A Multimodal Machine Learning Model Combining Dosimetric and Radiomic Shape Features for Predicting Moderate‑to‑Severe Xerostomia After Radiation Therapy for Nasopharyngeal Carcinoma

Abstract

Purpose

Radiation-induced xerostomia remains a frequent and debilitating complication in patients with nasopharyngeal carcinoma undergoing radiation therapy, severely impacting quality of life. Accurate prediction of xerostomia severity is crucial for personalized treatment planning. This study aimed to develop and validate a machine learning model that integrates dosimetric parameters with radiomic shape features from planning computed tomography images to predict moderate‑to‑severe xerostomia.

Methods and Materials

In this retrospective study, we included 100 patients with nasopharyngeal carcinoma treated with intensity modulated radiation therapy. Radiomic shape features were extracted from the overlap volume between the planning target volume and the left parotid gland. These were combined with standard dosimetric parameters (eg, V20 Gy and V30 Gy). A rigorous correlation-based feature selection was implemented to address multicollinearity. A random forest classifier was developed to predict xerostomia grade (grade 1 vs ≥grade 2), and its stability was rigorously assessed through 100 iterations of random train-test splitting. Model performance was evaluated using the area under the receiver operating characteristic curve, accuracy, and confusion matrix analysis.

Results

The model demonstrated moderate performance with modest variability in predicting moderate‑to‑severe xerostomia, achieving a mean area under the receiver operating characteristic curve of 0.78 (SD, 0.09) and an accuracy of 0.71 (SD, 0.09) across 100 random train‑test splits. Feature selection stability analysis identified 6 consistently important predictors: “Elongation,” “Maximum2DDiameterSlice,” “V20,” “Flatness,” “LeastAxisLength,” and “MajorAxisLength.” Radiomic shape features predominated among the most frequently selected and most influential parameters. Robust importance quantification further identified “V30,” “MeshVolume,” and “Sphericity” as the top contributors to the model’s predictions.

Conclusions

This study suggests a robust machine learning framework for predicting moderate‑to‑severe xerostomia, indicating the complementary value of radiomic shape features to traditional dosimetric metrics. The integration of these features may contribute to more comprehensive toxicity risk assessment and warrants further investigation in clinical decision-making contexts.

Introduction

Nasopharyngeal carcinoma (NPC) is endemic in southern China and Southeast Asia, with incidence rates 20 to 50 times higher than in Western countries and a 2.5:1 male predominance.1, 2, 3, 4, 5, 6 In recent decades, NPC incidence and mortality have declined because of advances in diagnostic technologies and individualized chemoradiation therapy.7, 8, 9, 10, 11

Because of the complex anatomic structure of the nasopharynx and the radiosensitivity of NPC, radiation therapy remains the primary treatment modality for nonmetastatic NPC.¹² Intensity modulated radiation therapy (IMRT) enables precise dose delivery to tumor targets while sparing adjacent normal tissues such as the parotid glands.^13,14 Nevertheless, radiation-induced xerostomia remains a frequent and challenging complication. Literature indicates that approximately 60% of patients with NPC treated with IMRT experience xerostomia, with around 30% developing late-stage and clinically significant symptoms.^15,16 Persistent xerostomia can severely impact patients’ quality of life, affecting speech, swallowing, nutritional status, and overall well-being.¹⁷

The accurate prediction of xerostomia risk is crucial for optimizing radiation therapy planning and implementing personalized protective strategies. Traditional approaches have primarily relied on dosimetric parameters, with previous studies demonstrating that when the mean dose to a single parotid gland is <20 Gy, or the mean dose to both glands is <25 Gy, the risk of xerostomia can be effectively reduced.¹⁷ However, these dosimetric criteria alone often provide incomplete prediction because patient-specific anatomic factors and tissue radiosensitivity may significantly influence toxicity development.

In recent years, machine learning (ML) algorithms have shown increasing promise in radiation oncology, particularly in predicting normal tissue complications.^18,19 Concurrently, the emergence of radiomics has provided a rich source of quantitative imaging features that can enhance predictive modeling.^20,21 The integration of radiomics with ML offers the potential to develop more accurate and individualized toxicity prediction models by capturing complex patterns in medical images that may not be apparent to the human eye.

