Influencing factors analysis and prediction model construction of medical students′ suicidal ideation based on machine learning algorithm
-
摘要:
目的 了解医学生自杀意念影响因素，并探索机器学习算法对医学生自杀意念的预测效果。 方法 于2021年11月―2022年3月对山东省医学生进行随机分层整群抽样，进行问卷调查。使用χ2检验、Fisher确切概率法、logistic回归分析模型探讨医学生自杀意念影响因素。训练集使用机器学习算法构建预测模型，测试集使用准确度、灵敏度等指标评估模型预测能力。 结果 医学生自杀意念检出率为12.80%。影响因素分析结果显示：住在农村(OR=1.523，95% CI：1.023~2.271，P=0.039)、存在抑郁症状(OR=3.874，95% CI：2.676~5.621，P＜0.001)等是自杀意念危险因素，近2年没有谈过恋爱(OR=0.601，95% CI：0.427~0.841，P=0.003)等是自杀意念的保护因素。4种模型预测准确率、灵敏度均＞0.90，Kappa值均＞0.80，阳性、阴性预测值均＞0.90。 结论 4种预测模型表现均良好，基于支持向量机(support vector machine，SVM)的自杀意念预测模型更具优势，可有效预测自杀意念风险，更有助于高危人群的早期识别与干预。 Abstract:Objective The purpose of this study is to understand the factors influencing suicidal ideation among medical students and to explore the predictive effect of machine learning (ML) algorithms. Methods A random stratified whole-group sampling of medical students in Shandong Province was conducted from November 2021 to March 2022 to conduct the questionnaire survey. The Chi-square test, Fisher′s exact probability method, and Logistic regression were used to explore the factors influencing suicidal ideation among medical students. The prediction models were constructed based on a machine learning algorithm in the training set. The predictive ability of the model was evaluated in the test set based on accuracy, sensitivity, and so on. Results The suicidal ideation rate in medical students was 12.80%. The results showed that living in rural areas (OR=1.523, 95% CI: 1.023-2.271, P=0.039), and depression symptoms (OR=3.874, 95% CI: 2.676-5.621, P < 0.001) were risk factors for suicidal ideation. Having not been in love in the past two years (OR=0.601, 95% CI: 0.427-0.841, P=0.003) and so on were protective factors for suicidal ideation. The prediction accuracy and sensitivity of the four models were higher than 0.90, the Kappa values were higher than 0.80, and the positive and negative predictive values were higher than 0.90. Conclusions All four prediction models perform well, among which the suicide ideation prediction model based on support vector machine (SVM) has more advantages. The SVM model can effectively predict the risk of suicidal ideation, which is more helpful for early identification and intervention of high-risk groups. -
Key words:
- Medical students /
- Machine learning /
- Suicidal ideation /
- Prediction model
-
表 1 一般人口学特征及自杀意念单因素分析结果
Table 1. Results of univariate analysis of general demographic characteristics and suicidal ideation
研究因素Study factor 无自杀意念
No suicidal ideation(n=1 598)有自杀意念
Suicidal ideation(n=235)t/t′/χ2
值valueP值
value年龄/岁Age/years 19.39±1.25 19.41±1.31 -0.303 0.762 受教育年限Education years 13.48±1.42 13.44±1.52 0.336 0.737 抑郁得分Depression score 12.19±8.25 22.43±10.83 -16.988 ＜0.001 抑郁情绪因子Depressive mood factor 2.84±2.71 12.07±3.89 -45.777 ＜0.001 积极情绪因子Positive emotion factor 4.75±3.34 7.50±2.45 -12.165 ＜0.001 躯体症状与活动迟滞因子Somatic symptoms and activity retardation factors 2.82±2.35 8.69±2.78 -34.810 ＜0.001 人际关系因子Interpersonal relationship factor 0.52±0.11 2.74±1.68 30.477 ＜0.001 性别Gender 4.001 0.045 男Male 625 108 女Female 973 127 民族Nation ＜0.001 1.000 汉族Han nationality 1 564 230 少数民族National minority 34 5 身体健康状况Physical condition ― ＜0.001 