A hybrid random forest model for stroke risk prediction

Objectives Our study aims to develop a stroke risk prediction model by multiple machine learning algorithms and optimize the model as a stroke risk prediction tool. Methods This retrospective multicenter study derived the original dataset from a high-quality health database. The dataset was incomplete and class imbalanced. Firstly, we eliminated extreme outliers and noises and imputed missing values by appropriate algorithms. We further used Synthetic Minority Over-sampling Technique to generate a balanced dataset. Secondly, we fitted seven algorithms to develop a machine learning-based prediction tool for clinical practice. Results Overall, 35,859 participants were included, of whom 781 (2.2%) experienced a stroke. The random forest model demonstrated the best performance with high predictive value and discrimination ability. For stroke risk prediction, the AUC of the best-performing model was 0.97. Conclusion A new random forest algorithms-based stroke risk prediction model using easily obtainable data was developed and outperformed established models. Future studies should further validate and optimize the current model to assess its generalizability and promote the wide application. The utilization of proposed random forest algorithms as an individualized risk prediction model could facilitate the application of clinical practice guidelines and shared decision-making.