Previous research in xerostomia prediction has primarily focused on dosimetric parameters and clinical factors. However, several recent studies have demonstrated the potential of radiomics features from the parotid gland. Van Dijk et al²² showed that computed tomography (CT)–based textural features, magnetic resonance imaging–derived intensity features related to fat concentration,²³ and positron emission tomography–derived metabolic features²⁴ can significantly improve prediction of late xerostomia beyond traditional dosimetric models. Despite these advances, the integration of radiomic shape characteristics derived from medical images remains limited. Furthermore, many existing models suffer from issues of overfitting and limited generalizability because of inadequate feature selection methods and insufficient validation of model stability.²⁵

Therefore, this study aims to develop a robust ML framework for predicting moderate‑to‑severe xerostomia in patients with NPC by integrating dosimetric parameters with radiomic shape features derived from planning CT images. We implement a rigorous correlation-based feature selection approach to identify the most predictive features and employ comprehensive stability assessment through repeated random sampling. The primary objectives are to (1) identify key predictors of xerostomia severity; (2) develop a stable classification model for predicting moderate‑to‑severe xerostomia; and (3) evaluate the relative importance of radiomic shape versus dosimetric features in toxicity risk assessment. This work may serve as an initial step toward developing a tool to support clinical decision-making.

Methods and Materials

Patient selection

This retrospective study included 100 consecutive patients with NPC who received definitive radiation therapy between 2022 and 2024. The inclusion criteria were as follows: (1) histologically confirmed diagnosis of NPC; (2) no evidence of distant metastasis at the time of diagnosis; (3) treatment with definitive IMRT; (4) age ≥14 years; (5) Karnofsky performance status score ≥70; and (6) primary tumor predominantly located on the left side (ie, left-dominant laterality).

Patients were excluded if they: (1) had other active malignancies (excluding cured basal cell carcinoma of the skin); or (2) were unable to complete radiation therapy or chemotherapy because of serious comorbidities or a poor general condition.

Patient demographics

Demographic and clinical characteristics of the 100 patients with NPC included in this study are summarized below. The sex distribution was significantly skewed, with male patients (75 cases) outnumbering female patients (25 cases) by a ratio of 3:1. The average age was 58.0 years (SD, 12.4 years). The average height and weight were 166.6 cm (SD, 8.2 cm) and 65.8 kg (SD, 10.0 kg), respectively.

Clinical staging was recorded according to the American Joint Committee on Cancer (AJCC) eighth edition. Among the included patients, 11% were stage II, 67% were stage III, and 22% were stage IV.

All selected patients had left-dominant tumors, defined as primary lesions predominantly located on the left side of the nasopharynx or showing clear left-sided asymmetry on imaging. This selection criterion was applied to control for confounding effects of tumor laterality on parotid gland dose distribution. In NPC, tumor location substantially influences the spatial relationship between the target volume and the ipsilateral parotid gland, thereby affecting both dosimetric parameters and radiomic shape features. By restricting the cohort to a homogeneous group with consistent tumor‑gland spatial relationships, we aimed to minimize laterality‑related confounding, isolate the true predictive signal of the extracted features, and enhance internal validity during this exploratory, hypothesis‑generating phase. Details are provided in Table 1.

Table 1.

Baseline characteristics of the study population

Variable	Value (n or mean ± SD)
sex (male/female)	75/25
Age (y)	58.0 ± 12.4
Height (cm)	166.6 ± 8.2
Weight (kg)	65.8 ± 10.0
Clinical stage	II: 11 (11%); III: 67 (67%); IV: 22 (22%)
Tumor laterality	All left-dominant

Xerostomia assessment

To evaluate the clinical relevance of the radiomic and dosimetric features used in our predictive model, xerostomia severity was retrospectively assessed in the cohort of 100 patients. Xerostomia grades were extracted from medical records at a median of 3.1 months (range, 2.5-3.8 months) after radiation therapy and were based on the Radiation Therapy Oncology Group toxicity criteria.²⁶ The severity was scored by experienced clinicians during routine follow-up. Among the evaluated cohort, the distribution of xerostomia grades was as follows: 53 patients (53%) experienced grade 1 xerostomia, 42 patients (42%) had grade 2, and 5 patients (5%) developed grade 3 toxicity.

