Machine Learning Finds Mixed Dietary Signals in Older Adults’ Predicted Skin Cancer

An interpretable machine-learning analysis of two National Health and Nutrition Examination Survey cycles found that dietary measures contributed to predictions of prevalent, self-reported non-melanoma skin cancer among adults aged 60 and older, but the signals were mixed rather than uniformly protective or harmful. The final model combined demographic, lifestyle and dietary variables, including individual flavonoids, caffeine, zinc and niacin. Age was its strongest predictor, while the dietary contributions varied by compound. These findings describe associations captured by a model; they do not show that eating, avoiding or supplementing with any nutrient causes or prevents skin cancer.

The researchers analyzed data from the 2007–2010 and 2017–2018 NHANES cycles, which together contained 29,940 participants. These nonconsecutive periods were selected because they included both the cancer questionnaire information required to identify the outcome and detailed dietary data. After excluding people younger than 60, those without information on non-melanoma skin cancer status and those with incomplete dietary or covariate data, the final sample contained 3,729 older adults. Of these, 3,607 reported no non-melanoma skin cancer and 122 reported having received such a diagnosis. The outcome came from a Medical Conditions questionnaire item and was not confirmed by histopathology; the questionnaire also supplied no information for distinguishing basal cell carcinoma, squamous cell carcinoma or other histological subtypes.

The groups differed in several demographic, socioeconomic, metabolic and dietary characteristics. People in the non-melanoma skin cancer group were older on average than controls, 73.57 versus 69.45 years, and 94.3% were non-Hispanic White, compared with 48.6% of controls. The skin-cancer group also included larger proportions with education above high school, a poverty income ratio above 3 and current alcohol consumption. Its mean Oxidative Balance Score was higher and its mean Dietary Inflammatory Index lower. Average reported intakes of vitamin E, zinc, calcium, magnesium, total anthocyanidins and several individual flavonoids were also higher. These are measured group differences, not effects of the nutrients: people already diagnosed with skin cancer may have changed their diets, and the cross-sectional survey cannot establish which came first.

Dietary exposure was estimated from two 24-hour dietary recall interviews conducted three to 10 days apart, and average daily intake was calculated across both recalls. The candidate variables included macronutrients, micronutrients, caffeine, flavonoid subclasses and four composite indices: the Composite Dietary Antioxidant Index, Oxidative Balance Score, Healthy Eating Index-2015 and Dietary Inflammatory Index. Missing values were imputed with missForest, a nonparametric iterative method based on random forests that accommodates mixed-type data. The researchers first compared logistic regression, least absolute shrinkage and selection operator, or LASSO, and random forest for preliminary feature selection. They then ranked predictors with random-forest importance and removed variables with pairwise absolute Pearson correlations above 0.70. Six correlated variables—total flavonols, copper, total flavonoids, vitamin B12, vitamin B6 and thiamin—were excluded, leaving 24 relatively independent features. Synthetic minority oversampling, or SMOTE, was applied only to the training set to address the marked imbalance between cases and controls; the held-out test set received no oversampling, undersampling or other resampling.

Twelve algorithms were evaluated on the untouched test set: logistic regression, support vector machine, random forest, XGBoost, LightGBM, multilayer perceptron, kernel nearest neighbors, Gaussian naïve Bayes, extra trees, adaptive boosting, gradient-boosting decision tree and linear discriminant analysis. The researchers assessed area under the receiver operating characteristic curve, 95% confidence intervals, accuracy, precision, sensitivity, specificity, balanced accuracy, the Matthews correlation coefficient, F1 score, calibration and decision-curve analysis. Linear discriminant analysis and logistic regression tied for the highest AUC at 0.742, with 95% confidence intervals of 0.666–0.815 and 0.668–0.815, respectively. Linear discriminant analysis was selected for the final screening-oriented model because its sensitivity was 0.730, balanced accuracy was 0.703 and Matthews correlation coefficient was 0.153. Its accuracy was 0.677, precision was 0.071 and specificity was 0.676.

The contrast between discrimination and raw accuracy was striking. Extra trees and random forest achieved overall accuracies above 0.93, yet their sensitivities were only 0.081 and 0.135, indicating that they missed most non-melanoma skin cancer cases in the test set. Their high accuracy largely reflected the overwhelming number of controls. Linear discriminant analysis instead identified 73% of cases while maintaining 67.6% specificity, producing a better balance between case detection and false positives. Its precision remained low at 0.071, consistent with the low prevalence of the outcome in the sample. Calibration curves suggested that its predicted probabilities were generally consistent with observed outcomes, while decision-curve analysis indicated greater net benefit than the other models across a broad range of threshold probabilities. The model’s moderate discrimination and very low precision also mean that its outputs should not be interpreted as diagnostic evidence.

