Development and Validation of an Early Recurrence Prediction Model for High-Grade Glioma Integrating Temporalis Muscle and Tumor Features: Exploring the Prognostic Value of Temporalis Muscle

Abstract

This study aimed to develop and validate a predictive model for early recurrence of high-grade glioma (HGG) within 180 days, assess the prognostic value of preoperative and postoperative temporalis muscle metrics (area and thickness), and explore their significance in postoperative follow-up. Seventy-one molecularly confirmed HGG patients were included, with data sourced from local data and TCIA (The Cancer Imaging Archive) RHUH-GBM (Río Hortega University Hospital Glioblastoma) dataset. Tumor segmentation was performed using deep learning, and radiomic features were extracted following comparison with manual segmentation. Feature selection was conducted using mutual information and recursive feature elimination. A comprehensive model integrating 3D tumor radiomics and temporalis muscle metrics was developed and compared with a tumor-only model to identify the optimal predictive framework. SHAP analysis was used to evaluate model interpretability and feature importance. The TM_Tumor_HistGradientBoosting model, incorporating 16 features including temporalis muscle metrics, outperformed the tumor-only model in accuracy (0.89), recall (0.87), and F1 score (0.88). SHAP analysis highlighted that preoperative temporalis muscle cross-sectional area was strongly associated with early recurrence risk, while postoperative temporalis muscle thickness significantly contributed to recurrence prediction. Combining temporalis muscle metrics with preoperative tumor MRI substantially improved the accuracy of early recurrence prediction in HGG. Temporalis muscle metrics serve as objective and sustainable prognostic indicators with significant clinical value in postoperative follow-up.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-025-01491-w.

Introduction

High-grade gliomas (World Health Organization grades III and IV) are the most frequent and fatal brain tumors [1–3]. Despite advances in surgery, radiotherapy, and chemotherapy, the prognosis of HGG patients remains poor, grade III with median overall survivals of 24–72 and grade IV with 14–16 months [4, 5]. One of the key challenges in treating HGG is the high rate of early recurrence, which severely impacts patient outcomes [6, 7].

Understanding early recurrence in high-grade gliomas is critical, as it is often associated with more aggressive tumor biology and increased resistance to therapy, which necessitates early intervention [7, 8]. Identifying the factors that contribute to early recurrence can help pinpoint high-risk patients who may benefit from more aggressive or alternative treatment strategies [9]. In this context, the prediction of early recurrence is vital for improving clinical outcomes. High-risk patients may benefit from more intensive therapeutic approaches, such as heightened postoperative monitoring, earlier initiation of radiotherapy, or adjustments to chemotherapy regimens to mitigate the risk of tumor recurrence [10, 11]. Conversely, low-risk patients can avoid unnecessary treatment burdens, thus reducing potential adverse effects and preserving quality of life. Furthermore, timely intervention in high-risk patients can extend progression-free survival (PFS) and improve overall prognosis, leading to enhanced survival rates [6, 12].

Tumor and peritumoral features have been extensively used to enhance radiomics analysis, such as for prognostic prediction. However, accurate tumor radiomics feature extraction relies on accurate delineation. High-quality delineation usually requires time-consuming manual segmentation by professional radiologists, and the consistency of delineation by doctors of different qualifications is difficult to ensure. To address the shortcomings of manual segmentation, we trained four deep learning segmentation models, including UNet (U-Net deep learning segmentation model) and its variants AttUNet (Attention U-Net), TransUNet (Transformer U-Net), and UNetPlusPlus (U-Net + + (improved version of U-N)) [13–17]. The best-performing network, selected based on Intersection over Union (IoU), was used as the automatic segmentation component in our model.

Furthermore, studies have shown that temporalis muscle thickness is associated with the prognosis of glioma patients [18]. Previous studies have established a close association between cancer and sarcopenia, with the psoas muscle often being used as a standard measure for assessing sarcopenia. In glioma patients, routine brain MRI scans provide an opportunity to evaluate the temporalis muscle, which has been demonstrated in several studies to serve as an effective indicator for sarcopenia. This accessibility of temporalis muscle information in glioma patients makes it a valuable imaging marker for prognostic assessment. Building upon existing research linking temporalis muscle and glioma prognosis, our study further investigates the predictive value of temporalis muscle measurements in forecasting early recurrence of high-grade gliomas (HGG) [19, 20]. We incorporated both preoperative and postoperative imaging metrics of the temporalis muscle into our analysis, alongside preoperative tumor and postoperative enhancement data. After feature engineering mutual information and RFE (recursive feature elimination) selection, multiple recurrence prediction models were established using the selected features, using algorithms such as AdaBoost (Adaptive Boosting), CatBoost (Categorical Boosting), AttUNet, HistGradientBoosting (Histogram-based Gradient Boosting), LightGBM (Light Gradient Boosting Machine), RandomForest, and XGBoost [17, 21–26]. Our objective was to develop an early recurrence prediction model incorporating temporalis muscle information alongside tumor data, and compare it with a model using preoperative tumor information alone to identify the best-performing model. Finally, we utilized SHAP to enhance the interpretability of the machine learning model, making it more suitable for clinical deployment, while confirming the predictive value of temporalis muscle quantitative metrics for early recurrence.

