Prediction Models for Diabetes in Children and Adolescents: A Review

Abstract

This review aims to present the latest advancements in prediction models for diabetes mellitus, with a particular focus on children and adolescents. It highlights models for predicting both type 1 and type 2 diabetes in this population, emphasizing the inclusion of risk factors that facilitate the identification of potential occurrence and early detection of diabetes in young individuals. Newly identified factors for differentiating between types of diabetes are discussed, alongside an overview of various machine learning and deep learning algorithms specifically adapted for diabetes prediction in children and adolescents. The advantages and limitations of these methods are critically examined. The review underscores the necessity of addressing challenges posed by incomplete datasets and emphasizes the importance of creating a comprehensive data repository. Such developments are essential for enabling artificial intelligence tools to generate models suitable for broad clinical application and advancing early diagnostic and preventive strategies for diabetes in children and adolescents.

1. Introduction

The Global Burden of Diseases Study (GBD) has provided critical projections for disease prevalence through 2050. Based on the diabetes burden from 1990 to 2021 across 204 countries, combined with proportions of type 1 and type 2 diabetes recorded in 2021, it is estimated that by 2050, diabetes will emerge as one of the leading global causes of death and disability, affecting all demographic groups [1]. The findings reveal that in 2021, 529 million people worldwide were living with diabetes. This figure is anticipated to rise to 1.31 billion by 2050. Regions such as North Africa and the Middle East are projected to experience the highest prevalence, while the prevalence rates are expected to surpass 10% in over 43.6% of countries of the world [2]. The diabetes-related mortality was analyzed, and it is expected that the mortality burden of both T1D and T2D will gradually decline. However, achieving a reduction in T2D mortality remains a significant challenge, requiring substantial and continuous efforts in prevention and management. Zhu et al. [3] and Liu et al. [4] projected the economic burden of diabetes. According to their study, the total economic costs of diabetes are expected to rise from $250.2 billion in 2020 to $460.4 billion in 2030, with an annual growth rate of 6.32%. Additionally, per capita costs are predicted to grow from $231 to $414, with an annual increase of 6.02%. The findings outlined above emphasize that diabetes represents a growing public health and economic challenge, necessitating enhanced resource allocation and robust prevention strategies.

Diabetes is a rapidly spreading disease not only in adults but also in young people. Since 1990, when the first publications on type 2 diabetes (T2D) in children and adolescents were published [5], researchers began noticing a growing incidence of diabetes worldwide [6,7], and the disease has firmly established as a global health issue. Youth-onset type 2 diabetes (T2D) was first identified in Pima Indian children and adolescents [8]. By the mid-1990s, various clinics across the United States began reporting T2D cases in children from diverse ethnic backgrounds, particularly among non-Hispanic Blacks and Hispanics [9,10]. At that time, the researchers noticed that in some regions of the United States, T2D is as common in adolescents as type 1 (T1D) [11,12,13]. In 2000, the SEARCH for Diabetes Study was launched in the US to explore the epidemiology of both types of diabetes: T1D and T2D in children [14,15,16]. It was followed by the initiation of the “Treatment Options for Type 2 Diabetes in Adolescents and Youth” (TODAY) study in 2004 [17]. Data from the SEARCH study indicate that between 2003 and 2018, the adjusted incidence of T2D among children and adolescents in the US nearly doubled [18]. While the US leads in the T2D epidemic among children and adolescents [19], similar trends are evident worldwide [20], for example, in Austria, Britain, Germany [21], Canada [22], Hong Kong [23], China [24], Japan [25], India [16,26], Israel [27], etc. However, the overall incidence of T2D in Europe remains lower than in the US. Thus, in Germany, there was a threefold increase in T2D prevalence among 10- to 19-year-olds between 2002 and 2020 [21]. A comparable threefold rise was also seen in children during 2008–2018 compared to 1997–2007. In the UK, the number of children registered as having T2D and treated in pediatric diabetes units increased by more than 50% from 2015 to 2020 [28]. The Chinese data reveal an average annual increase of 26.6% in T2D among youth aged 10–19 years [29].

