MRI-based clinical-radiomics nomogram to predict early neurological deterioration in isolated acute pontine infarction: a two-center study in Northeast China

Abstract

Objective: To predict the appearance of early neurological deterioration (END) among patients with isolated acute pontine infarction (API) based on magnetic resonance imaging (MRI)-derived radiomics of the infarct site.

Methods: 544 patients with isolated API were recruited from two centers and divided into the training set (n = 344) and the verification set (n = 200). In total, 1702 radiomics characteristics were extracted from each patient. A support vector machine algorithm was used to construct a radiomics signature (rad-score). Subsequently, univariate and multivariate logistic regression (LR) analysis was adopted to filter clinical indicators and establish clinical models. Then, based on the LR algorithm, the rad-score and clinical indicators were integrated to construct the clinical-radiomics model, which was compared with other models.

Results: A clinical-radiomics model was established, including the 5 indicators rad-score, age, initial systolic blood pressure, initial National Institute of Health Stroke Scale, and triglyceride. A nomogram was then made based on the model. The nomogram had good predictive accuracy, with an area under the curve (AUC) of 0.966 (95% confidence interval [CI] 0.947-0.985) and 0.920 (95% [CI] 0.873-0.967) in the training and verification sets, respectively. According to the decision curve analysis, the clinical-radiomics model showed better clinical value than the other models. In addition, the calibration curves also showed that the model has excellent consistency.

Conclusion: The clinical-radiomics model combined MRI-derived radiomics and clinical metrics and may serve as a scoring tool for early prediction of END among patients with isolated API.

Keywords: Acute pontine infarction; Early neurological deterioration; Magnetic resonance imaging; Nomogram; Radiomics.

Fig. 1

Patient recruitment criteria and process

Fig. 2

The flowchart for radiomics analysis (A) and research analysis (B). ICC = intraclass correlation coefficient, ROC = receiver operating characteristic, DCA = decision curve analysis, LASSO = least absolute shrinkage and selection operator, SVM = support vector machine

Fig. 3

Screening of radiomics features based on LASSO regression analysis. (A) Distribution of coefficients of the LASSO regression. Each line represents a radiomics feature. (B) Application of 10-fold cross-verification for tuning optimal parameters in LASSO regression. Finally, the optimal lambda (λ) 0.044984 was obtained, and a total of 9 radiomics features were filtered. (C) Histogram of coefficients for 9 radiomics features

Fig. 4

Plots of the clinical-radiomics model’s nomogram development, performance evaluation, and model comparison. (A, B) ROC curves of the models in the training set and verification set; (C) Nomogram developed based on the clinical-radiomics model; (D, E) DCA of nomogram in the training set and verification set. (F, G) calibration curves of the nomogram in the training set and verification set. SBP = systolic blood pressure, NIHSS = National Institute of Health Stroke Scale, TG = triglyceride, ROC = receiver operating characteristic, DCA = decision curve analysis