Machine Learning May Be Useful in Predicting AEs in Head and Neck Cancer


Meredith Cummings, BSN, RN, OCN, discusses the potential benefit of machine learning in predicting symptoms for patients with head and neck cancer.

Meredith Cummings, BSN, RN, OCN

Meredith Cummings, BSN, RN, OCN

A machine learning approach may prove useful in predicting treatment-related adverse events (AEs) in patients with head and neck cancer, and consequently offer nurses more time to mitigate symptoms, according to Meredith Cummings, BSN, RN, OCN.

“Trying to step in before [symptoms] become hugely problematic could possibly lead to better adherence to therapy and it could mean fewer long-term symptoms, which [are] a problem for these patients,” she told Oncology Nursing News®.

Data presented by Cummings at the 48th Annual Oncology Nursing Society Congress demonstrated that in a 25-article literature review, machine learning successfully predicted treatment-related symptoms in 14 of the articles. According to investigators, this meant that machine learning models can predict symptoms with an accuracy ranging from an area under the curve value of 65 to 85.

The most predicted toxicities in the review included mucositis, radiation induced temporal lobe injuries, and xerostomia. The machine learning methods the review included were a gradient boosting machine, LASSO regression, likelihood fuzz analysis, random forest, neural networks, support vector machine, and k-nearest neighbor.

Some of the other successfully predicted symptoms included dysphagia, and parotid shrinking. In addition, the machine learning also demonstrated the capability to predict surgery complications, radiation induced brain injuries, hospital admissions, and treatment delays.

However, only 2 studies examined used true external validation of their predictive models—in these studies, the machine learning predicted major postoperative AEs and radiation necrosis of the brain.

In an interview with Oncology Nursing News, Cummings, a doctoral student at the University of Pittsburgh School of Nursing, highlighted how machine learning may impact care and the importance of nurses being at the forefront of it.

“Moving forward, equipping people with ways to anticipate symptoms before they happen, whether that’s with an algorithm or looking at genetic or genomic predictors,” she said.

Oncology Nursing News: What methods were used to conduct this research?

Cummings: It was a scoping review [and] we used PRISMA which is a standard used for systematic reviews; there’s a PRISMA subset called the first scoping and it gives a breakdown of what to include [and] how to include things appropriately so that you don’t miss what’s going on. There is a well-recognized PRISMA flow sheet that we also used [in our] our search process [to identify] when things were excluded and why they were excluded. It forces you to break that down so people can see why things were and were not in.

First, we read through titles and abstracts to see if there was a mention of machine learning [or] head and neck cancer, [and whether] they were predicting any type of symptom or treatment-related AE or toxicity from chemotherapy, radiation, or surgery. If all 3 things occurred we included the article in the review.

[Of the] 331 articles populated from PubMed, 25 ended up being included because some were not using machine learning [or were] not predicting symptoms. Most were either predicting doses, response to treatment, type of tumor, or things that are more related to medical predictions as opposed to nursing outcomes. We ended up excluding esophageal cancer because it’s not technically designated as a head and neck cancer even though it is very close. Because of the definition of head and neck cancer, it felt safer and more complete to exclude it.

Most of the articles were about radiation, which makes sense because most patients who have head and neck cancer often receive radiation. We also looked at the breakdown of the machine learning modalities [and] usually, it was using a type of regression. There are several types of regression [such as] lasso and ridge regression—most of these were lasso which is a type of regression that is not too different than linear regression.

In terms of nursing implications and how we could use this clinically, most of the things that were predicted are things that nurses might not be able to specifically modify. [For example], there was some brain injury from radiation: although nurses can’t adjust doses, if we can anticipate patients who are going to have cognitive decline because of radiation induced brain injury, we can help work to develop interventions to mitigate those issues before they occur. That’s where I see where we can use this in the future.

What were the findings overall and did anything surprise you?

