Aug 2026· Acta Psychologica· Vol 269, pp.
107602
· 0 citations· 43 references
Medicine
TL;DR
Current performance estimates of LLMs with respect to depression screening are most likely optimistic, but when restricted to smaller models that could be locally deployed (for privacy protection) in a clinical setting, LLMs do not detect depression with sufficient accuracy, sensitivity, or specificity to be used in a screening programme.
Abstract
Background
Large language models have been evaluated on many healthcare tasks, including depression screening. However, it is unclear whether estimates of performance are accurate, especially in a setting with realistic clinical constraints.
Methods
We use the publicly-available Distress Analysis Interview Corpus - Wizard of Oz (DAIC-WOZ) to give performance estimates of LLMs that respect patient privacy and could be feasibly deployed in a clinical setting. These models are locally run and under 15 billion parameters.
Results
Accuracy, sensitivity, and specificity of the models we evaluated ranged from 0.233-0.677, 0.041-0.929, 0.000-0.729, respectively. There are significant differences in the performance we observed versus other studies that evaluate commercial models. We also demonstrate poor agreement amongst different LLMs.
Conclusion
Current performance estimates of LLMs with respect to depression screening are most likely optimistic. When restricted to smaller models that could be locally deployed (for privacy protection) in a clinical setting, LLMs do not detect depression with sufficient accuracy, sensitivity, or specificity to be used in a screening programme.
The extent to which people use general-purpose large language models (LLMs) for their mental health is unknown. Information about use patterns is important for clinicians, developers, and regulators. We surveyed U.S. adults (n = 1871) between August and October 2024 using stratified sampling across age, sex, and race/ethnicity to approximate national demographics. We found that 24% of participants use LLMs for mental health; they are disproportionately young, male, and Black, and have poor mental health. Participants reported difficulty accessing traditional treatment and using LLMs because they are free, convenient, and available. They report using LLMs for emotional support, learning therapy skills, and supplementing existing therapy. Using Pew-reported estimates of population LLM use, we conservatively estimate that as of 2024, 14–18 million U.S. adults may have been using LLMs for mental health. This work highlights the need for monitoring and evaluation to understand the potential harms and benefits of such use.
Elizabeth C. Stade, Zoe M. Tait, Samuel T. Campione et al.· npj Digital Medicine· 2 citations
Overall, advanced prompting markedly improved model performance, and top-tier LLMs demonstrated robust interpretive capability, while caution is needed for variables with complex clinical semantics such as blood pressure.
Jiwon You, Hangsik Shin· npj Digital Medicine· 0 citations
Introduction Large language models (LLMs) have been proposed as decision support tools in medicine, yet their role in forensic cause of death analysis remains unexplored. Methods In this study, we used 118 real-world cases spanning diverse categories of death to systematically evaluate the performance of four representative LLMs (GPT-4o, OpenAI o3, Gemini-2.5pro, and DeepSeek-R1) in forensic cause of death analysis. Two senior forensic pathologists independently evaluated each model’s decision-making capabilities regarding inference quality and conclusion accuracy. These metrics were assessed using an expert scoring system with a 5-point Likert scale, with original analytical statements and legally valid expert opinions serving as objective gold standards. In a sub-study, we examined the application potential of the locally deployed open-source model DeepSeek-R1:32b. Additionally, a targeted retrospective analysis was conducted to quantify the incidence and typologies of AI hallucinations. Results DeepSeek-R1 demonstrated a statistically significant advantage in inference quality scores over GPT-4o (p = 0.015, rrb = 0.28) and Gemini-2.5pro (p = 0.000003, rrb = 0.46), while no statistically significant differences were observed among the four models in terms of conclusion accuracy scores. The locally deployed DeepSeek-R1:32b model also showed no statistically significant difference from GPT-4o in conclusion accuracy scores. However, hallucinations persistently appear in the response reports of all LLMs. Discussion LLMs can provide limited auxiliary value in cause of death analysis but should not replace the final judgment of forensic experts. LLMs still require expert oversight to ensure evidence integrity and mitigate risks such as hallucination. Open source LLMs can further mitigate data privacy concerns and provide practical support for cause of death analysis.
Enhao Fu, Haojie Qin, Z. Tian et al.· Frontiers in Artificial Inte...· 0 citations
Open LLMs can extract clinical findings from Finnish pediatric records with accuracy comparable to published English benchmarks, and uncertainty-based triage substantially reduces required expert workload.
J. Leinonen, J. Knuutila, S. Kurki et al.· medRxiv· 0 citations
Automated detection of errors in clinical documentation is a promising application of large language models (LLMs), yet decisions to deploy such models rest on benchmarks that evaluate each clinical note in isolation. Error-detection benchmarks are typically constructed by injecting errors into notes, such that each erroneous note has a natural counterpart. Aggregate discriminative metrics (e.g., balanced accuracy or F1) do not exploit this structure. We show that this omission is consequential. In particular, evaluating 15 diverse LLMs on 4 standardized clinical error-detection test sets across 3 languages, we find that 13 of 15 models fall below the level of random pairwise discrimination, even while achieving F1 scores that standard practice would read as moderate. We also observe that the underlying bias patterns differ across languages: the same model can default to"no error"on one language and over-flag errors on another. To diagnose where discrimination breaks down, we further introduce a procedure to score the evidence models cite in their outputs. We find that while models consistently locate error-relevant content, they fail to produce the corresponding correct verdict on the clean counterpart. Finally, we show that F1 and pairwise accuracy are driven in opposite directions by the same underlying bias, so that ranking models by F1 may systematically promote the weakest discriminators. For safety-critical clinical NLP applications, we advocate for supplementing aggregate metrics with paired evaluations in benchmark reporting. Code and analysis scripts are available at https://github.com/healthylaife/paired-clinical-eval.
Abstract Objectives To develop a pipeline for evaluating large language models (LLMs) on the task of capturing symptoms from clinical encounters. Materials and Methods We created a gold standard dataset of symptom annotations from simulated doctor-patient encounter excerpts (264 encounters; 16 symptoms; double-coded and adjudicated). Nine different LLMs from 4 vendors (OpenAI, Meta, DeepSeek, Moonshot AI) were used as examples to test our evaluation pipeline; outputs were assessed for correct structure and symptom information. Results Of 3085 excerpts, 2087 (68%) contained symptoms. Pain, cough, and shortness of breath were most common; LLMs achieved F1 scores ranging 0.66-0.88 for these symptoms with minimal prompt engineering. Of tested models, GPT-4.1 demonstrated the best overall performance. Discussion Our evaluation pipeline and benchmarking dataset are publicly available and applicable to various LLMs, including open-source models. Conclusion This work supports the development and optimization of models that seek to improve patient symptom understanding.
Joshua Davis, B. Durieux, Chloe Van Dongen et al.· JAMIA Open· 0 citations