Methods

Data

A total of 71 adult patients (aged > 18 years) with primary newly diagnosed high-grade gliomas who underwent surgical treatment were included in this study. Thirty-seven patients were sourced from The Cancer Imaging Archive (TCIA) RHUH-GBM (Río Hortega University Hospital Glioblastoma) dataset (https://doi.org/10.7937/4545-c905), with data collected between January 2018 and December 2022. Additionally, 44 patients were collected from the local hospital PACS system between January 2018 and December 2023, adhering to specific inclusion and exclusion criteria. The use of local data was approved by the institutional review board, with informed consent waived. The public database, which does not contain patient identifiers, did not require institutional review board approval [27]. Both datasets provide early postoperative data and detailed resection information, and both sources were collected in a time-continuous manner. The imaging protocol is shown in Supplementary Table S1. The patients included in the sample underwent gross total resection (GTR) or near-total resection (NTR), defined as no residual tumor enhancement and a resection volume exceeding 95% of the initial enhancing tumor volume, respectively. Patients were treated with systemic temozolomide according to the Stupp protocol.

Data used include the following: (1) preoperative (within 30 days) and postoperative (within 72 h) MRI scans, including T1-weighted imaging (T1WI), T1-weighted contrast-enhanced imaging (T1CE), and fluid-attenuated inversion recovery imaging (T2 FLAIR); (2) available clinical data (age, sex); (3) IDH (isocitrate dehydrogenase) mutation status, WHO (World Health Organization) grade; and (4) primary prognostic indicators (PFS). Exclusion criteria were as follows: (1) incomplete imaging sequences; (2) history of brain trauma, other brain tumors, hematologic disorders, or current infectious diseases; and (3) previous brain treatments. Our inclusion and exclusion criteria are shown in Fig. 1.

Fig. 1.

Patient cohort selection and model development workflow. T1WI, T1-weighted imaging;T1CE, T1-weighted contrast-enhanced imaging; T2 FLAIR, fluid-attenuated inversion recovery imaging; PFS, progression-free survival. Description: This flowchart outlines the patient screening process based on inclusion and exclusion criteria. Patients from two datasets (Local Database and Río Hortega University Hospital, Spain) were evaluated for imaging completeness, medical history, and eligibility according to the study’s objectives. The final cohort was divided into a development cohort and a test cohort

Image Acquisition, Segmentation, and Radiomic Feature Extraction

The local dataset was collected from the local database’ PACS (picture archiving and communication system), with images acquired on 1.5 T or 3.0 T MRI scanners using a standard head coil in coronal, sagittal, and axial planes. The public dataset, similarly sourced from PACS, was obtained on a 1.5 T scanner. CE-T1 (contrast-enhanced T1-weighted imaging) images were acquired within 200 s post-injection of gadolinium-based contrast agents (0.2 mmol/kg). For analysis, CE-T1, T1WI, and T2FLAIR images were used.

Two researchers delineated regions of interest (ROIs) on 3D Slicer. Researcher A measured pre- and postoperative temporalis muscle thickness and maximum cross-sectional area on T1WI, tumor enhancement volume on CE-T1, and high-signal volume on T2FLAIR, and delineated the preoperative tumor ROI on CE-T1. To minimize information loss from 2D temporalis segmentation on T1WI, the largest cross-section and adjacent slices were measured and averaged. Researcher B performed independent measurements, and interobserver reliability was assessed on 34 patients using the intraclass correlation coefficient (ICC). If there is a disagreement, a third researcher, a senior doctor, resolved it.

For efficiency, we utilized automatic tumor segmentation, including UNet (U-Net deep learning segmentation model) and its variants AttUNet (Attention U-Net), TransUNet (Transformer U-Net), and UNetPlusPlus (U-Net + + , an improved version of U-Net), comparing its accuracy to manual delineation to select the best model [8]. The model was trained on 2D slices from 3D NRRD images, limited by sample size. Automatic segmentation was performed on each patient’s preoperative T1-weighted images (T1w) to generate tumor masks, with predicted labels aligned to manual segmentations, processed with edge smoothing, hole filling, and dilation, and reconstructed into 3D images for further analysis. Radiomic features, including default, wavelet, and logarithmic filter features, were extracted using PyRadiomics [9]. The feature extraction was implemented using several Python packages: SimpleITK for loading and processing 3D NRRD images [28–30], PyRadiomics for extracting radiomic features such as wavelet and logarithmic features [31], and OpenCV for image preprocessing tasks like thresholding and morphological operations [32]. NumPy was used for numerical operations on 3D image arrays [33], while Pandas handled the organization and storage of extracted features in a DataFrame. re (Regular Expressions) was used to extract numerical indices from filenames, and Pillow was used for basic image loading and saving. Although all patients underwent postoperative imaging, most had gross total resection (GTR) or near total resection (NTR) [9], making it challenging to accurately delineate residual lesions or microtumors on postoperative MRI, so feature extraction was not performed on postoperative CE-T1 tumor residual lesions. The dataset’s consistency was confirmed with an ICC > 0.8 based on temporalis muscle area, tumor imaging characteristics, and temporalis thickness. The ICC values for ROIs delineated by different radiologists are provided in Supplementary Table S2.

Radiomics Feature Selection

During the feature selection phase, radiomic features were extracted from preoperative tumor 3D regions of interest (ROIs) using PyRadiomics. Initially, a total of 949 radiomic features were extracted from each patient’s preoperative T1CE sequence (T1-weighted contrast-enhanced sequence). To reduce dimensionality, 50 meaningful features were selected using mutual information. Subsequently, recursive feature elimination (RFE) was applied to further refine the feature set, ultimately selecting 30 features that were used for machine learning model construction [10].

In addition to the radiomic features, a comprehensive dataset was synthesized, which included preoperative and postoperative temporal muscle imaging measurements, preoperative and postoperative abnormal signal volumes on T2 FLAIR and T1CE sequences, and patient baseline information. To ensure model practicality and simplicity, the final feature set was selected from the training data using RFE, ensuring that only the most relevant features were retained for the predictive model.

