Background Currently available cardiovascular disease (CVD) risk prediction tools may underestimate the risk in individuals with schizophrenia. Objective To develop and externally validate 5-year CVD risk prediction models for people with schizophrenia using large-scale register data in Sweden and Denmark with a machine learning (ML) approach. Methods Individuals with a diagnosis of schizophrenia, aged 30 and older and without prior CVD, were followed for up to 5 years. We investigated whether adding additional health-related and socio-demographic predictors to the established CVD risk factors improved predictions and compared ML models with logistic regression. External validation was performed across countries. Findings A lasso penalised logistic regression including additional predictors achieved the highest predictive performance, both on Swedish and Danish data, while complex ML models with interaction terms did not provide additional improvements. The area under the receiver operating characteristic curve (AUC) on the internal validation data was 0.745 (95% CI (0.742 to 0.749)) in the Swedish model, and 0.722, 95% CI (0.719 to 0.726) in the Danish model. External validation showed similar performance, yielding an AUC of 0.746, 95% CI (0.741 to 0.751) using the Danish model on the Swedish data, and an AUC of 0.720, 95% CI (0.712 to 0.726) using the Swedish model on the Danish validation data. Conclusions Incorporating additional health-related information, such as psychiatric comorbidities and medication use, improved 5-year CVD risk prediction for people with schizophrenia in both countries. Clinical implications The models can be deployed between Denmark and Sweden without loss of performance compared with training a model on each country.
Nielsen et al. (2026) studied this question.