Early Prediction of Peg Feeding Tube Placement In Head and Neck Cancer Using Dosiomic Features: Beyond Clinical Factors
Abstract
Purpose
Patients with head and neck (H&N) cancer frequently require percutaneous endoscopic gastrostomy (PEG) tube placement to maintain adequate nutrition during treatment. Current PEG placement decisions are primarily based on patient-specific clinical factors, although dosiomic features derived from radiation dose distributions may provide additional predictive value. This study aims to develop a predictive model for PEG tube placement by integrating dosiomic and clinical features.
Methods
We developed a novel Physics-informed Neural-Network (PNN) decision-making tool considering clinical and dosiomic features from 65 H&N patients who received radiation therapy dose ranging from 60-70Gy in 33-35 fractions. Thirty-two patients underwent PEG tube placement reactively. We incorporated 13 impactful clinical features: Age, Performance status, Weight loss at first OTV, Weight loss at last OTV, Tumor (T-stage), Neck Involvement, Level1 to Level5 (index), Bi-neck (index), Pre-treatment Dysphagia. Additionally, we extracted six dosiomic features from dose-volume histograms: Mean doses to oropharynx, Hypopharynx, Oral Cavity, Cervical Esophagus, Pharyngeal Constrictor, and Parotid Gland combined.
Results
Designed PNN model was evaluated using clinical features, dosiomic features, and their combination. Model based on clinical features alone achieved mean AUC of 0.70±0.02. Incorporation of dosiomic features improved discriminatory performance (mean AUC = 0.75±0.15). Importantly, the combined clinical-dosiomic model demonstrated the highest performance, with mean AUC of 0.78±0.09, sensitivity of 0.85±0.18, and specificity of 0.72±0.22. Results indicate that integrating dosiomic features with routinely available clinical data provides a measurable improvement in predictive performance, yielding a 4-11% increase in AUC compared with models using individual feature groups.
Conclusion
This study demonstrates the added value of dosiomic features in the development of a decision-support tool for PEG tube placement. The integration of dosiomic information with clinical data has the potential to improve patient-specific risk stratification. Nevertheless, further validation using large, multicenter cohorts is required to confirm the robustness and generalizability of the proposed approach.