These data show that diabetes, especially T2D, is not a disease of adults as it was once believed. Its frequency among children and adolescents is on the rise and follows, first of all, the growth of the obesity rate in childhood [30,31]. Early-onset T2D represents a significant global public health challenge [32]. The prognosis of this disease with early onset is significantly worse compared to T2D in adults [33], so urgent and effective measures are needed. However, the treatment of this disease faces numerous problems, including limited and insufficient therapeutic options. For this reason, it is necessary to take preventive measures to protect against the disease in young people [34]. The timely detection of the tendency to diabetes, even in early childhood or adolescence, is the topic of this review.

The aim of this review is to present the latest advancements in prediction models for diabetes mellitus in children and adolescents. The early detection strategies for the prediction of diabetes of type 1 and type 2 are considered. Approaches for distinguishing between these two types of diabetes in children and adolescents are also discussed. Advantages of inclusion of machine learning and deep learning into recent prediction models are shown. Based on the results reviewed in the paper, recommendations for future research, emphasizing the application of advanced AI tools and technologies, are given.

2. Prediction Models of Type 1 Diabetes (T1D)

Research for 215 countries in 2022 on T1D incidence in children and adolescents shows that it varies widely, and has increased in many nations [35] over the last decades. Based on data, the estimated annual rise over 20 years is expected to be 3%, which will increase incidence by 81%. In order to overcome and reduce the difficulties of this disease, it is necessary to recognize and treat it in a timely manner.

Clinical T1D arises from the progressive destruction of β-cells, ultimately leading to end-stage insulitis, often following a prolonged asymptomatic period. Identifying individuals at risk is crucial for delaying or preventing its onset. Since the 1980s, significant advancements have been made in predicting T1D [11]. Prevention strategies are categorized into primary prevention (reducing risk before β-cell damage), secondary prevention (halting β-cell destruction), and tertiary prevention (restoring β-cell function or preventing complications). Predictive accuracy has improved by incorporating risk factors such as family history and genetic, immunological, and metabolic markers.

2.1. Immunological Markers in Prediction

Immunological biomarkers are effective in predicting T1D in children and adolescents [12]. For the sake of immune regulation and protection of β-cells in T1D, the reduction in the need for insulin and the improvement of glycemic control is suggested. Laboratory parameters such as serum levels of calcium, magnesium, zinc, copper, iron, and selenium have to be incorporated into predictive models. The levels of these parameters provide insights into oxidative stress conditions and improve early detection strategies for managing diabetic ketoacidosis (DKA), which is the indicator of disease progression and also a leading cause of death in T1D. To evaluate predictive performance as disease severity increases the lambda-mu-sigma method for T1D [13], it is suggested to be utilized. The method is suitable for the prediction of T1D progression based on identifying the degree of metabolic acidosis progression.

2.2. Metabolic Markers as Risk Factors

For advancing the prevention of T1D, the identification and widespread use of reliable metabolic markers is of essential importance. For T1D, traditional risk factors are poor glycemic control, hypertension, dyslipidemia, obesity, albuminuria, unhealthy diet, and physical inactivity [36]. Novel biomarkers, like chronic inflammation, also contribute to risk.

2.3. Blood Glucose Marker

Eating disorder is predicted to have an impact on T1D prevalence among children and adolescents [37]. The results revealed significant gender differences: while males showed no associations between eating disorder psychopathology and factors like body mass index (BMI) or insulin restriction, in females, illness perceptions and insulin restriction accounted for 48% of the variance in eating disorder psychopathology. These findings highlight the need for increased clinical awareness, particularly for female adolescents, to reduce morbidity and mortality associated with T1D and eating disorder comorbidities. Kim et al. [38] investigated the link between skipping breakfast and T1D prevalence among adolescents. A randomized controlled trial revealed that a rice-based breakfast improves fasting insulin levels and insulin resistance (HOMA-IR) in adolescents prone to skipping breakfast, suggesting a positive impact on metabolic health and a potential reduction in diabetes risk. Conversely, skipping breakfast was linked to an increased risk of diabetes. T1D prediction requires precise blood glucose (BG) control, which is closely tied to dietary choices. Nutritional factors have been integrated into BG prediction models [39], providing longer prediction horizons for nutrient effects on current and future meals. This approach enhances clinical relevance, particularly in outpatient care for adolescents with T1D.

