A systematic review of the status and methodological considerations for estimating risk of first ever stroke in the general population

Aims: The methodological quality of development, validation, and modification of those models have not been evaluated via a thoroughly literature review. This study aims to describe the overall status and evaluate the methodological quality of risk prediction models for stroke incidence in the general population. Methods: We searched the database of EMBASE and MEDLINE by the combination of subject words and key words to collect the research on stroke risk prediction model in the general population. The retrieval time was from the establishment of the database to September 2019. It should be mentioned that risk of bias for each model was assessed, and data on population characteristics and model performance was also extracted. Results: The search screened 11,386 peer-reviewed publications and 57 citation searching, of which 48 were included in the review, describing the development of 51 prediction models, 47 external validation models, and 12 modification models. Among 51 development models, the predicted outcome concentrated on fatal or non-fatal stroke (n = 37, 73%). Thirty-nine development models (76%) were without internal validation. C-statistic or AUC was adopted for discrimination in 80% models, and Hosmer-Lemeshow test (n = 25, 49%) was also performed for calibration. Twenty-six development models (53%) were externally validated, among which only 2 (8%) were validated by independent researchers. Risk prediction performance was improved when models were modified by adding novel risk factors, such as the internal carotid artery plaque and intima-media thickness. Conclusion: Models for predicting stroke occurrence need further external validation, recalibration, or modification in different populations, to help interpret those models in the practice of stroke prevention.