We evaluate cognitive impairment (CI) classification from transcripts of speech in English, Slovene, and Korean. We compare zero-shot large language models (LLMs) used as direct classifiers under three input settings—transcript-only, linguistic-features-only, and combined—with supervised tabular approaches trained under a leave-one-out protocol. The tabular models operate on engineered linguistic features, transcript embeddings, and early or late fusion of both modalities. Across languages, zero-shot LLMs provide competitive no-training baselines, but supervised tabular models generally perform better, particularly when engineered linguistic features are included and combined with embeddings. Few-shot experiments focusing on embeddings indicate that the value of limited supervision is language-dependent, with some languages benefiting substantially from additional labelled examples while others remain constrained without richer feature representations. Overall, the results suggest that, in small-data CI detection, structured linguistic signals and simple fusion-based classifiers remain strong and reliable signals.

Recent work on clinical applications of language technology shows considerable potential for people with speech and language symptoms and disorders, including for the diagnosis and monitoring of diseases and disorders as well as the development of novel communication aids. This has resulted in a variety of digital health tools becoming accessible, including personalized automatic speech recognition for disordered speech and the monitoring of disease progression in neurodegeneration through language samples. Currently, these tools are almost exclusively accessible to speakers of high-resource languages. A major hurdle for small, lower-resourced language communities in this context is the creation of clinical language corpora. We describe ongoing efforts to build the necessary infrastructure for clinical speech and language data collection in Iceland through the Icelandic Language Biobank, a resource that leverages collaboration with clinicians and robust linguistically-informed data collection against data scarcity.

Automatic Speech Recognition (ASR) offers a scalable and cost-efficient alternative to manual transcription and is becoming increasingly relevant in clinical contexts, particularly for the detection of cognitive decline and mental health assessment. However, current ASR-systems still struggle with spontaneous speech, particularly when processing disfluencies, pauses, and speaker variability that often carry diagnostic value. This study evaluates state-of-the-art open ASR models targeting Swedish using recordings from the "Trip to Stockholm" discourse narrative task which elicits ecologically valid, cognitively demanding speech. Recognition quality is assessed using various metrics, alongside an analysis of linguistic and technical sources of error focused on disfluencies. Our findings show that disfluency-related phenomena degrade recognition performance. Possible post-processing strategies can improve specific error patterns emerging for filled pauses, word repetitions, and self-corrections. The results illustrate both the advances and ongoing limitations of ASR for spontaneous Swedish speech, emphasizing the need for models explicitly trained, or fine-tuned, on disfluent data to ensure robustness in clinical and research applications.

Patients with diverse neurocognitive conditions frequently exhibit measurable language deficits that serve as biomarkers for differential diagnosis and therapy decision making. Discourse analysis can offer reliable ecological measures of human communication, yet manual discourse analysis is cumbersome. Recent advances in automated analysis software provide quick and easy extraction of raw language metrics in the clinic. Nevertheless, transforming these measures into actionable clinical insights remains a significant challenge. The aim of this paper is to present the Automated Language Biomarker Application (ALBA), an integrated framework developed within the Open Brain AI ecosystem to bridge the gap between feature extraction and clinical interpretation. ALBA provides clinicians with a robust statistical infrastructure to benchmark individual patient measures against standardized, large-scale clinical corpora. By utilizing a shared elicitation and processing pipeline, the application ensures that user-provided data are directly comparable to population norms for conditions including Aphasia, Mild Cognitive Impairment (MCI), Dementia, and other neurological conditions. The system implements adaptive statistical logic, employing one-sample t-tests and robust non-parametric alternatives to provide real-time significance testing and dynamic visualizations (box, bar, and violin plots). By automating the comparison of "Language Signatures" against healthy controls and specific clinical phenotypes, ALBA facilitates rapid, evidence-based decision-making in both research and rehabilitation contexts.