2.4. Glycemic Control

Despite the relatively short duration of T1D and fair to good glycemic control in children and adolescents, significant cardiovascular (CVD) changes can occur [40]. Compared to healthy children, patients with T1D show increased epicardial fat thickness, enlarged left atrial diameter and left ventricular mass, fractional shortening, as well as elevated triglyceride and irisin hormone levels. These findings underscore the importance of regular echocardiographic screening in T1D patients for the early detection of subclinical cardiac dysfunction. Early identification can help mitigate the risk of future complications, including morbidity and mortality. Further research is needed to better understand the role of the irisin hormone in T1D and its potential effects on cardiovascular health in this population. Schweiger et al. [41] highlighted sex-related differences in CVD risk factors among children and adolescents with T1D. A review of the current literature reveals distinct disparities between boys and girls in their CVD risk profiles. Evidence shows that girls with T1D tend to have a more adverse CVD risk profile than boys, with greater exposure to multiple risk factors simultaneously. Girls often experience higher burdens of hyperglycemia, dyslipidemia, obesity, inflammation, physical inactivity, and poor diet compared to boys, placing them at a significantly higher risk of long-term CVD complications. By prioritizing early, tailored prevention strategies, healthcare providers can improve long-term cardiovascular outcomes in pediatric T1D populations and reduce the disproportionate burden of CVD risk among females. Hand tremor-based hypoglycemia detection and prediction in children and adolescents with T1D [42].

In conclusion, further research is needed to enhance early detection, enable personalized treatments, and prevent severe complications in T1D among children and adolescents. Such efforts are critical for reducing morbidity and mortality and improving the quality of life for young patients.

3. Prediction Models for Type 2 Diabetes (T2D)

As already mentioned, T2D is the leading cause of diabetes worldwide and is rapidly increasing, especially among children and adolescents. In the past ten years, the prevalence of T2DM has increased by more than two-fold [43]. The research on diabetes in adolescents, after adjustment for sex, race, or ethnic group, shows that the relative annual increase in incidence of T2D is 4.8% and is almost three-fold to T1D [44].

Research by Saeed et al. [33] has shown that there is an inverse relationship between the age of T2D onset and complication risk and mortality: T2D in children and adolescents has a worse prognosis compared to T2D that develops in adults. Therefore, timely diagnosis is crucial to preventing or delaying complications and improving health outcomes.

The study of Xie et al. [45] assessed the global burden of T2D among adolescents from 1990 to 2019 using data from the GBD 2019, encompassing 204 countries and territories. The findings reveal a significant increase in the global incidence and disability-adjusted life years DALY rates for T2D, with mortality rates showing a modest rise. In [46], it is stated that youth-onset T2D is becoming increasingly common globally and is characterized by a more severe and aggressive course compared to adult-onset diabetes. It is associated with limited treatment options, poorer responses to therapy, and an increased risk of complications at an early age. Therefore, research into the prevalence, pathogenesis, treatment challenges, and phenotypes of T2D in children and adolescents is of critical importance [47]. The pathophysiology of T2D in children and adolescents differs significantly from adult-onset cases, necessitating early identification of at-risk individuals, understanding unique phenotypes, and developing targeted prevention and treatment strategies. The study [48] showed that the effects of treatment and lifestyle interventions in children and adolescents with pre-diabetes or newly diagnosed T2D yielded limited health outcomes compared to those in adults.

