Objectives
Prediction models for post-stroke mortality can support medical decision-making. Although numerous models have been developed, external validation studies determining the models’ transportability beyond the original settings are lacking. We aimed to assess the performance of two prediction models for post-stroke mortality in Berlin, Germany.
Design
We used data from the Berlin-SPecific Acute Treatment in Ischaemic or hAemorrhagic stroke with Long-term follow-up (B-SPATIAL) registry.
Setting
Multicentre stroke registry in Berlin, Germany.
Participants
Adult patients admitted within 6 hours after symptom onset and with a 10th revision of the International Classification of Diseases discharge diagnosis of ischaemic stroke, haemorrhagic stroke or transient ischaemic attack at one of 15 hospitals with stroke units between 1 January 2016 and 31 January 2021.
Primary outcome measures
We evaluated calibration (calibration-in-the-large, intercept, slope and plot) and discrimination performance (c-statistic) of Bray et al’s 30-day mortality and Smith et al’s in-hospital mortality prediction models. Information on mortality was supplemented by Berlin city registration office records.
Results
For the validation of Bray et al’s model, we included 7879 patients (mean age 75; 55.0% men). We observed 763 (9.7%) deaths within 30 days of stroke compared with 680 (8.6%) predicted. The model’s c-statistic was 0.865 (95% CI: 0.851 to 0.879). For Smith et al’s model, we performed the validation among 1931 patients (mean age 75; 56.2% men), observing 105 (5.4%) in-hospital deaths compared with the 92 (4.8%) predicted. The c-statistic was 0.891 (95% CI: 0.864 to 0.918). The calibration plots of both models revealed an underestimation of the mortality risk for high-risk patients.
Conclusions
Among Berlin stroke patients, both models showed good calibration performance for low and medium-risk patients and high discrimination while underestimating risk among high-risk patients. The acceptable performance of Bray et al’s model in Berlin illustrates how a small number of routinely collected variables can be sufficient for valid prediction of post-stroke mortality.