To construct a multi-class classification model of MCI outcomes (reversion, stabilization and progression) and to explore the possible associated factors of these outcomes.

Data were collected from the National Alzheimer's Disease Coordinating Center Unified Data Set, involving 397 patients who were initially diagnosed with MCI during 2005—2019 with at least two follow-ups and complete clinical and follow-up data. Patients' demographic information, physical examination, disease history, and measurement results by assessment scales in MCI were selected for analysis. Boruta was used for feature selection. Random forest was used for supporting the classification of MCI outcomes. A stepwise multinomial logistic regression was used to explore the associated factors of different MCI outcomes.

Of the 397 cases, 124 (31.23%) reverted to normal cognition, 77 (44.58%) were in stable condition, and 96 (24.18%) progressed to dementia. In the multi-classification task, the accuracy of direct three-class classification was 67.58%. In multiple binary classifications, the accuracies of reversion and stabilization, and reversion and progression were above 90%. Among the hierarchical three-class classification, the accuracy based on stabilization→ reversion → progression was 84.38%. Stepwise multinomial Logistic regression analysis showed that age ≥80 years 〔OR=0.260, 95%CI (0.117, 0.574) 〕, self-reported cognitive impairment 〔OR=0.295, 95%CI (0.139, 0.623) 〕, and decreased Clinical Dementia Rating (CDR) scores〔OR=0.220, 95%CI (0.092, 0.525) 〕 were associated with decreased risk of MCI progressing to dementia, while history of stroke 〔OR=2.896, 95%CI (1.370, 6.122) 〕, fecal incontinence 〔OR=6.556, 95%CI (1.787, 24.047) 〕, and higher Functional Activities Questionnaire score〔OR=1.048, 95%CI (1.003, 1.095) 〕were associated with increased risk of MCI progressing to dementia. Decreased probability of reversion from MCI to normal cognition was related to age ≥80 years 〔OR=0.289, 95%CI (0.091, 0.914) 〕, obesity 〔OR=0.236, 95%CI (0.075, 0.740) 〕, self-reported cognitive impairment 〔OR=0.289, 95%CI (0.111, 0.757) 〕, and higher CDR scores 〔OR=0.015, 95%CI (0.003, 0.089) 〕, while increased probability of reversion from MCI to normal cognition was related to higher MMSE score 〔OR=1.708, 95%CI (1.428, 2.043) 〕 and higher numbers of correct naming of animals 〔OR=1.139, 95%CI (1.046, 1.240) 〕.

The outcome of MCI patients is affected by multiple factors. Advanced age, obesity, history of stroke, fecal incontinence, self-reported cognitive impairment and reduced functional activity were important factors influencing MCI progression and reversion.

构建MCI患者不同转归的多分类模型，寻找MCI患者双向转归（逆转、稳定、进展）的影响因素。

本研究数据来源于美国国家阿尔茨海默病症协调中心统一数据集（NACC UDS），筛选2005—2019年初次诊断为MCI并至少随访2次且数据无缺失的397例患者，采集其基本信息、身体检查指标、既往史及相关评估量表得分。基于Boruta算法进行特征选择，采用随机森林进行MCI不同转归的多分类辅助诊断；采用多分类逐步Logistic回归分析探究MCI患者不同转归的影响因素。

纳入的397例MCI患者中，124例逆转为NC，逆转率为31.23%；177例稳定，稳定率为44.58%；96例进展为痴呆，进展率为24.18%。将397例MCI患者的不同转归进行分层三分类，总准确度（ACC）为67.58%。在多个二分类中，进展和逆转、逆转和稳定的分类性能较优，ACC均在90%以上；分层三分类中稳定→逆转→进展的总ACC为84.38%。多分类逐步Logistic回归分析结果显示，对于MCI患者进展为痴呆而言，年龄≥80岁〔OR=0.260，95%CI（0.117，0.574）〕、自我报告认知障碍〔OR=0.295，95%CI（0.139，0.623）〕、临床痴呆评定量表（CDR）得分降低〔OR=0.220，95%CI（0.092，0.525）〕为保护因素，既往有卒中〔OR=2.896，95%CI（1.370，6.122）〕、既往有大便失禁〔OR=6.556，95%CI（1.787，24.047）〕、功能活动问卷（FAQ）得分升高〔OR=1.048，95%CI（1.003，1.095）〕为危险因素。对于MCI患者逆转为正常而言，年龄≥80岁〔OR=0.289，95%CI（0.091，0.914）〕、肥胖〔OR=0.236，95%CI（0.075，0.740）〕、自我报告认知障碍〔OR=0.289，95%CI（0.111，0.757）〕、CDR得分升高〔OR=0.015，95%CI（0.003，0.089）〕为危险因素，简易精神状态量表（MMSE）得分升高〔OR=1.708，95%CI（1.428，2.043）〕和动物命名正确数升高〔OR=1.139，95%CI（1.046，1.240）〕为保护因素。

MCI患者的转归受多个因素影响，高龄、肥胖、既往发生过卒中、既往发生过大便失禁、自我报告认知障碍和日常功能障碍是MCI双向转归的重要影响因素。