非常健康Very healthy 508 49 比较健康Relatively healthy 869 120 一般General 195 54 差Poor 26 12 慢性疾病Chronic disease 26.201 ＜0.001 有Yes 252 69 无No 1 346 166 性格Personality ― ＜0.001 偏内向Partial introvert 424 85 偏外向Partial extrovertive 304 31 内外兼有Both internal and external 849 109 不清楚Unclear 21 10 心理健康状况Mental health status 97.345 ＜0.001 非常健康Very healthy 536 40 比较健康Relatively healthy 813 99 一般Normal 220 77 差Relatively poor 29 19 抑郁症状Epressive symptoms 208.953 ＜0.001 无No 1 310 92 有Yes 288 143 户口类别Account type 3.315 0.069 农村户口Rural household registration 1 025 165 城市户口Urban household registration 573 70 常住地Permanent residence 7.590 0.022 城市City 712 88 乡镇Villages and towns 332 44 农村Country 554 103 父亲受教育程度Father′s level of education 3.629 0.304 大专Junior college 209 31 高中/中专High school/Technical secondary school 446 52 初中及以下Junior high school and below 730 118 母亲受教育程度Mather′s level of education 6.410 0.093 本科及以上Bachelor′s degree or above 131 28 大专Junior college 179 20 高中/中专High school/Technical secondary school 363 44 初中及以下Junior high school and below 925 143 家庭类型Homestyle 8.827 0.032 核心家庭Nuclear family 1 088 159 大家庭Big family 400 51 单亲家庭Single parent family 73 12 重组家庭Reconstituted family 37 13 独生子女Only child 1.308 0.253 是Yes 591 96 否No 1 007 139 留守子女Guarded children 11.640 0.001 是Yes 53 16 否No 1 545 219 家庭地位Family status 11.955 0.018 很高Very high 275 39 较高Higher 680 85 中等Intermediate 593 94 较低Lower 26 9 最低Minimum 24 8 家庭经济状况Family economic status 27.675 ＜0.001 很好Great 47 18 比较好Relatively good 228 25 中等Intermediate 1 037 128 差Bad 286 64 家庭教育方式Family education mode 38.383 ＜0.001 民主型Democratic type 1 040 107 权威型Authoritative 166 47 放任型Lax style 291 55 其他other 101 26 家庭和睦程度Degree of family harmony 53.489 ＜0.001 很好Great 862 77 比较好Relatively good 525 91 一般Normal 184 55 差Bad 27 12 与父母关系Relationship with parents ― ＜0.001 很好Great 861 81 比较好Relatively good 551 88 一般Normal 167 51 差Bad 19 15 家庭精神疾病史Family history of mental illness 4.583 0.032 有Have 108 25 无Not have 1 490 210 年级Grade ― 0.020 大一Freshman 467 55 大二Sophomore year 834 119 大三Junior 210 44 大四Senior 20 8 大五Senior year 38 5 硕博研究生Postgraduate 29 4 住校Live on campus or not 0.264 0.608 是Yes 1 547 226 否No 51 9 班委Class committee 1.062 0.303 是Yes 509 67 否No 1 089 168 学习成绩Academic record 29.264 ＜0.001 差Poor 128 44 中偏下Middle lower 312 48 中等Intermediate 745 95 中偏上Above average 313 36 优Optimal 100 12 与师长关系Relationship with teachers ― ＜0.001 很好Great 455 45 比较好Relatively good 543 77 一般Normal 585 102 差Bad 15 11 与同学关系Relationship with classmates ― ＜0.001 很好Great 546 43 比较好Relatively good 778 108 一般Normal 267 78 差Bad 7 6 社会关系Social relations 47.631 ＜0.001 很好Great 101 21 比较好Relatively good 514 47 中等Medium 915 134 差Bad 68 33 得到帮助Can get help or not 96.698 ＜0.001 能Yes 1 499 175 否No 99 60 知心朋友Intimate friend 48.136 ＜0.001 没有None 94 34 有1个Have 1 236 55 有2个There are two 433 70 有3个及以上There are 3 or more 835 76 是否恋爱Have a love affair or not 4.391 0.036 是Yes 740 126 否No 858 109 现实与理想差距The gap between reality and ideal 52.260 ＜0.001 没有差距No gap 97 13 有些差距Some gap 1 156 121 差距很大Wide gap 278 78 差距非常大Very wide gap 67 23 心理健康服务Mental health services 6.065 0.014 是Yes 205 44 否No 1 393 191 周围自杀行为Peripheral suicidal behavior 44.944 ＜0.001 有Have 138 54 无Not have 1 460 181 注：“―”使用Fisher确切概率法。
Note: "―" Fisher′s exact test.
