Health

Computer Model Predicts Radiation Toxicity Before it Occurs

CHICAGO — A new approach using machine learning has accurately predicted two of the most challenging side effects associated with receiving radiation therapy for head and neck cancer — significant weight loss and need for feeding tube placement.

There are many possible implications for its use, explained lead author Jay Reddy, MD, PhD, an assistant professor of radiation oncology at The University of Texas MD Anderson Cancer Center in Houston.

“We may be approached by a patient at our clinic with head and neck cancer and the patient gets dispositioned to a course of definitive radiation,” he said. “The patient factors can be put into the machine learning model and that would enable personalized predictions for that patient. We can then use these to aid in decision making in the clinic.”

For example, Reddy noted, it can help with questions that include: “Should we place a feeding tube prior to radiation therapy, or in the middle, or afterwards? How aggressive should we be in nutritional supplementation? Or should we wait and monitor more closely.”

“These are just a few of the potential uses,” he said.

Reddy spoke here at the 61st Annual Meeting of the American Society for Radiation Oncology (ASTRO).

May Improve Outcomes

In a discussion of the paper, Sanjay Aneja, MD, an assistant professor of Therapeutic Radiology at Yale School of Medicine, New Haven, Connecticut, said this study is important because it shows a growing “sentiment in healthcare that we should try to leverage all the sources we have to improve outcomes in patients.”

“Companies like Google and Amazon learn about their customers with every click and it’s about time that clinicians leverage that same information to learn more about their patients,” he said.

One of the highlights of this study is that the researchers figured out a way to employ machine learning to improve the quality of care. Most current machine learning program, in contrast, are focused on ways to mimic a physician’s diagnostic ability and to decide treatment courses.

“This study is also unique in that it attempts to attack a problem that is particularly challenging for physicians, and to predict adverse events before they occur,” Aneja said.

Another important highlight is the actual system that was created. “The model that was generated was drawn completely from clinical practice, so it doesn’t require abstractors and a research team to pull variables,” Aneja emphasized. “In theory, this type of system can be employed in a community practice or a large academic center.”

“As we move to quality-based metrics, I think solutions like this will be very important in efforts to understand how we can identify high-risk patients before adverse events occur,” he added.

Predictive of Weight Loss and Feeding Tube

Radiation therapy plays an integral role in head and neck cancer and nearly all patients experience toxicity in some manner. One of these is dysphagia, which can lead to weight loss and placement of a feeding tube. Other times it can lead to hospitalization, rehydration, and the need for nutritional support and pain management, Reddy explained during his presentation.

“When and how to intervene represents a common clinical decision in the management of head and neck cancer,” he said.

In this study, Reddy and colleagues hypothesized that employing a machine learning approach could permit accurate prediction of:

  • unplanned hospitalizations — defined as 3 months from the start of radiation;

  • significant weight loss — defined as more than 10% from baseline, which is the point that usually prompts the placement of a feeding tube; and

  • placement of a feeding tube.

Data was merged from an internal web-based charting tool (known as Brocade), the electronic health record (Epic), and the record/verify system (Mosaiq) to develop predictive models of these toxicities.

The study ran from May 2016 to August 2018, during which period 2121 consecutive radiation therapy courses for head and neck cancer patients were administered across five practice sites at MD Anderson. For each course, more than 700 clinical and treatment variables were collected, including demographics, tumor characteristics, prior treatment, and details about radiation treatment.

The courses of radiation were divided into two groups. The first was a training set, and the three models — random forest, gradient boosted decision trees, and logistic regression — were trained on 1896 radiation therapy courses to predict for each outcome.

The best performing model was then evaluated on an independent validation set comprising 225 consecutive courses of radiation therapy, and models that reached an area under the curve (AUC) of more than 0.70 were considered clinically valid.

The cohort was predominantly male (75.2%), with a median age of 62 years. Patients received a median radiation dose of 60 Gy (interquartile range [IQR], 30 – 69.3).

The most prevalent primary disease sites were the oropharynx (35%), oral cavity (14.8%), and salivary gland (6.1%).

The analysis showed that the incidence and AUC of the best-performing model for each outcome in the training set showed: 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)].

“These outcomes were similar in the validation cohort,” Reddy explained.

Random forest, the best performing model for unplanned hospitalization, did not rise to the threshold. In the validation set, the incidence was 14.2% with an AUC of 0.64. The other two measures, however, did meet the prespecified threshold of clinical validity. The incidence of feeding tube placement was 23.1% with an AUC of 0.755, and significant weight loss occurred in 14.2% with an AUC of 0.75.

“The application of three machine learning models to a structured dataset enabled the development of predictive models for acute radiation toxicities for head and neck cancer patients,” Reddy concluded.

“This study demonstrates the feasibility of employing precision oncology to predict acute radiation toxicities and it may facilitate the identification of patients for whom early intervention is warranted,” he added.

“Enormous Burden” for Patients

Approached for comment, Thomas J. Galloway, MD, associate professor, Department of Radiation Oncology, Fox Chase Cancer Center, Philadelphia, Pennsylvania, said that acute radiation toxicity is “an enormous burden for head and neck cancer patients, and since every patient is unique, the onset, severity, and duration of these toxicities is variable.”

This new study suggests that using a machine learning approach to predict acute radiation toxicity may allow these toxicities to be managed proactively, he commented.

“This is an exciting abstract and, as with many machine learning initiatives, there is a real possibility that the accuracy of the model will improve with increased training,” he said.

“However it is not known if this represents an improvement over physician evaluation,” he continued. “Neither feeding tube placement nor 10% body weight loss happen overnight. These are issues that typically develop over weeks and can — or should — be appreciated by an astute physician.”

“Determining what, if any, advantage this new technique provides for physician management of patients receiving head and neck radiation will be eagerly anticipated,” Galloway added.

Reddy previously received travel expenses from VisionRT. Galloway disclosed a relationship with Varian speaker bureau and UpToDate royalties.

61st Annual Meeting of the American Society for Radiation Oncology (ASTRO): Abstract 141. Presented September 17, 2019.

For more Medscape Oncology news, join us on Facebook and Twitter.




Source link

Tags
Back to top button
close
Thanks !

Thanks for sharing this, you are awesome !

Pin It on Pinterest

Share This

Share this post with your friends!

Close