Target delineation and expansion

In this study, the delineation of target volumes was performed on contrast-enhanced CT images by experienced radiation oncologists, following national guidelines and institutional protocols. The gross tumor volumes (GTVs) included the primary nasopharyngeal tumor (GTVnx) and metastatic cervical lymph nodes (GTVnd).

Clinical target volumes (CTVs) were defined using a risk-adapted approach based on our institutional practice. For the purpose of subsequent analysis, the CTVs were categorized in a manner consistent with the stratified framework described by Wang et al.²⁷ Specifically, CTV1 encompassed GTVnx with a 5-mm margin, the entire nasopharyngeal mucosa, and adjacent high-risk anatomic structures (comparable to level 1 and part of level 2 regions). A prescribed dose of 66 Gy was delivered to this volume, which is 10% higher than the 60 Gy recommended by Wang et al,²⁷ reflecting our institutional preference for enhanced local control in high-risk regions.

CTV2 included CTV1 and extended to cover medium- and low-risk areas (mainly level 2 to level 4), with a prescription dose of 60 Gy. Additionally, a low-risk subregion within CTV2, corresponding primarily to the level 4 area, was labeled as CTV3 (prophylactic irradiation area) and received 54 Gy. Although CTV3 is not formally defined in national guidelines, this subdivision was prospectively delineated as part of our institutional protocol and was used for the stratified analysis in this retrospective study.

All target volumes, including GTVnx and GTVnd, were uniformly expanded by 3 mm to generate the corresponding planning target volumes: planning gross tumor volume of the nasopharynx (PGTVnx), planning gross tumor volume of the neck nodes (PGTVnd), PTV1, PTV2, and PTV3. The prescribed doses to PGTVnx, PGTVnd, PTV1, PTV2, and PTV3 were 70 Gy, 68 Gy, 66 Gy, 60 Gy, and 54 Gy, respectively, delivered in 32 fractions.

Radiation therapy plan design

Using volumetric-modulated arc therapy and the Monaco 5.51 treatment planning system, physicists designed individualized radiation therapy plans for each patient. All plans were reviewed and approved by clinical oncologists, meeting the clinical treatment requirements.

Data preparation

The region of interest for feature extraction was defined as the spatial overlap between PTV2 and the left parotid gland, hereafter referred to as PTV2 + PL (Fig. 1). This overlap volume was generated using the Margins module in Monaco (version 5.51, Elekta). The corresponding CT images, along with the PTV2 + PL structure, were exported from Monaco in DICOM format and subsequently imported into 3D Slicer (version 5.6.2, Surgical Planning Laboratory, Brigham and Women's Hospital)²⁸ for radiomic analysis.

Figure 1.

Representative axial computed tomography image of a left‑dominant patient with nasopharyngeal carcinoma showing target volumes and parotid glands. The following color‑coded contours are displayed: PGTVnx (yellow), PTV1 (green), PTV2 (blue), left parotid gland (mint green), right parotid gland (brick red), and the PTV2 + PL overlap volume (red). The figure clearly illustrates the spatial relationship between PTV2 and the left parotid gland, with their overlap volume (PTV2 + PL) highlighted in red.

Abbreviations: GTV = gross tumor volume; PTV = planning target volume; PGTVnx = planning gross tumor volume of the nasopharynx.

The overlap volume between PTV2 and the left parotid gland (PTV2 + PL) was chosen to isolate the parotid subregion receiving intermediate‑to‑high dose irradiation, which is hypothesized to be more functionally relevant for xerostomia development than the whole gland. This region is directly exposed to the prescribed dose and may better reflect radiation‑induced geometric deformation and damage. This rationale is supported by prior studies demonstrating that the parotid-PTV overlap volume is a key determinant of parotid dose and xerostomia risk.29, 30, 31

Radiomic feature extraction was performed using the Radiomics module in 3D Slicer, which uses PyRadiomics (version 3.1.0, Dana-Farber Cancer Institute), an open-source platform developed in accordance with the Image Biomarker Standardisation Initiative guidelines. Consequently, the calculation of all morphologic features—including MeshVolume, SurfaceArea, Sphericity, Elongation, and Flatness—follows the standardized definitions and reference values established by the Image Biomarker Standardisation Initiative, ensuring transparency and reproducibility. Within the extraction customization panel, we selected both the shape (3-dimensional) and 2-dimensional feature classes to comprehensively capture shape-based characteristics.