Type 2 diabetes (T2DM) is a complex and chronic metabolic disorder characterized by insulin resistance, impaired insulin secretion, and hyperglycemia [49]. The disease has a heterogeneous etiology and risk factors at the social level and behavioral, environmental, and genetic susceptibility [50]. It is associated with serious complications, but the early diagnosis and initiation of therapy may prevent or delay the onset of long-term complications. Chronic complications of diabetes include the development of cardiovascular disease [50], end-stage kidney disease, retinopathy leading to blindness [51], and limb amputations. All these complications contribute to excess morbidity and mortality in patients with diabetes mellitus, with mortality risk increasing over time and with age [52]. Recent studies show that pre-diabetes also increases the risk of cancer [53]. Given the rising incidence and complications of type 2 diabetes, early identification and prevention are crucial. The timely detection of susceptibility to diabetes, better diagnostic criteria, and early identification of individuals with T2D, preferably during younger adolescent years, and as a new management strategy, are essential. There is an urgent need for targeted, gender- and age-specific prevention strategies, healthier lifestyle promotion, and improved management of T2D and obesity, which will help children with T2D to live a long and healthy life [54]. Proactive preventive strategies and aggressive management are critical for minimizing complications and improving long-term outcomes [55].

The study of Collins et al. [56] reviewed and evaluated the methodology and reporting practices used to develop risk prediction models for detecting undiagnosed (prevalent) type 2 diabetes and predicting future (incident) type 2 diabetes. In the review, it is stated that poor methodological practices, including univariate pre-screening, categorization of continuous predictors, and inadequate handling of missing data, were prevalent and undermined the reliability of many prediction models developed since that date. The limitations compromise the accuracy of estimated probabilities for identifying undiagnosed diabetes and predicting future risk. To provide a foundation for creating new models, external validation and recalibration of existing models presented in [45] were performed. The newly developed models incorporate novel risk factors and advanced traditional statistical methods [45]. To enhance the accuracy of new T2D prediction models, the integration of personalized, comprehensive risk factors was suggested, as these are key to improving predictive performance and clinical utility.

In the paper [57], factor analysis was used to explore the underlying structure of metabolic syndrome in T2D. Three key factors were identified: a “metabolic” factor, which includes body mass index (BMI), waist circumference, glucose levels, and triglycerides; an “inflammatory” factor [57] defined by BMI and fibrinogen levels; and a “blood pressure” factor. It should be noted that these factors are crucial for predicting T2D not only in adults but also in adolescents and children, making them significant for predicting T2D. The results suggest that chronic inflammation is associated with insulin resistance, highlighting its importance as a component of metabolic syndrome.

The increased risk of developing T2D in children and adolescents is associated with genetics, an obesogenic environment, and obesity, along with the complex interactions among these factors and gender differences, economic status, ethnic groups, etc. [55].

3.1. Obesity—A Global Risk Factor in T2D

Severe obesity stands out as one of the most significant risk factors for youth-onset T2D [58,59]. Obesity is closely linked to the worsening of insulin resistance, a primary feature of the pathophysiology of T2D, along with progressive β-cell and α-cell dysfunction [60]. In children and adolescents, parallel to the increase in obesity, the prevalence of T2D also increases. At the time of diagnosis, it can reach even a value of 77% [60]. For the elimination of obesity in young people with T2D, education is very important, it has to focus on behavioral changes like diet and activity [61]. However, the study [62,63,64] shows that there is non-insulin-dependent diabetes mellitus among children and adolescents. Approximately 20–25% of adolescents have T2D, even without obesity [60].

3.2. Genetic Predisposition

The majority of children and adolescents with T2D have a family history of the disease. It is a highly heritable condition, with 90% of children and youth affected having a first- or second-degree relative who also has T2D [65]: in the TODAY study [17], it is stated that 59.6% of adolescents with T2D had a first-degree relative with diabetes, while 89% had a grandparent affected by the condition. This elevated risk highlights the role of genetic predisposition, environmental factors, and intrauterine influences in the development of T2D. T2D is often diagnosticated in the additional influence is with children born small for gestational age (<2500 g), newborn macrosomia of diabetic mothers (>4000 g), and premature adrenarche in girls (pubic hair appearing before the age of 8 years). An increase in this metabolic disease and also hypertension was reported in boys with a lower birth weight, which supports the theory of programming conditions from the intrauterine life [66].

