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

To investigate the occurrence of traumatic stress symptoms (TSS) among healthcare workers active during the COVID-19 pandemic and to obtain insight as to which pandemic-related stressful experiences are associated with onset and persistence of traumatic stress.

This is a multicenter prospective cohort study. Spanish healthcare workers (N = 4,809) participated at an initial assessment (i.e., just after the first wave of the Spain COVID-19 pandemic) and at a 4-month follow-up assessment using web-based surveys. Logistic regression investigated associations of 19 pandemic-related stressful experiences across four domains (infection-related, work-related, health-related and financial) with TSS prevalence, incidence and persistence, including simulations of population attributable risk proportions (PARP).

Thirty-day TSS prevalence at T1 was 22.1%. Four-month incidence and persistence were 11.6% and 54.2%, respectively. Auxiliary nurses had highest rates of TSS prevalence (35.1%) and incidence (16.1%). All 19 pandemic-related stressful experiences under study were associated with TSS prevalence or incidence, especially experiences from the domains of health-related (PARP range 88.4–95.6%) and work-related stressful experiences (PARP range 76.8–86.5%). Nine stressful experiences were also associated with TSS persistence, of which having patient(s) in care who died from COVID-19 had the strongest association. This association remained significant after adjusting for co-occurring depression and anxiety.

TSSs among Spanish healthcare workers active during the COVID-19 pandemic are common and associated with various pandemic-related stressful experiences. Future research should investigate if these stressful experiences represent truly traumatic experiences and carry risk for the development of post-traumatic stress disorder.

Benzodiazepines are extensively used in primary care, but their long-term use is associated with adverse health outcomes and dependence.

To analyse the efficacy of two structured interventions in primary care to enable patients to discontinue long-term benzodiazepine use.

A multicentre three-arm cluster randomised controlled trial was conducted, with randomisation at general practitioner level (trial registration ISRCTN13024375). A total of 532 patients taking benzodiazepines for at least 6 months participated. After all patients were included, general practitioners were randomly allocated (1:1:1) to usual care, a structured intervention with follow-up visits (SIF) or a structured intervention with written instructions (SIW). The primary end-point was the last month self-declared benzodiazepine discontinuation confirmed by prescription claims at 12 months.

At 12 months, 76 of 168 (45%) patients in the SIW group and 86 of 191 (45%) in the SIF group had discontinued benzodiazepine use compared with 26 of 173 (15%) in the control group. After adjusting by cluster, the relative risks for benzodiazepine discontinuation were 3.01 (95% CI 2.03–4.46, P<0.0001) in the SIW and 3.00 (95% CI 2.04–4.40, P<0.0001) in the SIF group. The most frequently reported withdrawal symptoms were insomnia, anxiety and irritability.

Both interventions led to significant reductions in long-term benzodiazepine use in patients without severe comorbidity. A structured intervention with a written individualised stepped-dose reduction is less time-consuming and as effective in primary care as a more complex intervention involving follow-up visits.