When Can We Trust LLMs in Mental Health? Large-Scale Benchmarks for Reliable LLM Evaluation

Evaluating Large Language Models (LLMs) for mental health support is challenging due to the emotionally and cognitively complex nature of therapeutic dialogue. Existing benchmarks are limited in scale, reliability, often relying on synthetic or social media data, and lack frameworks to assess when automated judges can be trusted. To address the need for large-scale dialogue datasets and judge reliability assessment, we introduce two benchmarks that provide a framework for generation and evaluation. MentalBench-100k consolidates 10,000 one-turn conversations from three real scenarios datasets, each paired with nine LLM-generated responses, yielding 100,000 response pairs. MentalAlign-70k}reframes evaluation by comparing four high-performing LLM judges with human experts across 70,000 ratings on seven attributes, grouped into Cognitive Support Score (CSS) and Affective Resonance Score (ARS). We then employ the Affective Cognitive Agreement Framework, a statistical methodology using intraclass correlation coefficients (ICC) with confidence intervals to quantify agreement, consistency, and bias between LLM judges and human experts. Our analysis reveals systematic inflation by LLM judges, strong reliability for cognitive attributes such as guidance and informativeness, reduced precision for empathy, and some unreliability in safety and relevance. Our contributions establish new methodological and empirical foundations for reliable, large-scale evaluation of LLMs in mental health. We release the benchmarks and codes at: https://github.com/abeerbadawi/MentalBench/

Leveraging Self-Training and Variational Autoencoder for Agitation Detection in People with Dementia Using Wearable Sensors

Dementia is a neurodegenerative disorder that has been growing among elder people over the past decades. This growth profoundly impacts the quality of life for patients and caregivers due to the symptoms arising from it. Agitation and aggression (AA) are some of the symptoms of people with severe dementia (PwD) in long-term care or hospitals. AA not only causes discomfort but also puts the patients or others at potential risk. Existing monitoring solutions utilizing different wearable sensors integrated with Artificial Intelligence (AI) offer a way to detect AA early enough for timely and adequate medical intervention. However, most studies are limited by the availability of accurately labeled datasets, which significantly affects the efficacy of such solutions in real-world scenarios. This study presents a novel comprehensive approach to detect AA in PwD using physiological data from the Empatica E4 wristbands. The research creates a diverse dataset, consisting of three distinct datasets gathered from 14 participants across multiple hospitals in Canada. These datasets have not been extensively explored due to their limited labeling. We propose a novel approach employing self-training and a variational autoencoder (VAE) to detect AA in PwD effectively. The proposed approach aims to learn the representation of the features extracted using the VAE and then uses a semi-supervised block to generate labels, classify events, and detect AA. We demonstrate that combining Self-Training and Variational Autoencoder mechanism significantly improves model performance in classifying AA in PwD. Among the tested techniques, the XGBoost classifier achieved the highest accuracy of 90.16\%. By effectively addressing the challenge of limited labeled data, the proposed system not only learns new labels but also proves its superiority in detecting AA.