Finally, in [67], it is stated that obesity combined with genetic predisposition and/or family history is the main risk factor for T2DM in children and adolescents.

3.3. Gender Differences

In review [68], the role of gender in T2D in adolescents is discussed. As sex hormones significantly impact glucose homeostasis, insulin secretion, and action, they have an influence on diabetes progression. It was found that girls at a younger age (up to 15 years), reached a higher stage of puberty compared to boys, and girls at this age had a lower prevalence of diabetic ketoacidosis and higher cholesterol levels compared to boys [69]. However, the SEARCH study [16], which was conducted in the US among individuals 15–19 years of age, gives the opposite result: the incidence of T2D in girls is 21.6 per 100,000, and in boys, it is 14.2. Similarly to the US, the estimated standardized prevalence of T2D in Germany was 1.4 times higher among girls (12.8) than boys (9.0) in the period between 2002 and 2020 [21]. It is worth saying that this finding is opposite to the results obtained for obesity in T2D adolescents: boys exhibited higher odds of obesity than girls—the odds ratio was 2.10 [70]. However, the definition of obesity has to be reexamined—does it adequately consider body composition variations in genders and certain populations [71]. The research produced in China shows that there is no statistically significant difference in the incidence between male and female adolescents [24]. The same result was also obtained in Hong Kong [23].

3.4. Various Ethnic Groups

There are diverse risk factors for T2D among children across the world. In the [65] study, the highest rates of T2D were found in American Indian, African American, Asian/Pacific Islander, and Hispanic youth (in descending order), and the lowest incidence occurred in non-Hispanic, white youth dependently on the specificity of regions/territory and populations. In the review [72] focusing on European countries, it was observed that rates of overweight and obesity among children and adolescents of non-native population are more frequent and render them more susceptible to developing T2D and other metabolic abnormalities than the native ones. The pediatric obesity pandemic of the past few decades has been accompanied by an increase in the incidence and prevalence of T2D in childhood, with a disproportionate disease burden in children of minority ethnic groups [73] (the prevalence of T2D in non-natives is up to two-and-a-half times higher than that of the native-born boys and girls, for example, in Israel [27]). It is found that it is the consequence that adolescents are often obese, lead sedentary lifestyles, and belong to lower socioeconomic groups. Due to T2D, this group faces a heightened risk of early and frequent microvascular and macrovascular complications. To overcome the problem, there is an urgent need to provide special attention for the prevention and early detection of T2D in minority ethnic groups.

3.5. Socioeconomic Status

Research indicates significant socioeconomic disparities in children and adolescent T2D prevalence depending on the country. In the United States, 23.6% of adolescents with T2D belong to a high socioeconomic status (SES) group, while in India, the figure reaches 88.5% [26]. A study conducted in England and Wales between 2009 and 2016 on children and youth under 19 with T2D found that less than half lived in the most disadvantaged areas [36]. Low SES is a well-documented risk factor for obesity and T2D, primarily due to limited access to nutritious foods and lower levels of physical activity. In contrast, in developing countries, higher SES groups tend to engage in less physical activity and consume greater amounts of fat, salt, and processed foods compared to those with lower SES [51]. It is especially seen in populations in urban areas in comparison to rural areas: the annual incidence of T2D in urban areas is twice as high as in rural areas [12]. In economically thriving regions undergoing industrialization and urbanization, dietary shifts and declining physical activity have contributed to increased pediatric obesity rates, which is a recognized risk factor for T2D. In 2019, research concluded that low–middle and middle Socio-Demographic Index (SDI) countries had the highest incidence and disability-adjusted life year (DALY) rates, while low-SDI countries exhibited a lower incidence but higher mortality rates [14,44]. More recently, the incidence and burden of T2D have risen among children and adolescents, and young adults, particularly in lower socioeconomic countries [74]. A significant proportion of youth with T2D live below the poverty line or come from low-resourced homes [65]. Socioeconomically disadvantaged youth are disproportionately affected by T2D, further exacerbating health inequalities [47].