Model Development and Validation

Local and public datasets were merged and divided into training and validation sets with an 80:20 ratio. Missing values in direct measurements were imputed using mean values. Feature selection was performed using recursive feature elimination (RFE). Internal validation and hyperparameter tuning of the models were conducted using Optuna [12].

Radiomic features of manually and automatically segmented 3D tumors were then incorporated into the models. A comprehensive model was developed by integrating preoperative and postoperative temporalis muscle and tumor imaging measurements with preoperative tumor MRI radiomics features. Additionally, tumor-only features were used for modeling as a comparative approach. Finally, the best-performing model was selected from multiple machine learning models based on automatic tumor segmentation. The interpretability of the model and the ranking of feature importance were analyzed using the SHAP (SHapley Additive exPlanations) package. The overall workflow for model construction is summarized in Fig. 2.

Fig. 2.

Combining temporal muscle and tumor imaging for early recurrence prediction in glioma. Description: This flowchart illustrates the main steps of the study. ROI delineation: extracting regions of interest (ROI) from preoperative and postoperative imaging (e.g., FLAIR), including tumors and temporalis muscles. Radiomic feature extraction: features were extracted using an automated segmentation model, incorporating imaging-derived measurements and demographic data. Feature selection: important features were selected using methods such as mutual information and recursive feature elimination (RFE). Model training and evaluation: a predictive model was constructed, with ROC curves generated and a correlation heatmap created to interpret feature relationships. Model interpretation: feature importance (e.g., SHAP values) was analyzed to aid understanding of the model and provide clinical insights

Statistical Analysis

For continuous variables, the Wilcoxon rank-sum test is used when the p-value is less than 0.001, which is suitable for data that do not meet the normality assumption. Conversely, when the p-value is equal to or greater than 0.001, the t-test is employed, provided that the data satisfy the normality assumption. For categorical variables, the chi-square test is the preferred method. Mean imputation was performed for missing values in imaging direct measurements, with the proportion of missing values kept below 10%. Samples with missing radiomic data were excluded. The dataset without missing values was then standardized. Drawings and all analyses were performed using Python (v.3.10), R (v.4.2.1), PowerPoint, and XMind. A two-tailed p-value < 0.05 was considered statistically significant.

Results

Table 1 summarizes the clinical and imaging characteristics of 71 patients with high-grade gliomas, stratified by recurrence status within 180 days. In the non-recurrent group (N = 49), IDH wild-type tumors were observed in 85.7% (N = 42) of patients, while IDH-mutant tumors accounted for 14.3% (N = 7). Similarly, WHO grade 4 tumors were predominant, comprising 87.8% (N = 43), whereas WHO grade 3 tumors were identified in 12.2% (N = 6). In the recurrent group (N = 22), the prevalence of IDH wild-type was higher at 95.5% (N = 21), with only 4.55% (N = 1) classified as IDH-mutant. Likewise, the proportion of WHO grade 4 tumors was 95.5% (N = 21), leaving only 4.55% (N = 1) as WHO grade 3. Statistical analysis revealed no significant differences between groups in terms of IDH status (p = 0.420) or WHO grade distribution (p = 0.423). Statistical analysis revealed a significant difference in progression-free survival (PFS) between the non-recurrence and recurrence groups. The non-recurrence group (N = 49) had a mean PFS of 482 ± 311 days, which was significantly longer compared to the recurrence group (N = 22), with a mean PFS of 118 ± 52.0 days (p < 0.001). The mean age of the study population was 59.5 years (± 11.6), with non-recurrent patients averaging 57.9 years (± 12.0) and recurrent patients averaging 63.0 years (± 9.92). The difference approached statistical significance (p = 0.068). The gender distribution was comparable, with males comprising 69.0% of the cohort (p = 0.860). Among clinical parameters, preoperative temporal muscle cross-sectional area (PRE_CSA), preoperative temporal muscle thickness (PRE_TMT), and the first measurement of temporal muscle cross-sectional area within 72 h postoperatively (BL1_CSA) showed no significant differences between groups (p > 0.2). However, wavelet features, such as wavelet.HHH_glcm_Imc2 (p = 0.009) and wavelet.LHH_firstorder_InterquartileRange (p = 0.033), exhibited significant textural and intensity differences, reflecting tumor heterogeneity. Additionally, log.sigma.4.0.mm.3D_glszm_ZonePercentage (p = 0.003) demonstrated significant differences, potentially correlating with tumor invasiveness.

Table 1.

Summary of clinical and imaging characteristics of high-grade gliomas patients stratified by recurrence status within 180 days