下载: 导出CSV
表 2 医学生自杀意念的多因素logistic回归分析模型分析
Table 2. Multivariate logistic regression analysis of suicidal ideation among medical students
变量Variable β sx Z值value OR值value (95% CI) P值value 常住地Permanent residence 城市City 1.000 乡镇Villages and towns 0.400 0.221 1.811 1.492(0.964~2.295) 0.070 农村Country 0.420 0.203 2.067 1.523(1.023~2.271) 0.039 学习成绩Academic record 差Poor 1.000 中偏下Middle-lower -0.612 0.296 -2.072 0.542(0.304~0.969) 0.038 中等Intermediate -0.353 0.275 -1.284 0.703(0.413~1.213) 0.199 中偏上Center above -0.600 0.325 -1.844 0.549(0.289~1.039) 0.065 优Optimal -0.334 0.419 -0.796 0.716(0.306~1.599) 0.426 与父母关系Relationship with parents 很好Great 1.000 比较好Relatively good -1.228 0.680 3.255 0.293(0.077~1.112) 0.071 一般Normal -1.040 0.684 2.310 0.354(0.093~1.351) 0.129 差Bad -1.206 0.725 2.770 0.299(0.072~1.239) 0.096 得到帮助Can get help or not 是Yes 1.000 否No 0.514 0.241 2.326 1.753(1.088~2.806) 0.020 恋爱Have a love affair or not 是Yes 1.000 否No -0.115 0.172 -2.952 0.601(0.427~0.841) 0.003 慢性疾病Chronic disease 有Have 1.000 无Not have -0.124 0.197 -2.948 0.559(0.381~0.826) 0.003 心理健康状况Mental health status 非常健康Very healthy 1.000 比较健康Relatively healthy -0.710 0.633 1.255 0.492(0.142~1.702) 0.263 一般Normal -0.496 0.619 0.642 0.609(0.181~2.049) 0.423 差Poor -0.090 0.621 0.021 1.094(0.324~3.696) 0.885 周围自杀行为Peripheral suicidal behavior 有Have 1.000 无Not have -0.665 0.228 -1.031 0.514(0.326~0.925) 0.005 抑郁症状Depress symptom 无Not have 1.000 有Have 1.354 0.189 7.159 3.874(2.676~5.621) ＜0.001
下载: 导出CSV
表 3 4种预测模型混淆矩阵
Table 3. Confusion matrix of four prediction models
模型Model 实际值Actual value 预测值Predicted value 有自杀意念=1
Have suicidal ideation=1无自杀意念=0
No suicidal ideation=0Logistic 有自杀意念组Suicidal ideation group 82 1 无自杀意念组No suicidal ideation group 2 164 RF 有自杀意念组Suicidal ideation group 82 1 无自杀意念组No suicidal ideation group 23 143 NBM 有自杀意念组Suicidal ideation group 78 5 无自杀意念组No suicidal ideation group 5 161 SVM 有自杀意念组Suicidal ideation group 83 0 无自杀意念组No suicidal ideation group 0 166 注：RF, 随机森林; NBM, 朴素贝叶斯模型; SVM, 支持向量机。
Note: RF, fandom forest; NBM, naive Bayesian model; SVM, Support vector machine.