Beyond obesity as the leading global risk factor for T2D [45], regional socio-demographic factors also play a role, which is influenced by SES and SDI. Household air pollution and low fruit intake are prominent in low-SDI regions, while ambient air pollution and smoking are significant in high-SDI regions. The rising burden of early-onset T2D underscores the need for targeted interventions based on socio-demographic contexts. Without action, disparities will worsen, leading to poorer outcomes. Urgent implementation of predictive models and holistic prevention strategies is crucial, alongside better youth engagement and addressing intergenerational risk factors to reduce the growing burden.

3.6. Other Pre-Diabetes Diagnosis Markers

Future research should focus on sensitive biomarkers, optimized thresholds, and dynamic monitoring in T2D diagnostics [75]. In [74], it is stated that pre-diabetes based on glucose thresholds and glycated hemoglobin A1c level can possibly be predicted. However, it is challenging due to inconsistent fasting glucose and HbA1c glycohemoglobin cutoffs [76]. Pancreatic beta-cell dysfunction and insulin resistance are suitable as they are the key factors contributing to the development of T2D [43]. The screening of hypertension in children would also be a good marker for the prediction and early diagnosis of T2D [19].

4. Models for Differentiating T1D from T2D

Differentiating T1D from T2D remains a challenging task, as the two conditions share numerous overlapping characteristics. Traits such as obesity, insulin resistance, and metabolic syndrome are strongly associated with T2D; however, evidence suggests that beta-cell dysfunction may begin even before impaired glucose tolerance is observed [55]. Patients with T2D typically present with elevated C-peptide levels, the absence of autoantibodies, a strong family history of diabetes, obesity, and markers of insulin resistance, such as hypertension, acanthosis nigricans, or other related features. Understanding the underlying mechanisms of diabetes and testing for its causes are crucial in guiding treatment strategies tailored to the individual needs of patients.

To address this diagnostic challenge, Thomas et al. [77] developed a method for classifying insulin-treated diabetes types using two datasets: the UK Biobank (UKBB) and the Diabetes Alliance for Research in England (DARE). In the UKBB, the proportions of T1D were estimated using a genetic risk score specifically designed for T1D, though it is important to note that genetic predisposition does not entirely rule out the possibility of T2D. In contrast, the DARE dataset defined T1D as severe insulin deficiency and T2D as insulin-treated diabetes. The combination of these datasets resulted in a method that demonstrated higher accuracy, particularly in patients under 20 years of age. While this approach showed promise, further research and optimization are recommended, particularly through the inclusion of additional biomarkers to improve diagnostic precision.

Reitzle et al. [78] also contributed to the effort of differentiating between T1D and T2D by developing an algorithm that uses key characteristics of both types of diabetes. These characteristics include diagnostic parameters, age of onset, and prescribed medications. The algorithm facilitates the classification of diabetes types, while enabling the analysis of important epidemiological indicators and the prevalence of comorbidities. The authors recommended integrating this algorithm into future diabetes surveillance programs to enhance the understanding, monitoring, and management of this condition.

Together, these studies highlight the importance of precise classification methods in diabetes care. Accurate differentiation between T1D and T2D is essential for ensuring appropriate treatment plans and for advancing research into diabetes-related complications [79]. Further development of predictive models, incorporation of novel biomarkers, and refinement of diagnostic tools are necessary to improve differentiation strategies and ultimately enhance clinical outcomes for patients with diabetes.