	[ALL]	Non-recurrence	Recurrence	p.overall
	N = 71	N = 49	N = 22
Age	59.5 ± 11.6	57.9 ± 12.0	63.0 ± 9.92	0.068
Gender				0.860
Male	49 (69.0%)	33 (67.3%)	16 (72.7%)
Female	22 (31.0%)	16 (32.7%)	6 (27.3%)
PRE_CSA	539 ± 110	541 ± 110	534 ± 114	0.808
PRE_TMT	15.2 ± 2.35	15.1 ± 2.18	15.3 ± 2.76	0.813
BL1_CSA	558 ± 109	560 ± 109	553 ± 112	0.811
BL1_TMT	15.5 ± 2.41	15.2 ± 2.45	16.0 ± 2.30	0.217
IDH:				0.420
0	63 (88.7%)	42 (85.7%)	21 (95.5%)
1	8 (11.3%)	7 (14.3%)	1 (4.55%)
WHO grade				0.423
3	7 (9.86%)	6 (12.2%)	1 (4.55%)
4	64 (90.1%)	43 (87.8%)	21 (95.5%)
Preoperative_contrast_enhancing_tumor_volume_mm3	34,468 ± 27,162	33,106 ± 27,016	37,500 ± 27,874	0.539
Postoperative_contrast_enhancing_residual_tumor_mm3	633 ± 2884	759 ± 3441	354 ± 732	0.436
Preoperative_T2_FLAIR_abnormality_mm3	90,411 ± 52,400	94,451 ± 54,739	81,414 ± 46,700	0.308
Postoperative_T2_FLAIR_abnormality_mm3	36,957 ± 28,406	37,119 ± 27,908	36,597 ± 30,153	0.945
wavelet.HHL_glrlm_RunVariance	1.92 ± 0.46	1.87 ± 0.51	2.02 ± 0.29	0.112
wavelet.HHL_firstorder_Median	− 0.02 ± 0.09	− 0.03 ± 0.10	0.00 ± 0.02	0.072
wavelet.HHH_glszm_SizeZoneNonUniformityNormalized	0.14 ± 0.10	0.14 ± 0.11	0.14 ± 0.07	0.811
wavelet.HHH_glcm_Imc2	0.07 ± 0.02	0.07 ± 0.02	0.06 ± 0.02	0.009
wavelet.LHH_glrlm_ShortRunHighGrayLevelEmphasis	3.33 ± 1.33	3.42 ± 1.43	3.14 ± 1.06	0.352
wavelet.LHH_firstorder_InterquartileRange	4.31 ± 1.36	4.49 ± 1.51	3.89 ± 0.83	0.033
wavelet.LLL_glcm_ClusterProminence	12,145 ± 9953	12,784 ± 10,399	10,722 ± 8943	0.398
wavelet.HLH_glrlm_RunPercentage	0.53 ± 0.04	0.54 ± 0.05	0.52 ± 0.02	0.078
wavelet.LHL_glrlm_RunEntropy	3.56 ± 0.21	3.53 ± 0.23	3.61 ± 0.16	0.082
wavelet.HLH_gldm_LargeDependenceEmphasis	185 ± 27.3	182 ± 31.2	191 ± 14.3	0.116
wavelet.HLH_gldm_SmallDependenceEmphasis	0.01 ± 0.00	0.01 ± 0.00	0.01 ± 0.00	0.023
wavelet.LHH_firstorder_MeanAbsoluteDeviation	3.10 ± 0.92	3.23 ± 1.01	2.83 ± 0.61	0.042
wavelet.HHH_glcm_Id	0.75 ± 0.00	0.75 ± 0.00	0.75 ± 0.00	0.391
wavelet.LHL_glrlm_RunPercentage	0.54 ± 0.06	0.55 ± 0.07	0.52 ± 0.05	0.057
wavelet.HHL_glcm_Imc2	0.19 ± 0.05	0.19 ± 0.05	0.18 ± 0.04	0.183
wavelet.LHH_glcm_ClusterTendency	0.54 ± 0.01	0.54 ± 0.02	0.54 ± 0.01	0.830
wavelet.LHH_glcm_Contrast	0.48 ± 0.02	0.48 ± 0.02	0.47 ± 0.02	0.481
wavelet.HHH_glcm_InverseVariance	0.50 ± 0.01	0.50 ± 0.01	0.50 ± 0.01	0.394
wavelet.LHL_firstorder_Skewness	− 0.57 ± 0.28	− 0.57 ± 0.28	− 0.57 ± 0.27	0.963
wavelet.HLH_gldm_DependenceNonUniformityNormalized	0.09 ± 0.01	0.09 ± 0.01	0.09 ± 0.01	0.610
log.sigma.4.0.mm.3D_glszm_ZonePercentage	0.00 ± 0.00	0.00 ± 0.00	0.00 ± 0.00	0.003
wavelet.HHH_glszm_GrayLevelNonUniformity	5.62 ± 3.49	5.89 ± 3.57	5.01 ± 3.30	0.317
wavelet.HHH_glcm_DifferenceAverage	0.50 ± 0.01	0.50 ± 0.01	0.50 ± 0.01	0.389
wavelet.LLL_firstorder_Range	585 ± 80.3	584 ± 88.9	587 ± 58.5	0.871
wavelet.LLH_glrlm_ShortRunHighGrayLevelEmphasis	12.8 ± 6.21	13.2 ± 5.98	11.9 ± 6.77	0.467
wavelet.LHH_glcm_DifferenceEntropy	0.99 ± 0.02	0.99 ± 0.02	0.99 ± 0.01	0.615
wavelet.HLL_glszm_SizeZoneNonUniformityNormalized	0.14 ± 0.03	0.14 ± 0.03	0.15 ± 0.03	0.617
wavelet.HHL_glcm_Contrast	0.53 ± 0.07	0.53 ± 0.07	0.52 ± 0.05	0.311
wavelet.HHH_glcm_MCC	0.06 ± 0.02	0.06 ± 0.02	0.06 ± 0.02	0.272
progression.free.survival_PFS_days	370 ± 309	482 ± 311	118 ± 52.0	< 0.001

To enhance the extraction of tumor information, we trained an automatic segmentation model to improve radiomic feature acquisition. As shown in Fig. 3, the AttUNet model achieved an IoU of 0.832 on the test set, emerging as the best-performing segmentation model. AttUNet demonstrated robust segmentation performance, facilitating the extraction of radiomic features.

Fig. 3.

The results of automatic tumor segmentation are shown in the figure. The figure shows the results of several automatic tumor segmentation models. AttUnet achieved a good intersection-over-union ratio of 0.83 and was therefore selected as the best segmentation model to replace manual segmentation

We initially developed several predictive models using tumor information extracted through an automatic segmentation model. Subsequently, we constructed additional models by combining temporal muscle information with radiomic features derived from the tumor segmentation model. The performance of these models is presented in Tables 2 and 3, respectively.

