Machine Learning Points to Chemoradiotherapy Benefit for Older Head and Neck Cancer Patients

For older patients diagnosed with advanced head and neck cancer, one of the most consequential questions in oncology has long gone unanswered: does the addition of chemotherapy to radiotherapy actually improve survival, or does the toxicity of aggressive treatment outweigh any benefit in a population often excluded from clinical trials? A new machine learning study published in the journal Medical Oncology offers one of the most data-driven answers to date, and its conclusion is striking. Among patients aged 65 and older with stage III-IV head and neck squamous cell carcinoma who were not treated surgically, concurrent chemoradiotherapy was associated with a roughly 30 percent reduction in the hazard of dying from the disease compared with radiotherapy alone.

The research team, led by Rasheed Omobolaji Alabi of the University of Helsinki together with collaborators at the University of Vaasa, Uppsala University, and Karolinska Institute, drew on 1,525 patient records from the Surveillance, Epidemiology, and End Results (SEER) Program, the National Cancer Institute’s vast population-based registry covering a substantial fraction of the United States. All included patients had histologically confirmed squamous cell carcinoma arising in the lip and oral cavity, oropharynx, hypopharynx, or larynx, had not undergone surgery, and had received either definitive radiotherapy alone or concurrent chemoradiotherapy. The cohort was overwhelmingly male, with 1,221 men and 304 women, and the mean age at diagnosis was 73.1 years, ranging up to 90.

What distinguishes this analysis from previous comparisons of the two treatment modalities is its methodological architecture. The researchers combined classical survival statistics with machine learning techniques in a way that had not previously been applied to this specific clinical question. Kaplan-Meier curves and multivariable Cox proportional hazards regression provided the traditional backbone, while an extreme gradient boosting model, trained alongside logistic regression and easy ensemble algorithms in Azure Machine Learning Studio and Jupyter notebooks, was used to probe how each patient characteristic contributed to disease-specific survival. The model was validated with five-fold cross-validation, a technique that repeatedly splits the data into training and testing subsets to guard against overfitting in the absence of an external validation cohort.

The survival results were unambiguous. Kaplan-Meier curves for the two treatment groups separated early and remained distinct throughout follow-up, which ranged from zero to 167 months, with the log-rank test yielding a p-value below 0.001. In the Cox regression, chemoradiotherapy carried a hazard ratio of 0.699, with a 95 percent confidence interval of 0.598 to 0.818, making it the strongest protective factor in the entire model. Conversely, the presence of distant metastasis emerged as the most powerful adverse prognostic factor, with a hazard ratio of 2.520, while hypopharyngeal tumor site and increasing age also independently predicted worse disease-specific survival.

The machine learning layer of the study added a different kind of insight. The extreme gradient boosting model achieved a performance accuracy of 55.5 percent in predicting disease-specific survival, with an area under the receiver operating characteristics curve of 0.57, figures the authors themselves acknowledge as modest. But predictive accuracy was never the primary goal. Instead, the team deployed permutation feature importance, a model-agnostic ranking method that works by randomly shuffling each input variable one at a time and measuring how much the model’s performance degrades. Features whose shuffling causes the largest drop in accuracy are deemed the most influential. In this analysis, the four most important features for survival were age, metastatic stage, overall TNM stage, and, critically, the treatment option itself.

To make the model’s inner workings transparent, the researchers also applied SHapley Additive exPlanations, or SHAP, a technique borrowed from cooperative game theory that assigns each predictor a quantified contribution to individual predictions. The SHAP beeswarm plot confirmed the global picture: higher age, advanced stage, and distant metastasis pushed predicted mortality risk upward, while treatment with chemoradiotherapy contributed substantially toward improved survival predictions, and being married was also associated with better outcomes. A local SHAP waterfall plot for a single illustrative patient showed how treatment status and tumor grade were the strongest contributors to reducing that individual’s predicted risk, demonstrating the kind of patient-specific interpretability that clinicians increasingly demand from artificial intelligence tools.

