Surface-based brain age index reflects hematological impacts on cognitive development in children with beta-thalassemia major Authors: Shumin Xu, Zhuoshuo Li, Xinyi Liu, Sixi Liu, Xiaodong Wang, Pengjie Yang, Xuan Cai, Hongwu Zeng, Mengting Liu, Yaowen Li Published: 2026 ABSTRACT SummaryChildren with beta-thalassemia major (β-TM) are at risk of neurodevelopmental or cognitive impairment. In this study, we developed SurfGNN, a surface-based graph neural network model, to estimate brain age from cortical morphological features extracted via structural MRI, to a cross-sectional cohort of 25 β-TM patients and 40 age- and sex-matched healthy controls (ages 6–15 years). A brain age index (BAI)—the difference between predicted and chronological age—was computed and analyzed in relation to cognitive performance (Wechsler Intelligence Scale scores) and hematological parameters. Results showed that β-TM children had significantly lower cognitive scores and more delayed brain age (lower BAI) compared to healthy controls. Furthermore, BAI significantly correlated with hemoglobin and ferritin levels and mediated their association with cognitive performance, as demonstrated by mediation analysis. The SurfGNN model outperformed conventional machine learning baselines and provided interpretable insights into regional cortical alterations. Our findings suggest that BAI may serve as a sensitive biomarker for detecting early brain developmental delays in β-TM.Subject areas: Health sciences, Medicine, Hematology, Neurology, Psychiatry Introduction β-Thalassemia major (β-TM) is among the most severe hereditary hemolytic anemias, affecting approximately 1.5% of the global population as carriers.1 The standard treatment for β-TM involves regular blood transfusions, which, while essential for survival, result in chronic iron overload. This accumulation of iron adversely affects multiple organs, including the heart, liver, pancreas, and notably, the brain.2,3 Recent findings have emphasized the impact of brain iron deposition on cognitive function in β-TM, leading to the recommendation that cognitive evaluations become a routine part of β-TM management.4 The combined effects of chronic transfusions, ineffective erythropoiesis, and increased intestinal iron absorption exacerbate brain iron overload. Human neuroimaging studies link elevated brain iron to deficits in executive function, language/speech, working memory, and visuospatial abilities.5,6,7 These problems can be especially pronounced in school-aged children, where learning and memory difficulties impede cognitive development.7,8 Complementing the human evidence, thalassaemia mouse models show that chronic iron overload impairs neural function via iron-induced blood-brain barrier compromise.9 Traditional evaluations, such as the Wechsler Intelligence Scale for children, fourth edition (WISC-IV),10 are widely used to assess cognitive abilities in pediatric populations11 and have also been applied in studies of β-TM to characterize neurocognitive outcomes.4 The WISC-IV is a standardized assessment, individually administered to measure a child’s overall intellectual functioning and specific cognitive abilities. It includes a variety of subtests that capture a broad range of cognitive strengths and weaknesses, making it especially useful for identifying neurodevelopmental impacts in conditions like β-TM. While these standardized tools provide a measure of cognitive performance, they fail to uncover the underlying neural mechanisms, lack the ability to localize specific brain regions involved, and rely on static, single-time-point assessments that do not capture the progressive nature of β-TM-related cognitive impairments. These limitations underscore the pressing need for advanced technologies capable of providing more nuanced, longitudinal insights into the neural and cognitive impacts of β-TM. Magnetic resonance imaging (MRI) has emerged as an invaluable tool for investigating brain structure and pathology, offering high-resolution insights into the neural correlates of cognitive deficits in β-TM.12,13,14,15,16 In humans with β-TM, iron-sensitive MRI contrasts—such as T2/R2 relaxometry, susceptibility-weighted imaging (SWI), and quantitative susceptibility mapping (QSM)—have revealed signal changes consistent with elevated brain iron, with effects that track systemic iron burden. In some reports, cognitive performance, supporting a framework in which iron dysregulation is accompanied by tissue and microstructural changes rather than being an isolated biochemical phenomenon.3,16 Consistently, preclinical imaging work in a β-thalassemia mice further supports MRI’s sensitivity to brain iron: Yatmarka et al.17 showed that brain T2∗ values were reduced with iron overload and increased after deferiprone chelation, correlating with histological iron measures, thereby validating MRI as a noninvasive readout of brain iron-overload-related tissue alteration in thalassemia. Building on this rationale, structural MRI—particularly cortical surface-based morphometry—provides regionally specific markers of iron-overload-induced structural alteration, such as changes in cortical thickness and surface area that map onto networks subserving executive function, memory, language, and visuospatial processing. Subtle deviations (e.g., cortical thinning or altered folding) are measurable at millimetric resolution and have been linked to attentional and executive deficits in developmental cohorts.18 Accordingly, we focused on morphometric indices as primary endpoints to quantify brain structural deviations in pediatric β-TM. Recently, the advent of brain age prediction techniques has further expanded the utility of structural MRI in understanding cognitive impairments.19,20,21 Brain age refers to an estimated biological age of the brain, derived from neuroimaging features that capture structural and functional maturation or degeneration. The difference between predicted brain age and an individual’s chronological age reflects deviations in neurodevelopment or aging processes, which can be quantified as the brain age index (BAI).22,23 Brain age estimation typically relies on machine-learning models trained on large datasets of healthy individuals,24 enabling the detection of subtle, spatially distributed patterns of cortical and subcortical morphology that may precede overt clinical symptoms.25,26 In β-TM children, brain age prediction may serve as a sensitive biomarker for identifying early deviations in neurodevelopment caused by iron overload and anemia, potentially preceding clinically apparent symptoms. This approach provides a dynamic and individualized measure of brain health, enabling early intervention and tailored-therapeutic strategies. Cortical morphology analysis has highlighted the value of investigating topological features in brain age prediction.27,28 Unlike conventional methods that apply machine learning directly to MR images, surface-based deep learning (DL) emphasizes features critical to neural development, such as cortical folding and volume changes, providing enhanced interpretability.29 Additionally, cortical surface meshes encode valuable topological relationships, capturing spatial adjacencies