Cognitive decline refers to the gradual loss of thinking abilities, including memory, attention, reasoning, and problem-solving. It can be a normal part of ageing or a symptom of conditions like dementia or Alzheimer’s disease when it significantly interferes with daily life. Early diagnosis is crucial, as timely intervention can slow progression and improve quality of life. Emerging approaches such as Digital Linguistic Biomarkers, subtle changes in speech and language patterns captured through digital tools, offer a promising, non-invasive way to detect early signs of cognitive decline before more obvious symptoms appear and perform massive population screening. In this position paper, we contend that the prevailing paradigm for the automatic detection of cognitive decline, primarily relying on classifiers that analyse subjects’ linguistic productions at a single point in time, is not the most effective approach. Instead, we advocate for a paradigm shift toward longitudinal analyses that track linguistic patterns over decades. To support this perspective, we present an experiment in which we compile and analyse a long-term corpus of spontaneous speech productions from well-known individuals, enabling insights into cognitive changes across extended time spans.

Language sample analysis (LSA) is a process that complements standardized psychometric tests for diagnosing, for example, developmental language disorder (DLD) in children. However, its labor-intensive nature has limited its use in speech-language pathology practice. We introduce an approach that leverages natural language processing (NLP) methods not based on commercial large language models (LLMs) applied to transcribed speech data from 119 children in the German-speaking part of Switzerland with typical and atypical language development. This preliminary study aims to identify optimal practices that support speech-language pathologists in diagnosing DLD more efficiently with active involvement of human specialists. Preliminary findings underscore the potential of integrating locally deployed NLP methods into the process of semi-automatic LSA.

Depression is the leading cause of global disability and early detection is crucial for effective intervention. Recent advances in large language models (LLMs) offer potential for analyzing text to identify depression symptoms. This work investigates the zero-shot capability of LLMs to recognize nine DSM5 depression symptoms from short-text inputs. We evaluated eight open LLMs with model sizes ranging from 1.5B to 14B parameters using a clinically annotated dataset and assessed both overall agreement and symptom-level performance. Results indicate that while smaller models exhibit limited clinical accuracy, the Qwen 2.5-7B model achieves substantial performance with a Cohen’s Kappa of 0.603 and a Macro F1 score of 0.648. Notably, a performance plateau between the 7B and 14B Qwen variants suggests that model scaling alone does not guarantee improved symptom-level classification, establishing Qwen 2.5-7B as a resource-efficient model. Further analysis of the best-performing model revealed strengths in identifying salient symptoms like suicidal thoughts, but limitations in recognizing core symptoms such as depressed mood and anhedonia. Misclassification analysis reveals that the model frequently misclassifies posts expressing ’depressed mood’ as ’no symptom’ or vice versa, often overlooking indicators of irritability or social withdrawal. These findings suggest that resource-efficient LLMs can support preliminary symptom screening in zero shot settings, but there is risk of overlooking clinically important symptoms without fine-tuning.

Depression is a pressing global public health issue, yet publicly available Chinese-language resources for depression risk detection remain scarce and largely focus on binary classification. To address this limitation, we release CNSocialDepress, a benchmark dataset for depression risk detection on Chinese social media. The dataset contains 44,178 posts from 233 users; psychological experts annotated 10,306 depression-related segments. CNSocialDepress provides binary risk labels along with structured, multidimensional psychological attributes, enabling interpretable and fine-grained analyses of depressive signals. Experimental results demonstrate the dataset’s utility across a range of NLP tasks, including structured psychological profiling and fine-tuning large language models for depression detection. Comprehensive evaluations highlight the dataset’s effectiveness and practical value for depression risk identification and psychological analysis, thereby providing insights for mental health applications tailored to Chinese-speaking populations.

