The answer, according to a team of researchers from China Three Gorges University and Yichang hospitals led by Yixing Li, Dehui Liu, Ruixue Qin, Xingchang Li, and Mei Li, is uncomfortable for anyone hoping that diet holds the key. Their models could explain a moderate share of the between-country variation in thyroid cancer incidence—the best-performing stacking ensemble reached an out-of-fold R² of 0.613 with a rank correlation of 0.901 between predicted and observed rates—but almost all of that predictive power came from socioeconomic indicators, not food. Nine development variables alone, drawn from the Human Development Index and its life expectancy, education, and income sub-indices, achieved an R² of 0.609. A diet-only model built from 140 dietary variables managed just 0.284. And when the researchers added the entire dietary block to the socioeconomic baseline, the improvement was statistically indistinguishable from zero in every model family they tested.

The technical rigor behind that negative result is what makes it credible. The team integrated four open databases—GLOBOCAN 2022 cancer estimates, FAOSTAT Food Balance Sheets covering 113 food items from 2010 to 2022, the UNDP Human Development Report 2025 statistical annex, and WHO obesity prevalence—into a single harmonized dataset spanning 163 countries and 155 variables. Thirteen regression algorithms were evaluated inside a fully nested cross-validation framework: multiple imputation, standardization, four-method consensus feature selection, and hyperparameter tuning were all refitted from scratch within the training partition of every outer fold, so no held-out country ever influenced the model that predicted it. Confidence intervals were obtained by resampling countries with replacement across 2,000 bootstrap replicates, propagating the true sampling variability of the 163 units of analysis rather than the quirks of arbitrary fold boundaries.

The researchers also tested whether their models could generalize beyond geographic similarity. Because neighboring countries resemble one another in diet, registry infrastructure, and screening practice, random cross-validation folds are not truly independent. When entire world regions were held out in turn, the models preserved the rank ordering of countries reasonably well—pooled rank correlation of 0.813—but lost much of their ability to reproduce absolute rates. The models transport the ordering of countries across regions better than their level, which the authors interpret as a sign that part of the apparent accuracy under random validation comes from regional peers in the training set. Cyprus and South Korea, the two highest-incidence countries and both documented examples of screening-driven overdiagnosis, were consistently underpredicted, suggesting that unmeasured diagnostic intensity explains the shortfall better than dietary composition does.

But the study is not purely a story about wealth. When the researchers moved from prediction to confounder-adjusted association analysis, a specific dietary signal did survive scrutiny. Under an extended adjustment set containing the HDI, its three sub-indices, obesity prevalence, and region fixed effects, and after Benjamini–Hochberg correction across all comparisons, five variables retained positive partial correlations with incidence: seafood supply (partial r = 0.196, q = 0.031), dairy supply (0.198, q = 0.031), the iodine-rich food availability index (0.224, q = 0.026), the animal-to-plant protein ratio (0.260, q = 0.011), and mollusks (0.203, q = 0.031). Plant protein, which showed an inverse unadjusted association, did not survive multiplicity correction—a null result the authors state plainly rather than dress up as a protective effect.

The iodine-rich food availability index is the study’s most robust finding, and the team interrogated it with unusual care. The index combines national per-capita supply of seafood, dairy, and eggs—the three food groups identified by WHO/FAO as the principal dietary sources of iodine worldwide—weighted as seafood plus 0.10 times dairy plus 0.05 times eggs. Because any such weighting is an analytical choice rather than a measurement, the researchers rebuilt the composite under four alternative schemes, including equal-weighted, standardized, and iodine-content-weighted variants. The association held across all of them, with extended-adjustment partial correlations ranging from 0.196 to 0.256. Entered separately, seafood and dairy each retained a significant association while eggs did not, despite a comparable unadjusted correlation. And in a specificity test, an iodine-poor animal-origin food index built from bovine, pig, poultry, and mutton and goat meat had the largest unadjusted correlation of any variable examined (ρ = 0.309) but failed to survive adjustment (partial r = 0.102, q = 0.198 under the extended set). The animal-origin foods poorest in iodine are precisely the ones whose association vanishes once confounders are removed.

