Development and validation of an insulin resistance prediction model in children and adolescents using machine learning algorithms

Introduction

Insulin resistance (IR) is a critical precursor to several metabolic disorders, including type 2 diabetes mellitus (T2DM), hypertension, and polycystic ovary syndrome (PCOS), particularly among children and adolescents (1,2). The prevalence of IR in pediatric populations has increased alongside rising obesity rates, with strong associations between obesity, IR, and metabolic syndrome components (3). Early detection and intervention for IR are essential to prevent its progression to more severe metabolic complications (4). Therefore, developing accurate and practical predictive models for IR in children and adolescents is of great clinical significance.

A variety of elements contribute to IR, including dietary habits, disruptions in lipid metabolism, oxidative stress, issues with the function of mitochondria, inflammation, and genetic predispositions (1). Research has consistently shown strong associations between hematological parameters and IR across different populations. Studies have revealed positive correlations between IR and red blood cell (RBC) count, hemoglobin (HGB), hematocrit, and white blood cell (WBC) count (5-7). Lifestyle factors, including physical inactivity and stress, were found to be significantly correlated with waist circumference, an indicator of IR, among female medical students in Saudi Arabia. Additionally, anthropometric measurements, such as body mass index (BMI), waist circumference, blood pressure, and postprandial blood sugar levels, are significantly correlated with IR (8). These factors have been incorporated into various prediction models, including logistic regression (LR) models and machine learning (ML) models, which have shown promising results in predicting IR risk. For instance, Hall et al. developed an ML model for children with 78% overall accuracy using anthropometric features (9), while Zhang & Wan focused on children aged 6–12 years, with their extreme gradient boosting (XGBoost) model achieving an area under the curve (AUC) of 0.85 and identifying glucose, waist circumference, and age as key predictors (10). However, these studies often focus on limited populations (e.g., children with obesity or diabetes) or specific age groups and may overlook the importance of dietary factors in predicting IR (11-13). Additionally, few studies have compared the performance of different ML algorithms in predicting IR among children and adolescents. This gap limits our understanding of which models may be most effective in clinical settings.

ML offers a powerful tool for predicting complex conditions like IR by identifying patterns and relationships in large datasets. In this study, we selected four widely used ML algorithms—the exhaustive Chi-squared automatic interaction detector (E-CHAID), back propagation neural network (BPNN), support vector machine (SVM), and LR—each with distinct advantages. E-CHAID is effective in handling categorical data and identifying interactions between variables (14). BPNN has strong nonlinear mapping capabilities and adaptability, making it suitable for complex datasets (15). SVM is renowned for its robustness in high-dimensional spaces and ability to handle nonlinear relationships through kernel functions (16). LR, on the other hand, provides clear and interpretable models, which are crucial for clinical decision-making (17). Given the diverse characteristics of these algorithms, comparing their accuracy in predicting IR is essential to identify the most effective model for clinical use. By evaluating these models using a comprehensive dataset that includes both basic health information and dietary factors, we aim to provide a more accurate and practical tool for early IR prediction in children and adolescents. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2024-502/rc).

Methods

Study participants

In this research, we collected information on individuals aged 6–17 years from the China Health and Nutrition Survey (CHNS) database, employing this dataset as our training sample. The CHNS is a collaborative project between the Carolina Population Center at the University of North Carolina and the National Institute for Nutrition and Health at the Chinese Center for Disease Control and Prevention. It was established to track the primary care nutritional and health conditions of individuals across 15 Chinese provinces from 1989 to 2015. Notably, the 2009 CHNS dataset includes biomarker information for children and adolescents, which is crucial for our study; hence, we utilized this specific dataset for our analysis. Finally, a total of 827 qualified participants aged 6–17 years from nine provinces in China were included in the training dataset.