Table 2.

Analysis of model performance using tumor information

Model	Accuracy	Precision	Recall	F1 Score	AUC-ROC
AdaBoost	0.67	0.50	0.2	0.28	0.86
CatBoost	0.73	1.00	0.20	0.33	1.00
HistGradientBoosting	0.80	1.00	0.40	0.57	1.00
LightGBM	0.80	1.00	0.40	0.57	0.96
RandomForest	0.80	1.00	0.40	0.57	1.00
XGBoost	0.73	1.00	0.2	0.33	1.00

Table 3.

Analysis of model performance using temporal muscle and tumor information

Model	Accuracy	Precision	Recall	F1 Score	AUC-ROC
AdaBoost	0.8	1.0	0.40	0.57	0.76
CatBoost	0.73	1.00	0.20	0.33	0.72
HistGradientBoosting	0.86	1.00	0.6	0.75	0.92
LightGBM	0.80	0.75	0.60	0.67	0.82
RandomForest	0.67	0.00	0.00	0.00	0.98
XGBoost	0.73	1.00	0.20	0.33	1.00

Table 2 presents the performance analysis of tumor-based models for predicting early glioma recurrence. The table includes metrics such as accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUC-ROC). Among the models, HistGradientBoosting, LightGBM, and RandomForest demonstrated superior performance, each achieving an accuracy of 0.80, a precision of 1.00, a recall of 0.40, and an F1 score of 0.57, with AUC-ROC values of 1.00. CatBoost and XGBoost models performed slightly lower, though their AUC-ROC values remained at 1.00. In contrast, the AdaBoost model showed the lowest recall (0.20) and comparatively weaker overall performance. These results highlight the notable performance variations among different machine learning models when using tumor features to predict early recurrence.

Table 3 presents the performance analysis of various machine learning models, including AdaBoost, CatBoost, HistGradientBoosting, LightGBM, RandomForest, and XGBoost, using a combination of temporal muscle and tumor information. The evaluation metrics include accuracy, precision, recall, F1 score, and AUC-ROC. HistGradientBoosting emerges as the best-performing model, achieving the highest accuracy (0.86), F1 score (0.75), and AUC-ROC (0.92). This model demonstrates strong overall performance in terms of both classification accuracy and model discrimination. AdaBoost also performed well, with an accuracy of 0.80 and an AUC-ROC of 0.76, but its precision (1.0) was much higher than its recall (0.40). RandomForest, despite achieving an impressive AUC-ROC of 0.98, exhibited poor classification performance with an accuracy of only 0.67 and both precision and recall of 0.00. The performance of CatBoost, XGBoost, and LightGBM was similar, with CatBoost and XGBoost achieving a precision of 1.00 but low recall values (0.20).

To evaluate the effectiveness of temporal muscle features in predicting early recurrence, we compared a model incorporating both temporal muscle and tumor features (TM_Tumor_HistGradientBoosting) with a model using only tumor features, as illustrated in Fig. 4. The results indicated that the TM_Tumor_HistGradientBoosting model outperformed others in accuracy, recall, and F1 score, achieving values of 0.89, 0.87, and 0.88, respectively, demonstrating superior predictive performance.

Fig. 4.

Comparison of AUC of 6 machine learning models based on automatic tumor segmentation data

We selected the best predictive model for interpretability analysis, conducted on the validation set. These interpretations provide insights into the mechanisms underlying the model’s predictions of early recurrence in the population. As shown in Fig. 5, this figure presents a SHAP (SHapley Additive exPlanations) summary plot, providing a global interpretation of how each feature influences the model’s prediction of early recurrence across all patients in the study. The x-axis displays the SHAP values, which quantify the direction and magnitude of each feature’s effect on the predicted outcome. Features shifting the model output toward a higher risk of early recurrence have positive SHAP values, whereas those reducing the risk have negative SHAP values.

Fig. 5.

SHAP beeswarm explains the predictive model based on the automatically segmented data. The previous shortcoming of machine learning and deep learning is the lack of interpretability, so we use SHAP to explain the model. The left side of the SHAP honeycomb diagram shows the importance ranking of each feature, and the importance gradually increases from bottom to top. The horizontal axis shows the impact of each feature value on the outcome from low to high, with red representing positive outcomes and blue representing negative outcomes. For example, the smaller the value of PRE_CSA, the more likely the outcome predicted by the model will tend to be early recurrence, which also indirectly confirms the previous research conclusions

The y-axis ranks the features by their overall importance, with the most influential features at the top. Each dot corresponds to an individual patient’s data point, and the color scale (blue to red) represents the actual feature values, from low to high. Thus, the horizontal spread of dots indicates how variation in each feature contributes to the prediction across different patients.

For instance, preoperative temporal muscle cross-sectional area (PRE_CSA) appears near the top, signifying its critical role in predicting early recurrence risk. Several radiomic features, such as wavelet-LLL_glcm_ClusterProminence, also demonstrate wide SHAP value distributions, underscoring their importance in capturing tumor and peritumoral heterogeneity. Overall, this summary plot highlights key clinical and radiomic predictors, offering insight into the mechanisms by which the model identifies patients at higher risk of early recurrence in high-grade glioma.

To further highlight the importance of the selected variables, Fig. 6 shows that preoperative temporal muscle cross-sectional area contributed the most to the model, accounting for 13.9% of the total contribution, while postoperative temporal muscle thickness ranked fifth, contributing 8.1%.

Fig. 6.