between neighboring vertices, which can inform more robust brain age predictions.24,28 In this context, the surface graph neural network (SurfGNN) demonstrates a significant methodological suitability.30 Compared with traditional assessments like the WISC-IV, which measure overt behavioral performance, brain age prediction probes underlying neural substrates and can reveal anatomical alterations before they manifest as measurable cognitive deficits. While WISC-IV scores can be influenced by motivation, cultural factors, and situational variability, the BAI reflects structural brain characteristics that are less susceptible to transient influences, making it a powerful and objective complement to conventional neuropsychological testing. In this study, we test the hypothesis that brain age prediction using SurfGNN can provide a sensitive and dynamic biomarker of neurodevelopmental deviation in β-TM patients. Specifically, we propose that chronic anemia and iron overload associated with β-TM accelerate cortical alteration and induce region-specific morphological changes, which contribute to cognitive impairment. By integrating clinical data, including hemoglobin concentration and measures of iron burden, we aim to establish robust correlations between β-TM-related physiological factors and brain morphological alterations. Furthermore, we explore the interpretability of the SurfGNN model to identify key brain regions implicated in cognitive dysfunction, with the goal of elucidating the neural mechanisms underlying these impairments and informing tailored therapeutic interventions. Results Demographic information and clinical variables In total, 27 individuals with β-TM were included in the study, consisting of 15 males with an average age of 9.83 years (range: 6.33–15.08, standard deviation ±2.09 years). For normal control group, 40 individuals were included in the study, consisting of 22 males with an average age of 9.72 years (standard deviation ±1.81 years). Comprehensive demographic and clinical details of these groups are provided in Table 1. Additionally, to provide a more comprehensive clinical profile of the β-TM cohort, we have included detailed patient characteristics in the supplemental information (Table S1), encompassing splenectomy status and time since splenectomy, neurological symptoms involvement, transfusion frequency, and volume per session, as well as the type of iron chelation therapy. Table 1Demographic and clinical characteristics for β-TM and NC subjectsDemographicTMNCp-valuesSubjects (n)2540–Sex: male/female (n)14/1122/18–Age (years, mean ± SD, range)9.83 ± 2.09 (6.33–15.08)9.72 ± 1.81 (6.42–13.08)0.891Education (years, mean ± SD, range)3.68 ± 2.39 (0–9)3.96 ± 2.37 (1–9)0.658Clinical measurementsCognitive test scores (mean ± SD, range)71.57 ± 16.25 (35–106)107.07 ± 11.23 (88–127)<0.0001Red blood cell count (×1012/L)3.71 ± 0.51 (2.85–4.74)N/A–Hemoglobin level (g/L)104.28 ± 13.31 (82–126)N/A–Mean corpuscular volume (fL)80.72 ± 3.75 (70.6–89.3)N/A–Mean corpuscular hemoglobin (pg)27.86 ± 1.19 (25.2–29.8)N/A–Mean corpuscular hemoglobin concentration (g/dL)345.53 ± 9.92 (328–364)N/A–Red cell distribution width-coefficient of variation (RDW-CV)(%)15.08 ± 2.86 (11.4–22.8)N/A–Red cell distribution width–standard deviation (RDW-SD) (fL)42.06 ± 9.21 (13.7–63.1)N/A–Serum iron (μg/dL)36.32 ± 8.79 (23.1–51.3)N/A–Serum ferritin level (ng/mL)2843.7 ± 1,547.5 (227.2–5,467.5)N/A– Demographic and clinical characteristics for β-TM and NC subjects Differences in cognitive test between group Overall, individuals with thalassemia major (TM) exhibited significantly lower WISC-IV scores (p < 0.001) compared to healthy controls, suggesting impaired cognitive function in β-TM patients. Detailed information on the behavioral scores for these groups is presented in Table 1. Blood parameters such as hemoglobin level and serum ferritin level were exclusively measured in β-TM patients only in this study. To enhance clinical interpretability, we further categorized the WISC-IV full-scale IQ scores into standard classification ranges (very superior, superior, high average, average, low average, borderline, and extremely low). As summarized in Table 2, the majority of β-TM children were classified in the extremely low (<70) or border line (70–79) low average (80–89) ranges, with 22 out of 25 participants (88%) scoring below 90. In contrast, most healthy controls clustered in the average (90–109), high average (110–119), or superior (120–129) ranges, with 37 out of 40 participants (92.5%) scoring ≥90. This categorical distribution highlights the clinically meaningful cognitive deficits observed in the β-TM group, extending beyond the continuous score comparisons. Table 2Distribution of WISC-IV full-scale scores by standard classificationClassification (WISC-IV)Score rangeHealthy controls (n = 40)β-TM (n = 25)Very superior (Gifted)≥13000Superior120–12990High average110–11970Average90–109213Low average80–8938Borderline70–7902Extremely low<70013 Distribution of WISC-IV full-scale scores by standard classification Evidence of the harmonization for cortical features In this study, we applied ComBat harmonization, a statistical method commonly used in medical imaging to correct for batch effects. The rationale for performing this harmonization was to minimize site- and scanner-related variability between the in-house cohort and the external Nathan Kline Institute (NKI) dataset, thereby ensuring that cortical features were directly comparable across datasets. This step was essential to reduce potential confounding effects of acquisition differences and to improve the reliability of subsequent SurfGNN training and evaluation. A comparison of cortical features before and after the harmonization can be found in Figure 1. Before harmonization, the distributions of the four cortical features—cortical thickness, curvature, sulcal depth, and mean curvature—varied significantly across datasets (NKI and our dataset), with substantial inter-group differences. After applying ComBat harmonization,31 the distributions of these features became more aligned, showing reduced site-specific variability. Specifically, cortical thickness, which initially displayed skewed distribution, exhibited a more uniform distribution with diminished discrepancies in both central tendency and spread across groups. Similarly, curvature, mean curvature, and sulcal depth measurements all showed certain extent of variability before harmonization, also demonstrated more centralized and comparable distribution across groups after ComBat. These results demonstrate the effectiveness of ComBat in reducing confounding biases and aligning the cortical feature distributions, ensuring that the data were more comparable and reliable across groups. Distribution of cortical feature values across surface vertices before and after ComBat harmonization The x axis represents the cortical feature values e.g., (A) curvature; (B) sulcal depth; (C) cortical thickness; (D) mean curvature), and the y axis represents the number of cortical surface vertices exhibiting that value. To generate these plots, cortical surfaces from all participants within each dataset were averaged and projected onto a