下载: 导出CSV
表 4 4种预测模型间AUC两两比较
Table 4. A pairwise comparison of the AUC of four prediction models
模型Model Logistic RF NBM SVM Logistic RF Z=-10.079, P＜0.001 NBM Z=1.730, P=0.083 Z=-11.993, P＜0.001 SVM Z=0.155, P=0.607 Z=-9.396, P＜0.001 Z=122.340, P＜0.001 注：RF, 随机森林; NBM, 朴素贝叶斯模型; SVM, 支持向量机; AUC，曲线下面积。
Note: RF, fandom forest; NBM, naive Bayesian model; SVM, Support vector machine; AUC，area under the curve.
下载: 导出CSV
表 5 4种模型预测效能
Table 5. Four model prediction efficiency
模型
ModelAUC(95% CI) 准确率
Precision
rate灵敏度
Sensitivity特异度
SpecificityKappa值
value阳性预测值
Positive predictive value阴性预测值
Negative predictive valueLogistic 0.988(0.923~0.999) 0.987 0.991 0.988 0.971 0.976 0.988 RF 0.798(0.768~0.828) 0.904 0.994 0.861 0.797 0.781 0.993 NBM 0.989(0.956~0.999) ① 0.960 0.939 0.969 0.910 0.940 0.970 SVM 0.971(0.921~0.986) 0.999 ① 0.996 ① 0.998 ① 0.998 ① 0.999 ① 0.998 ① 注：注：RF, 随机森林; NBM, 朴素贝叶斯模型; SVM, 支持向量机; AUC，曲线下面积。
①代表每项最大值。
Note: RF, fandom forest; NBM, naive Bayesian model; SVM, Support vector machine; AUC，area under the curve.
① Means the maximum value of each item.
下载: 导出CSV
-
[1] WHO. Suicide[EB/OL]. (2019-09-02)[2020-03-05]. http://www.who.int/zh/news-room/fact-sheets/detail/suicide .[2] Meyer IH, Luo F, Wilson BDM, et al. Sexual orientation enumeration in state antibullying statutes in the United States: associations with bullying, suicidal ideation, and suicide attempts among youth[J]. LGBT Health, 2019, 6(1): 9-14. DOI: 10.1089/lgbt.2018.0194. [3] Ruderfer DM, Walsh CG, Aguirre MW, et al. Significant shared heritability underlies suicide attempt and clinically predicted probability of attempting suicide[J]. Mol Psychiatry, 2020, 25(10): 2422-2430. DOI: 10.1038/s41380-018-0326-8. [4] Sun L, Zhou C, Xu L, et al. Suicidal ideation, plans and attempts among medical college students in China: The effect of their parental characteristics[J]. Psychiatry Res, 2017, 247: 139-143. DOI: 10.1016/j.psychres.2016.11.024. [5] 谭琪钰, 栾烨, 徐超, 等. 中国大学生自杀意念检出率的Meta分析[J]. 现代预防医学, 2022, 49(7): 1269-1274. DOI: 1003-8507(2022)07-1269-06.Tan QY, Luan Y, Xu C, et al. Meta-analysis of the detection rate of suicidal ideation among Chinese college students[J]. Modern Prevent Med, 2022, 49(7): 1269-1274. DOI: 1003-8507(2022)07-1269-06. [6] 关素珍, 刘向阳, 郑平, 等. 医学生自杀意念与家庭因素关系的研究[J]. 中国预防医学杂志, 2015, 16(1): 36-38. DOI: 10.16506/j.1009-6639.2015.01.012.Guan SZ, Liu XY, Zheng P, et al. A study on the relationship between suicidal ideation and family factors among medical students[J]. Chin J Prev Med, 2015, 16(1): 36-38. DOI: 10.16506/j.1009-6639.2015.01.012. [7] 刘裕兴, 聂光辉, 梁好. 医学生儿童期创伤与自杀意念: 述情障碍和反刍思维的链式中介作用[J]. 