SHAP feature importance analysis for early glioma recurrence prediction. The figure demonstrates the results of the SHAP feature importance analysis, which quantifies the contribution of each selected feature to the model’s prediction of early glioma recurrence. Notably, the cross-sectional area of the temporalis muscle before surgery emerged as one of the most influential features, contributing 13.9% to the prediction model. The post-surgical temporalis muscle thickness also displayed substantial importance, contributing 8.1%. These findings underscore the predictive value of temporalis muscle measurements, both pre- and post-surgery, in forecasting early recurrence of gliomas, further enhancing the robustness of the model

The best predictive model also enables personalized predictions for individual patients, illustrating the contribution of each variable to the outcome. As depicted in Fig. 7, it shows SHAP (SHapley Additive exPlanations) interpretation of a single example (patient 1) in a binary classification model. Baseline information of patient 1 included 42 years old, female, IDH mutant, and WHO grade 4. This “force plot” (also called a “waterfall plot”) illustrates how the prediction for a single example in a binary classification model was arrived at. It decomposes the model’s output into the contribution of individual features (SHAP values) and shows how each feature incrementally moves the model’s initial prediction (baseline) toward the final output. Red bars represent features that push the model’s prediction toward the “positive” class (increase the log odds of a positive outcome), while blue bars represent features that push the model’s prediction toward the “negative” class (decrease the log odds of a positive outcome). The horizontal axis represents the log odds of the model’s output, which ranges from approximately − 8 to − 4 and can be interpreted as the model’s overall tendency to predict an example as negative or positive. The further to the right (higher values), the more the prediction tends to be in the positive class; the further to the left (lower values, including negative values), the more the prediction tends to be in the negative class. In the SHAP plot, the series of bars moving from left to right culminate at the far right, f(x) = − 3.321, which is the final log-odds prediction for this sample.

Fig. 7.

SHAP waterfall plot explains the predictive model based on the automatically segmented data. The plot presents the SHAP values for various features contributing to the model’s prediction. The baseline information for patient 1 includes 42 years old, female, IDH mutant, and WHO grade 4. The horizontal axis represents the log odds of the model’s output, ranging from approximately − 8 to − 4, with the final log-odds prediction for this sample being f(x) = − 3.321. Each feature’s contribution is represented by a bar, with red bars indicating features that push the prediction towards the positive class (increase log-odds) and blue bars indicating features that push the prediction towards the negative class (decrease log-odds). The most influential features for this patient are listed on the vertical axis, including wavelet-LLL_glcm_ClusterProminence, wavelet-HLH_gldm_DependenceNonUniformityNormalized, PRE_CSA, and age. The final probability of a positive outcome is approximately 3.4%, calculated by applying the sigmoid function to the log-odds value: p = 1 / (1 + exp(− f(x))). The baseline prediction corresponds to a probability of approximately 0.56%, and the feature contributions drive the prediction towards a higher probability of a positive outcome

This plot represents the personalized prediction explanation for patient 1 in the dataset. The vertical axis lists the most influential features for this sample, sorted by their contribution to the model output. These features include wavelet-LLL_glcm_ClusterProminence, wavelet-HLH_gldm_DependenceNonUniformityNormalized, PRE_CSA, wavelet-LLL_firstorder_Range, age, wavelet-LHH_firstorder_InterquartileRange, BL1_TMT, wavelet-LHH_glrlm_ShortRunHighGrayLevelEmphasis, Postoperative_T2_FLAIR_abnormality_mm3, and a group of “7 other features” grouped together. Each feature shows its value for this sample and the corresponding SHAP value (red/blue bar length). The red bars represent features that push the prediction towards the positive class (increase the log odds), while the blue bars represent features that push the prediction towards the negative class (decrease the log odds). Each red/blue bar represents the SHAP value of a feature, showing its contribution to the final prediction.

For example, wavelet-LLL_glcm_ClusterProminence (eigenvalue, − 0.53) contributes + 3.22, significantly pushing the prediction towards the positive class (increase in log odds by 3.22 units). Wavelet-HLH_gldm_DependenceNonUniformityNormalized (eigenvalue, 1.201) contributes + 2.18, further strengthening the model's preference for the positive class. PRE_CSA (eigenvalue, 0.405) contributes − 2.08, pulling the prediction towards the negative class (decreased log odds by 2.08 units). Wavelet-LLL_firstorder_Range (eigenvalue, 0.858) contributes − 1.98, also pushing the prediction towards the negative class. Age contributes − 1.34, slightly reducing the log odds and thus leading to a negative prediction. Wavelet-LHH_firstorder_InterquartileRange (eigenvalue, − 0.969) contributes + 1.22, promoting a positive prediction. BL1_TMT (eigenvalue, 1.192) contributes + 0.95 (red). Wavelet-LHH_glrlm_ShortRunHighGrayLevelEmphasis (eigenvalue, 0.115) contributes + 0.58. Postoperative_T2_FLAIR_abnormality_mm3 (eigenvalue, − 1.148) contributes + 0.5 (red). The other seven feature groups contribute − 1.38 (blue, aggregated). The cumulative sum of the individual feature contributions (positive and negative) and the baseline output of the model, E[f(x)] = − 5.18, gives a final log odds of f(x) = − 3.321. In binary classification models such as logistic regression or tree-based models, f(x) represents the final log probability of a positive outcome. The formula for log probability is $f\left(x\right)=\text{log}\left(p,\frac{}{},\left(1,-,p\right)\right)$, where p is the probability of a positive prediction. To obtain the probability of a positive outcome, the log probability is transformed using the sigmoid function: $p=1/\left(1,+,e^{\left(-,f,\left(x\right)\right)}\right)$. For this example, f(x) = − 3.321, which gives p ≈ 0.034, meaning that the model predicts a 3.4% probability of a positive outcome for this sample. The baseline log probability E[f(x)] = − 5.18 corresponds to a baseline probability of about 0.56%. After the feature contribution, the final probability for this particular sample increases to 3.49%. Patient 1 did not develop early recurrence in the real-world clinical scenario, confirming the model’s prediction. Finally, we use a heatmap to show the correlation between different features in Fig. 8. The heatmap utilizes a color-coding scheme where blue denotes negative correlations and red signifies positive correlations. The color intensity reflects the strength of the correlation, with 1.00 representing the strongest positive correlation and − 1.00 indicating the strongest negative correlation. The features listed on the axes include various medical imaging metrics, such as “PRE_CSA” and “BL1_TMT.” Each cell in the heatmap represents the correlation between two features, with the values within the cells quantifying the strength of these relationships.