standardized template surface with 81,924 vertices. Blue bars indicate the vertex-wise distribution from the external NKI dataset, while red represents the in-house dataset. After harmonization, the distributions of cortical features between the two datasets became more consistent, reducing site-specific variability and improving comparability for subsequent SurfGNN analysis. Performance of SurfGNN-based brain age prediction Table 3 compares SurfGNN with both traditional and DL models for brain-age prediction on the NKI dataset. All models were trained using cortical surface features (cortical thickness, sulcal depth, curvature, and mean curvature) and evaluated with 5-fold cross-validation followed by age-bias correction. Performance was assessed using mean absolute error (MAE), Pearson’s correlation coefficient (PCC), and number of trainable parameters (Params). As shown, SurfGNN achieved the lowest MAE and highest PCC (Figure 2A) while maintaining a moderate parameter size, indicating both superior accuracy and computational efficiency. In contrast, classical methods such as random forest (RF) and support vector machine (SVM) performed less well, and while graph-based models, such as graph convolutional network (GCN),32 graph attention network (GAT),33 and mixture model network (MoNet),29 and advanced architectures, such as spherical convolutional neural network (SphericalCNN),34 and surface vision transformer (SiT),35 improved accuracy, they required substantially larger parameter sizes. These results highlight that explicitly modeling cortical surface topology gives SurfGNN a distinct advantage in brain-age prediction. Table 3Comparison of brain age prediction performance across different regression models using the NKI datasetMethodMAEPCCParamRF1.76 ± 1.630.58 ± 0.0180.06MSVM1.81 ± 1.600.54 ± 0.0160.00MGCN1.75 ± 1.570.59 ± 0.0160.34MGAT1.62 ± 1.410.67 ± 0.0150.34MSphericalCNN1.49 ± 1.210.72 ± 0.0136.82MMoNet1.60 ± 1.360.71 ± 0.0140.24MSiT1.58 ± 1.280.69 ± 0.0127.44MSurfGNN(ours)1.39±1.120.80±0.0124.61MFor all models, cortical surface features (cortical thickness, sulcal depth, curvature, and mean curvature) were used as inputs, and model performance was evaluated with 5-fold cross-validation. SurfGNN achieved the lowest MAE and the highest PCC among all models, while maintaining a moderate parameter size, highlighting its balance of predictive accuracy and computational efficiency.MAE, mean absolute error; PCC, Pearson’s correlation coefficient; Param, number of trainable parameters; RF, random forest; SVM, support vector machine; GCN, graph convolutional network; GAT, graph attention network; SphericalCNN, spherical convolutional neural network; MoNet, mixture model network; SiT, surface vision transformer; M, million. Comparison of brain age prediction performance across different regression models using the NKI dataset For all models, cortical surface features (cortical thickness, sulcal depth, curvature, and mean curvature) were used as inputs, and model performance was evaluated with 5-fold cross-validation. SurfGNN achieved the lowest MAE and the highest PCC among all models, while maintaining a moderate parameter size, highlighting its balance of predictive accuracy and computational efficiency. MAE, mean absolute error; PCC, Pearson’s correlation coefficient; Param, number of trainable parameters; RF, random forest; SVM, support vector machine; GCN, graph convolutional network; GAT, graph attention network; SphericalCNN, spherical convolutional neural network; MoNet, mixture model network; SiT, surface vision transformer; M, million. Performance of SurfGNN-based brain age prediction (A) Scatterplot illustrating PBA predictions using the SurfGNN model on the NKI dataset. (B) Comparative scatterplot showing PBA in our dataset, highlighting differences between the healthy group and the β-TM group. Test experiments on our in-house dataset achieved a MAE of 1.431, PCC = 0.64, p < 0.001 for healthy controls and 1.897, PCC = 0.52, p = 0.002 for β-TM children (Figure 2B). Statistical comparison of the absolution between NKI and healthy controls in our dataset indicates there was no significant difference in absolute errors between the two groups (p = 0.78), indicating the two groups of data were well-aligned. Clinical implications of the BAI To account for the age dependence inherent in the training process, we performed age correction prior to analyzing the BAI. Figure S3 illustrates the relationship between the age-corrected BAI and age. As shown in the figure, the correlation between the corrected BAI and age is minimal. Subsequently, the analysis of the BAI revealed significant differences between the healthy controls (HC) group and the β-TM group, suggesting potential variations in brain aging patterns between the two cohorts, as shown in Figure 3A. Specifically, the HC group demonstrated a higher mean BAI (mean = −0.127, std = 0.798) compared to the β-TM group (mean = −0.624, std = 0.952), with the difference reaching statistical significance (p = 0.015). Brain age index (BAI) and its association with cognitive function in β-TM children (A) Children with β-TM exhibited significantly lower BAI compared to healthy children, indicating impaired brain development in the β-TM group. BAI was corrected for age prior to analysis (see Figure S3). (B) The association between WISC-IV scores and BAI was assessed using a general linear model (GLM) that controlled for age, sex, and intracranial volume (ICV). Under this model, WISC-IV scores were significantly correlated with BAI in children with β-TM, whereas no significant association was observed in healthy controls. This suggests that BAI may serve as a sensitive imaging biomarker for identifying children at risk of cognitive deficits in β-TM. Further correlation analysis, after grouping the samples into two age-based groups, showed consistent findings (see Figures S4 and S5). The relationship between the BAI and cognitive performance, as measured by WISC-IV scores, was evaluated across different groups to understand the potential role of brain aging patterns in cognitive development. The association between WISC-IV scores and BAI was assessed using a general linear model (GLM) that controlled for age, sex, and intracranial volume (ICV). For β-TM-children, association of BAI with WISC-IV scores (Figure 3B) showed that a lower BAI at for children was significantly associated with lower WISC-IV performance (Pearson’s correlation = 0.49; p = 0.009). In contrast, no such association was detected in healthy controls (Pearson’s correlation = −0.09, p = 0.62), suggesting that the typical brain development trajectory in this group is not strongly influenced by BAI in relation to WISC-IV scores. Univariate analyses examining hemoglobin and ferritin levels in relation to WISC-IV scores indicated no direct significant association between these hematological parameters and cognitive performance (all p > 0.37; Figures 4A and 4B), after controlling for age, gender, and brain volume. In contrast, univariate analyses showed that higher hemoglobin levels (Pearson’s correlation = 0.43; corrected p = 0.017) and lower serum ferritin level (Pearson’s correlation = −0.49; corrected p = 0.009) were