SHapley Additive Explanations, or SHAP, was used to show how each feature contributed to the linear discriminant analysis predictions. Age was the dominant influence, with a mean absolute SHAP value of 0.71; greater age increased predicted risk, and its contribution became more pronounced around age 70. Among dietary variables, the most important were luteolin at 0.49, total flavones at 0.44, caffeine at 0.36, zinc at 0.32 and niacin at 0.29. Higher modeled values of luteolin, caffeine and zinc generally pushed predictions upward, while higher total flavones, quercetin and kaempferol generally pushed them downward. Niacin was predominantly associated with a negative model contribution, although the detailed analysis also identified possible interactions among nutrients, including an apparent inverse pattern involving selenium and niacin. At the individual level, one representative high-risk prediction of 0.998 was driven most strongly upward by caffeine, with a SHAP value of +4.66, and alcohol, at +4.19, while kaempferol, at −1.50, and niacin, at −1.06, contributed most strongly in the opposite direction.

The apparently counterintuitive signals may reflect the complexity of diet, aging and cumulative sun exposure rather than harmful effects of antioxidant-rich foods. Zinc supports antioxidant enzymes and immune function, but the researchers noted that excessive intake may impair copper-dependent redox balance through competitive inhibition of intestinal copper absorption. In the dataset, people with non-melanoma skin cancer reported mean zinc intake of 11.29 milligrams per day, compared with 9.99 milligrams per day among controls. Caffeine was also linked to higher predicted risk in the model, contrary to some earlier observational reports. Luteolin has shown anti-inflammatory and anti-aging activity in experimental systems, but concentrations used in laboratory studies may be much higher than those achieved from food, while metabolism and bioavailability limit systemic exposure. Conversely, the inverse model contributions from quercetin and kaempferol accord with findings in experimental skin systems involving UV-associated signaling, inflammation and carcinogenesis. Those mechanisms supply hypotheses, not proof that ordinary dietary intake protects against human skin cancer.

The study also demonstrates why accuracy can be misleading in an imbalanced medical prediction problem. With 122 cases among 3,729 participants, a model that labels nearly everyone as a control can appear highly accurate while being poor at identifying rare cases. The investigators therefore emphasized sensitivity, AUC, balanced accuracy and the Matthews correlation coefficient alongside calibration and clinical utility. SHAP explanations make the selected model easier to inspect, but they do not transform associative prediction into causal evidence. A feature can influence a prediction because it marks an unmeasured factor—such as socioeconomic circumstances, outdoor work, sun exposure or behavior after diagnosis—rather than because it independently produces disease. The ranking of caffeine and individual flavonoids as risk-increasing model signals likewise does not contradict the broader literature by establishing adverse dietary effects; it identifies associations that require prospective testing.

Several limitations constrain the findings. The cross-sectional design precludes causal inference and is vulnerable to reverse causation, including dietary changes made after a skin-cancer diagnosis. Self-reported cancer status may contain recall or reporting errors and cannot support clinical diagnosis or subtype distinction. The survey cycles may not reflect current diets, supplement use, sun-protection practices or dermatologic screening. Crucially, the available data lacked key UV-related variables, including lifetime sun exposure, occupational and recreational exposure, sunscreen use and related behaviors. Age, race, socioeconomic position, smoking, alcohol use, body mass index and other measured characteristics could not fully compensate for those omissions. Larger longitudinal studies across more diverse populations, incorporating UV exposure, genetic information and biomarkers, would be needed to test whether the identified dietary patterns persist and whether they have causal or clinical significance. For now, the analysis is best understood as a map of subtle associations in an observational dataset, not a dietary prescription for skin-cancer prevention.

Subject of Research: Interpretable machine learning of dietary antioxidant profiles and self-reported non-melanoma skin cancer in older adults

Article Title: From Nutritional Epidemiology to Clinical Prediction: Interpretable Machine Learning Reveals Subtle Associations Between Dietary Antioxidant Profiles and Non‐Melanoma Skin Cancer in Older Adults

Article References: Yu, B. H., Kang, J. Y., & Kim, K. E. (2026). From Nutritional Epidemiology to Clinical Prediction: Interpretable Machine Learning Reveals Subtle Associations Between Dietary Antioxidant Profiles and Non‐Melanoma Skin Cancer in Older Adults. Food Science & Nutrition, 14(10), Article e72458. https://doi.org/10.1002/fsn3.72458

Image Credits: AI Generated

DOI: 10.1002/fsn3.72458

Keywords: NHANES, non-melanoma skin cancer, machine learning, dietary antioxidants, older adults, flavonoids, SHAP analysis, nutritional epidemiology, risk prediction, Nutritional, Epidemiology, Clinical