基于XGBoost模型的珠海市严重精神障碍患者服药依从性影响因素研究

Study on medication adherence factors among patients with severe mental disorders in Zhuhai city based on XGBoost model

中文关键词:  严重精神障碍  服药依从性  影响因素  XGBoost模型

英文关键词:Severe mental disorders  Medication adherence  Influencing factors  XGBoost model.

中文摘要:

      【摘要】背景 严重精神障碍患者服药依从性低会增加患者家庭和社会的疾病负担，且服药依从性存在较多影响因素，如Logistic回归等传统方法难以量化影响因素重要性，引入极限梯度提升（XGBoost）结合沙普利可加性解释（SHAP）可有效突破这一局限，量化各因素的相对贡献权重，为识别核心影响因素提供支撑。目的 探讨影响珠海市严重精神障碍患者服药依从性的影响因素，为优化患者管理策略提供参考。方法 提取2023年1月1日—2025年3月31日在珠海市精神卫生系统平台登记在册的严重精神障碍患者数据，最终对9 329例患者的数据进行分析。通过单因素分析和多因素Logistic回归分析筛选影响因素，构建XGBoost模型并结合SHAP算法量化各影响因素的重要性。结果 在9 329例患者中，服药依从的患者8 446例，服药依从率为90.53%。未婚（OR=1.237，95% CI：1.019~1.502）或离异（OR=1.389，95% CI：1.038~1.832）、诊断为精神发育迟滞伴发精神障碍（OR=3.025，95% CI：2.402~3.796）或偏执性精神病（OR=5.117，95% CI：3.086~8.299）、病程2~4年（OR=1.355，95% CI：1.085~1.696）、病程4~6年（OR=2.143，95% CI：1.671~2.747）、病程>6年（OR=1.681，95% CI：1.365~2.079）、未办理监护人补助（OR=1.412，95% CI：1.099~1.801）、未领取残疾人证（OR=1.900，95% CI：1.588~2.282）、非关爱帮扶对象（OR=1.384，95% CI：1.183~1.617）、非社区服务对象（OR=1.313，95% CI：1.042~1.645）及未与监护人同住（OR=1.257，95% CI：1.048~1.501），均为服药依从性的危险因素。已办理门特（OR=0.716，95% CI：0.609~0.842）以及有精神疾病家族史（OR=0.713，95% CI：0.503~0.982）是服药依从性的保护因素。XGBoost模型预测性能良好，灵敏度为0.433，特异度为0.944，准确度为0.891，AUC为0.837，F1值为0.449。影响因素重要性排序显示，重要性排名前三的影响因素分别为：病程、诊断以及残疾人证领取情况。结论 政策性保障（领取残疾人证、办理门特）与临床疾病特征（病程、诊断类型）是影响珠海市严重精神障碍患者服药依从性的关键因素。

英文摘要:

      Study on medication adherence factors among patients with severe mental disorders in Zhuhai city based on XGBoost model Ye Zhongshu1,2, Teng Yongyong1,2, Quan Jingju1,2, Sun Yajun1,2, Huang Jiaju1,2, Wu Yixuan1,2, Han Changlin1,2, Zhang Guangchuan1,2* (1.The Third People′s Hospital of Zhuhai, Zhuhai 519000, China; 2.Zhuhai?Engineering?Technology?Research?Center?of?Occupational and Mental Health, Zhuhai 519000, China *Corresponding author: Zhang Guangchuan, E-mail: [e-mail]) 【Abstract】 Background Low medication adherence in patients with severe mental disorders (SMDs) increases the disease burden on patients' families and society. Medication adherence is influenced by numerous factors, and traditional methods (such as Logistic regression) struggle to quantify the importance of these factors. Extreme gradient boosting (XGBoost) combined with Shapley's additivity explanation (SHAP) is less commonly used in this field.. Objective To explore the factors influencing medication adherence among SMDs patients in Zhuhai and provide a reference for optimizing patient management strategies. Methods A cross-sectional study was conducted using data from SMD patients managed by the Zhuhai Mental Health System Platform between January 1, 2023, and March 31, 2025. A total of 9,329 patients diagnosed with six types of SMDs (including schizophrenia and bipolar disorder) were included. Influencing factors were screened using univariate analysis and multivariate logistic regression analysis,, and an XGBoost model combined with the SHAP algorithm was constructed to quantify the importance of each influencing factor. Results Among 9,329 patients, 8,446 demonstrated medication adherence, yielding an adherence rate of 90.53%. Multivariable analysis identified several risk factors significantly associated with medication non-adherence: being unmarried (OR=1.237, 95% CI: 1.019~1.502) or divorced (OR=1.389, 95% CI: 1.038~1.832); a diagnosis of mental retardation with psychiatric disorders (OR=3.025, 95% CI: 2.402~3.796) or paranoid psychosis (OR=5.117, 95% CI: 3.086~8.299); a disease duration of 2~4 years (OR=1.355, 95% CI: 1.085~1.696), 4~6 years (OR=2.143, 95% CI: 1.671~2.747), or >6 years (OR=1.681, 95% CI: 1.365~2.079); lack of guardian subsidies (OR=1.412, 95% CI: 1.099~1.801); absence of a disability certificate (OR=1.900, 95% CI: 1.588~2.282); not being enrolled in care and support groups (OR=1.384, 95% CI: 1.183~1.617) or community services (OR=1.313, 95% CI: 1.042~1.645); and not cohabiting with a guardian (OR=1.257, 95% CI: 1.048~1.501). Conversely, the enrollment in special outpatient disease programs (OR=0.716, 95% CI: 0.609~0.842) and a family history of mental illness (OR=0.713, 95% CI: 0.503~0.982) were identified as protective factors.The XGBoost model exhibited robust predictive performance, with a sensitivity of 0.433, specificity of 0.944, accuracy of 0.891, an Area Under the Curve (AUC) of 0.837, and an F1 score of 0.449. Feature importance ranking indicated that the three most influential predictors were disease duration, clinical diagnosis, and the acquisition of disability certificates. Conclusion Policy-based support (acquisition of disability certificates, special outpatient disease enrollment) and clinical disease characteristics (disease duration, diagnosis type) are key factors affecting medication adherence among patients with severe mental disorders in Zhuhai City. [Funded by Zhuhai Medical Research Project (number, 2220009000281)]