Scholarlink

Importance <p>This study employs artificial intelligence methods to predict mood phases in patients with bipolar disorder, addressing the issue of poor prognosis caused by recurrent episodes and uncertainty in mood phases.</p> Objective <p>To explore the patterns of mood transitions in patients with bipolar disorder, we developed a mood phase transition model using a Transformer model to investigate whether predicting mood changes can improve prognosis.</p> Design <p>We conducted cohort follow-up assessments of patients with bipolar disorder. At each visit, patients were evaluated using the Hamilton Depression Rating Scale (HAMD) and the Young Mania Rating Scale (YMRS) as clinician-rated assessments, along with the BDCC self-rating scale. We then input these data into several different AI models for training and validated the models’ performance using the data.</p> Setting <p>The study was conducted through online medical platforms and offline follow-up evaluations.</p> Participants <p>The study included 812 patients diagnosed with bipolar disorder according to DSM-5 criteria, who had at least one BDCC assessment result and at least one depressive episode meeting HAMD criteria and one manic/hypomanic episode meeting YMRS criteria.</p> Result <p>In the experiment utilizing current self-assessment scales for rapid identification of affective states, the best performance was observed with the Transformer and RF models, with AUCs for affective state identification of 0.83 (95% CI 0.77–0.89) for the Transformer model, and 0.88 (95% CI 0.83–0.93) for the RF model. In experiments predicting the next affective state, the AUC for the Transformer prediction was 0.76 (95% CI 0.65–0.87), and for the RF model, it was 0.78 (95% CI 0.68–0.88). In predictions of affective states 90&#xa0;days later, the Transformer model performed best, with accuracies of 82.76%, 79.31%, 58.62%, and 41.38% for the Transformer, CNN, RF, and SVM models, respectively.</p> Conclusions and relevance <p>To some extent, the AI model can predict patients’ mood phase transition patterns, achieving an AUC greater than 0.8. Decision Curve Analysis (DCA) indicates that patients may obtain clinical benefits based on this predictive model. Additionally, the model demonstrates optimal predictive performance at 180&#xa0;days.</p> Trial registration <p><a href="http://ClinicalTrials.gov">http://ClinicalTrials.gov</a> under the identifier NCT02015143</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Utilizing AI models to identify and predict phase transition patterns of bipolar disorder patients

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年4月24日

查看全文
  • Xiaoyang Feng,
  • Yantao Ma,
  • Sikai Ruan,
  • Liangkun Guo,
  • Jun Ji,
  • Yuyanan Zhang,
  • Yaoyao Sun,
  • Zhe Lu,
  • Zhewei Kang,
  • Yundan Liao,
  • Junyuan Sun,
  • Guorui Zhao,
  • Rui Yuan,
  • Yunqin Zhu,
  • Zhiying Li,
  • Yue Zhu,
  • Shuzhe Zhou,
  • Lili Guan,
  • Gary S. Sachs,
  • Fava Maurizio,
  • Xin Yu,
  • Weihua Yue

摘要

Importance

This study employs artificial intelligence methods to predict mood phases in patients with bipolar disorder, addressing the issue of poor prognosis caused by recurrent episodes and uncertainty in mood phases.

Objective

To explore the patterns of mood transitions in patients with bipolar disorder, we developed a mood phase transition model using a Transformer model to investigate whether predicting mood changes can improve prognosis.

Design

We conducted cohort follow-up assessments of patients with bipolar disorder. At each visit, patients were evaluated using the Hamilton Depression Rating Scale (HAMD) and the Young Mania Rating Scale (YMRS) as clinician-rated assessments, along with the BDCC self-rating scale. We then input these data into several different AI models for training and validated the models’ performance using the data.

Setting

The study was conducted through online medical platforms and offline follow-up evaluations.

Participants

The study included 812 patients diagnosed with bipolar disorder according to DSM-5 criteria, who had at least one BDCC assessment result and at least one depressive episode meeting HAMD criteria and one manic/hypomanic episode meeting YMRS criteria.

Result

In the experiment utilizing current self-assessment scales for rapid identification of affective states, the best performance was observed with the Transformer and RF models, with AUCs for affective state identification of 0.83 (95% CI 0.77–0.89) for the Transformer model, and 0.88 (95% CI 0.83–0.93) for the RF model. In experiments predicting the next affective state, the AUC for the Transformer prediction was 0.76 (95% CI 0.65–0.87), and for the RF model, it was 0.78 (95% CI 0.68–0.88). In predictions of affective states 90 days later, the Transformer model performed best, with accuracies of 82.76%, 79.31%, 58.62%, and 41.38% for the Transformer, CNN, RF, and SVM models, respectively.

Conclusions and relevance

To some extent, the AI model can predict patients’ mood phase transition patterns, achieving an AUC greater than 0.8. Decision Curve Analysis (DCA) indicates that patients may obtain clinical benefits based on this predictive model. Additionally, the model demonstrates optimal predictive performance at 180 days.

Trial registration

http://ClinicalTrials.gov under the identifier NCT02015143