We deliberately focused on shape features only because our primary hypothesis concerned the geometric configuration of the parotid‑PTV overlap volume and because shape features are preprocessing‑independent, which enhances reproducibility in this exploratory study. All CT images were acquired using a Philips CT Big Bore scanner with consistent institutional protocols, ensuring high geometric consistency across the data set. Because only shape-based radiomic features were extracted—which are inherently invariant to intensity normalization, gray-level discretization, and voxel resampling—no additional preprocessing (eg, resampling to isotropic voxels, discretization, or interpolation) was applied. PyRadiomics internally accounts for voxel dimensions when computing physical shape metrics (eg, volume in mm³ and axis lengths in mm), making shape features comparable across patients despite minor variations in native voxel size.

The extracted radiomic shape features included Elongation, Flatness, LeastAxisLength, MajorAxisLength, MeshVolume (representing the volume of the PTV2 + PL region), Sphericity, and other geometric descriptors from both 2-dimensional and 3-dimensional perspectives, yielding a total of 14 radiomic shape features.

In addition to radiomic shape features, dosimetric and volumetric parameters were incorporated based on previous studies indicating that combining radiomic features with dosimetric data enhances predictive model performance in radiation therapy.³² Therefore, 4 additional dosimetric features were extracted to complement the shape-based characteristics: (1) V20 of the left parotid gland, defined as the percentage of the left gland volume receiving at least 20 Gy; (2) V30 of the left parotid gland, defined as the percentage of the left gland volume receiving at least 30 Gy; (3) V50 of the left parotid gland, defined as the percentage of the left gland volume receiving at least 50 Gy; and (4) the mean dose to the left parotid gland. In total, 18 features (14 radiomic shape and 4 dosimetric) were used for modeling.

Statistical analysis

To ensure model robustness and address feature redundancy, a comprehensive statistical workflow was implemented. First, Spearman correlation coefficients were calculated among all extracted features to identify highly correlated pairs (|r| > 0.8). In cases of strong interfeature correlation, redundant features were systematically removed during each training iteration to mitigate multicollinearity and retain only representative variables for subsequent analysis.

Following feature selection, the remaining features were standardized using z-score normalization to ensure consistent scaling across different measurement units. To prevent data leakage, all feature selection (based on Spearman correlation) and z-score normalization were performed exclusively on the training data in each iteration, and the derived parameters were then applied to the corresponding test set. When 2 features exhibited a correlation coefficient >0.8, the feature with the higher index was removed. No additional statistical threshold was used for preselecting predictors.

The data set was then subjected to 100 iterations of random stratified splitting (80% training, 20% testing) to evaluate model performance stability across different data partitions.

The predictive performance was assessed using the area under the receiver operating characteristic curve (AUC) and accuracy metrics. Calibration of the model was evaluated using calibration slopes and intercepts derived from linear regression of the calibration curve (10 bins) across the 100 iterations. The Brier score was also computed for each test set as the mean squared difference between predicted probabilities and observed outcomes to quantify overall prediction error. Additionally, comprehensive feature importance analysis was conducted to identify the most influential predictors of xerostomia grade classification. Feature importance was quantified using the Gini importance metric from the random forest classifier, which measures the total reduction in node impurity attributable to each feature.

The final importance score for each feature was calculated using the following formula:

$$\text{Robust}\text{Importance}\text{Score}=\frac{1}{N}\times \sum _{j=1}^N{I}_{ij}$$

where N = 100 (total number of iterations), I_ij = importance of feature i in iteration j, and I_ij = 0 when feature i was not selected in iteration j.

This approach penalizes features with inconsistent selection and provides a unified metric reflecting both predictive power and selection stability.

The robust importance scores thus reflect the comprehensive impact of each feature across the entire validation framework, offering enhanced reliability for identifying clinically relevant biomarkers. All statistical analyses and ML implementations were performed using Python (version 3.6.15, Python Software Foundation; Scikit-learn developers) and the Scikit-learn library.³³

Random forest classification model

A random forest classifier was implemented using the selected radiomic and dosimetric features as input variables. The feature set comprised radiomic shape features extracted from the PTV2 + PL overlap region combined with established dosimetric parameters, as detailed in the Data Preparation section.

