Introduction
Neurological symptom worsening, also known as early neurological deterioration (END), is common in isolated acute pontine infarction (API), with a prevalence ranging from 14 to 35%, based on the diagnostic standard for END and the time from onset to symptom assessment in individual studies [
1,
2]. END can lead to poor patient outcomes, poor prognosis, and even death [
3]. Therefore, identifying END at an early stage and prompt treatment initiation is crucial. Previous studies have been conducted to forecast the occurrence of END among patients with isolated API, revealing that the National Institutes of Health Stroke Scale (NIHSS) score (a scale that can be used as a measure of stroke severity) [
4], age, fasting blood glucose, and high-density lipoprotein ratio correlated with the appearance of END [
5]. In addition, the relationship between magnetic resonance imaging (MRI) and END has been studied in depth due to its high sensitivity for the early detection of lesions. The topographic location [
1,
2,
6] and the size of the infarct lesion [
7,
8] were found to be predictors of END.
Radiomics is an emerging technology for analyzing medical imaging data and can capture multiple characteristics that cannot be identified by the naked eye [
9,
10]. Analyzing and filtering these features allows for the construction of classification models to support clinicians in assessing diseases in a timely, comprehensive, and non-invasive manner [
11,
12]. Radiomics has been shown to be an effective image analysis method for depicting ischemic penumbra(area of hypoperfusion around the infarct core within the same vessel supply as the infarct core) [
13], predicting malignant cerebral edema (mainly characterised by a malignant course accompanied by severe cerebral oedema, which can lead to the occurrence of brain herniation causing death or severe neurological dysfunction) [
14], and clinical results among patients with acute ischemic stroke (AIS) [
15]. Previously, the mean relative apparent diffusion coefficient (ADC) value has been reported to be a predictor of END in patients with isolated API [
16]. However, further research is needed due to the research limitations, such as not extracting high-throughput information from the images, integrating multiple sequences in MRI images, and building the model from a relatively small dataset (
n = 63) without external verification. Diffusion-weighted imaging (DWI) is the most sensitive technology for the early detection of AIS, but relatively few studies have been performed using DWI in isolated API. Furthermore, a single MRI modality was applied in most of the studies. Therefore, this research aimed to construct a clinical-radiomics model for the identification of END based on multiple sequences of radiomics features, and identifying independent clinical factors, with external validation.
Discussion
In this study, we successfully developed a nomogram model to predict functional outcomes in patients with API by integrating radiomics and clinical information. Simultaneously, the model shows better prediction accuracy in both the internal training set and the external validation set, making the results more convincing. Furthermore, to our knowledge, this is the largest sample of studies using radiomics to predict END occurrence.
END is associated with a poor prognosis and even death. The cold climate of northern China leads to a high incidence of END. Therefore, predicting END onset at an early phase and adjusting patients’ treatment plans in a timely manner is essential. MRI provides important information about the size of the lesion, the appearance of bleeding, and the vascular status in patients with acute ischemic stroke (AIS). Up to now, studies have focused on forecasting the appearance of END based on conventional imaging features, such as topographic location and the extent of infarct lesions [
21]. However, some biological features cannot be assessed by visual inspection, but are closely correlated with the appearance and growth of diseases. Such features could be beneficial for END prediction and have received increasing attention in recent years.
Radiomics is an emerging field of study based on the quantitative analysis of medical images and the automatic extraction of subtle features in images that help improve the accuracy of diagnosis, prognosis, and prediction of various diseases [
22]. Currently, radiomics has been used to assess recurrence rates [
23], hemorrhagic transformation [
24], and prognosis [
25] in patients with AIS in various ways. Oge et al. demonstrated that the mean ADC value can be used as a predictor for the occurrence of END in isolated API patients [
16]. However, previous studies have used a single MRI sequence for feature extraction, while integrating multiple sequences may yield more valid information. In addition, previous studies have not been validated with external centers, making their generalizability somewhat limited. This research extracts the radiomics characteristics from both DWI and ADC sequences for model construction and validation, in order to better reflect the heterogeneity of END. The study findings confirm our proposed hypothesis and may provide new directions to support END prediction.
This research finally extracted 9 radiomics characteristics to construct the rad-score, with the features extracted from DWI accounting for the largest coefficient weights. Among these features, the best predictors were the Gray Level Size Zone Matrix (GLSZM) and Neighbouring Gray Tone Difference Matrix (NGTDM), which were extracted based on the wavelet transform algorithm. The former mainly provides information on the distance between different pixels or voxels in a 2-dimensional or 3-dimensional spatial neighborhood, while the latter mainly provides information on the difference between the average gray level of each pixel or voxel and that of the neighboring pixels or voxels. They both reflect the heterogeneity of the lesion site, with higher values indicating higher heterogeneity in the lesion signal and a greater risk of developing END.
Meanwhile, the four clinical indicators age, initial SBP, initial NIHSS, and TG were recognized as independent predictors of the occurrence of END. Therefore, these four clinical indicators and the rad-score were used to build a clinical-radiomics model, which was presented as a nomogram to predict the probability of END occurrence. Consistent with previous studies [
26‐
29], advanced age, higher NIHSS scores, higher SBP levels at admission, and higher TG were risk elements for the development of END. Notably, the nomogram showed better predictive power and clinical usefulness than the radiomics signature model and the clinical model, both in the training set and in the verification set. In addition, we have developed a new type of prediction method that integrates multiple clinical predictors and plots them on scaled lines to show the relationships between the variables in the prediction model, making the results more readable and easier to evaluate [
30]. More importantly, it allows an individual risk assessment for each patient with relatively good accuracy and relevance [
31]. For example, a 64-year-old patient with isolated API with an initial SBP of 162 mmHg, an initial NIHSSH of 11, TG levels of 2.77 mmol, and a rad-score of 0.88 received a final score of 248, predicting a 99.5% risk of developing END. This score prompts clinicians to be more vigilant and adjust the medication regimen according to the condition (Supplementary Fig.
1). To the best of our knowledge, this is the largest multicenter-validated radiomics research on predicting the occurrence of END in patients with isolated API, providing an intuitive, reliable, and convenient tool to differentiate between END and non-END. Importantly, the rad-score in the clinical-radiomics model can be obtained from routine MRI examinations, and the clinical indicators in the model are blood tests routinely performed in API patients. Therefore, this score does not involve additional cost, so the model has good generalizability and practicality.
Nevertheless, the limitations of the current study should be acknowledged. First, the representativeness and generalisability of the dataset is somewhat limited, as the data for this study come from multi-centre hospitals in a single geographical area. Second, although deep learning-based image segmentation achieves good clinical results in terms of both precision and accuracy, the large amount of data required and the expense involved limit its application in this study. Finally, due to the complexity of image processing, feature extraction, and data processing, the model is not currently well generalised, and in the future we will work on packaging the machine learning model into an interface (e.g. by “containerising” the model) that can be used by anyone as an off-the-shelf system or tool.
In conclusion, a clinical-radiomics model including rad-score, age, initial SBP, initial NIHSS, and TG was developed and validated based on the radiomics characteristics and clinical elements of multimodal MRI of the infarct region. Compared to the radiomics signature and clinical model, the clinical-radiomics model provided a more reliable and rapid forecast of END, which may help optimize disease management and develop personalized medication regimens.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.