Predicting early neurological deterioration in acute branch atheromatous disease without reperfusion therapy: a
1Department of Neurology, The First People's Hospital of Anqing Affiliated to Anhui Medical University, Anqing, China.
Insights
Machine learning accurately predicts early neurological deterioration (END) in acute branch atheromatous disease (BAD) patients without reperfusion. Key factors include infarct size, LDH, and SBP, enabling risk stratification.
Area of Science:
- Neurology
- Artificial Intelligence
- Medical Informatics
Background:
- Acute branch atheromatous disease (BAD) is a major cause of stroke in Asia.
- Early neurological deterioration (END) frequently occurs in BAD patients.
- Predicting END is crucial for managing BAD patients who do not receive reperfusion therapy.
Purpose of the Study:
- To develop and validate machine learning models for predicting END risk in acute BAD patients.
- To identify key clinical variables associated with END in this population.
- To create a simple scoring system for END risk stratification.
Main Methods:
- Retrospective enrollment of 369 acute BAD patients without reperfusion therapy.
- Feature selection using LASSO regression and bootstrap stability assessment.
- Development and validation of seven machine learning models, with XGBoost selected for optimal performance.
- SHAP analysis for model interpretability and development of a scoring system.
- Evaluation of model discrimination (AUC) and clinical utility (DCA).
Main Results:
- Five key predictors identified: maximum infarct area, lactate dehydrogenase (LDH), number of infarct slices, admission systolic blood pressure (SBP), and neutrophil count.
- The XGBoost model demonstrated high predictive accuracy (AUC training: 0.927, validation: 0.846, nested cross-validation: 0.866).
- A simple scoring system effectively stratified patients into low, intermediate, and high END risk groups.
- SHAP analysis highlighted maximum infarct area and LDH as the most significant predictors.
Conclusions:
- An XGBoost-based prediction model and a simple scoring system reliably predict END risk in acute BAD patients not receiving reperfusion therapy.
- The model integrates key clinical variables: maximum infarct area, LDH, infarct slice count, admission SBP, and neutrophil count.
- This tool aids in identifying high-risk patients for closer monitoring and tailored management.
Background:
Acute branch atheromatous disease (BAD) is one of the leading contributors to morbidity and disability in Asia, and early neurological deterioration (END) is common in affected patients. This study aimed to establish machine learning models to predict the risk of END in patients without reperfusion therapy.
Methods:
Patients with acute BAD who did not receive reperfusion therapy were retrospectively enrolled. Core predictive features were selected by LASSO regression with bootstrap stability assessment, and we used seven machine learning algorithms to build models. XGBoost was selected based on validation performance, nested cross-validation, and 1,000-iteration bootstrap validation. A spline logistic regression model served as the non-linear baseline. SHAP analysis was used to explain the model and develop a simple scoring system. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC), and clinical utility was evaluated using decision curve analysis (DCA).
Results:
A total of 369 patients were included in our research. We screened predictive factors with LASSO regression and ultimately identified five key variables. These included maximum infarct area, lactate dehydrogenase (LDH), number of infarct slices, admission systolic blood pressure (SBP), and neutrophil count. The XGBoost model achieved the best overall performance, with AUC of 0.927 in the training set and 0.846 in the validation set. Nested cross-validation yielded an unbiased AUC of 0.866 (95% CI: 0.817-0.925), and bootstrap validation produced a mean OOB AUC of 0.855 (95% CI: 0.760-0.941). The scoring system stratified patients into low (0-6 points), intermediate (7-13 points), and high (14-20 points) risk groups. DCA demonstrated favorable clinical utility. SHAP analysis also indicated that maximum infarct area and LDH were the top two predictors of END.
Conclusion:
An XGBoost-based prediction model and a simple scoring system, integrating maximum infarct area, LDH, number of infarct slices, admission SBP, and neutrophil count, provide reliable END risk prediction for acute BAD patients without reperfusion therapy.