There are people who have been using machine learning algorithms in research to predict treatment related symptoms in patients who have received treatment for head and neck cancer. Only 2 of the studies that we reviewed used the gold standard of external validation where they use an external data set to test the model that they have and only 1 was perspective in design so everybody else was looking retrospectively at data.

Something that’s an implication for nursing [is] these models are often predicting 1 symptom. When nurses look at symptoms, we look at things in groups or clusters, so having nurse clinicians and nurse researchers on board when these algorithms are being developed is essential. Then [it is key to] teach them the language, how [to] use these models, what ways can we [use] it to help patients, and [identify whether] the incidence of this 1 symptom related to these other symptoms.

Pulling it back together is essential because knowing this one thing is helpful, but patients don’t ever experience just 1 symptom. A lot of research has indicated they experience 10 to 14 symptoms at a time. Looking at 1 [symptom only] could leave a leave a large gap for these patients, but having nurses and clinicians involved [is critical because] nurses have such a unique lens looking at the patient as a person in terms of what’s going on with them and symptoms.

It was the most surprising that there were only 2 [studies] using external validation, which calls into question how well those models would work when you are to use them on an unseen dataset.

What is the main takeaway from this research?

Nurses play such an integral role in using [and] developing models [so] that we can be active participants at all steps of the development [process] of these tools for our patients. [Nurse input makes these] clinically relevant because we bring such a unique lens. I am by no means a machine learning expert, [but] I’ve learned a lot by doing the scoping review and I would encourage other people to read about it.

My biggest takeaway would probably be that nurses are an important facet of this research, even if we’re not fully in it yet. It could have a high potential to help patients if we can develop things that are clinically relevant and address problems.

Looking forward, what are the greatest challenges in helping address the symptoms for this patient population?

There are several challenges in addressing patients in this population. One is where their disease is: it is a vulnerable tissue area, so receiving radiation to the head and neck leaves you with a possibility of having many AEs and symptoms in that area specifically. Think about it like burning your mouth with hot coffee, you feel that for days after, and that’s not something [as] intense [as] radiation.

Encouraging patients to be forthcoming about what their symptoms are, learning how to best address what their needs are, and asking them about what they’re experiencing [is important]. Sometimes they’re not forthcoming about what’s going on, sometimes we only see them in the clinic once a week or once a month, and frequently checking in with them [would make a difference].

Moving forward what are the next steps for this research?

We’ve continued to work at identifying what the experiences are of this patient population. Fortunately, treatment has moved quickly and they’re having better outcomes, [but] that means that they’re having longer term symptoms because the treatment is intense.

Those are all things that nurses participate in [a lot], so even if you’re not a researcher, anticipating what patient is at risk for this is something we talk about all the time [with] how do we know who’s at risk for XYZ? That’s where nurses fit in saying, ‘I know this could happen so what can we do to mitigate that before it happens?’


Cummings M, Nilsen M, Bender C, Al-Zaiti A. Predicting anti-cancer treatment-related symptoms in patients with head and neck cancer using a machine learning approach: A scoping review. Poster presented at: 48th Annual Oncology Nursing Society Congress; April 26-30, 2023; San Antonio, TX. Accessed May 25, 2023.

Related Videos
Colleen O’Leary, DNP, RN, AOCNS, EBP-C, LSSYB, in an interview with Oncology Nursing News.
Michelle H. Johann, DNP, RN, PHN, CPAN, WTA, in an interview with Oncology Nursing News explaining surgical path cards
Jessica MacIntyre, DNP, MBA, APRN, NP-C, AOCNP, in an interview with Oncology Nursing News
Andrea Wagner, M.S.N., RN, OCN, in an interview with Oncology Nursing News discussing her abstract on verbal orders for CRS.
John Rodriguez in an interview with Oncology Nursing News discussing his abstract on reducing nurse burnout
Alison Tray, of Hartford Healthcare, discusses her team's research on a multidisciplinary team approach to manage the cancer drug shortage
Cancer-Related Cognitive Impairment
Related Content
© 2024 MJH Life Sciences

All rights reserved.