associated with increased BAI, controlling for age, gender, and brain volume. No other hematological parameter remained significant after adjustment (p > 0.27). The significant findings are visualized in Figures 4C and 4D with model-adjusted regression lines and in-figure statistics. The detailed statistics can be found in supplemental information (Table S2). Hematological factors and brain age index (BAI) in β-TM children: Univariate and mediation analyses Univariate analyses revealed that higher (A) hemoglobin levels and (B) lower ferritin level was linked to increased BAI, as identified by SurfGNN. (C) Mediation analysis suggests that the BAI induced by SurfGNN mediates the impact of hematological abnormalities on WISC-IV scores. In addition, we conducted a mediation analysis controlling for age, gender, and brain volume (Figure 4C). The results revealed that hemoglobin level had a significant positive effect on BAI (a = 0.47, SE = 0.10, p = 0.0061). The results also showed that serum ferritin levels had a significant negative effect on BAI (a = −0.45, SE = 0.09, p = 0.0043), indicating that higher ferritin levels were associated with a lower (more negative) BAI. In turn, BAI significantly predicted the WISC-IV scores (b = 0.73, SE = 0.12, p < 0.001). The direct effect of hemoglobin level (c′ = 0.12, SE = 0.08, p = 0.37) and ferritin level (c′ = 0.02, SE = 0.07, p = 0.78) on WISC-IV scores was not significant, suggesting that the direct relationship between hemoglobin level, ferritin level and WISC-IV scores is weak. Critically, the indirect effect of hemoglobin level (a × b = 0.44, 95% CI [0.32, 0.60]) and ferrintin level (a × b = −0.43, 95% CI [−0.58, −0.29]) on WISC-IV scores through BAI were significant, indicating full mediation. These findings suggest that BAI functions as an intermediate biomarker linking systemic hematological status to cognitive outcomes. In other words, the detrimental impact of anemia and iron overload on cognition appears to operate primarily through their influence on brain development, as captured by the BAI. This underscores the potential clinical utility of BAI as a sensitive imaging-derived marker for identifying children at risk of cognitive impairment due to hematological abnormalities. We acknowledge that sample size in our cohort is relatively limited, and excessive stratification would reduce statistical power and may lead to unstable estimates. For these reasons, we did not place stratified analyses as a primary conclusion in the main text. Nevertheless, to provide additional insights, we have now included the results of age-stratified analyses (6–10 and 11–16 years) here in the supplemental information (Figures S4 and S5; Tables S3–S5), which show consistent findings. Examine the relationship between WISC-IV, BAI, and growth parameters To further evaluate the developmental context of neurocognitive and brain morphological alterations, we examined the relationships among the BAI, cognitive performance (WISC-IV), and anthropometric growth measures (height, weight, and BMI) in terms of pediatric developmental standards. Because height, weight, and BMI were not available for the healthy control group, this analysis was restricted to TM patients. After controlling for age, sex, and ICV, GLM analyses revealed a significantly positive association between BAI and WISC-IV scores (r = 0.484, p_FDR = 0.035), In contrast, BAI showed no significant relationship with height, weight, or BMI (p > 0.239). Likewise, WISC-IV scores were not significantly correlated with any growth parameters (p > 0.240). The detailed statistics can be found in supplementary materials (Table S6). Interpretability Figure 5 presents interpretability scores identifying the cortical features that most strongly influence brain age prediction in both HC and β-TM groups. Based on four cortical features, the heatmap analysis indicates that cortical thickness is the most influential predictor, followed by sulcal depth and Gaussian curvature, while mean curvature contributes minimally. These findings suggest that specific regional morphometric features play a critical role in brain age prediction and highlight distinct structural differences in cortical organization between the two groups. Regionally, consistent patterns are observed across all four cortical feature heatmaps in both HC and β-TM groups. Specifically, cortical changes in the central, temporal, and occipital lobes are strongly associated with brain age increases, with notable contributions from the limbic regions, particularly the inferior temporal cortices. Interpretability maps of SurfGNN and between-group statistical comparisons Left: average node-wise interpretability (saliency) maps for cortical features (cortical thickness, sulcal depth, Gaussian curvature, and mean curvature) across healthy controls (HC) and β-TM children. Warmer colors denote greater local contribution of a feature to SurfGNN’s brain-age prediction. Right: group comparisons of interpretability scores across the cortical surface, performed using the SurfStat toolbox in MATLAB. General linear models (GLMs) were applied, controlling for sex, postmenstrual age at scan, and brain volume. A random field theory (RFT)-based correction was used to adjust p values for multiple comparisons. The T statistics corresponding to the adjusted p values are displayed, with warm colors indicating β-TM > HC and cool colors indicating HC > β-TM. In the comparison maps (right), significant regional differences are observed, with certain regions exhibiting higher or lower interpretability scores in the β-TM group compared to the HC group. For example, cortical thickness in the inferior temporal lobe and cingulate cortex shows higher interpretability scores, as does sulcal depth in the same regions. Additionally, higher interpretability scores for mean curvature are evident in the inferior frontal, temporal, and occipital lobes. These findings suggest that specific regional morphometric features play a critical role in brain age prediction and highlight distinct structural patterns between the two groups. Overall, the interpretability scores in Figure 5 not only validate the biological plausibility of the model’s predictions but also identify cortical regions most vulnerable to β-TM and iron-overload related neurodevelopmental disruption. Discussion In this study, we explored brain age prediction in children with β-TM using an innovative surface-based predictive model, SurfGNN. This model demonstrated exceptional accuracy in predicting brain age by leveraging the organization of cortical surface and cortical features, outperforming all state-of-the-art methods. Additionally, we utilized the BAI to assess deviations in brain developmental trajectories in β-TM children. Our findings indicate that the BAI derived from SurfGNN effectively explains the impact of hematological factors, such as hemoglobin and ferritin levels, on cognitive performance. Cognitive and functional impairments in β-TM In this study, we observed a significant reduction in cognitive abilities among children with β-TM, as assessed using the Wechsler intelligence scales. Importantly, the categorical classification of WISC-IV scores provided