Fig. 8.

SHAP heatmaps explain the predictive model based on the automatically segmented data. The heatmap utilizes a color-coding scheme where blue denotes negative correlations and red signifies positive correlations. The color intensity reflects the strength of the correlation, with 1.00 representing the strongest positive correlation and − 1.00 indicating the strongest negative correlation. The features listed on the axes include various medical imaging metrics, such as “PRE_CSA” and “BL1_TMT.” Each cell in the heatmap represents the correlation between two features, with the values within the cells quantifying the strength of these relationships

Discussion

Identifying early postoperative recurrence is crucial for optimizing treatment decisions. Although maximal safe resection, radiotherapy, and temozolomide chemotherapy can prolong survival [34–36], some patients still experience tumor recurrence within 6 months after surgery [37]. The timing of recurrence is closely associated with prognosis, necessitating individualized management strategies for high-risk patients. Accurate identification of early recurrence risk can enhance clinical decision-making, inform treatment selection, and improve patients’ physical and psychological well-being as well as overall quality of life [38]. Early identification of high-grade glioma recurrence risk is of significant clinical importance. MRI imaging features hold predictive value for recurrence, but there is still a lack of reliable and accurate prediction tools [39–41]. While previous studies have identified factors such as age, sex, and Karnofsky score (KPS) as being associated with recurrence, the predictive accuracy of these indices is limited. Some studies have developed prediction models for glioma recurrence based on clinical information, but these models lack radiomics data [42]. Other studies have focused on automatic MRI segmentation techniques and primarily on survival analysis [43, 44]. Although IDH-mutant gliomas accounted for only a small proportion of our Study population, the majority of patients were classified as WHO grade 4 and IDH wild-type (90.1% and 88.7%, respectively), which is consistent with clinical reality. Moreover, while IDH status and MGMT methylation are critical biomarkers for treatment response [45], they are derived from postoperative pathological analysis. In contrast, our study utilizes preoperative temporal muscle measurements and tumor radiomic features to predict early recurrence. By excluding postoperative factors (such as IDH and MGMT status), our approach mitigates potential biases associated with postoperative information while enabling recurrence prediction within 72 h after surgery, independent of pathological findings. This design not only improved the clinical interpretability of the model, but also provided a more sufficient time window for individualized treatment decisions.

This study further explores the potential of integrating tumor and temporalis muscle data for early recurrence prediction. The inclusion of temporalis muscle measurements significantly enhanced the predictive performance of the model. Consistent with existing literature, we found that preoperative and postoperative MRI features play a key role in predicting recurrence, especially by capturing tumor radiomic feature complexity and heterogeneity [46]. In addition, postoperative temporalis muscle thickness was identified as a key indicator of the patient’s overall health status and was closely associated with the risk of recurrence [42, 47–50]. However, postoperative thickening of the temporalis muscle was associated with a higher risk of recurrence, which may reflect edema resulting from a more complex surgical procedure and longer operation time. This could be related to the greater difficulty in resecting more aggressive tumors, as the imaging was performed within 72 h post-surgery. In addition, we found that the cross-sectional area of the temporalis muscle before surgery was significantly associated with early recurrence. In our study population, age ranked as the tenth most important factor influencing prognosis, generally exhibiting a trend where younger patients had a lower likelihood of recurrence. However, this trend was not absolute and did not follow a linear pattern. In the individualized interpretation of our model, the impact of age on outcomes varied for each patient, reinforcing the non-linear relationship between age and glioma prognosis, as previously reported in the literature [51]. These findings challenge the applicability of current age groupings for gliomas and advocate for the adoption of precision medicine strategies guided by exact age. Several radiomic features derived from the T1CE sequence were highly influential in predicting glioma recurrence. These features capture tumor heterogeneity and complexity, which are known to be associated with more aggressive tumor behavior and recurrence risk [52, 53]. Texture-based radiomic features further underscore their value in assessing tumor tissue heterogeneity, a crucial factor in understanding tumor aggressiveness and recurrence potential. Furthermore, the postoperative T2 FLAIR abnormality volume (Postoperative_T2_FLAIR_abnormality_mm3) was a significant predictor, emphasizing the role of postoperative MRI changes as indicators of tumor progression or complications. This aligns with previous findings that T2 FLAIR abnormalities are correlated with poor prognosis and recurrence [54]. In summary, the combination of radiomic features and clinical parameters, particularly temporal muscle thickness and postoperative changes, provides a robust framework for predicting early recurrence in high-grade gliomas. The model offers valuable insights into the factors influencing glioma progression, supporting its potential clinical application in early recurrence risk prediction.

