Joint modeling strategy for using electronic medical records data to build machine learning models: an example of intracerebral hemorrhage

Abstract Background Outliers and class imbalance in medical data could affect the accuracy of machine learning models. For physicians who want to apply predictive models, how to use the data at hand to build a model and what model to choose are very thorny problems. Therefore, it is necessary to consider outliers, imbalanced data, model selection, and parameter tuning when modeling. Methods This study used a joint modeling strategy consisting of: outlier detection and removal, data balancing, model fitting and prediction, performance evaluation. We collected medical record data for all ICH patients with admissions in 2017–2019 from Sichuan Province. Clinical and radiological variables were used to construct models to predict mortality outcomes 90 days after discharge. We used stacking ensemble learning to combine logistic regression (LR), random forest (RF), artificial neural network (ANN), support vector machine (SVM), and k-nearest neighbors (KNN) models. Accuracy, sensitivity, specificity, AUC, precision, and F1 score were used to evaluate model performance. Finally, we compared all 84 combinations of the joint modeling strategy, including training set with and without cross-validated committees filter (CVCF), five resampling techniques (random under-sampling (RUS), random over-sampling (ROS), adaptive synthetic sampling (ADASYN), Borderline synthetic minority oversampling technique (Borderline SMOTE), synthetic minority oversampling technique and edited nearest neighbor (SMOTEENN)) and no resampling, seven models (LR, RF, ANN, SVM, KNN, Stacking, AdaBoost). Results Among 4207 patients with ICH, 2909 (69.15%) survived 90 days after discharge, and 1298 (30.85%) died within 90 days after discharge. The performance of all models improved with removing outliers by CVCF except sensitivity. For data balancing processing, the performance of training set without resampling was better than that of training set with resampling in terms of accuracy, specificity, and precision. And the AUC of ROS was the best. For seven models, the average accuracy, specificity, AUC, and precision of RF were the highest. Stacking performed best in F1 score. Among all 84 combinations of joint modeling strategy, eight combinations performed best in terms of accuracy (0.816). For sensitivity, the best performance was SMOTEENN + Stacking (0.662). For specificity, the best performance was CVCF + KNN (0.987). Stacking and AdaBoost had the best performances in AUC (0.756) and F1 score (0.602), respectively. For precision, the best performance was CVCF + SVM (0.938). Conclusion This study proposed a joint modeling strategy including outlier detection and removal, data balancing, model fitting and prediction, performance evaluation, in order to provide a reference for physicians and researchers who want to build their own models. This study illustrated the importance of outlier detection and removal for machine learning and showed that ensemble learning might be a good modeling strategy. Due to the low imbalanced ratio (IR, the ratio of majority class and minority class) in this study, we did not find any improvement in models with resampling in terms of accuracy, specificity, and precision, while ROS performed best on AUC.