The model was developed using the Scikit-learn library in Python (version 3.6.15) to predict xerostomia severity grades (grade 1 vs grade ≥2). Random forest was selected for its ability to capture complex, nonlinear relationships between features and outcome, while providing robust performance across different data distributions. Additionally, the ensemble nature of random forest reduces the risk of overfitting and enables feature importance analysis, enhancing model interpretability.

This ML approach supports data-driven decision-making in treatment planning optimization and demonstrates the potential of combining radiomic and dosimetric features for predicting radiation-induced toxicities.

Baseline model for comparison

To isolate the added value of radiomic shape features, we constructed a baseline random forest model using only the 4 dosimetric parameters (V20, V30, V50, and mean dose). All data preprocessing, feature selection (Spearman correlation with ρ > 0.8), standardization, and the 100‑split internal validation framework were kept identical to those used for the full model. The only difference between the baseline and the full model is the feature set.

Model evaluation

Following model development, we employed a robust evaluation approach to assess the performance of the random forest classifier. The data set was subjected to 100 iterations of random stratified splitting, with 80% of the data used for training and 20% reserved for testing in each iteration.

For each iteration, we computed the AUC on the held‑out test set as the primary performance metric, given its robustness to class imbalance and ability to measure overall class discrimination. Additionally, we reported conventional accuracy as a secondary performance metric. The mean AUC and accuracy across all 100 iterations, along with their standard deviations, were calculated to provide a comprehensive assessment of model stability and predictive capability.

Results

Feature correlation analysis

To reduce feature redundancy, we performed correlation-based feature selection within each training iteration. Figure 2 displays the pairwise correlation matrix of the initial 18 candidate features, revealing substantial interfeature correlations among several radiomic shape parameters.

Figure 2.

Spearman correlation heatmap of radiomic shape and dosimetric features.

Feature selection stability

The stability of feature selection across 100 random sampling iterations revealed a clear hierarchy of feature importance for xerostomia grade prediction (Fig. 3). Six features demonstrated perfect selection frequency (100%): “Elongation,” “Maximum2DDiameterSlice,” “V20,” “Flatness,” “LeastAxisLength,” and “MajorAxisLength,” indicating their fundamental and nonredundant value in the prediction model.

Figure 3.

Feature selection frequency across 100 random sampling iterations.

The dosimetric parameter “V30” showed near-perfect retention (95%), followed by radiomic shape features “Sphericity” (90%), “Maximum2DDiameterColumn” (89%), “MeshVolume” (85%), and “MinorAxisLength” (84%). These high-frequency features collectively represent the core set of predictors that consistently contributed to accurate xerostomia grade classification across different data partitions.

In contrast, several features exhibited substantially lower selection frequencies: “SurfaceArea” (15%), “Maximum2DDiameterRow” (8%), “V50” (5%), and “SurfaceVolumeRatio” (1%). The remaining features—“Maximum3DDiameter,” “VoxelVolume,” and “Mean dose - left parotid”—were never selected in any iteration, suggesting that these parameters provided redundant information that was adequately captured by other correlated features in the data set.

Robust feature importance analysis

To complement the feature selection frequency analysis and provide a quantitative assessment of predictor contributions, we calculated robust feature importance scores using the conservative aggregation method across 100 random sampling iterations. The average feature importance is presented in Fig. 4, which displays the 15 features that demonstrated nonzero importance values.

Figure 4.

Robust feature importance scores across 100 random sampling iterations.

The robust importance analysis revealed a hierarchical structure of predictor contributions to xerostomia grade classification. The dosimetric parameter “V30” emerged as the most influential predictor with an average importance score of 0.134, followed closely by radiomic shape features “MeshVolume” (0.120) and “Sphericity” (0.117).

The top 5 features—"V30,” “MeshVolume,” “Sphericity,” “MajorAxisLength” (0.097), and “LeastAxisLength” (0.089)—collectively accounted for 55.7% of the total feature importance. Radiomic shape features demonstrated particularly strong representation, occupying 8 of the top 10 positions in the importance ranking.

The importance scores exhibited a clear descending trend, with a substantial drop observed after the top 11 features. Parameters including “MinorAxisLength” (0.046), “SurfaceArea” (0.014), “V50” (0.009), “Maximum2DDiameterRow” (0.008), and “SurfaceVolumeRatio” (0.002) demonstrated minimal contributions to the prediction model, while 3 features—“VoxelVolume,” “Maximum3DDiameter,” and “Mean dose - left parotid”—received 0 importance scores across all iterations.