Given our research scope—developing and validating an early IR prediction model using routine physical examination data and dietary factors—a cross-sectional approach allowed us to meet our objectives efficiently with the available resources. To mitigate limitations of cross-sectional data, we validated the model using an external dataset from two hospitals. We collected data on children and adolescents aged 6–17 years from the physical examination Department and the Department of Pediatrics at the Nanchong Central Hospital and the Nanchong City Jialing District People’s Hospital (a tertiary hospital and a secondary hospital, respectively) from January 2019 to October 2024. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the Institutional Review Boards of the Nanchong Central Hospital (No. 2024-033) and the Nanchong City Jialing District People’s Hospital (No. 2024-002). The participants’ legal guardians provided written informed consent to participate in this study. This compilation included fundamental demographic details, lifestyle practices, and blood sample analyses. In conclusion, a dataset encompassing 207 individuals within the age bracket of 6–17 years was created. Due to CHNS not collecting data on children and adolescents in Sichuan province, and all external test sets come from two hospitals in Sichuan province, China, there may be some regional differences in the population of the training and test sets.

Outcome

The Homeostatic Model Assessment for IR (HOMA-IR) calculation method is simple, has a good correlation with the hyperinsulinemic-euglycemic clamp technique, and is more widely used in clinical applications for the diagnosis of IR. Children and adolescents with a HOMA-IR score >3.0 were defined as having IR in this study. The HOMA-IR score was calculated as [fasting insulin (mU/L)] × [fasting glucose (mmol/L)]/22.5. While HOMA-IR calculation is straightforward, it requires measuring fasting insulin and glucose levels (18). In resource-limited areas or primary care settings, insulin testing may be impractical or costly. Our study aims to identify high-risk individuals for IR using early warning indicators and ML models that integrate clinical factors and biomarkers from routine physical exams. This allows preliminary IR risk assessment without direct insulin measurement, prioritizing those needing further testing in resource-constrained environments.

Data preprocessing and feature selection

The dataset obtained from the CHNS contained missing values. Initially, any row or column with more than 50% missing data was removed; subsequently, a multivariate imputation technique was applied to address the remaining gaps in the data. We strive to ensure that the selected variables can be obtained during routine physical examinations in most primary healthcare hospitals, which is child-friendly. Combining the data provided by the CHNS database, we selected 41 features for the study, including 17 concerning basic information [age, gender, nationality, province, stratum, urban, height, weight, BMI, systolic blood pressure (SBP), diastolic blood pressure (DBP), upper arm circumference, hip circumference, waist circumference, waist-height ratio, waist-to-hip ratio, and arm-height ratio (AHtR)], 12 about daily habits (sleep duration, currently in school, physical exercise, physical exercise per week, physical exercise in school, physical exercise per week, doing homework time, energy intake, carbohydrate intake, fat intake, protein intake), and 12 laboratory tests {total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), apolipoprotein A (APO_A), apolipoprotein B (APO_B), lipoprotein(a) [Lp(a)], WBC, RBC, HGB, platelet (PLT), and glucose}. The outcome variable IR is assigned a value of 1, while non-IR is assigned a value of 0. Among the other independent variables, categorical variables were assigned values according to the rule that 1 represents positive and 0 represents negative. The continuity variable remains unchanged.

Model construction

All available data present in the CHNS database were utilized to enhance the statistical power and generalizability of the outcomes. Using IBM SPSS Modeler 18.0 software, four ML algorithms that are widely used in clinics, namely, E-CHAID, BPNN, SVM, and LR, were trained for model building.

