Assessment of outcomes in postaneurysmal subarachnoid bleed patients admitted to the intensive care unit utilizing the subarachnoid haemorrhage international trialist clinicoradiological prediction model for dichotomised functional outcome and mortality

Objective

The objective of this study was to assess the Subarachnoid Haemorrhage International Trialist (SAHIT) prediction model in a tertiary adult intensive care unit (ICU) cohort when assessing patient outcomes against predicted outcomes, firstly by assessing the discrimination and validation of the model in the Princess Alexandra Hospital (PA) intensive care cohort and secondly comparing the predicted outcomes using the SAHIT model to the actual cohort outcomes using a Monte Carlo simulation.

Methods

Six logistic regression models designed by the SAHIT Collaboration Group were applied to the PA cohort considering early predictive factors such as clinical grade and treatment modality to predict the risk of both mortality and unfavourable outcome at 6 months according to the Glasgow Outcome Score. The six SAHIT logistic regression models were applied to a retrospectively collected cohort of aneurysmal subarachnoid patients who were admitted to the ICU, generating individual risk scores for mortality and poor functional outcome. Area under the curve (AUC) and calibration slope/intercept and Brier score were used to assess the strength of the model in interpreting the current data set. A Monte Carlo analysis was used to compare the actual mortality outcomes to the predicted outcomes to determine if the cohort performance was better or worse than predicted by the mortality model.

Results

Overall, the PA cohort actual mortality was higher than the predicted mortality rate based on the risk scores generated by the SAHIT models, demonstrated by Monte Carlo simulation using the SAHIT model risk scores. The core, neuroimaging, and full models for functional outcome produced AUCs of 0.719 (95% confidence interval [CI]: 0.55–0.84), 0.709 (95% CI: 0.55–0.83), and 0.738 (95% CI: 0.58–0.85). Regarding mortality, the respective AUCs were 0.684 (95% CI: 0.57–0.78), 0.678 (95% CI: 0.56–0.77), and 0.749 (95% CI: 0.64–0.84). Regarding calibration, there was modest calibration in general, with higher degrees of calibration in the fully functional outcome model.

Conclusion

The cohort outcomes for mortality occurred at a rate higher than the risk predictions suggested using the logistic regression created by the SAHITs. Applying the externally trained model provided adequate discrimination and modest calibration, yet underestimated risk when applied to the intensive care cohort, reflected in the probability density function analysis. Using the SAHIT models in this cohort may result in underestimation of mortality for the individual patient, and the accuracy of the model is not sufficient for individual patient prediction. These results challenge the appropriateness of using admission-based models for dynamic ICU populations and highlight the urgent need for critical care–specific prognostic tools.