Garcia Danilo, Granjard Alexandre, Vanhée Loïs, Berg Matilda, Andersson Gerhard, Lasota Marta, Sikström Sverker
Department of Social Studies, University of Stavanger, Stavanger, Norway; Department of Behavioral Sciences and Learning, Linköping University, Linköping, Sweden; Promotion of Health and Innovation for Well-Being (PHI-WELL), Department of Social Studies, University of Stavanger, Stavanger, Norway; Lab for Biopsychosocial Personality Research (BPS-PR), International Network for Well-Being; Promotion of Health and Innovation (PHI) Lab, International Network for Well-Being; Centre for Ethics, Law and Mental Health (CELAM), University of Gothenburg, Gothenburg Sweden; Department of Psychology, University of Gothenburg, Gothenburg, Sweden.
Promotion of Health and Innovation for Well-Being (PHI-WELL), Department of Social Studies, University of Stavanger, Stavanger, Norway; Lab for Biopsychosocial Personality Research (BPS-PR), International Network for Well-Being; Promotion of Health and Innovation (PHI) Lab, International Network for Well-Being; Department of Psychology, University of Gothenburg, Gothenburg, Sweden.
J Affect Disord. 2025 Jul 15;381:659-668. doi: 10.1016/j.jad.2025.04.003. Epub 2025 Apr 3.
Although patients prefer describing their problems using words, mental health interventions are commonly evaluated using rating scales. Fortunately, recent advances in natural language processing (i.e., AI-methods) yield new opportunities to quantify people's own mental health descriptions. Our aim was to explore whether responses to open-ended questions, quantified using AI, provide additional value in measuring intervention outcomes compared to traditional rating scales.
Swedish adolescents (N = 44) who received Internet-based Cognitive Behavioral Therapy (ICBT) for eight weeks completed (pre/post) scales measuring anxiety and depression and three open-ended questions (related to depression, anxiety and general mental health). The language responses were quantified using a large language model and quantitative methods to predict mental health as measured by rating scales, valence (i.e., words' positive/negative affectivity), and semantic content (i.e., meaning).
Similar to the rating scales, language measures revealed statistically significant health improvements between pre and post measures such as reduced depression and anxiety symptoms and an increase in the use of words conveying positive emotions and different meanings (e.g., pre-intervention: "anxious", depressed; post-intervention: "happy", "the future"). Notably, the health changes identified through semantic content measures remained statistically significant even after accounting for the changes captured by the rating scales.
Language responses analyzed using AI-methods assessed outcomes with fewer items, demonstrating effectiveness and accuracy comparable to traditional rating scales. Additionally, this approach provided valuable insights into patients' well-being beyond mere symptom reduction, thus highlighting areas of improvement that rating scales often overlook. Since patients often prefer using natural language to express their mental health, this method could complement, and address comprehension issues associated fixed-item questionnaires.
尽管患者更喜欢用文字描述自己的问题，但心理健康干预措施通常使用评定量表进行评估。幸运的是，自然语言处理（即人工智能方法）的最新进展为量化人们自己的心理健康描述带来了新机遇。我们的目的是探讨与传统评定量表相比，使用人工智能量化的开放式问题的回答在测量干预效果方面是否能提供额外价值。
44名接受为期八周的基于互联网的认知行为疗法（ICBT）的瑞典青少年完成了（治疗前/后）测量焦虑和抑郁的量表以及三个开放式问题（与抑郁、焦虑和总体心理健康相关）。使用大语言模型和定量方法对语言回答进行量化，以预测评定量表所测量的心理健康、效价（即词语的积极/消极情感）和语义内容（即含义）。
与评定量表类似，语言测量显示治疗前和治疗后在健康方面有统计学上的显著改善，如抑郁和焦虑症状减轻，以及传达积极情绪和不同含义的词语使用增加（例如，干预前：“焦虑”、“抑郁”；干预后：“快乐”、“未来”）。值得注意的是，即使在考虑了评定量表所捕捉的变化之后，通过语义内容测量确定的健康变化在统计学上仍然显著。
使用人工智能方法分析的语言回答用更少的项目评估了效果，显示出与传统评定量表相当的有效性和准确性。此外，这种方法除了单纯的症状减轻之外，还为患者的幸福感提供了有价值的见解，从而突出了评定量表经常忽略的改善领域。由于患者通常更喜欢使用自然语言来表达他们的心理健康，这种方法可以补充并解决与固定项目问卷相关的理解问题。
