A study by Reddy et al investigating the use of a machine-learning model to predict which patients with head and neck cancer being treated with radiation may experience significant weight loss, feeding tube placement, and unplanned hospitalization has found that the model accurately identified the patients most likely to experience significant weight loss and the need for feeding tube placement. The findings could help determine which patients might benefit from early interventions to avoid these treatment side effects. The study was presented at the 61st Annual Meeting of the American Society for Radiation Oncology (ASTRO) and simultaneously published in the International Journal of Radiation Oncology · Biology · Physics.

Study Methodology

The researchers collected data on consecutive radiation therapy courses performed from 2016–2018 on 2,121 patients with head and neck cancer from three sources: electronic health records from Epic, an internal web-based charting tool (Brocade), and the record/verify system Mosaiq. The data included more than 700 clinical and treatment variables for these patients, including demographics, tumor characteristics, prior treatment, and radiation therapy details. The incidence of unplanned hospitalizations, feeding tube placement, and significant weight loss were extracted from Epic.

Random forest, gradient boosted decision trees, and logistic regression models were trained on the initial 1,896 radiation therapy courses to predict for each outcome. Tenfold cross validation was used to select model type and hyperparameters, using area under the receiver operating characteristic curve (AUC) to measure performance. The best-performing model was then evaluated on an independent validation set consisting of the subsequent 225 consecutive radiotherapy courses for head and neck cancer. Models with an AUC > 0.70 were considered clinically valid.

Among the 2,121 patients, 75.2% were male, median age was 62 years old, and median radiotherapy dose was 60 Gy. The most prevalent primary disease sites were the oropharynx (35%), oral cavity (14.8%), and salivary gland (6.1%).

Results

The researchers found that the incidence and AUC of the best-performing model for each outcome in the training set were: unplanned hospitalization (13.2%, 0.676 [random forest]), feeding tube placement (17.8%, 0.787 [boosted decision trees]), and significant weight loss (16.9%, 0.843 [boosted decision trees]). In the validation set, unplanned hospitalization incidence was 14.2%, with an AUC of 0.64. Feeding tube placement occurred in 23.1% of patients with an AUC of 0.755, and significant weight loss occurred in 14.2% of patients with an AUC of 0.751, both of which met the prespecified threshold of clinical validity.

“Application of this machine-learning approach yielded clinical valid models for predicting feeding tube placement and significant weight loss … though the model for unplanned hospitalization did not reach our prespecified clinical validity threshold, this may change with increasing training data. Further refinement of precision oncology approaches could be transformative by identifying patients who may benefit from early intervention to avert significant weight loss and the need for feeding tube placement due to radiotherapy for head and neck cancer,” concluded the study authors.

Clinical Relevance

“Being able to identify which patients are at greatest risk [for side effects] would allow radiation oncologists to take steps to prevent or mitigate these possible side effects,” said lead study author Jay Reddy, MD, PhD, of The University of Texas MD Anderson Cancer Center, in a statement. “If [patients have] an intermediate risk, and they might get through treatment without needing a feeding tube, we could take precautions, such as setting them up with a nutritionist and providing them with nutritional supplements. If we know their risk for feeding tube placement is extremely high—a better than 50% chance they would need one—we could place it ahead of time, so they wouldn’t have to be admitted to the hospital after treatment. We’d know to keep a closer eye on that patient.”

This type of machine-based learning approach may also help radiation oncologists personalize treatment for their patients, according to Dr. Reddy. “Machine learning can make doctors more efficient and treatment safer by reducing the risk of error. It has the potential for influencing all aspects of radiation oncology today—anything where a computer can look at data and recognize a pattern,” he concluded.

Disclosure: For full disclosures of the study authors, visit redjournal.org.