Application of machine learning to create a prediction model for the need of aggressive nutritional intervention during head & neck cancer radiotherapy.

BACKGROUND

Curative-intent radiotherapy (RT) or chemoradiotherapy (CRT) for head and neck squamous cell carcinoma (HNSCC) frequently leads to mucositis, dysphagia, and odynophagia, resulting in nutritional compromise, weight loss, and the need for feeding tube (FT) insertion. This study aimed to develop and internally validate an artificial neural network (ANN)-based machine learning model to predict patients at risk of requiring reactive FT insertion, >10 % weight loss, or grade ≥3 dysphagia during or after treatment.

METHODS

This prospective, single-institutional observational study included 100 patients with HNSCC treated between August 2022 and June 2024. Clinical and dosimetric data were collected, including prescription planning target volume (prescPTV), total PTV, and mean doses to the oral cavity, superior-middle pharyngeal constrictor (SMPC), inferior pharyngeal constrictor (IPC), cervical esophagus, esophageal inlet muscle (EIM), larynx, pharyngeal mucosa (PM), and base of tongue (BOT). Toxicities were graded using CTCAE v5.0. An "event" was defined as reactive FT placement, >10 % weight loss, or grade ≥3 dysphagia. Binary logistic regression analyses and a multilayer perceptron ANN model were applied, using a 67:33 training-testing split.

RESULTS

Of 100 patients (median age = 54 years; 74 % male), 47 % received concurrent chemotherapy. Grade ≥3 dysphagia or FT placement occurred in 62 % of cases. Mean oral cavity dose >50.3 Gy and concurrent chemotherapy were significantly associated with the event (AUC = 0.646). The ANN model achieved 81.1 % accuracy with an AUC = 0.896. The most influential predictors (normalized importance) were concurrent chemotherapy (91.5 %), mean oral cavity dose (83.3 %), mean BOT dose (79.8 %), mean EIM dose (65.3 %), and mean glottic larynx dose (58.6 %).

CONCLUSION

The ANN-based model demonstrated strong predictive performance for identifying patients at risk of aggressive nutritional intervention during RT/CRT for HNSCC. Incorporating mean oral cavity dose and chemotherapy status into treatment planning may facilitate early nutritional assessment and proactive feeding support. External validation in larger cohorts is warranted.