Wang Xianghe, Gao Tianqi, Guo Xiaodong, Huang Bingjie, Ji Yunfei, Hu Wanheng, Yin Xiaolin, Zheng Yue, Pu Chengcheng, Yu Xin
Peking University Sixth Hospital, Beijing, China.
Peking University Institute of Mental Health, Beijing, China.
Schizophrenia (Heidelb). 2025 Jul 22;11(1):105. doi: 10.1038/s41537-025-00640-y.
Cognitive impairments are generally observed in patients with schizophrenia. However, it is unclear whether neurocognitive dysfunction can predict the efficacy of antipsychotics for first-episode schizophrenia (FES). Machine learning methods provide a relatively unbiased approach when evaluating heterogeneous data, especially when building multifactor prediction models. This study conducted a secondary analysis based on the Chinese FES Trial (CNFEST), which was a 1-year study involving a randomized controlled trial for the first eight weeks followed by a 48-week open-label observation. The current study aimed to build a prediction model of eight-week antipsychotic response based on baseline clinical and demographic features. Six machine learning algorithms, including random forest, eXtreme gradient boosting (XGBoost), logistic regression, linear support vector machine (SVM), radial basis function SVM and poly SVM were applied and compared to draw the prediction model. The predictive effects were evaluated by balanced accuracy, sensitivity and specificity. The predictive factors were compared with F scores. A total of 450 qualified subjects contributed to the model. The prediction model constructed via XGBoost algorithm had the highest accuracy (68.8%) and prognostic certainty (44.3%) among all the algorithms. The baseline neurocognitive tests with strong predictive significance were the Grooved Pegboard Test, Trail Making Test Part A, Paced Auditory Serial Addition Test, Brief Visuospatial Learning Test, Hopkins Verbal Learning Test and Color Trails Test. This study emphasizes the importance of fine motor skills, verbal learning, visual learning, working memory and attention for the response of drug-naïve FES patients to antipsychotics. The model generated by XGBoost, which shows preferable accuracy, provides psychiatric practitioners with a possible way to predict efficacy for FES patients.
精神分裂症患者通常会出现认知障碍。然而，神经认知功能障碍是否能预测首发精神分裂症（FES）患者使用抗精神病药物的疗效尚不清楚。在评估异质性数据时，机器学习方法提供了一种相对无偏的方法，尤其是在构建多因素预测模型时。本研究基于中国首发精神分裂症试验（CNFEST）进行了二次分析，该试验为期1年，前八周为随机对照试验，随后是48周的开放标签观察。本研究旨在基于基线临床和人口统计学特征建立一个为期八周的抗精神病药物反应预测模型。应用并比较了六种机器学习算法，包括随机森林、极端梯度提升（XGBoost）、逻辑回归、线性支持向量机（SVM）、径向基函数SVM和多项式SVM，以绘制预测模型。通过平衡准确率、敏感性和特异性评估预测效果。将预测因素与F分数进行比较。共有450名合格受试者参与了该模型。在所有算法中，通过XGBoost算法构建的预测模型具有最高的准确率（68.8%）和预后确定性（44.3%）。具有强预测意义的基线神经认知测试包括沟槽钉板测试、连线测验A、听觉连续加法测验、简易视觉空间学习测试、霍普金斯词语学习测试和彩色连线测试。本研究强调了精细运动技能、言语学习、视觉学习、工作记忆和注意力对未使用过药物的FES患者对抗精神病药物反应的重要性。XGBoost生成的模型显示出较好的准确率，为精神科医生预测FES患者的疗效提供了一种可能的方法。
