| 肖轩,戴西件,李艺晖,杨培,龚良庚.基于机器学习模型预测抑郁症患者痴呆风险[J].中国医学影像技术,2024,40(9):1309~1313 |
| 基于机器学习模型预测抑郁症患者痴呆风险 |
| Machine learning model for predicting the risk of dementia in patients with depression |
| 投稿时间：2024-02-28 修订日期：2024-03-25 |
| DOI：10.13929/j.issn.1003-3289.2024.09.007 |
| 中文关键词: 痴呆 抑郁症 机器学习 |
| 英文关键词:dementia depressive disorder machine learning |
| 基金项目:江西省自然科学基金(20224BAB216077)。 |
| 作者 | 单位 | E-mail | | 肖轩 | 南昌大学第二附属医院医学影像中心, 江西 南昌 330006
智能医学影像江西省重点实验室, 江西 南昌 330006 | | | 戴西件 | 南昌大学第二附属医院医学影像中心, 江西 南昌 330006
智能医学影像江西省重点实验室, 江西 南昌 330006 | | | 李艺晖 | 南昌大学第二附属医院医学影像中心, 江西 南昌 330006
智能医学影像江西省重点实验室, 江西 南昌 330006 | | | 杨培 | 南昌大学第二附属医院医学影像中心, 江西 南昌 330006
智能医学影像江西省重点实验室, 江西 南昌 330006 | | | 龚良庚 | 南昌大学第二附属医院医学影像中心, 江西 南昌 330006
智能医学影像江西省重点实验室, 江西 南昌 330006 | gong111999@126.com |
|
| 摘要点击次数: 719 |
| 全文下载次数: 358 |
| 中文摘要: |
| 目的 观察机器学习模型预测抑郁症患者痴呆风险的价值。方法 回顾性收集UK Biobank数据库中31 587例抑郁症患者,根据随访显示是否罹患痴呆分为痴呆组(n=896)与非痴呆组(n=30 691),并按6 ∶ 4比例分为训练集(n=18952)及测试集(n=12635);根据访谈和问卷调查共纳入包括人口特征、生活方式、健康状况、身体指标及影像学表现等共190个因素并进行筛选;基于筛选所获因素分别以轻量级梯度提升机(LightGBM)、岭回归(Ridge)和自适应提升(AdaBoost)建立模型,观察其预测抑郁症患者痴呆风险的价值。结果 最终纳入10个因素,包括年龄、腰围、就业状态、日间小憩、日间瞌睡或嗜睡、使用手机时长、家庭成员数量、患抑郁症时长、愧疚感及因心理问题就诊;由此建立的LightGBM、Ridge和AdaBoost模型预测训练集抑郁症患者痴呆风险的曲线下面积(AUC)分别为0.914、0.832和0.889,两两比较差异均有统计学意义(P均＜0.05),而在测试集AUC分别为0.866、0.842和0.859,除LightGBM与AdaBoost外,其余两两比较差异均有统计学意义(P均＜0.05)。校准曲线显示LightGBM模型拟合度最佳。结论 LightGBM机器学习模型有助于预测抑郁症患者痴呆风险。 |
| 英文摘要: |
| Objective To observe the value of machine learning model for predicting the risk of dementia in patients with depression. Methods Totally 31 587 depression patients from UK Biobank database were retrospectively enrolled and divided into dementia group (n=896) or non-dementia group (n=30 691) based on follow-up data showed developed dementia or not, also divided into training set (n=18 952) or test set (n=12 635) at the ratio of 6 ∶ 4. Based on interviews and questionnaire surveys, a total of 190 factors including demographic characteristics, lifestyle, health status, physical indicators and imaging data were included and screened to establish models with light gradient boosting machine (LightGBM), ridge regression (Ridge) and adaptive boosting (AdaBoost), and the value for predicting the risk of dementia in patients with depression was observed. Results A total of 10 factors were ultimately enrolled, including age, waist circumference, employment status, daytime rest, daytime doze or drowsiness, duration of mobile phone use, number of family members, duration of depression, guilt and seeking medical attention due to psychological issues. Based on the above factors, the models were established. In training set, the area under the curve (AUC) of LightGBM, Ridge and AdaBoost model for predicting dementia risk in patients with depression was 0.914, 0.832 and 0.889, respectively, and the differences between each 2 models were significant (all P＜0.05); while in test set, the AUC was 0.866, 0.842 and 0.859, respectively, except for LightGBM and AdaBoost, the other with significant differences between each two (both P＜0.05). The calibration curve showed that the LightGBM model had the best fit. Conclusion LightGBM machine learning model was helpful for predicting the risk of dementia in patients with depression. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