additional interpretability by mapping group differences onto clinically recognized ranges of intellectual functioning. The finding that over four-fifths of β-TM children fell below the “average” range (<90), whereas the vast majority of controls scored in the normal-to-above normal ranges, reinforces the substantial cognitive impact of β-TM. This categorical perspective underscores the clinical relevance of our findings and may help clinicians better appreciate the magnitude of impairment in this population. The decline was evident across multiple cognitive domains, including verbal comprehension, working memory, and processing speed, highlighting the pervasive impact of the disease on neurodevelopmental outcomes. The use of standardized intelligence scales provided robust and quantitative evidence of these deficits, emphasizing the need for targeted interventions to address the cognitive challenges faced by this population. The underlying mechanisms contributing to cognitive impairment in β-TM remain inadequately understood and require further investigation.4 While several studies have highlighted potential factors such as chronic anemia,36 iron overload37 due to repeated blood transfusions, and associated oxidative stress,38 the precise pathways through which these factors impact neurodevelopment and cognitive function are yet to be fully elucidated. Anemia-induced hypoxia may impair oxygen delivery to the brain, affecting both neuronal activity and synaptic plasticity.36 Systemic inflammation and altered metabolic profiles often observed in β-TM patients might further exacerbate neurodevelopmental challenges.39 Furthermore, most of the evidence to date suggests that iron overload in specific brain regions may play a critical role in the cognitive impairments observed in individuals with β-TM. Excess iron, resulting from repeated blood transfusions and insufficient chelation, is known to accumulate in various tissues, including the brain, where it can induce oxidative stress, disrupt cellular homeostasis, and cause neuroinflammation.37 Specific brain regions such as the basal ganglia, hippocampus, and prefrontal cortex, which are particularly susceptible to iron-induced damage, are often implicated in cognitive processes.3,40,41 Studies using advanced imaging techniques, such as QSM and R2∗ relaxometry, have revealed increased iron deposition in these regions,16,42,43,44 correlating with cognitive deficits. However, the exact pathways by which iron overload influences neural function and whether it interacts with other factors, such as anemia-induced hypoxia, or systemic inflammation, remain underexplored.45 In this study, we demonstrated that significant morphological alterations occur in the brains of children with β-TM. Our findings provide compelling evidence that these morphological changes are closely associated with hematological factors such as hemoglobin and ferritin levels. Chronic anemia and iron overload, hallmarks of β-TM, may contribute to these structural differences through mechanisms such as hypoxia-induced neurodegeneration, iron toxicity, and systemic inflammation.46,47 Additionally, our results suggest that these alterations are not merely markers of disease pathology but may directly impact cognitive performance, as indicated by the correlations observed between BAI and neuropsychological outcomes. These insights underscore the importance of early and targeted management of hematological abnormalities in β-TM to mitigate their potential impact on brain development and cognitive function. BAI mediates hematological abnormalities and cognitive underperformance This study uses causal mediation analysis to demonstrate the causal effects of hematological factors on cognitive impairments in β-TM patients.21 Low hemoglobin and high ferritin are well-recognized systemic risk factors that may contribute to vascular complications and endocrine-related growth retardation in β-TM. These systemic mechanisms could, in principle, also affect cognitive outcomes. However, our findings highlight that hematological s alone were not directly associated with cognitive performance in this cohort, whereas the BAI were strongly linked to cognition. Importantly, BAI fully mediated the relationship between hemoglobin/ferritin levels and WISC-IV scores, suggesting that systemic abnormalities exert their effect on cognition primarily through their impact on brain development rather than directly via growth delay or vascular factors. Other factors, such as endocrine dysfunction and growth retardation could also contribute to neurodevelopmental deficits in this patient population, and our study lacked the data to fully evaluate these effects. To reflect this limitation, we have tempered our interpretation and avoided attributing cognitive impairment solely to hematological abnormalities. Nevertheless, the inclusion of BAI adds a necessary and complementary dimension, providing a brain-derived biomarker that bridges hematological abnormalities and neurocognitive outcomes and offers mechanistic insights that cannot be obtained from systemic or growth measures alone. In addition to hematological factors, we further explored whether neurocognitive and brain developmental alterations were influenced by general somatic growth in β-TM. Our analyses showed that the BAI and cognitive performance were largely independent of anthropometric parameters, including height, weight, and BMI. While β-TM is often accompanied by growth retardation due to chronic anemia, endocrine dysfunction, and chelation-related metabolic effects; our findings suggest that deviations in brain morphology and cognition reflect distinct neurodevelopmental mechanisms rather than a simple consequence of delayed physical maturation. This dissociation between somatic and neural growth underscores the need to consider brain-specific vulnerability in pediatric β-TM, where systemic hypoxia and iron overload may exert direct neurotoxic effects even in the absence of overt stunting. Longitudinal studies integrating endocrine, metabolic, and neuroimaging assessments will be essential to disentangle the trajectories of physical versus neural development and to guide targeted interventions aimed at preserving cognitive and brain health in this population. Reliability of interpretability The results reveal that cortical thickness emerges as the most significant contributor to brain age estimation during this developmental period. This finding aligns with prior literature demonstrating that cortical thickness undergoes a gradual decrease starting from the childhood period until the mortality, reflecting the natural process of synaptic pruning and continued myelination. Particularly, the cortical thinning process is faster in childhood period than in adulthood and aging period.48,49 The decrease in cortical thickness generally starts in primary sensory and motor regions, such as the occipital and parietal cortices, before progressing to association areas like the prefrontal cortices. These association areas are involved in higher order cognitive functions, and their structural maturation continues into late adolescence and early adulthood. Studies such as those by Sowell et