Predicting Feeding Tube Placement in Head and Neck Cancer Patients Receiving Radiation Therapy With Machine Learning

M. Dohopolski, K. Wang, H. E. Morgan, L. Chen, D. J. Sher, J. Wang

Research output: Contribution to journalArticlepeer-review

Abstract

PURPOSE/OBJECTIVE(S): During radiation therapy (RT) for head and neck malignancies, patients often experience weight loss secondary to odynophagia, dysphagia, and dysgeusia, which may necessitate enteral nutrition. Prophylactic feeding tube (FT) placement in every patient is not cost effective, and the procedure is not without complications. Additionally, FT placement has been associated with significant delay in a return to a normal diet. Machine learning models may help facilitate early identification of patients requiring early FT placement, augmenting outcomes and quality of life in these patients. MATERIALS/METHODS: A retrospective cohort of 283 patients from a single institution were included that received definitive or adjuvant RT for head and neck cancer from 2016-2020, and that did not receive > 3 Gy/fraction or FT < 10 days after start of RT. The outcome predicted was FT placement or > 10% weight loss after start of RT if a patient declined FT recommendation. The majority of patients were ECOG 0-1 (51.5%), former/current smokers (67.5%), white (65.7%) and had oropharyngeal/nasopharyngeal disease (60.4%). Definitive chemoRT or RT was received by 91.5% of patients. One hundred twenty-seven patients either had a FT placed (majority) or had > 10% weight loss following the start of RT. Sixty percent of patients were used as training; 20% and 20% were used for validation and testing. Over 50 clinical features were obtained for each patient. Recursive feature analysis was performed to select top performing features in the validation dataset. Hyperparameters were selected for logistic regression (LR), support vector machine (SVM), and explainable boosting machine (EBM) models using Optuna. Separate models were created that used five, six, seven, and eight selected clinical features. RESULTS: The top performing 5-8 feature models for LR, SVM, and EBM are reported using test data as follows. The LR models obtained areas under the receiver operating curve (AUCs) of 0.653-0.685 with the following clinical variables: chemotherapy regimen, ECOG status, RT laterality (unilateral vs bilateral), ethnicity, T stage, smoking status, and pre-RT weight. The SVM models obtained AUCs of 0.530-0.669 with T/N stage, ethnicity, RT laterality, pre-RT pain medications, RT regimen, chemotherapy regimen, primary site, ECOG status, pre-RT weight. The EBM models obtained AUCs of 0.630-0.723 with chemotherapy regimen, T/N stage, RT laterality, RT regimen, gender, age, and ethnicity. CONCLUSION: The explainable EBM models outperformed LR and SVM models on the test set with consistent use of similar features within the 5-8 feature models. Clinical features most commonly utilized in the models were chemotherapy regimen, RT laterality, and ethnicity.

Original languageEnglish (US)
Pages (from-to)e410
JournalInternational journal of radiation oncology, biology, physics
Volume111
Issue number3
DOIs
StatePublished - Nov 1 2021

ASJC Scopus subject areas

  • Radiation
  • Oncology
  • Radiology Nuclear Medicine and imaging
  • Cancer Research

Fingerprint

Dive into the research topics of 'Predicting Feeding Tube Placement in Head and Neck Cancer Patients Receiving Radiation Therapy With Machine Learning'. Together they form a unique fingerprint.

Cite this