ionwide dataset with almos

forced to be less precise

d for a maximum period of 90 days.

Swedish prescription regulations

The association of various patient- and disease characteristics and post-aSAH depression have been investigated in a wide range of studies with varying results. Occasionally, female sex has been associated with post-aSAH depression8,9, along with younger age10, ethnicity and non-fluency in the majority language3. Multiple studies have shown association between prior psychiatric conditions and post-aSAH depression3,8,11, along with certain somatic comorbidities such as diabetes and chronic obstructive pulmonary disease being weakly associated12. Outcome tools such as the Modified Rankin Scale (mRS) and the Glasgow Outcome Scale (GOS) have been consistently demonstrated to be associated with depression9,10,12, as has post-aSAH epilepsy10. Certain evidence also indicates an association between depression and aSAH-related infarcts and infarct or aneurysm location3,13,14, and at least two studies show a correlation between cortisol levels and depression15,16. Finally, indicators of more severe disease (e.g. increased Hunt-Hess score) have shown an association with post-aSAH depression9,10.

SAH. Importantly, we also show that this increased risk is approximately 40% higher than what would be expected as a consequence of intensive care alone, supporting an inherent pathophysiologic effect of aSAH on the development om mood disorders. Additionally, we are able to translate this increased risk into a statistically significant increase in sick leave, highlighting the socio-economic impact of the disease.

uasi-causal in

ore pronounced in patients treated with AD after aSAH. Since we have elected to view SA and DP as outcomes resulting from mood disorders, we have chosen to evaluate a time frame beginning one year after the onset of disease. While this approach allows for a correct temporal analysis, we are still unable to assess causality and our results should therefore be interpreted with caution. Regardless, our findings do highlight the association between mood disorders and socio-economic consequences with profound patient-centred implications.

ly improved covariate balance (Supplementary Materials) and was then use

he significant differences between the study and control cohorts, especially with regards to markers of disease severity, a propensity score matching process was undertaken to evaluate the marginal effect of aSAH on the risk of post-ICU AD-treatment versus the risk conferred by exposure to intensive care alone. Using a the Random Fores

a cohort consisting of all ICU-patients with non-neurovascular primary causes of admission in Sweden between 2012-2018. A comparison of baseline demographics, comorbidities, and ICU-variables is shown in Table 2. Of note, patients in the control cohort were predominantly male, and both older and more comorbid tha

years post-aSAH when each indiv

bootstrap resample and the resulting metrics presented as risk ratios along with their bootstrap distribution.

rtion of AD-treated patients requiring immediate invasive ventilation and having a poorer initial consciousness level.

In total, 78985 unique patients were extracted from the NPR (Figure 1). Following the selection strategy described in the Methods section, a study cohort of 7378 patients with a presumed aSAH was identified. Of these patients, 1423 had dispensed a prescription of antidepressants in the 180 days prior to ICU-admission for aSAH and were thus excluded from final analysis, along with 1652 patients who died within 1 year following ICU-admission or where the ICU admission occurred less than 1 year prior to the end of the study. The remai

Propensity scores denoting the probability of cohort membership (study or control cohorts) were subsequently calculated for each admission in the combined data set by employing a Random Forests classification model. First, 20% of the data was allocated to a holdout set while the model was trained on the remaining 80%. Model preprocessing included a log-transformation of the ICU length of stay variable along with dummy coding of nominal variables and normalization of covariates . Hyperparameter tuning across a 10-fold cross-validation resample of the training subset was performed by using a space-filling grid search with the Latin Hypercube method and a grid size of 20. The hyperparameter settings maximizing the Area Under the Receiver Operating Characteristic curve (AUROC) were selected and the final model trained. Subsequent model testing on the 20% holdout set was performed, analysing accuracy, Brier score, and AUROC, with further validation performed using a 10-fold cross-validation resample and plotting the ROC for each fold individually. Finally, the propensity scores were predicted on the entire data set.

on of the Δleave was clearly bimodal (see Su

end of the study.

s of socio-economic impact

Collinearity b

Raw register data were thoroughly preprocessed in order to create continuo

mitigate the competing risk

180

Figure

Using the nationwide study cohort, selection of aSAH cases were subsequently performed using the steps seen in Figure 1 (SQL source code available in Supplementary Materials).

send

LATE was estimated using a model implemented in rstan with non-informative Jeffreys priors. A detailed description of the estimand and estimator is provided in the methodological supplement. The rstan code is available in the article notebook available online.

performance class 1 or 2.