al.50 and Shaw et al.51 have illustrated this sequence of cortical thinning, showing that primary areas responsible for sensory and motor processing mature earlier, while regions associated with executive functions, language, and social cognition mature later.52,53 This regional specificity reflects the hierarchical organization of brain development, with basic functions stabilizing first and complex, integrative functions undergoing prolonged maturation.54 These observations are consistent with our results, where regions in the occipital, motor, and temporal cortices exhibited greater saliency in our interpretation maps. Regarding brain regions most vulnerable to β-TM, our findings reveal that the cortical thickness in the inferior temporal lobe, inferior frontal cortex, and cingulate cortex demonstrates higher interpretability scores. Similarly, sulcal depth in the temporal and cingulate regions also exhibits elevated interpretability scores. Furthermore, mean curvature in the inferior frontal and temporal cortices is associated with increased interpretability scores. β-TM is known to significantly affect various brain regions,9,41 particularly those susceptible to chronic hypoxia, iron overload, and microvascular changes induced by the disease and its treatments. While subcortical regions such as the basal ganglia, cerebellum, and thalamus are the most sensitive,37,55,56 cortical areas, particularly the frontal and temporal cortices, are also frequently implicated.3,46,47 These cortical regions are often compromised due to chronic iron accumulation and oxidative stress, leading to disruptions in cognitive and executive functions. MRI studies, including SWI and T2∗ relaxation time measurements, frequently identify iron deposition, and microstructural abnormalities in these areas.46,47 Furthermore, diffusion tensor imaging (DTI) has demonstrated altered white matter integrity,57,58 underscoring the extensive neurological impact of β-TM. These results emphasize the critical need for strategies to monitor and mitigate neurotoxic effects in affected individuals. Brain age prediction using SurfGNN By capturing age-related morphological alterations, we propose that SurfGNN, a surface-based DL model leveraging cortical morphological features, offers enhanced precision in predicting individual brain age and neurodevelopmental outcomes. Previously, brain age prediction has been achieved using machine learning models based on volumetric images59 or structural connectome metrics.60 However, image-based convolutional neural networks (CNNs) often struggle to detect morphological variations along the cortical manifold, which serve as sensitive indicators of early neurodevelopment.24,61,62,63 To address these limitations, we integrated morphological features derived from cortical surface topology into SurfGNN. Our findings demonstrate that SurfGNN achieves superior predictive accuracy compared to state-of-the-art methods, including image-fed CNN models. SurfGNN is a surface-oriented DL model tailored for phenotype prediction. SurfGNN incorporates two core learning modules—topological scale learning (TSL) and regional scale learning (RSL)—both grounded in graph neural networks (GNNs) to enable comprehensive surface analysis. Brain function is known to follow a hierarchical, multiscale processing paradigm,64,65 where localized information is progressively integrated into global representations. SurfGNN aligns with this paradigm: the TSL module processes fine-grained, localized data, while the RSL module focuses on coarser, global features. The TSL module, inspired by spherical CNNs,34 transfers brain topology from densely connected meshes at fine scales to coarser representations at broader scales. Although this approach may have limitations in detecting regional heterogeneities, it excels at capturing foundational graph structures during early learning stages. The RSL module, in contrast, performs detailed graph analysis by applying convolution operations to specific vertices, enabling them to absorb higher level information. Additionally, its region-specific pooling mechanism prioritizes critical vertices while excluding less relevant ones, enhancing predictive focus. The strength of this GNN-based framework lies in its capacity to target brain regions closely associated with specific phenotypes, reflecting natural regional variability in brain responses. The TSL module ensures consistent integration of localized information, facilitating unbiased data transmission to higher levels. Meanwhile, the RSL module emphasizes regional feature selection and manipulation, highlighting critical information for analysis. This region-focused, coarse-grained approach aligns with the complexity of brain processes and has proven effective in brain research.66,67 By seamlessly integrating local information and refining graph structures, SurfGNN enables detailed and accurate analyses. Results presented in Table S5 underscore the critical roles of the TSL and RSL modules in achieving these outcomes. Limitations of the study This study has several limitations. While we identified that cortical morphological alterations induced BAI associated with decreased brain age impair cognitive performance, the underlying mechanisms, particularly the impact of iron accumulation on morphological impairment, remain unclear. Future research should incorporate in vivo and quantitative assessments of cerebral iron levels using advanced MRI techniques such as transverse relaxation rate, SWI, and QSM.37 Moreover, the limited availability of β-TM data prevents external validation of our findings. Additionally, the lack of blood sample collection from the control group restricts our ability to compare hematological metrics, such as iron levels and red blood cell counts, between β-TM patients and healthy individuals, thereby limiting insights into how these parameters influence cognitive function. Another limitation is the absence of a disease-control group with iron overload but normal hemoglobin typing, such as patients with hereditary hemochromatosis. Including such a cohort in future studies would help isolate the specific neuroanatomical and cognitive effects of iron overload independent of β-TM-related anemia and other disease processes. Prior studies have highlighted that age and educational exposure strongly shape brain development.68,69 Although stratified analyses across narrower age groups can be informative, in our study the BAI was computed after age-bias correction, ensuring independence from chronological age and comparability across children aged 6–15 years. Given our limited sample size, further stratification would reduce statistical power; therefore, stratified results were not emphasized in the main text. Instead, we have provided these analyses in the supplemental information. Future studies with larger cohorts will allow stratified analyses by narrower age bands to further validate the robustness of our findings. Our data-driven approach also presents challenges, particularly in interpreting the role of brain regions not directly associated with the cognitive tasks studied. The relatively small sample size may exaggerate effect sizes, especially in the