The propensity-score matching analysis added an important layer of caution. Using logistic regression to estimate each patient’s propensity for receiving chemoradiotherapy, and then matching patients one-to-one by nearest-neighbor matching without replacement, the team sought to reduce the treatment-selection bias that plagues any non-randomized comparison. The analysis revealed substantial differences in nearly all covariates between the two treatment groups, with patients receiving chemoradiotherapy generally exhibiting higher propensity scores and more concentrated distributions. In other words, healthier or differently profiled patients were systematically more likely to receive the combined treatment, a reality that means the observed survival advantage, while consistent across every analytical method, cannot be interpreted as definitively causal.

The clinical context makes these findings consequential. Head and neck cancer is among the ten most common malignancies worldwide, with an estimated 54,000 new cases and about 11,230 deaths in the United States in 2022, and between 25 and 40 percent of patients are diagnosed at age 70 or older, most already at an advanced stage. Yet older patients remain underrepresented in clinical trials, leaving clinicians with a dearth of evidence about whether they can tolerate aggressive treatment. Guidelines from the National Comprehensive Cancer Network, the American Society of Clinical Oncology, and the European Society for Medical Oncology recommend concurrent chemoradiotherapy for advanced disease, but the evidence base for older patients specifically has been thin, and conflicting reports have appeared, with some studies favoring radiotherapy alone. The new machine learning analysis aligns with the randomized trial literature suggesting that chemoradiotherapy yields better disease-specific survival, and it echoes the authors’ argument that treatment decisions should rest on patient-related factors rather than chronological age alone.

The study also situates itself within a broader technological transformation of head and neck oncology. The authors discuss how FDG-PET/CT-based radiotherapy planning can refine staging and enable biologically informed dose escalation to metabolically active tumor regions, how Internet of Things devices enable continuous remote monitoring of microvascular free-flap reconstructions after surgery, and how three-dimensionally printed patient-specific surgical plates have measurably improved the accuracy of bony reconstruction. These developments, they argue, illustrate how data-driven and digitally enabled medicine is reshaping every stage of cancer care, from imaging and treatment planning to postoperative surveillance.

The limitations of the study are substantial and the authors are candid about them. The SEER database lacks information on HPV status, a key prognostic factor in oropharyngeal cancer; on comorbidities and frailty, both of which strongly influence survival and treatment tolerance in older patients; on the specific chemotherapy regimen, its dose, and its completeness; and on the type and intent of radiotherapy. The retrospective design introduces selection bias, and the uneven distribution of tumor sites, dominated by laryngeal and oral cavity cancers, may skew the analysis. The modest accuracy of the machine learning model means it cannot yet be used to predict survival for new patients under treatment consideration, and the authors emphasize that the findings should be regarded as exploratory. Still, the convergence of Kaplan-Meier curves, Cox regression, permutation feature importance, and SHAP analysis on the same conclusion gives the result unusual internal consistency. The team calls for multicenter randomized controlled trials and age-specific prospective studies, with separate analyses for different tumor sites, to confirm the benefit. Until then, the study offers older patients and their physicians something they have rarely had: large-scale, algorithmically interrogated evidence that age alone should not disqualify anyone from the most effective nonsurgical treatment available.

Subject of Research: Machine learning comparison of chemoradiotherapy versus radiotherapy alone for disease-specific survival in older patients with advanced-stage, non-surgically treated head and neck squamous cell carcinoma

Article Title: Effectiveness of radiotherapy or chemoradiotherapy in older patients with advanced-stage, non-surgically treated head and neck cancer: A machine learning analysis

Article References: Alabi, R. O., Almangush, A., Elmusrati, M., Tiblom Ehrsson, Y., Laurell, G., & Mäkitie, A. A. (2026). Effectiveness of radiotherapy or chemoradiotherapy in older patients with advanced-stage, non-surgically treated head and neck cancer: A machine learning analysis. Medical Oncology, 43(11), Article 315. https://doi.org/10.1007/s12032-026-03420-5

Image Credits: AI Generated

DOI: 10.1007/s12032-026-03420-5

Keywords: head and neck cancer, chemoradiotherapy, radiotherapy, machine learning, SEER database, disease-specific survival, elderly patients, propensity score matching, SHAP, permutation feature importance, Cox proportional hazards, HPV status