Model performance for predicting moderate‑to‑severe xerostomia

To assess the predictive capability of the model, we evaluated its performance across the 100 random train‑test splits. The random forest classifier demonstrated moderate performance in predicting moderate‑to‑severe (≥grade 2) xerostomia (Fig. 5).

Figure 5.

Stability of xerostomia prediction performance across 100 random data splits.

Across 100 random train-test splits, the model achieved a mean AUC of 0.78 (SD, 0.09; 95% CI, 0.58-0.92), a mean accuracy of 0.71 (SD, 0.09; 95% CI, 0.55-0.85), with sensitivity of 0.64 (95% CI, 0.33-0.89) and specificity of 0.77 (95% CI, 0.45-1.00). To assess potential overfitting, we also computed the training set AUC for each of the 100 random splits. For the full model, the mean training AUC was 0.984 (SD, 0.012; 95% CI, 0.962-0.999), while the mean test AUC remained 0.78 (SD, 0.09; 95% CI, 0.58-0.92). To further assess performance under class imbalance, we also evaluated precision and F1 score. The model achieved a mean precision of 0.71 (SD, 0.13; 95% CI, 0.50-1.00) and a mean F1 score of 0.66 (SD, 0.11; 95% CI, 0.43-0.83). The confusion matrix from a representative split (Table 2) illustrates this performance pattern, showing the model’s ability to maintain balanced classification across iterations.

Table 2.

Confusion matrix for xerostomia grade prediction

Actual grade	Predicted grade 1	Predicted grade ≥2
Actual grade 1	8	3
Actual grade ≥2	3	6

Calibration of the model was assessed using calibration slopes and intercepts across the 100 random splits. The mean calibration slope was 0.938 (SD, 0.341), and the mean intercept was 0.023 (SD, 0.166). The mean Brier score across the 100 test sets was 0.195 (SD, 0.037; 95% CI, 0.128-0.271). For comparison, the dosimetric‑only baseline achieved a mean Brier score of 0.253 (SD, 0.051; 95% CI, 0.160-0.343), confirming the improved predictive accuracy of the full model. As shown in Fig. 6, the calibration curve of the full model lies closer to the ideal diagonal than that of the dosimetric‑only baseline.

Figure 6.

Calibration curves of the full model and the dosimetric‑only baseline. The diagonal dashed line represents perfect calibration.

Comparison with a dosimetric‑only baseline model

To quantify the added value of radiomic shape features, we compared the full model against a dosimetric‑only baseline. The full model (dosimetric + radiomic shape features) demonstrated superior predictive performance compared with the dosimetric‑only baseline. Across 100 random train‑test splits, the full model achieved a mean AUC of 0.78 (95% CI, 0.58-0.92) versus 0.66 (95% CI, 0.48-0.86) for the baseline, and a mean accuracy of 0.71 (95% CI, 0.55-0.85) versus 0.61 (95% CI, 0.40-0.80). Sensitivity improved from 0.58 (95% CI, 0.22-0.89) in the baseline model to 0.64 (95% CI, 0.33-0.89) in the full model, and specificity improved from 0.64 (95% CI, 0.32-0.91) to 0.77 (95% CI, 0.45-1.00). Detailed performance metrics for both models are summarized in Table 3.

Table 3.

Comparison of model performance between the dosimetric‑only baseline and the full model (radiomics + dosimetric) across 100 random train‑test splits

Model	Mean AUC (95% CI)	Mean accuracy (95% CI)	Mean sensitivity (95% CI)	Mean specificity (95% CI)
Dosimetric‑only	0.66 (0.48-0.86)	0.61 (0.40-0.80)	0.58 (0.22-0.89)	0.64 (0.32-0.91)
Full model (radiomics + dosimetric)	0.78 (0.58-0.92)	0.71 (0.55-0.85)	0.64 (0.33-0.89)	0.77 (0.45-1.00)

Abbreviation: AUC = area under the receiver operating characteristic curve.