Limitations and Future Directions

One limitation of our study is that we did not use T2 FLAIR sequences to extract tumor and peritumoral information because the slice thickness in both local and public databases was 5 mm or 6 mm. This may lead to reduced spatial resolution, increased artifacts, volume measurement errors, and inaccurate label information [55, 56]. Although the T2 FLAIR sequence is considered an independent factor in glioma prognosis prediction, its advantage is that it reflects peritumoral lesions, which requires high spatial resolution [57, 58]. Thicker T2 FLAIR sequences affect the extraction of peritumoral information. Therefore, although we outlined the ROI of this sequence in the early stages of the study, we ultimately chose not to include it. However, the models still achieved good predictive performance.

Another limitation is that we used 2D segmentation to quantify temporalis muscle measurements. Future studies should adopt 3D segmentation technology to obtain more comprehensive and accurate temporalis muscle data.

In the future, we hope to collect a larger dataset to perform complete 3D deep learning segmentation to avoid the loss of information.

Conclusion

The integration of preoperative and postoperative temporalis muscle quality metrics with preoperative tumor MRI imaging significantly enhances the accuracy of early recurrence prediction in high-grade gliomas. Furthermore, during postoperative follow-up, the temporalis muscle serves as a sustainable and reliable predictor, offering significant clinical value for early detection of recurrence and supporting timely intervention to optimize patient outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AdaBoost

Adaptive Boosting

AUC

Area under the receiver operating characteristic curve

AttUNet

Attention U-Net (deep learning segmentation model)

BL1_CSA

First postoperative cross-sectional area of the temporal muscle within 72h

BL1_TMT

First postoperative temporal muscle thickness within 72h

CatBoost

Categorical Boosting

CE-T1

Contrast-enhanced T1-weighted imaging

GTR

Gross total resection

HGG

High-grade glioma

HistGradientBoosting

Histogram-based Gradient Boosting

ICC

Intraclass correlation coefficient

IDH

Isocitrate dehydrogenase

IOU

Intersection over Union

KPS

Karnofsky Performance Status

LightGBM

Light Gradient Boosting Machine

ML

Machine learning

MRI

Magnetic resonance imaging

NTR

Near total resection

PACS

Picture archiving and communication system

PFS

Progression-free survival

PRE_CSA

Preoperative cross-sectional area of the temporal muscle

PRE_TMT

Preoperative temporal muscle thickness

Postoperative_contrast_enhancing_residual_tumor_mm3

Volume of contrast-enhancing residual tumor post-surgery in cubic millimeters

Preoperative_contrast_enhancing_tumor_volume_mm3

Volume of contrast-enhancing tumor pre-surgery in cubic millimeters

Preoperative_T2_FLAIR_abnormality_mm3

Volume of T2 FLAIR abnormality pre-surgery in cubic millimeters

Postoperative_T2_FLAIR_abnormality_mm3

Volume of T2 FLAIR abnormality post-surgery in cubic millimeters

RFE

Recursive feature elimination

RHUH-GBM

The Río Hortega University Hospital Glioblastoma dataset

ROI

Region of interest

SHAP

SHapley Additive exPlanations

TCIA

The Cancer Imaging Archive

T1CE

T1 contrast-enhanced imaging

T1WI

T1-weighted imaging

T2 FLAIR

T2 fluid attenuated inversion recovery

TransUNet

Transformer U-Net (segmentation model)

Unet

U-Net (deep learning segmentation model)

UnetPlusPlus

U-Net + + (improved version of U-N)

XGBoost

EXtreme Gradient Boosting

WHO

World Health Organization

Author Contribution

Qianni Zhu and Xiaocong Hu co-designed the research, conducted data analysis, built models, For Review Only, developed the interpretable framework, and contributed equally to this work. Qihui Ye contributed to study design, data extraction, and writing. Chunxiang Wu and Xuewen Dong provided clinical insights and assisted in data integration and supported data preprocessing. Weihua Li and Fan Lin handled proofreading and provided feedback.

Funding

This work was supported by the Shenzhen Science and Technology Innovation Commission, Basic Research Program (Key Project) [Grant No. JCYJ20200109120205924] for the project titled “Study on Photothermal Tumor Vaccine.” Additionally, it received support from the Shenzhen University curriculum reform initiative under the project titled “Study on Enhancing the Curriculum for Master of Medicine Programs Availability of data and material.”

For any additional inquiries regarding the data and materials used in this study, please contact the corresponding author.

Data Availability

This study utilizes imaging and clinical data from Shenzhen Second People's Hospital. Researchers with relevant study needs may request access by contacting the corresponding author with a description of their intended use. Additionally, a portion of the data was obtained from the RHUH public dataset, which can be accessed by readers through the referenced link, subject to the dataset's access policies.

Declarations

Statistics and Biometry

No complex statistical methods were necessary for this paper.

Informed Consent

The use of local data was approved by the institutional review board, with informed consent waived. In contrast, TCIA, being a publicly accessible database without patient identifiers, did not require institutional review board approval.

Ethical Approval

The use of local data was approved by the institutional review board, with informed consent waived. In contrast, TCIA, being a publicly accessible database without patient identifiers, did not require institutional review board approval.

Study Subjects or Cohorts Overlap

The RHUH dataset has already been reported in the following publication:

The authors recommended this paper as the best source of additional information about the dataset: Cepeda, S., García-García, S., Arrese, I., Herrero, F., Escudero, T., Zamora, T., & Sarabia, R. (2023). The Río Hortega University Hospital Glioblastoma dataset: A comprehensive collection of preoperative, early postoperative, and recurrence MRI scans (RHUH-GBM). In Data in Brief (Vol. 50, p. 109617). Elsevier BV. 10.1016/j.dib.2023.109617.

Methodology

Methodology: retrospective.

Competing Interest

The authors declare that they have no conflicts of interest related to this study. There are no financial relationships with any organizations that could have an interest in the submitted work.