Alayo Itxaso, Pujol Oriol, Alonso Jordi, Ferrer Montse, Amigo Franco, Portillo-Van Diest Ana, Aragonès Enric, Aragon Peña Andrés, Asúnsolo Del Barco Ángel, Campos Mireia, Espuga Meritxell, González-Pinto Ana, Haro Josep Maria, López-Fresneña Nieves, Martínez de Salázar Alma D, Molina Juan D, Ortí-Lucas Rafael M, Parellada Mara, Pelayo-Terán José Maria, Forjaz Maria João, Pérez-Zapata Aurora, Pijoan José Ignacio, Plana Nieves, Polentinos-Castro Elena, Puig Maria Teresa, Rius Cristina, Sanz Ferran, Serra Cònsol, Urreta-Barallobre Iratxe, Bruffaerts Ronny, Vieta Eduard, Pérez-Solá Víctor, Mortier Philippe, Vilagut Gemma
Hospital del Mar Research Institute, Barcelona, Spain.
Biosistemak Institute for Health Systems Research, Bilbao, Bizkaia, Spain.
Epidemiol Psychiatr Sci. 2025 May 8;34:e28. doi: 10.1017/S2045796025000198.
Studies conducted during the COVID-19 pandemic found high occurrence of suicidal thoughts and behaviours (STBs) among healthcare workers (HCWs). The current study aimed to (1) develop a machine learning-based prediction model for future STBs using data from a large prospective cohort of Spanish HCWs and (2) identify the most important variables in terms of contribution to the model's predictive accuracy.
This is a prospective, multicentre cohort study of Spanish HCWs active during the COVID-19 pandemic. A total of 8,996 HCWs participated in the web-based baseline survey (May-July 2020) and 4,809 in the 4-month follow-up survey. A total of 219 predictor variables were derived from the baseline survey. The outcome variable was any STB at the 4-month follow-up. Variable selection was done using an L1 regularized linear Support Vector Classifier (SVC). A random forest model with 5-fold cross-validation was developed, in which the Synthetic Minority Oversampling Technique (SMOTE) and undersampling of the majority class balancing techniques were tested. The model was evaluated by the area under the Receiver Operating Characteristic (AUROC) curve and the area under the precision-recall curve. Shapley's additive explanatory values (SHAP values) were used to evaluate the overall contribution of each variable to the prediction of future STBs. Results were obtained separately by gender.
The prevalence of STBs in HCWs at the 4-month follow-up was 7.9% (women = 7.8%, men = 8.2%). Thirty-four variables were selected by the L1 regularized linear SVC. The best results were obtained without data balancing techniques: AUROC = 0.87 (0.86 for women and 0.87 for men) and area under the precision-recall curve = 0.50 (0.55 for women and 0.45 for men). Based on SHAP values, the most important baseline predictors for any STB at the 4-month follow-up were the presence of passive suicidal ideation, the number of days in the past 30 days with passive or active suicidal ideation, the number of days in the past 30 days with binge eating episodes, the number of panic attacks (women only) and the frequency of intrusive thoughts (men only).
Machine learning-based prediction models for STBs in HCWs during the COVID-19 pandemic trained on web-based survey data present high discrimination and classification capacity. Future clinical implementations of this model could enable the early detection of HCWs at the highest risk for developing adverse mental health outcomes.
NCT04556565.
在新冠疫情期间进行的研究发现，医护人员中自杀念头和行为（STB）的发生率很高。本研究旨在：（1）利用来自西班牙医护人员大型前瞻性队列的数据，开发一种基于机器学习的未来STB预测模型；（2）确定对模型预测准确性贡献最大的最重要变量。
这是一项对在新冠疫情期间工作的西班牙医护人员进行的前瞻性多中心队列研究。共有8996名医护人员参与了基于网络的基线调查（2020年5月至7月），4809人参与了4个月的随访调查。从基线调查中得出了总共219个预测变量。结果变量是4个月随访时的任何STB。使用L1正则化线性支持向量分类器（SVC）进行变量选择。开发了一个具有5折交叉验证的随机森林模型，并测试了合成少数过采样技术（SMOTE）和多数类欠采样平衡技术。通过接受者操作特征（AUROC）曲线下面积和精确召回率曲线下面积对模型进行评估。使用夏普利加性解释值（SHAP值）来评估每个变量对未来STB预测的总体贡献。结果按性别分别得出。
4个月随访时医护人员中STB的患病率为7.9%（女性=7.8%，男性=8.2%）。L1正则化线性SVC选择了34个变量。在不使用数据平衡技术的情况下获得了最佳结果：AUROC=0.87（女性为0.86，男性为0.87），精确召回率曲线下面积=0.50（女性为0.55，男性为0.45）。基于SHAP值，4个月随访时任何STB的最重要基线预测因素是存在被动自杀意念、过去30天内有被动或主动自杀意念的天数、过去30天内有暴饮暴食发作的天数、惊恐发作的次数（仅女性）和侵入性思维的频率（仅男性）。
基于机器学习的新冠疫情期间医护人员STB预测模型，在基于网络的调查数据上进行训练，具有很高的区分度和分类能力。该模型未来的临床应用可以早期发现发生不良心理健康结果风险最高的医护人员。
NCT04556565。