Discussion

In this study, we developed a robust ML framework for predicting moderate‑to‑severe xerostomia by integrating dosimetric parameters with radiomic shape features. Our approach demonstrated key advantages, including dynamic feature selection to address multicollinearity and comprehensive stability assessment through 100 iterations of random sampling. A notable finding was the clear hierarchy of feature importance: 6 features—Elongation, Maximum2DDiameterSlice, V20, Flatness, LeastAxisLength, and MajorAxisLength—demonstrated perfect selection frequency (100%), indicating their fundamental and nonredundant value in prediction. However, selection frequency and predictive importance capture different aspects of feature value. Although these 6 features showed perfect selection consistency, their predictive contributions varied considerably. V30 emerged as the most influential predictor (importance, 0.134) despite slightly lower selection frequency (95%), followed closely by MeshVolume (0.120) and Sphericity (0.117). This pattern suggests that selection frequency reflects stability across data partitions, whereas importance magnitude indicates predictive impact when included.

Our finding that shape features from the PTV2‑parotid overlap region are predictive complements the work of van Dijk et al,²⁴ who demonstrated that pretreatment positron emission tomography–derived metabolic features can significantly improve prediction of late xerostomia. Although their studies focused on textural and intensity features, our results suggest that geometric characteristics extracted from the high‑dose subvolume also carry independent predictive information.

The predictive value of these shape features may be related to 2 potential mechanisms: they could serve as spatial proxies for dose distribution (eg, elongation might reflect portions of the gland extending into lower-dose areas) and might capture intrinsic anatomic susceptibility to radiation damage (eg, gland geometry could correlate with tissue composition). Consistent with the observation that tumor shape influences dose distribution,³⁴ these geometric characteristics may offer predictive information beyond standard dosimetric parameters. However, these mechanistic interpretations remain speculative and require further investigation.

The predominance of radiomic shape features (8 of the top 10 importance rankings) suggests their complementary value to dosimetric metrics, aligning with previous observations that organ geometry influences dose distribution (Sümer et al³⁴). The distributed predictive structure may explain the moderate but reasonably consistent performance, with higher specificity (0.77) than sensitivity (0.64). The moderate AUC (0.78) and relatively low sensitivity may be partly attributable to the limited sample size (n = 100), as reflected by the wide CIs (eg, sensitivity, 0.33-0.89). The higher specificity indicates that the model is more reliable at ruling out moderate‑to‑severe xerostomia, which could be clinically valuable for de‑escalation strategies. Among dosimetric parameters, V30 contributed more than V20, consistent with known dose‑volume thresholds; V50 and mean dose were rarely selected, likely due to redundancy with V30/V20.

Several limitations of our study should be acknowledged. The sample size, although typical for radiomics studies, may have limited model performance and generalizability. The wide 95% confidence intervals observed for all performance metrics—particularly for sensitivity (0.33-0.89)—reflect the inherent variability because of the limited sample size, which necessitated dichotomization of the endpoint. This variability underscores the need for validation in larger, more balanced cohorts to improve model stability. The single-institution design and retrospective nature introduce potential biases that should be addressed in future validation.

We also acknowledge the lack of an external validation cohort. Our sample size (n = 100) is modest relative to the number of input features (n = 18), and although we used rigorous internal validation (100 random train‑test splits), this does not replace independent validation. The substantial discrepancy between the mean training AUC (0.984) and test AUC (0.78) suggests a degree of overfitting, which is not unexpected given the limited sample size. Although we employed correlation‑based feature selection and a random forest classifier designed to mitigate overfitting, this finding underscores the critical need for external validation in larger, independent cohorts. Without external validation, our reported performance metrics (eg, AUC, 0.78; SD, 0.09) may be optimistically biased, and the model’s reliability for clinical decision‑making remains unproven. We currently do not have access to an additional independent data set that meets the same strict inclusion criteria (left‑dominant tumors, consistent IMRT protocol, and identical radiomics pipeline). This is a major limitation that must be addressed by future multi‑institutional studies.