中国临床心理学杂志, 2022, 30(3): 683-687. DOI: 10.16128/j.cnki.1005-3611.2022.03.037.Liu YX, Nie GH, Liang H. Childhood trauma and suicidal ideation in medical students: the chain mediating role of narrative impairment and ruminative thinking[J]. Chin J Clin Psychol, 2022, 30(3): 683-687. DOI: 10.16128/j.cnki.1005-3611.2022.03.037. [8] Franklin JC, Ribeiro JD, Fox KR, et al. Risk factors for suicidal thoughts and behaviors: a meta-analysis of 50 years of research[J]. Psychol Bull, 2017, 143(2): 187-232. DOI: 10.1037/bul0000084. [9] Chan MK, Bhatti H, Meader N, et al. Predicting suicide following self-harm: systematic review of risk factors and risk scales[J]. Br J Psychiatry, 2016, 209(4): 277-283. DOI: 10.1192/bjp.bp.115.170050. [10] Liu X, Liu X, Sun J, et al. Proactive suicide prevention online (PSPO): machine identification and crisis management for Chinese social media users with suicidal thoughts and behaviors[J]. J Med Internet Res, 2019, 21(5): e11705. DOI: 10.2196/11705. [11] 马鸣, 刘欢, 刘润香. 机器学习在大学生自杀意念预测中的应用[J]. 中国学校卫生, 2022, 43(5): 763-767. DOI: 10.16835/j.cnki.1000-9817.2022.05.029.Ma M, Liu H, Liu RX. Application of machine learning in the prediction of suicidal ideation among college students[J]. Chin J Sch Health, 2022, 43(5): 763-767. DOI: 10.16835/j.cnki.1000-9817.2022.05.029. [12] 何作力, 张红亚, 李瑞, 等. 2019年大连市中学生抑郁症状现况及影响因素[J]. 预防医学论坛, 2021, 27(2): 88-91. DOI: 10.16406/j.pmt.issn.1672-9153.2021.02.005.He ZL, Zhang HY, Li R, et al. Current status of depressive symptoms and influencing factors among secondary school students in Dalian in 2019[J]. Prev Med Forum, 2021, 27(2): 88-91. DOI: 10.16406/j.pmt.issn.1672-9153.2021.02.005. [13] 李献云, 费立鹏, 童永胜, 等. Beck自杀意念量表中文版在社区成年人群中应用的信效度[J]. 中国心理卫生杂志, 2010, 24(4): 250-255. DOI: 10.3969/j.issn.1000-6729.2010.04.003.Li XY, Fei LP, Tong YS, et al. Reliability and validity of the Chinese version of Beck Suicidal Ideation Scale in a community-based adult population[J]. Chin J Mental Health, 2010, 24(4): 250-255. DOI: 10.3969/j.issn.1000-6729.2010.04.003. [14] 阿拉依·阿汗, 田翔华, 肖齐, 等. 关联规则与Logistic回归在维吾尔族健康体检人群代谢综合征数据挖掘中的应用[J]. 现代预防医学, 2018, 45(7): 1161-1165. DOI: 1003-8507(2018)07-1161-05.Alai A, Tian XH, Xiao Q, et al. Application of association rules and logistic regression in data mining of metabolic syndrome in Uyghur health checkup population[J]. Mode Prev Med, 2018, 45(7): 1161-1165. DOI: 1003-8507(2018)07-1161-05. [15] 董红瑶, 王弈丹, 李丽红. 随机森林优化算法综述[J]. 信息与电脑, 2021, 33(17): 34-37. DOI: 1003-9767(2021)17-034-04.Dong HY, Wang YD, Li LH. A review of random forest optimization algorithms[J]. China Computer & Communication, 2021, 33(17): 34-37. DOI: 1003-9767(2021)17-034-04. [16] 张航. 基于朴素贝叶斯的中文文本分类及Python实现[D]. 济南: 山东师范大学, 2018.Zhang H. Chinese text classification based on plain Bayes and Python implementation[D]. Jinan: Shandong Normal University, 2018. [17] 韩力群. 人工神经网络教程[M]. 北京: 北京邮电大学出版社. 2006, 125-127.Han LQ. Tutorial on Artificial Neural Networks[M]. Beijing: Beijing University of Posts and Telecommunications Press. 2006, 125-127. [18] 赵晓艳. 基于交叉验证的AUC度量的置信区间的研究[D]. 