 3.2, demonstrating the benefits of explicit clinical grounding for conceptual accuracy.
It is shown that con-textual transformer/LLM models yield more reliable macro-level performance under imbalance than TF–IDF baselines, particularly for semantically adja- cent classes.
Nehal Shah, Mehul Barot· International journal of com...· 0 citations
Annotation quality is a major bottleneck in building reliable and explainable artificial intelligence (XAI) systems for mental health research. In depression-related datasets, labels are often assigned without structured evidence, symptom-level justification, or traceable alignment with the criteria of the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR), limiting both transparency and downstream model interpretability. We propose a self-evolving, expert-in-the-loop annotation framework for Major Depressive Disorder (MDD) that combines large language model (LLM)-assisted labeling with expert verification. The framework is intended to support the construction of explainable, DSM-5-TR-aligned datasets rather than to perform clinical diagnosis. It operates in three stages: candidate evidence selection from textual records, criterion-level DSM-5-TR analysis, and case-level synthesis that produces label-level diagnostic and severity annotations. A dual-memory architecture, composed of Example Memory and Reflection Memory, is designed to internalize expert feedback and iteratively improve future annotations without retraining. We describe this mechanism and leave its evaluation across multiple feedback cycles to future work. In addition to final labels, the framework exports clinical evidence, reasoning traces, and edit histories, enabling comprehensive auditability. In a pilot study using expert-reviewed samples, the proposed approach improves annotation consistency and explainability while reducing manual revision effort.
Hoang-Loc Cao, Van Pham, T. Nguyen et al.· 0 citations
The system reduces supervisory triage latency from 72 hours to real time (~10 seconds per session), enabling proactive intervention in high-risk cases and addresses the cold-start problem through Bayesian priors and implements timestamp-based modality synchronization for robust multi-modal fusion.
Shreeya Sharma, Ravish Gupta, Saket Kumar et al.· 0 citations
The Large Language Model (LLM) provides a more efficient and convenient way for long-term medical dialogue by using its understanding and generation ability. However, LLM also has some core challenges, such as the forgetting of key medical history and the insufficient capture of disease trend. This paper proposes a dynamic memory retrieval framework based on dual-graph enhancement. The framework constructs a two-tier architecture of fine-grained event graph and macro portrait. Specifically, the event graph connects events across time through entity nodes, and saves the semantic relationships between events; The portrait part organizes the events in multiple macro dimensions to maintain the evolution trend summary of the patient’s condition, psychology and living habits. Experiments based on the Long-Term Health Monitoring Dataset (LTHM) show that the comprehensiveness of this framework is 0.708, which is superior to other baseline models, with the relevance of 0.973 and the faithfulness of 0.947, ensuring the accuracy and reliability of the answers.
BackgroundPosttraumatic stress disorder (PTSD) is common yet frequently underdiagnosed, in part due to barriers to systematic screening and the reliance on self-report instruments. Large language models (LLMs) have shown promise in extracting clinically relevant information from unstructured language, but their ability to infer item-level PTSD symptom severity from clinical interviews remains unclear.MethodsUsing the Distress Analysis Interview Corpus-Wizard of Oz (DAIC-WoZ), we analyzed 100 semi-structured clinical interview transcripts paired with item-level PTSD Checklist-Civilian Version (PCL-C) scores. Six LLMs (DeepSeek 3.1, Claude Sonnet 4, LLaMA 4 Scout, GPT-4o, GPT-5, and Gemini 2.5 Flash) used zero-shot prompting to predict all 17 PCL-C items. Performance was assessed for binary symptom endorsement (≥3 vs. < 3), 5-point Likert prediction, and DSM-IV symptom-cluster analyses using accuracy, F1 score, and Matthews correlation coefficient (MCC).ResultsFor binary prediction, Claude 4 achieved the highest mean accuracy (0.705; 95% CI, 0.681-0.728), followed by DeepSeek 3.1(0.699; 95% CI, 0.675-0.724) and Gemini 2.5 (0.698; 95% CI, 0.677-0.718). For Likert prediction, DeepSeek 3.1 performed best (accuracy = 0.438; 95% CI, 0.401-0.475), only modestly above the majority-class baseline (0.399; 95% CI, 0.355-0.443). Performance varied by symptom domain, with re-experiencing and hyperarousal symptoms generally predicted more accurately than avoidance/numbing symptoms. Across models, predicted item-level symptom patterns showed a meaningful alignment with observed PCL-C responses despite reduced accuracy in fine-grained severity estimation.ConclusionZero-shot LLMs' performance was insufficient for clinical application in predicting PTSD symptoms from semi-structured interview transcripts. While models showed some ability to capture overall symptom patterns, performance varied across domains and remained limited for fine-grained severity estimation. Given these constraints and the non-trauma-specific nature of the dataset, findings should be interpreted as preliminary, with only modest differences observed between models.Plain Language Summary TitleCan Artificial Intelligence Identify PTSD Symptoms from Conversations? A Study Using Clinical Interview TranscriptsPlain Language SummaryPost-traumatic stress disorder (PTSD) is a common mental health condition, but it is often missed in clinical settings. Screening usually relies on questionnaires that patients must complete themselves, which may not always happen due to time, stigma, or discomfort discussing trauma. Researchers are exploring whether artificial intelligence (AI) could help identify PTSD symptoms from conversations instead.In this study, we tested several advanced AI systems, known as large language models, to see if they could estimate PTSD symptoms based on written transcripts of clinical interviews. These interviews were not specifically designed to assess trauma, which makes the task more challenging but closer to real-world situations. We compared the AI predictions to participants' own questionnaire responses about their symptoms.We found that the AI models were somewhat able to recognize general patterns of PTSD symptoms, especially more visible ones like sleep problems or distressing dreams. However, they struggled with more internal or less obvious symptoms, such as avoidance or emotional numbness. Overall, their accuracy was moderate and not reliable enough for clinical use, particularly when trying to estimate how severe symptoms were.Importantly, differences between the AI models were small, and none performed well enough to replace existing screening methods. These findings suggest that while AI may have future potential as a supportive tool, it is not yet ready to be used for diagnosing or screening PTSD on its own.Further research using better data, improved methods, and real clinical settings is needed before this approach could be considered for practical use.
Bazen Gashaw Teferra, Christian Kevin Sidharta, Wei-Ni Hsiang et al.· Canadian journal of psychiat...· 0 citations
Multimodal behavioral analysis offers a scalable approach to assessing depression, anxiety, and stress, yet generic fusion models often ignore the psychometric structure of questionnaire labels. In DASS-21, risk labels are derived from ordered symptom items through fixed item-to-subscale mappings. We propose \textbf{DynaBridge}, a dynamic summary-guided cross-task multimodal framework for DASS-structured mental health assessment. DynaBridge encodes acoustic, visual, and textual cues across multiple sessions and augments them with frozen-LLM-generated DASS-aware summaries as participant-level semantic evidence. It predicts ordinal item distributions, reconstructs depression, anxiety, and stress risk evidence from item-level soft scores, and fuses this evidence with direct multimodal risk predictions. A confidence-aware refinement strategy further incorporates high-confidence semantic cues conservatively. On the official AdoDAS validation split, DynaBridge outperforms the official baseline and representative multimodal methods, achieving 0.5012 mean F1 for D/A/S risk prediction and 0.3216 mean QWK for DASS-21 item prediction. These results show the value of bridging multimodal cues, semantic summaries, and DASS-21 psychometric structure.
Shiyu Teng, H. Yu, Jiaqing Liu et al.· 0 citations