Furthermore, our cohort exclusively comprised patients with left‑dominant tumors to control for laterality confounding. Although this strengthens internal validity, it limits generalizability to patients with right‑sided or midline tumors. Thus, our findings should be considered as proof‑of‑concept in this well‑controlled laterality setting. The single‑institution, retrospective design further compounds this limitation. Additionally, the focus on structural and dosimetric features excludes potential contributions from clinical factors such as patient age, comorbidities, or concurrent chemotherapy, which may further enhance prediction accuracy. Furthermore, we did not compare the predictive value of features extracted from the PTV2 + PL overlap volume with those from the whole parotid gland. We also acknowledge that direct comparison with previously published xerostomia prediction models—such as those by Beetz et al³⁵ and Van den Bosch et al³⁶—was not feasible in the current study. These models were developed in mixed head and neck cancer populations using patient‑reported outcomes at different time points, whereas our cohort consisted exclusively of NPC patients with physician‑rated toxicity at 3 months. Given these substantial differences in study populations, endpoint definitions, and feature sets, direct quantitative comparison with these published models was not feasible. External validation and head‑to‑head comparisons under standardized conditions, therefore, remain essential next steps to position our model within the context of existing approaches. Furthermore, our analysis focused exclusively on the ipsilateral parotid gland, as the cohort was restricted to left‑dominant tumors to control for laterality confounding. Although this design choice was intentional, it may limit the generalizability of our findings to settings in which contralateral gland dose is an important predictor, and future studies incorporating bilateral dose parameters are warranted.

We also acknowledge that baseline xerostomia data were not available for inclusion in our model because of the retrospective nature of this study. Baseline xerostomia has been established as a strong predictor of posttreatment xerostomia in multiple studies³⁵; its omission may lead to underestimation or overestimation of risk for some patients, as pre‑existing salivary dysfunction is not accounted for. Prospective studies are, therefore, needed to incorporate baseline xerostomia assessments to improve model accuracy and clinical utility.

The absence of grade 0 xerostomia cases in our cohort is likely attributable to the high-dose exposure inherent to nasopharyngeal radiation therapy and the 3-month assessment time point, which falls within the acute toxicity phase. This distribution supports our dichotomization into grade 1 versus ≥grade 2. Because all patients could be unambiguously classified into 1 of these 2 groups, the absence of grade 0 cases does not compromise model validity.

Given the lack of independent validation in the present study, future research must urgently focus on external validation in multi‑institutional cohorts to verify the generalizability of our findings. Because of anatomic symmetry, the same features from the right parotid in right‑dominant tumors are expected to be similarly predictive; we hypothesize comparable performance after feature mapping (eg, image mirroring), which requires validation. For bilateral disease, contralateral dose parameters or a location‑specific model may be needed. These are hypotheses requiring formal testing. Specifically, future studies should include patients with right‑dominant and midline tumors to evaluate whether the model’s predictive performance generalizes across different tumor locations. Future studies should also investigate the role of contralateral parotid dose, particularly when applying the model to patients with right‑dominant or bilateral tumors. Incorporating bilateral dose parameters may further improve predictive performance and generalizability. If necessary, location‑specific models could be developed, or laterality could be incorporated as an additional predictive feature. Integration of clinical parameters with radiomic features may further improve predictive performance. Additionally, exploration of more advanced ML architectures and investigation of temporal changes in feature importance throughout treatment could provide deeper insights into xerostomia pathogenesis. Future studies should also compare the predictive performance of features extracted from the PTV2 + PL overlap volume with those from the whole parotid gland, in order to validate the clinical utility of this subvolume‑based approach.

Conclusions

This study demonstrates the feasibility of ML for predicting moderate‑to‑severe xerostomia using radiomic shape and dosimetric features. The random forest classifier achieved reasonably consistent performance, with a mean AUC of 0.78 (SD, 0.09) and mean accuracy of 0.71 (SD, 0.09). Feature selection analysis revealed 6 consistently selected predictors with perfect selection frequency, indicating their stability, while robust importance quantification identified “V30,” “MeshVolume,” and “Sphericity” as the most influential features based on their predictive contributions.

Notably, radiomic shape features showed predictive value, occupying 8 of the top 10 positions in importance ranking, while also predominating among the most frequently selected parameters. This dual evidence suggests their complementary value to traditional dosimetric metrics in toxicity prediction. The distributed predictive structure observed—with both dose-volume parameters and radiomic shape characteristics contributing substantially—suggests that further investigation into integrated approaches for comprehensive risk assessment is warranted.

The robust feature selection methodology and comprehensive stability assessment may serve as an initial step toward developing clinical decision support tools. These findings suggest the potential of ML approaches to identify patients at risk of moderate‑to‑severe xerostomia, though further prospective studies are needed to assess their clinical utility and impact on patient outcomes.

Disclosures

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.