The E-CHAID algorithm model is configured as follows. Maximum tree depth: 5. Importance values were adjusted via the Bonferroni method. Pearson’s χ² test was used as the test statistic. Maximum number of iterations: 100. Termination rule: splitting stops if the minimum record count in the parent branch is less than 2% of the total. Splitting stops if the minimum record count in the child node after splitting is less than 1% of the total. For the BPNN model, a multilayer perceptron is used for modeling. The hidden layer is set to automatically calculate, with a maximum training time of 15 minutes, a maximum number of training epochs of 250, and a minimum accuracy of 90%. The optimal network structure for the BPNN model is 11 → 6 → 1. For the SVM model, all probabilities were selected for inclusion, and the termination criterion is 1.0E−5, the regularization parameter is 10, the regression accuracy is 1.6, the kernel type is the recombinant fusion (RBF), and the RBF gamma is 1.0. Additionally, the calculation of the importance of the predictive variables is selected. For LR, the binary process uses an enter method, with a scale set to user-defined (value of 3.0), all probabilities were selected for inclusion, and the outlier tolerance is 1.0E−10.

After filling in missing data, we employed univariate LR analysis to identify variables that were statistically significant (α=0.05). Following this, we developed four distinct ML algorithm models. To further assess the accuracy and applicability of these models, they underwent external validation using datasets from two different hospitals.

Statistical analysis

SPSS version 26.0 (IBM Corporation, Armonk, NY, USA) was used for the Chi-squared test, t-test, univariate LR, and stepwise linear regression analysis. An independent t-test was used to compare continuous variables, and the Chi-squared test or Fisher’s exact test was used to compare categorical variables. All tests were two-tailed, and P<0.05 was considered to indicate statistical significance. To evaluate the performance of each IR risk prediction model for children and adolescents, we used the HOMA-IR calculator to assess the risk probability, with the calculation formula as described earlier. Four performance metrics, namely, sensitivity, specificity, Youden’s index, and AUC, were calculated to better understand the differences between the models.

Results

Subject characteristics

Data from a total of 827 children and adolescents from the CHNS database were gathered. Among them, 281 were identified with IR, defined by a HOMA-IR >3.0, and 546 were identified without IR. Accordingly, 207 children and adolescents were included in the external test set. We initially considered 41 variables, including demographic details, lifestyle habits, and blood test results. Data on weekly physical activity and homework duration were excluded because >50% of participants were missing. Ultimately, 39 variables were advanced to the feature selection phase.

A comparison of feature information between the training and test cohorts is presented in Tables 1,2. No significant disparities were observed between the two sets regarding sex, 3-day average intake of carbohydrates, 3-day average intake of fat, waist-to-hip ratio, sleep duration, number of sports activities currently in school, number of sports activities in school, number of sports activities, number of sports activities in school per week, homework, TG, TC, HDL-C, LDL-C, APO_A, Lp(a), WBC, RBC, PLT, or glucose (P>0.05). Conversely, significant differences were noted between the two sets in terms of the prevalence of IR, stratum, urban area, nationality, province, 3-day average intake of energy, 3-day average intake of protein, SBP, DBP, height, weight, BMI, upper arm circumference, hip circumference, waist circumference, AHtR, APO_B, and HGB, all of which were statistically significant (P<0.05).

Univariate LR

Through univariate LR analysis, the results revealed that age, province, 3-day average intake of carbohydrates, 3-day average intake of fat, 3-day average intake of energy, 3-day average intake of protein, SBP, DBP, height, weight, BMI, upper arm circumference, hip circumference, waist circumference, waist-height ratio, AHtR, currently in school, TG, TC, APO_B, RBC, HGB, and glucose were statistically significant variables in predicting IR, with a total of 23 variables.

E-CHAID model

The first optimal grouping variable in the decision tree is hip circumference, which forms a four-way branching tree (≤69.50, 69.51–76.00, 76.01–87.00, >87.00 cm). The stratified categorical variable nodes in the E-CHAID model include hip circumference, TG, glucose, height, and waist-height ratio, with a total of 20 nodes established, resulting in 14 terminal nodes. Among the five independent variables selected, the importance of IR clinical early warning indicators was ranked as follows: glucose (0.46), hip circumference (0.29), height (0.13), waist-height ratio (0.06), and TG (0.06). The model’s prediction accuracy can be found in Table 3, and the confusion matrix can be found in Table 4.