How much of the seafood–incidence correlation actually tracks national development? A descriptive path decomposition, following the Baron and Kenny framework in a purely ecological sense, partitioned the total association into an HDI-tracking component and a residual direct term. The HDI pathway accounted for 50.2 percent of the seafood–incidence correlation, with a bootstrap interval running from 16 to 90 percent—so the point estimate should be read as roughly half, with wide uncertainty, rather than as a precise partition. The residual component is consistent with the adjusted association analysis, in which seafood retains a positive partial correlation after adjustment for development, obesity, and geography. But the authors make no causal claim about it: the analysis is a cross-sectional ecological quantity, and unmeasured confounders including radiation exposure history, national iodization programs, and genetic susceptibility remain plausible explanations.

The secondary mortality analysis adds an important interpretive check. Thyroid cancer is highly survivable, and its incidence is strongly influenced by diagnostic intensity, while its mortality is comparatively insensitive to the detection of small, indolent, screen-detected tumors. The dietary variables followed the same ordering for mortality as for incidence, so the incidence associations are not a pattern that appears only in a detection-sensitive outcome. But associations with the mortality-to-incidence ratio were close to zero or negative—HDI at ρ = 0.053 and seafood at ρ = 0.065. Since that ratio is comparatively insensitive to detection intensity while incidence is not, this contrast indicates that the ecological correlations operate substantially through case detection rather than through case fatality. The authors are explicit that this is a consistency check on the primary findings, not an independent result.

The study also benchmarked a high-dimensional deep tabular model, DynaTab, against the conventional estimators under the identical nested protocol. DynaTab reached a pooled out-of-fold R² of 0.626, statistically comparable to the stacking ensemble and random forest, with intervals that overlap almost entirely. A purpose-built high-dimensional model therefore matches, rather than overturns, the predictive accuracy achievable on these data—and it does not alter the block-wise finding, which concerns which variable groups carry the between-country signal rather than which model extracts it. In a sensitivity classification analysis, Extra Trees reached an AUC of 0.932 at the median split, though the authors caution that dichotomizing a skewed continuous rate makes the contrast artificially easy and should not be read as evidence of clinical utility.

What the iodine-rich food availability index does and does not measure is the study’s most important caveat. The index is a measure of national food-group availability, not of iodine intake, iodine status, or iodine exposure in any individual. Food balance sheets describe supply rather than consumption and do not capture waste, spoilage, or unequal distribution. The iodine content of a given food group varies substantially with production method, soil and water iodine content, feed composition, and processing—a kilogram of milk in one country is not nutritionally equivalent to a kilogram in another. Most important, the largest single determinant of iodine intake in most countries is not any of these food groups but salt iodization, which FAOSTAT does not measure and which could not be adjusted for at the country level. The authors describe their result throughout as an association between the availability of iodine-rich food groups and thyroid cancer incidence, and never as an association with iodine exposure.

The methodological lesson is as valuable as the dietary hypothesis. In these data, adding 140 dietary variables to nine development indicators produced no improvement in prediction that the bootstrap intervals could distinguish from zero, while the model importance rankings were dominated by the income and life expectancy sub-indices. Ecological machine learning studies that report dietary predictors of cancer burden without an explicit socioeconomic baseline, a leakage-free validation design, and multiplicity control risk attributing to diet what belongs to development. The positive implication concerns the specific hypothesis the study was designed to test: the availability of iodine-dense animal-origin food groups retains a positive association with thyroid cancer incidence after adjustment for development, metabolic status, and geography, after correction for multiple comparisons, and after the construction of the index is varied and tested against an iodine-poor negative control. That is a hypothesis worth taking to designs that can test it—cohorts with validated dietary assessment and measured urinary iodine, direct measures of national thyroid imaging utilization and salt iodization coverage, Mendelian randomization to address confounding by design, and longitudinal panel models in which exposure is observed before outcome. Until such evidence exists, these findings remain ecological correlates and should not be translated into individual dietary advice.

Subject of Research: Ecological machine learning analysis of global thyroid cancer incidence in relation to dietary and socioeconomic factors

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

Article References: Li, Y., Liu, D., Qin, R., Li, X., & Li, M. (2026). Global thyroid cancer incidence in relation to dietary and socioeconomic factors: an ecological machine learning study. Frontiers in Oncology, 16, Article 1842508. https://doi.org/10.3389/fonc.2026.1842508

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1842508

Keywords: thyroid cancer, machine learning, ecological study, GLOBOCAN, FAOSTAT, iodine-rich food availability, seafood, Human Development Index, overdiagnosis, path decomposition, Frontiers in Oncology, cancer epidemiology