context of high-dimensional feature spaces, increasing the risk of false positives. Replication with larger, more diverse cohorts is crucial for validation.70 Specific limitations also pertain to the SurfGNN architecture. The TSL structure emphasizes local direction-invariant features, which may inadvertently overlook local shape characteristics, limiting low-scale feature extraction when high-resolution inputs are used. Future work should focus on improving the information extraction module to address these shortcomings. Additionally, the interpretability of the model’s outputs remains limited to higher scale features, necessitating efforts to enhance precision while maintaining prediction accuracy. The NKI dataset’s age annotations, precise only to the year, may introduce errors during model training. Despite the use of harmonization techniques, site variances between the NKI dataset and our dataset may contribute to further inaccuracies. Addressing these challenges in future studies will be essential to develop more reliable and robust patterns for understanding the neural and cognitive impacts of β-TM. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Yaowen Li (owengli@126.com). Materials availability This study did not generate new unique reagents. Data and code availability •Data reported in this paper will be shared by the lead contact upon request. Data reported in this paper will be shared by the lead contact upon request. •This paper does not report original code. This paper does not report original code. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments The authors thank all subjects involved in the study. This study was supported by Shenzhen Science and Technology Innovation Program JCYJ20240813112459030; Sanming Project of Medicine in Shenzen Municipality no. SZSM202011005; National Natural Science Foundation of China under grant no. U24A20340; and the special fund for High-level Hospital Construction of Guangdong Province no. ynkt2021-zz47. Author contributions Conceptualization, M.L. and Y.L.; methodology, S.X., Y.L., Z.L., and S.X.; investigation, S.X., S.L., and Y.L.; writing – original draft, S.X. and Z.L.; writing – review and editing, X.L., X.W., P.Y., M.L., and H.Z.; funding acquisition, H.Z.; resources, X.C.; supervision, M.L. and Y.L.; data curation, X.C.; project administration, X.C. Declaration of interests All authors declare no conflict-of-interest related to this work. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work the authors used ChatGPT 4 in order to correct the language and grammar issues in the manuscript. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERDeposited dataThe Nathan Kline Institute (NKI) datasetNooner et al.71https://fcon_1000.projects.nitrc.org/Clinical DatasetThis paperN/ACode of the SurfGNN ModelLi et al.30https://github.com/ZhuoshL/SurfGNNSoftware and algorithmsPython version 3.Python Softwarehttps://www.python.org/Pytorch version 2.1.0PyTorch Softwarehttps://pytorch.org/MATLAB version R2019bMATLAB versionhttps://www.mathworks.com/products/matlab.htmlSurfStat toolboxKeith J. Worsley, McGill Universityhttps://www.math.mcgill.ca/keith/surfstat/CIVET softwareMacDonald et al.72https://www.bic.mni.mcgill.ca/ServicesSoftware/CIVETIBM SPSS Statistics version 28IBM SPSS Statistics Softwarehttps://www.ibm.com/products/spss-statistics/PROCESS microAndrew F. Hayeshttps://processmacro.org/index.html Experimental model and study participant details Ethic This study was approved by the Ethics Committee of Shenzhen Children’s Hospital (Approval No. 202004002). Written informed consent was obtained from all participants’ legal guardians and assent was also obtained from all child participants, prior exsert influence to participation in the study. All data were fully de-identified to protect participant privacy and were handled in accordance with the protocols of our institution. Clinical dataset Participants were recruited between May 2020 and October 2023. After QC, 25 patients with β-TM (mean hemoglobin level: 104.28 ± 13.31 g/L; red cell distribution width: 42.06 ± 9.21 fL; mean ferritin level: 2843.7 ± 1547.5 ng/mL) and 40 healthy controls were recruited from institutional referral centers specializing in β-TM. The recruitment of healthy control subjects is described in Methods S1.1. Participants were excluded if they met any of the following criteria: (1) presence of mental disorders, personality disorders, or dependency on psychotropic substances; (2) inability to cooperate during MRI, leading to suboptimal image quality; (3) history of organic or metabolic brain conditions; (4) presence of cerebrovascular risk factors such as hypertension or diabetes, confirmed via clinical interview and medical history; or (5) contraindications for MRI procedures. External training dataset To train the SurfGNN model, we applied a publicly available dataset: The Nathan Kline Institute (NKI) dataset, specifically the Rockland Sample,71 which is a widely recognized resource in neuroscience and neuroimaging research. It is part of the NKI-Rockland Initiative,71 which aims to provide openly accessible, high-quality neuroimaging data to facilitate the study of brain function, structure, and behavior across the lifespan. After exclusion, we selected 738 children aged 6–16 years (mean ± SD: 11.5 ± 2.7; 404 males) with available T1-weighted (T1w) structural MRI; only T1w images were analyzed and these data are entirely independent of our in-house cohort. From the age-eligible pool, we excluded participants without a usable T1w scan and those failing structural image quality checks (e.g., gross motion/artifacts per NKI flags and visual review) and removed records with missing key demographics. To limit confounding, we restricted the NKI subset to the same developmental band as our cohort and adjusted all between-group models for age, sex, and intracranial volume (ICV). All were neurologically healthy, the standardized single-site acquisition further reduces site/scanner heterogeneity. A detailed description of the NKI cohort recruitment process can be found in Methods S1.2. Method details Image acquisition For our clinical dataset, structural brain imaging was conducted using a 3-Tesla MRI scanner (Magnetom Skyra, Siemens Medical Solutions, Erlangen, Germany) equipped with a 20-channel head coil at Shenzhen Children’s Hospital. T1-weighted images were acquired using a three-dimensional magnetization-prepared rapid gradient echo (3D-MPRAGE) sequence with the following parameters: 176 contiguous sagittal slices, isotropic voxel size of 1 mm3, repetition time (TR) of 2300 ms, echo time (TE) of 2.26 ms, field of view (FOV) of 256 mm, single average acquisition, and slice oversampling of 18.2%. To rule out organic brain lesions, a preliminary T2-weighted sequence (TR = 2300 ms, TE = 10.6 ms) was performed prior to the T1-weighted scan. For the NKI dataset, structural brain imaging was conducted using a 3-Tesla MRI scanner (Magnetom Skyra, Siemens Medical Solutions, Erlangen, Germany) equipped with a 20-channel head coil at Shenzhen Children’s Hospital. T1-weighted images were acquired using a three-dimensional magnetization-prepared rapid gradient echo (3D-MPRAGE) sequence with the following