BPNN model

The importance of the IR clinical early warning indicators is as follows: glucose (0.14), TG (0.07), waist circumference (0.07), APO_B (0.06), BMI (0.06), weight (0.06), 3-day average intake of protein (0.05), DBP (0.05), height (0.05), and waist-height ratio (0.04). The model’s prediction accuracy can be found in Table 3, and the confusion matrix can be found in Table 4.

SVM model

The importance of the IR clinical early warning indicators is as follows: glucose (0.10), SBP (0.09), age (0.09), TG (0.06), APO_B (0.06), height (0.06), upper arm circumference (0.05), province (0.05), currently in school (0.05), and 3-day average carbohydrate fat (0.04). The model’s prediction accuracy can be found in Table 3, and the confusion matrix can be found in Table 4.

LR model

The importance of the IR clinical early warning indicators is as follows: waist circumference (0.18), glucose (0.13), 3-day average intake of energy (0.09), BMI (0.08), 3-day average intake of carbohydrates (0.07), hip circumference (0.06), age (0.06), AHtR (0.05), height (0.05), and DBP (0.04). The model’s prediction accuracy can be found in Table 3, and the confusion matrix can be found in Table 4.

Comparison of the predictive ability of the four prediction models

Table 3 shows that for both the training and testing samples, the accuracy of the SVM model (91.90%, 90.34%) is greater than that of the E-CHAID model (77.75%, 72.95%), BPNN model (75.94%, 70.05%), and LR model (76.18%, 71.01%). Table 5 shows that in the training sample, the sensitivity (0.851), specificity (0.954), Youden’s index (0.805), and AUC (0.967) of the SVM model were all greater than those of the E-CHAID model (0.480, 0.930, 0.410, 0.814), the BPNN model (0.555, 0.864, 0.419, 0.803), and the LR model (0.512, 0.890, 0.402, 0.808). In the external testing sample, the sensitivity (0.870), specificity (0.930), Youden’s index (0.800), and AUC (0.958) of the SVM model were all higher than those of the E-CHAID model (0.728, 0.730, 0.458, 0.803), the BPNN model (0.565, 0.809, 0.374, 0.810), and the LR model (0.663, 0.748, 0.411, 0.804). In summary, the predictive ability of the SVM model is superior to that of the E-CHAID model, the BPNN model, and the LR regression model, whereas the IR predictive abilities of the E-CHAID model, BPNN model, and LR model were essentially comparable.

Discussion

Although IR might not cause any signs or discomfort in children or teens, it is a significant risk factor for numerous metabolic disorders. If left unchecked, IR can impact the growth and development of these young individuals. IR in young individuals can lead to numerous health complications, including cardiovascular risk factors, T2DM, and nonalcoholic fatty liver disease (19). Early recognition and intervention are crucial for preventing the progression of IR and its associated complications. Treatment strategies focus on lifestyle modifications, pharmacotherapy, and, in some cases, surgical interventions (3). Even small improvements in diet and lifestyle can positively impact measured IR in children and adolescents (19). Research on IR has focused mostly on the predictive ability of specific variables, such as waist circumference and BMI (20). However, only a few single variables showed some correlation with IR, and this approach is not comprehensive enough in predicting IR, making it difficult to develop accurate evaluation criteria. Stawiski et al. applied neural network techniques to construct a predictive model for IR in children and achieved good predictive performance (21). A more recent study developed multivariate regression models using demographic and clinical variables, which showed good discriminative ability (AUC: 0.834–0.868) for IR prediction in children and adolescents with obesity (22). However, in terms of the population to which it was adapted, their study focused mainly on children with type 1 diabetes mellitus (T1DM) or obesity, whereas our study focused on a broader population covering all children and adolescents aged 6–17 years. Multiple studies have been able to accurately predict IR in children, but the indicators used, such as fasting blood glucose and genetic test results, are difficult to obtain from physical examination results, resulting in poor practical usability (23,24). The results of this study confirm the previous hypothesis that ML techniques can indeed accurately predict IR at a very early stage in school-aged children and adolescents via routine physical examination results and dietary questionnaires.

