Machine learning-based diagnostic model for stroke in non-neurological intensive care unit patients with acute neurological manifestations

Abstract

Stroke is a neurological complication that can occur in patients admitted to the intensive care unit (ICU) for non-neurological conditions, leading to increased mortality and prolonged hospital stays. The incidence of stroke in ICU settings is notably higher compared to the general population, and delays in diagnosis can lead to irreversible neurological damage. Early diagnosis of stroke is critical to protect brain tissue and treat neurological defects. Therefore, we developed a machine learning model to diagnose stroke in patients with acute neurological manifestations in the ICU. We retrospectively collected data on patients' underlying diseases, blood coagulation tests, procedures, and medications before neurological symptom onset from 206 patients at the Chungbuk National University Hospital ICU (July 2020-July 2022) and 45 patients at Chungnam National University Hospital between (July 2020-March 2023). Using the Categorical Boosting (CatBoost) algorithm with Bayesian optimization for hyperparameter selection and k-fold cross-validation to mitigate overfitting, we analyzed model-feature relationships with SHapley Additive exPlanations (SHAP) values. Internal model validation yielded an average accuracy of 0.7560, sensitivity of 0.8959, specificity of 0.7000, and area under the receiver operating characteristic curve (AUROC) of 0.8201. External validation yielded an accuracy of 0.7778, sensitivity of 0.7500, specificity of 0.7931, and an AUROC of 0.7328. These results demonstrated the model's effectiveness in diagnosing stroke in non-neurological ICU patients with acute neurological manifestations using their electronic health records, making it valuable for the early detection of stroke in ICU patients.

Keywords: Clinical decision support system; Intensive care unit; Neurological manifestation; Stroke.

Fig. 1

Flowchart showing the inclusion and exclusion process for internal validation.

Fig. 2

Flowchart showing the inclusion and exclusion process for external validation.

Fig. 3

Results of 4-fold cross-validation: (A) Receiver operating characteristic (ROC) curves, (B) confusion matrices of our model on the test dataset for each fold, and (C) ROC curves.

Fig. 4

Confusion matrix (right) and ROC curve (left) for the external dataset. ROC, receiver operating characteristic.

Fig. 5

SHAP summary plot (top) and absolute summary plot (bottom) for the training dataset. SHAP, SHapley Additive exPlanations. In the summary plot, colors closer to red represent high feature values and colors closer to blue represent low feature values.

Fig. 6

SHAP force plots of non-stroke (A) and stroke (B, C) patients. SHAP, SHapley Additive exPlanations. The data information that predicted the patient as stroke is colored red, and the data information that predicted the patient as non-stroke is colored blue.