太原: 山西大学, 2021.Zhao XY. Research on confidence interval of AUC metric based on cross-validation[D]. Taiyuan: Shanxi University, 2021. [19] Wege N, Muth T, Li J, et al. Mental health among currently enrolled medical students in Germany[J]. Public Health, 2016, 132: 92-100. DOI: 10.1016/j.puhe.2015.12.014. [20] 王一帆, 孙龙, 魏真, 等. 医学院校学生自杀意念与父母拒绝相关性研究[J]. 伤害医学(电子版), 2021, 10(4): 17-24. DOI: 10.3868/j.issn.2095-1566.2021.04.004.Wang YF, Sun L, Wei Z, et al. A study on the correlation between suicidal ideation and parental rejection among medical school students[J]. Injury Medicine(Electronic Edition), 2021, 10(4): 17-24. DOI: 10.3868/j.issn.2095-1566.2021.04.004. [21] 沈可, 黄笛, 杨雨飞, 等. 2007-2020年中国大陆医学生自杀意念发生率的Meta分析[J]. 医学新知, 2020, 30(6): 431-441. DOI: 10.12173/j.issn.1004-5511.2020.06.03.Shen K, Huang D, Yang YF, et al. Meta-analysis of the incidence of suicidal ideation among medical students in mainland China from 2007 to 2020[J]. Medicine New, 2020, 30(6): 431-441. DOI: 10.12173/j.issn.1004-5511.2020.06.03. [22] 李海燕, 任梦飞, 鞠磊, 等. 潍坊市中小学生自杀意念与心理健康状况的相关性[J]. 中国健康心理学杂志, 2019, 27(5): 651-654. DOI: 10.13342/j.cnki.cjhp.2019.05.013.Li HY, Ren MF, Ju L, et al. Correlation between suicidal ideation and mental health status of primary and secondary school students in Weifang City[J]. China Journal of Health Psychology, 2019, 27(5): 651-654. DOI: 10.13342/j.cnki.cjhp.2019.05.013. [23] 严红虹. 大学生自杀意念影响因素的定量研究与定性研究[D]. 广州: 暨南大学, 2009.Yan HH. A quantitative and qualitative study of the factors influencing suicidal ideation among college students[D]. Guangzhou: Jinan University, 2009. [24] 吴茵琪. 粤港两地大学生心理健康的状况、影响因素及其比较研究[D]. 广州: 广州中医药大学, 2020.Wu YQ. A comparative study on the status, influencing factors and mental health of college students in Guangdong and Hong Kong[D]. Guangzhou: Guangzhou University of Chinese Medicine, 2020. [25] 况利, 徐小明, 曾琪. 机器学习用于自杀研究的综述[J]. 山东大学学报(医学版), 2022, 60(4): 10-16. DOI: 10.6040/j.issn.1671-7554.0.2021.0175.Kuang L, Xu XM, Zeng Q. A review of machine learning for suicide research[J]. Journal of Shandong University (Health Science), 2022, 60(4): 10-16. DOI: 10.6040/j.issn.1671-7554.0.2021.0175. [26] Just MA, Pan L, Cherkassky VL, et al. Machine learning of neural representations of suicide and emotion concepts identifies suicidal youth[J]. Nat Hum Behav, 2017, 1: 911-919. DOI: 10.1038/s41562-017-0234-y. [27] Walsh CG, Ribeiro JD, Franklin JC. Predicting suicide attempts in adolescents with longitudinal clinical data and machine learning[J]. J Child Psychol Psychiatry, 2018, 59(12): 1261-1270. DOI: 10.1111/jcpp.12916. [28] Metzger MH, Tvardik N, Gicquel Q, et al. Use of emergency department electronic medical records for automated epidemiological surveillance of suicide attempts: a French pilot study[J]. Int J Methods Psychiatr Res, 2017, 26(2): e1522. DOI: 10.1002/mpr.1522. -