Through comparative analysis, this study revealed that the SVM model outperforms several other models in terms of predictive performance, including the E-CHAID model, the BPNN model, and the LR model. This result may be closely related to the unique advantages of the SVM model. The SVM model excels at distinguishing different categories by finding the optimal hyperplane, effectively identifying complex patterns and relationships when dealing with high-dimensional data and nonlinear problems (25). Moreover, the SVM model possesses good generalizability, enabling accurate predictions for unknown data with limited training data (26). These characteristics have led the SVM model to demonstrate high predictive accuracy in this study, providing a powerful tool for research in related fields.

The IR warning indicators selected by the four ML models are basically consistent with those used in previous studies and include mainly glucose, waist circumference, age, BMI, waist-height ratio, TG, blood pressure, etc. (10). Notably, several physical examination indicators and dietary factors discovered in this study, including hip circumference, upper arm circumference, protein intake, energy intake, and carbohydrate intake, have received little attention in previous studies. Previous studies have shown that a high-carbohydrate diet in Western countries is associated with an increased risk of IR in children and adolescents (27,28). Our research further revealed that increased protein and energy intake are also associated with an increased risk of IR. This suggests that we should pay more attention to the balanced and moderate composition of children’s and adolescents’ diets, avoiding excessive intake of carbohydrates, proteins, and energy.

There are certain limitations in this research. First, the sizes of both the training and testing datasets were limited, which could potentially lead to overfitting. Nonetheless, the utilization of an external test set demonstrated that the extent of this issue was within acceptable bounds. To further address this concern, we have applied appropriate regularization parameters for models that support regularization. Specifically, for LR, we set the regularization scale to a user-defined value of 3.0; for SVM, we set the regularization parameter to 10 and the RBF gamma to 1.0. These settings help to balance the trade-off between bias and variance, ensuring that the model performs well on unseen data. In future work, we plan to expand our dataset by incorporating additional data sources and conducting multi-center studies. Secondly, regional differences exist between the CHNS dataset (excluding Sichuan province) and our Sichuan-based test set, potentially impacting result generalizability. However, this design allows us to assess the model’s generalization capability. For instance, the SVM model showed strong performance on the Sichuan test set (sensitivity 0.870, specificity 0.930, Youden’s index 0.800, AUC 0.958), comparable to the training set, indicating good generalization. Future work will integrate datasets from other regions to further enhance model robustness and generalizability. Additionally, the models were based on a predefined set of clinically relevant variables, which may not cover all potential factors influencing IR. Unmeasured or unconsidered variables (such as folate, vitamin B12 diets, and rosiglitazone) could also significantly impact IR development and progression (29,30). To enhance model comprehensiveness, we plan to integrate additional data sources and conduct multi-center studies in future work. This will expand our dataset and include a broader range of variables, providing a more holistic understanding of IR contributors in children and adolescents.

Conclusions

This study aims to provide a risk classification prediction method for early clinical warning of IR in children and adolescents. We integrated the basic information, lifestyle habits, blood tests, and dietary questionnaire results of children and adolescents, which are easily obtained during the physical examination process. The results showed that the SVM model performed better in predicting individual risk than did the E-CHAID model, BPNN model, and LR model. In practical applications, their respective advantages should be combined to better leverage their application strengths and serve clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tp.amegroups.com/article/view/10.21037/tp-2024-502/rc

Data Sharing Statement: Available at https://tp.amegroups.com/article/view/10.21037/tp-2024-502/dss

Peer Review File: Available at https://tp.amegroups.com/article/view/10.21037/tp-2024-502/prf