Despite advancements in NLP-based mental health screening, research remains predominantly English-centric, leaving under-resourced languages insufficiently explored. This study investigates depression detection in Modern Greek social media through a series of experiments. We benchmark traditional machine learning (ML) models against transformer architectures (GreekBERT, GreekSocialBERT, mBERT, and XLM-R) under two settings: a topic-oriented control corpus and a high-similarity stress-test contrasting a gold case of a depressed user with a matched control. Transformer models consistently outperform ML models (F1 = 0.95) but offer limited interpretability. To address this limitation, we incorporate LIWC-derived psycholinguistic features with SHAP explanations to examine model behavior in relation to established linguistic markers. The analysis reveals linguistic patterns consistent with depressive symptoms, such as reduced work-related engagement, social withdrawal, and the motivational deficits characteristically linked to anhedonia in clinical literature. Overall, the results provide a baseline for depression detection in Modern Greek and underscore the importance of grounding automated screening in clinically interpretable evidence.

Psychopathy is a complex personality disorder characterized by persistent deficits in empathy and manipulative behavior. Traditional diagnostic methods often rely on subjective clinical assessments, which are susceptible to deception. This research proposes an objective, non-invasive computational framework for profiling psychopathic traits using Natural Language Processing (NLP) and Machine Learning. We developed a systematic pipeline utilizing transcribed interviews from confirmed criminal psychopaths and a balanced control group. To address data sparsity and noise, we employed the Dynamic Variance Thresholding (DyVaT) algorithm to construct a semantically dense vocabulary of over 1,300 features. The methodology integrates advanced preprocessing, TF-IDF vectorization, and synonym-based data augmentation to enhance model generalization. Among the evaluated classifiers, a Linear Support Vector Machine (SVM) achieved the highest performance, with an accuracy of 0.8081 and an F1-score of 0.7957. Our findings demonstrate the efficacy of linguistic biomarkers and feature importance analysis in distinguishing psychopathic speech patterns. This study provides a scalable methodology for early screening and diagnostics, with significant implications for forensic psychology, security, and ethical AI deployment in mental health.

This work discusses sexual offending, specifically child sexual abuse material (CSAM), in the context of prevention. We introduce a domain-specific, span-level annotation scheme and guidelines to identify psychosocial risk and protective factors in therapist-led, anonymous chat interventions with voluntarily help-seeking individuals concerned about their pedophilic interests and the risk of CSAM use. The scheme is grounded in previous research and clinical experience, and intended for within-intervention guidance and longitudinal tracking, rather than actuarial risk scoring. Annotating a pilot subset (8 clients, 31 sessions), inter-annotator agreement was moderate but improved after calibration, which is consistent with the linguistic and clinical ambivalence present in the data. We track a session-wise Protective Ratio, i.e., the share of protective factors among all coded factors, and examine its behaviour over time during the intervention and around self-reported relapse within clients. In exploratory automation, LLM-based span extraction outperforms BERT baselines but overall performance remains limited by small data and mixed-evidence spans. While complete anonymisation of the corpus is in progress, we release the label scheme, guidelines, and non-sensitive artefacts of our analyses.

This study investigates the use of cosine similarity measures across syntactic, lexical, and semantic vector repre- sentations to detect repetitions in the spontaneous speech of autistic children. It focuses on direct repetitions (i.e., immediate verbatim repetitions of linguistic output produced by another individual) and self-repetitions (i.e., within-speaker recurrence). The performance of similarity-based methods is then compared with state-of-the-art black-box classification models based on BERT, trained on the same data. Using spontaneous speech data from French- and Dutch- speaking autistic children, the results show that lexical and semantic similarity provide reliable cues for identifying self-repetitions, achieving high precision and recall, with F1-scores exceeding 83%, comparable to those obtained by BERT-based models. In contrast, direct repetitions are more difficult to detect using similarity-based approaches, with BERT models clearly outperforming them and reaching F1-scores above 73%. Across all conditions, syntactic similarity consistently underperforms relative to lexical and semantic measures. These findings highlight the strengths and limitations of similarity-based approaches and suggest directions for future research, particularly in improving the detection of direct repetitions and assessing the cross-linguistic generalizability of these methods.