Machine-learning-based risk stratification for probability of dying in patients with basal ganglia hemorrhage

Abstract

To confirm whether machine learning algorithms (MLA) can achieve an effective risk stratification of dying within 7 days after basal ganglia hemorrhage (BGH). We collected patients with BGH admitted to Sichuan Provincial People's Hospital between August 2005 and August 2021. We developed standard ML-supervised models and fusion models to assess the prognostic risk of patients with BGH and compared them with the classical logistic regression model. We also use the SHAP algorithm to provide clinical interpretability. 1383 patients with BGH were included and divided into the conservative treatment group (CTG) and surgical treatment group (STG). In CTG, the Stack model has the highest sensitivity (78.5%). In STG, Weight-Stack model achieves 58.6% sensitivity and 85.1% specificity, and XGBoost achieves 61.4% sensitivity and 82.4% specificity. The SHAP algorithm shows that the predicted preferred characteristics of the CTG are consciousness, hemorrhage volume, prehospital time, break into ventricles, brain herniation, intraoperative blood loss, and hsCRP were also added to the STG. XGBoost, Stack, and Weight-Stack models combined with easily available clinical data enable risk stratification of BGH patients with high performance. These ML classifiers could assist clinicians and families to identify risk states timely when emergency admission and offer medical care and nursing information.

Figure 1

Distribution of time to death in patients with basal ganglia hemorrhage. The upper right corner shows deaths within 30 days of study subjects.

Figure 2

Flow diagram of study recruitment and exclusion.

Figure 3

Overall workflow summarizing model algorithm.

Figure 4

Receiver operating characteristic curves versus precision-recall curves for the conservative treatment group and the surgical treatment group. (a) Receiver operating characteristic curves for all models in the conservative treatment group. (b) Receiver operating characteristic curves for all models in the surgical treatment group. (c) Precision-recall curves for all models in the conservative treatment group. (d) Precision-recall curves for all models in the surgical treatment group.

Figure 5

Overall SHAP interpretation of XGBoost and Stack model risk stratification. XGBoost uses the tree model interpretation method and the Stack model uses the Kernel interpretation method. (a, c): Ranking of all features in order of importance and showing the top 20 features. (b, d): Distribution showing the correlation of features with risk stratification results. Each point is a sample, red dots represent a high value of features, blue dots represent a low value of features, a negative SHAP value indicates a decrease in the probability of high risk and a positive one indicates an increase in the probability of high risk.

Figure 6

SHAP force plots based on individual patient prediction scores. (a) Low- and high-risk instances in the CTG. (b) Low- and high-risk instances in the STG. the base value is the mean Shap values predicted by the model, and the red and blue directed bars indicate the risk and safety features, respectively; the length of the bars represents the importance of the feature, which together drive the prediction from the base to the final value. The Shap values shown in the graph are the log odds of the true predicted probability.