parameters: 176 contiguous sagittal slices, isotropic voxel size of 1 mm3, repetition time (TR) of 2300 ms, echo time (TE) of 2.26 ms, field of view (FOV) of 256 mm, single average acquisition, and slice oversampling of 18.2%. To rule out organic brain lesions, a preliminary T2-weighted sequence (TR = 2300 ms, TE = 10.6 ms) was performed prior to the T1-weighted scan. Image preprocessing Cortical surface reconstruction and subcortical volume segmentation were performed using CIVET pipeline, a widely used image processing software,72,73 based on structural brain imaging data. After reconstructing the cortical surface, we extracted cortical metrics—including cortical thickness, sulcal depth, mean curvature, and curvature—at 81,924 vertices. To reduce complexity, vertex down-sampling was performed using an icosahedron-based approach, generating four additional resolutions (20,484, 5,124, 1,284, and 324 vertices) suitable for input into the SurfGNN model. Detailed preprocessing procedures are provided in the Methods S2. Cortical features between NKI and our dataset were harmonized using the ComBat method31 prior to SurfGNN training. The NKI dataset was then divided into k=5 groups (folds) for cross-validation, with 20% of the data used for testing in each fold. The remaining 80% was further split iteratively into training (64%) to fit the models and validation (16%) to tune hyperparameters. Clinical cognitive function assessment Cognitive performance was administered to both healthy children and those with medical conditions using the Wechsler Intelligence Scale for Children, Fourth Edition (WISC- IV). This standardized test provides composite scores across four core domains: Verbal Comprehension, Perceptual Reasoning, Working Memory, and Processing Speed. Each domain score was derived from multiple subtests and age-normed according to standardized WISC-IV procedures. The Full-Scale score was then obtained by aggregating the domain scores to yield an overall index of intellectual functioning. For our analyses, we used both the domain-specific indices and the aggregated Full-Scale IQ scores as continuous variables. These standardized scoring procedures ensured comparability across participants and allowed us to quantitatively capture a broad spectrum of cognitive functioning in both β-thalassemia patients and healthy controls. Architecture of SurfGNN The SurfGNN model utilizes a multi-graph input framework, where each individual graph represents a specific cortical feature. As shown in Figure S1, the complete architecture includes specialized graph convolutional layers and pooling mechanisms designed for topology-sampling learning and region-specific feature extraction. These components are specifically tailored to capture both local cortical relationships and global structural dependencies. The architecture begins with the generation of a sparse graph, maintaining a consistent structure across various cortical features, which is achieved through topology-sampling learning structures. These graphs are then processed through region-specific learning structures, allowing the model to focus on extracting distinctive features from different cortical regions. To combine the learned representations from each cortical feature, an innovative score-weighted fusion framework is applied. This mechanism aggregates the features with learned weights that reflect their relevance to the final phenotype prediction. Subsequently, the fused features are passed through a readout layer that performs the final phenotype prediction. A comprehensive description of the SurfGNN model architecture, including specific details about the graph convolutional and pooling layers, as well as the score-weighted fusion mechanism, can be found in the supplemental information (Methods S3). Quantification and statistical analysis To address age dependency in the training residuals, we applied a bias correction method as described in.22,28 Specifically, we used a linear model PBA=α∗PBA′+β to obtain an unbiased estimate of predicted brain age PBA, calculated as: PBA′=(PBA−β)/α, where the parameters α and β were estimated during training (using both the training and validation sets) and subsequently applied to the test set. After computing PBA′ for each subject, we derived a metric called the Brain Age Index (BAI), which represents a subject’s relative brain health status. BAI was calculated by subtracting the subject’s postmenstrual age at the time of the MR scan from their unbiased PBA′, as described in.22,28 We assessed the impact of each variable on the Brain Age Index (BAI) using a univariate analysis approach. This was conducted through a general linear mixed-effects model, which accounted for both within- and between-subject variations. The model was adjusted for postmenstrual age (PMA) at the time of the scan and other relevant clinical factors, enabling us to isolate and evaluate the effect of the primary variable of interest on BAI. Cortical feature-related statistics between groups were performed across 324 vertices of the cortical surface, using the SurfStat toolbox in Matlab. General linear models (GLM) models were used in all cortical features comparisons between the HCs and the β-TM peers, controlling for sex, PMA at scan, and brain volume. In vertex-wise shape analysis, the random field theory (RFT)-based approach was applied to adjust P values for multiple comparison errors. In addition, we conducted a causal mediation analysis74 to examine whether BAI mediated the relationship between hypothesized hematological risk factors (i.e. hemoglobin level and serum ferritin level) to cognitive performances, while controlling for age, gender, and brain volume as covariates. Using Hayes’ PROCESS macro for SPSS (Model 4), we estimated the direct effect and indirect effect while adjusting for covariates in all paths. Bootstrapping with 1,000 resamples was employed to test the significance of the indirect effect, providing 95% confidence intervals (CIs) for the mediation effect. Age, gender, and brain volume were included to control for potential confounding influences on BAI to cognitive performance. All continuous variables were mean-centered to reduce multicollinearity. The p values from multiple comparisons were adjusted using the false discovery rate (FDR) method. The significance level was set at 0.05 for all analyses. A flow diagram summarizing participant recruitment, exclusions, imaging acquisition, preprocessing, SurfGNN analysis, and statistical testing is provided in Figure S2. Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.116072. Contributor Information Mengting Liu, Email: liumt55@mail.sysu.edu.cn. Yaowen Li, Email: owengli@126.com. Supplemental information Document S1. Figures S1–S5, Methods S1–S3, and Tables S1–S6 References Associated Data Supplementary Materials Document S1. Figures S1–S5, Methods S1–S3, and Tables S1–S6 Data Availability Statement •Data reported in this paper will be shared by the lead contact upon request. Data reported in this paper will be shared by the lead contact upon request. •This paper does not report original code. This paper does not report original code. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.