Clinical, radiological, and radiomics feature-based explainable machine learning models for prediction of neurological deterioration and 90-day outcomes in mild intracerebral hemorrhage

Abstract

Background: The risks and prognosis of mild intracerebral hemorrhage (ICH) patients were easily overlooked by clinicians. Our goal was to use machine learning (ML) methods to predict mild ICH patients' neurological deterioration (ND) and 90-day prognosis.

Methods: This prospective study recruited 257 patients with mild ICH for this study. After exclusions, 148 patients were included in the ND study and 144 patients in the 90-day prognosis study. We trained five ML models using filtered data, including clinical, traditional imaging, and radiomics indicators based on non-contrast computed tomography (NCCT). Additionally, we incorporated the Shapley Additive Explanation (SHAP) method to display key features and visualize the decision-making process of the model for each individual.

Results: A total of 21 (14.2%) mild ICH patients developed ND, and 35 (24.3%) mild ICH patients had a 90-day poor prognosis. In the validation set, the support vector machine (SVM) models achieved an AUC of 0.846 (95% confidence intervals (CI), 0.627-1.000) and an F1-score of 0.667 for predicting ND, and an AUC of 0.970 (95% CI, 0.928-1.000), and an F1-score of 0.846 for predicting 90-day prognosis. The SHAP analysis results indicated that several clinical features, the island sign, and the radiomics features of the hematoma were of significant value in predicting ND and 90-day prognosis.

Conclusion: The ML models, constructed using clinical, traditional imaging, and radiomics indicators, demonstrated good classification performance in predicting ND and 90-day prognosis in patients with mild ICH, and have the potential to serve as an effective tool in clinical practice.

Clinical trial number: Not applicable.

Keywords: Computed tomography; Intracerebral hemorrhage; Machine learning; Modified rankin scale; Neurological deterioration; Radiomics.

Fig. 1

The schema of study workflow

Fig. 2

Examples of each imaging sign of traditional imaging features: (a) Swirl sign; (b) Blend sign; (c) Black hole sign; (d) Low density sign; (e) Island sign; (f) Satellite sign

Fig. 3

LASSO regression analysis results. (a, b) Feature coefficient path diagram and mean squared error curve of the ND dataset; (c, d) Feature coefficient path diagram and mean squared error curve of the mRS-90 dataset

Fig. 4

Classification of the five ML models for predicting ND. (a) The ROC of five ML models in the validation set. (b) The DCA of five ML models in the validation set

Fig. 5

The key features of the optimal model for predicting ND identified using the SHAP algorithm. (a) The global bar plot of SHAP values, with features ordered by their contribution to the ML model in descending order. (b) The beeswarm summary plot of SHAP values, illustrating the impact of these features on predictions

Fig. 6

Classification of the five ML models for predicting poor prognosis. (a) The ROC of five ML models in the validation set. (b) The DCA of five ML models in the validation set

Fig. 7

The key features of the optimal model for predicting poor prognosis identified using the SHAP algorithm. (a) The global bar plot of SHAP values, with features ordered by their contribution to the ML model in descending order. (b) The beeswarm summary plot of SHAP values, illustrating the impact of these features on predictions. The colors represent feature values, ranging from high (red) to low (blue), and the horizontal position indicates whether the feature value contributes to a positive or negative prediction

Fig. 8

Two mild ICH cases were selected from the validation sets of the ND and mRS-90 prediction tasks, respectively. The decision-making process of the SVM model is illustrated using force plots. (a, c) Case #1, a 69-year-old male, experienced a hemorrhage in the left thalamus and subsequently developed ND during hospitalization. Despite a relatively low ICH score, due to factors including a low 24-hour FOUR score, high APACHE-II score, and other variables, the SVM model predicted that the patient had a high risk of ND. (b, d) Case #2, a 52-year-old male, experienced a hemorrhage in the right basal ganglia and temporal lobe and mRS-90 of 4. Due to factors including high NIHSS score at admission, high BMI, and the presence of the island sign, the SVM model predicted that this patient had a higher risk of poor 90-day prognosis. Each feature provides a SHAP value to the model’s base value. The final prediction value, f(x), is derived by the weighted sum of these features as processed by the model. When f(x) > 0, the model determines the condition to be ND or mRS > 2; otherwise, it is considered non-ND or mRS ≤ 2