Global thyroid cancer incidence in relation to dietary and socioeconomic factors: an ecological machine learning study

Background Thyroid cancer incidence has risen dramatically worldwide, with marked geographic heterogeneity suggesting dietary and environmental contributions beyond screening effects. The roles of specific dietary patterns in thyroid cancer remain poorly characterized at the global level. Objectives Age-standardized thyroid cancer incidence was the prespecified primary outcome. We aimed to (1) build machine learning models predicting country-level thyroid cancer incidence from dietary and socioeconomic features, (2) quantify confounder-adjusted dietary associations with incidence, including prespecified sensitivity analyses of the iodine-rich food availability index, and (3) identify key dietary predictors through explainability analyses and a descriptive decomposition of the development pathway. Mortality was analyzed only as a secondary outcome in order to evaluate whether the incidence patterns could be attributed to differential cancer detection rather than to differences in disease occurrence. Methods We integrated GLOBOCAN 2022 (cancer outcomes), FAOSTAT Food Balance Sheets (113 food items, 2010–2022; primary exposure 2015–2019 mean), the UNDP Human Development Report 2025 statistical annex (HDI 2023 and its life expectancy, education, and income sub-indices), and WHO obesity prevalence (2022) across 163 countries, yielding 155 harmonized variables and 150 numeric predictors. All modeling was performed inside a nested cross-validation framework: multiple imputation by chained equations, standardization, four-method consensus feature selection, and hyperparameter search were fitted exclusively on the training partition of each outer fold (10 folds × five repetitions), and performance was summarized from pooled out-of-fold predictions with 95% confidence intervals obtained by resampling countries (2,000 bootstrap replicates). Three analyses separated dietary signal from confounding: block-wise models comparing socioeconomic-only, diet-only and combined feature sets; leave-one-region-out validation across six world regions; and confounder-adjusted partial correlations under a basic (HDI, obesity) and an extended adjustment set (HDI, obesity, the three HDI sub-indices, and region fixed effects), with Benjamini–Hochberg and Bonferroni correction applied across all comparisons. Iodine-related exposure was represented by an iodine-rich food availability index constructed from seafood, dairy, and egg supply; its construction was interrogated in four sensitivity analyses in which the components were entered separately and jointly, four weighting schemes were compared, total animal protein supply was added to the adjustment set, and an iodine-poor animal-origin food index was used as a specificity control. Binary classification and all mortality analyses were secondary. Results Regression on incidence was the primary task. Stacking ensemble performed best, with a pooled out-of-fold R ² of 0.613 (95% CI [0.510, 0.726]), RMSE = 3.04, MAE = 1.82, and Spearman ρ = 0.901 (95% CI [0.862, 0.924]); random forest ( R ² = 0.609), bagging ( R ² = 0.599), and KNN ( R ² = 0.580) followed, and a high-dimensional deep tabular model (DynaTab) was statistically comparable ( R ² = 0.626). Predictive accuracy was, however, almost entirely socioeconomic: nine development variables alone gave R ² = 0.609, 140 dietary variables alone gave 0.284, and the incremental contribution of the dietary block was –0.048 (95% CI −0.169 to +0.034), +0.021, and +0.017 across model families. Under leave-one-region-out validation, the rank ordering of countries was largely preserved (pooled ρ = 0.813), but the level was not. Unadjusted associations with incidence were positive for the iodine-rich food availability index ( ρ = 0.234), total energy ( ρ = 0.230), seafood ( ρ = 0.219), and the animal-to-plant protein ratio ( ρ = 0.221). After adjustment and Benjamini–Hochberg correction, seafood supply, dairy supply, the iodine-rich food availability index, the animal-to-plant protein ratio, and mollusks remained significant under the extended adjustment set (seafood partial r = 0.196, q = 0.031; iodine-rich food availability index 0.224, q = 0.026), whereas plant protein did not (–0.095, q = 0.309). In sensitivity analyses, the index association was carried by seafood (0.196, q = 0.017) and dairy (0.198, q = 0.017) but not eggs (0.065, q = 0.411), was stable across equal-weighted, z-score, and iodine-content-weighted constructions (0.238–0.256), survived further adjustment for total animal protein supply (0.203, q = 0.032), and was not reproduced by an iodine-poor animal-origin food index (0.102, q = 0.198). Path decomposition attributed 50.2% of the seafood–incidence correlation to the HDI pathway (95% bootstrap CI 16%–90%). In the secondary mortality analysis, the dietary ordering was preserved, but associations with the mortality-to-incidence ratio were close to zero (HDI ρ = 0.053; seafood ρ = 0.065), and in the sensitivity classification analysis, Extra Trees reached AUC = 0.932 (95% CI [0.893, 0.964]) at the median split. Conclusions With all model construction confined to training folds, machine learning models built from national food supply and socioeconomic data explain a moderate share of the between-country variation in thyroid cancer incidence, but that predictive accuracy is carried by development indicators rather than by diet. The availability of iodine-rich food groups, seafood supply, and the animal-to-plant protein ratio nevertheless retained positive ecological associations with incidence after adjustment for development and metabolic confounders and after correction for multiple comparisons, and the index association proved robust to how the index was constructed and specific to iodine-rich rather than to animal-origin food groups in general; plant protein showed no association once multiplicity was controlled. Roughly half of the seafood–incidence correlation tracks the development pathway, and the secondary m