Funding: This research was funded by the Innovation Fund Youth Program of the National Science Library (Chengdu) of the Chinese Academy of Sciences (No. E3Z0000301).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tp.amegroups.com/article/view/10.21037/tp-2024-502/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the Institutional Review Boards of the Nanchong Central Hospital (No. 2024-033) and the Nanchong City Jialing District People’s Hospital (No. 2024-002). The participants’ legal guardians provided written informed consent to participate in this study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Luo Y, Luo D, Li M, et al. Insulin resistance in pediatric obesity: from mechanisms to treatment strategies. Pediatr Diabetes 2024;2024:2298306.
2. Huo Y, Ji S, Yang H, et al. Differential expression of microRNA in the serum of patients with polycystic ovary syndrome with insulin resistance. Ann Transl Med 2022;10:762. [Crossref] [PubMed]
3. Al-Beltagi M, Bediwy AS, Saeed NK. Insulin-resistance in paediatric age: Its magnitude and implications. World J Diabetes 2022;13:282-307. [Crossref] [PubMed]
4. Ziamanesh F, Mohammadi M, Ebrahimpour S, et al. Unraveling the link between insulin resistance and Non-alcoholic fatty liver disease (or metabolic dysfunction-associated steatotic liver disease): A Narrative Review. J Diabetes Metab Disord 2023;22:1083-94. [Crossref] [PubMed]
5. Koivula T, Lempiäinen S, Laine S, et al. Cross-Sectional Associations of Body Adiposity, Sedentary Behavior, and Physical Activity with Hemoglobin and White Blood Cell Count. Int J Environ Res Public Health 2022;19:14347. [Crossref] [PubMed]
6. Alvarez-Jimenez L, Morales-Palomo F, Moreno-Cabañas A, et al. Effects of statin therapy on glycemic control and insulin resistance: A systematic review and meta-analysis. Eur J Pharmacol 2023;947:175672. [Crossref] [PubMed]
7. Chen JY, Chen YH, Lee YC, et al. The Association Between White Blood Cell Count and Insulin Resistance in Community-Dwelling Middle-Aged and Older Populations in Taiwan: A Community-Based Cross-Sectional Study. Front Med (Lausanne) 2022;9:813222. [Crossref] [PubMed]
8. Badawy Y, Aljohani NH, Salem GA, et al. Predictability of the Development of Insulin Resistance Based on the Risk Factors Among Female Medical Students at a Private College in Saudi Arabia. Cureus 2023;15:e39112. [Crossref] [PubMed]
9. Hall AJ, Hussain A, Shaikh MG. Predicting insulin resistance in children using a machine-learning-based clinical decision support system. In: Advances in Brain Inspired Cognitive Systems: 8th International Conference, BICS 2016, Beijing, China, November 28-30, 2016, Proceedings 8. Springer International Publishing; 2016:274-83.
10. Zhang Q, Wan NJ. Simple Method to Predict Insulin Resistance in Children Aged 6-12 Years by Using Machine Learning. Diabetes Metab Syndr Obes 2022;15:2963-75. [Crossref] [PubMed]
11. Yan W, Wu S, Liu Q, et al. The link between obesity and insulin resistance among children: Effects of key metabolites. J Diabetes 2023;15:1020-8. [Crossref] [PubMed]
12. El Sehmawy AA, Diab FEA, Hassan DA, et al. Utility of Adipokines and IL-10 in Association with Anthropometry in Prediction of Insulin Resistance in Obese Children. Diabetes Metab Syndr Obes 2022;15:3231-41. [Crossref] [PubMed]
13. Rodríguez-Rodríguez E, Salas-González MD, Ortega RM, et al. Leukocytes and Neutrophil-Lymphocyte Ratio as Indicators of Insulin Resistance in Overweight/Obese School-Children. Front Nutr 2021;8:811081. [Crossref] [PubMed]
14. Altay Y. Prediction of the live weight at breeding age from morphological measurements taken at weaning in indigenous Honamli kids using data mining algorithms. Trop Anim Health Prod 2022;54:172. [Crossref] [PubMed]
15. Jiang S, Li Y, Jiao Y, et al. A back propagation neural network approach to estimate the glomerular filtration rate in an older population. BMC Geriatr 2023;23:322. [Crossref] [PubMed]
16. Teng S, Zheng N, Al-Huqail AA, et al. Effect of nanoparticle macroalgae in the treatment of fatty liver disease using logistic regression, and support vector machine. Environ Res 2023;224:115426. [Crossref] [PubMed]
17. Lee AM, Hu J, Xu Y, et al. Using Machine Learning to Identify Metabolomic Signatures of Pediatric Chronic Kidney Disease Etiology. J Am Soc Nephrol 2022;33:375-86. [Crossref] [PubMed]
18. Khalili D, Khayamzadeh M, Kohansal K, et al. Are HOMA-IR and HOMA-B good predictors for diabetes and pre-diabetes subtypes? BMC Endocr Disord 2023;23:39. [Crossref] [PubMed]
19. Chen ME, Hannon TS. Clinical manifestations of insulin resistance in youth. In: Zeitler P, Nadeau K. editors. Insulin Resistance: Childhood Precursors of Adult Disease. Cham: Springer International Publishing; 2019:3-17.
20. Hatami H, Montazeri SA, Hashemi N, et al. Optimal Cutoff Points for Anthropometric Variables to Predict Insulin Resistance in Polycystic Ovary Syndrome. Int J Endocrinol Metab 2017;15:e12353. [Crossref] [PubMed]
21. Stawiski K, Pietrzak I, Młynarski W, et al. NIRCa: An artificial neural network-based insulin resistance calculator. Pediatr Diabetes 2018;19:231-5. [Crossref] [PubMed]
22. Araújo D, Morgado C, Correia-Pinto J, et al. Predicting Insulin Resistance in a Pediatric Population With Obesity. J Pediatr Gastroenterol Nutr 2023;77:779-87. [Crossref] [PubMed]
23. Lin H, Tas E, Børsheim E, et al. Circulating miRNA Signatures Associated with Insulin Resistance in Adolescents with Obesity. Diabetes Metab Syndr Obes 2020;13:4929-39. [Crossref] [PubMed]
24. Zhang PP, Song JY, Li L, et al. Associations between genetic variants of HSD17B13 and fasting plasma glucose in Chinese children. Nutr Metab Cardiovasc Dis 2023;33:1778-84. [Crossref] [PubMed]
25. Bharadwaj, Prakash KB, Kanagachidambaresan GR. Pattern recognition and machine learning. In: Prakash KB, Kanagachidambaresan GR. editors. Programming with TensorFlow: Solution for Edge Computing Applications. Cham: Springer; 2021:105-44.
26. Du KL, Swamy MNS. Support vector machines. In: Neural Networks and Statistical Learning. London: Springer; 2014:469-524.
27. Romero-Polvo A, Denova-Gutiérrez E, Rivera-Paredez B, et al. Association between dietary patterns and insulin resistance in Mexican children and adolescents. Ann Nutr Metab 2012;61:142-50. [Crossref] [PubMed]
28. Gow ML, Garnett SP, Baur LA, et al. The Effectiveness of Different Diet Strategies to Reduce Type 2 Diabetes Risk in Youth. Nutrients 2016;8:486. [Crossref] [PubMed]
29. Qian L, Yin X, Lan T, et al. Peroxisome proliferator-activated receptor gamma preserves intracellular homeostasis of insulin-resistant periodontal ligament stem cells. Ann Transl Med 2022;10:580. [Crossref] [PubMed]
30. Zhang H, Wang X, Zhang J, et al. Early supplementation of folate and vitamin B12 improves insulin resistance in intrauterine growth retardation rats. Transl Pediatr 2022;11:466-73. [Crossref] [PubMed]