5. Machine Learning and Deep Learning Models to Predict Diabetes in Children and Adolescents

In 2021, 355,900 new cases of T1D were diagnosed in children and adolescents globally, with projections indicating an increase of over 100,000 cases by 2050 [1]. The delayed diagnosis of T1D is a significant problem, contributing to up to 70% of deaths in those under 25, with diabetic ketoacidosis (DKA) being a life-threatening complication in undiagnosed cases. Diagnosis is challenging because symptoms like fatigue, weight loss, and increased thirst are non-specific. Although blood glucose tests can confirm the condition, the lack of immediate lab results complicates timely diagnosis. Genetic risk scores are underutilized in healthcare due to integration challenges. On the other hand, T2D is a chronic disease that affects millions of people worldwide, and its early diagnosis is a challenge. As already mentioned, genetic, environmental, and lifestyle factors are key to the occurrence of T2D. There are more biological indicators that indicate T2D, but the priority is an elevated level of glucose in the blood as well as insulin resistance [32]. Traditional methods of T2D prediction often rely on subjective judgments, leading to delays and suboptimal outcomes. Recent advancements, such as machine learning (ML) algorithms, have shown promise in identifying T1D and T2D not only in adults but also in children and adolescents. It is believed that this tool could have significant potential to improve early detection of diabetes. As a result, it is expected that 72% of children with diabetes will be flagged 90 days prior to diagnosis. To ensure the benefits for all populations, dissemination over the world is necessary.

Machine learning (ML) and artificial intelligence (AI) are recognized in pathology and laboratory medicine as useful techniques in data preprocessing and as basic supervised concepts for disease prediction [80]. It is also the case for diabetes [81]. The advantage of the application of ML and AI is the possibility of processing a large amount of data and detecting diabetes in individuals [82]. Early identification of diabetes remains critical for enabling preventive actions and improving outcomes [83].

5.1. ML Classifier

Two decades ago, Sisodia [84] suggested the use of ML algorithms for diabetes prediction in children and adolescents. In the study by Giddings et al. [85], it is argued that ML-based prediction models have the potential for improvement. ML, for the prediction of T2D, utilizes tabular data, including demographic, biometric, laboratory, lifestyle, and dietary information [86]. Beyond traditional laboratory biomarkers, the non-invasive dietary data, particularly those which were optimized through grid searches, have shown potential as effective predictors. These models can support innovative strategies, particularly for children and adolescents with limited access to regular medical care, facilitating improved early diagnosis of T2D. However, the applicability of ML results remains constrained. This limitation is primarily due to the lack of inclusion of ethnic diversity and the variability inherent in different healthcare environments.

5.2. Comparison of ML Algorithms

In ref. [87], various ML algorithms for diabetes prediction are presented. The most often applied ML classifiers are as follows: Naive Bayes (NB), k-Nearest Neighbors (k-NN), Decision Tree (DT), Support Vector Machine (SVM), and Bernoulli Naïve Bayes (BNB). The classifiers have different properties. NB is a linear probabilistic classifier based on the simplest Bayesian network model, while BNB is the NB classifier for multivariate Bernoulli models. The k-NN algorithm is a non-parametric, supervised learning classifier, which uses proximity to make classifications or predictions about the grouping of an individual data point. In comparison, NB tends to be faster than k-NN when applied to large amounts of data. DT is an ML algorithm with a hierarchical tree structure. Due to the abundance of ML algorithms that have already been developed, making the appropriate choice is not an easy task. In [88], the comparison of ML algorithms for diabetes prediction is presented. The paper [89] examines the effectiveness of four of them. These are NB, k-NN, SVM, and DT. All these classifiers were evaluated using precision, recall, F1-score, and ROC scores. In the methods, the anthropometric and biochemical data were used. The dataset comprised T2D patients and healthy individuals, incorporating variables such as glycated hemoglobin (HbA1c) [76], BMI [71], fasting blood sugar (FBS), blood pressure [90], and serum lipids (TC, TG, HDL-c, and LDL-c). Among these, NB showed the highest predictive accuracy for diabetic cases, followed by SVM, while k-NN and DT demonstrated lower performance. Recursive feature elimination identified HbA1c, TC, and BMI as the most significant predictors, whereas HDL-c, and LDL-c were deemed less important. In [91], the other five ML models are compared: k-NN, BNB, DT, Logistic Regression (LR), and SVM. In analysis, the features of samples such as glucose levels, BMI, and family history are applied. Among the models, k-NN achieved the highest accuracy and was followed by BNB. Data balancing and dimensionality reduction techniques played a pivotal role in enhancing model performance. While the findings highlight ML’s potential in diabetes detection, the study acknowledges its limitations. The dataset’s small size and demographic specificity constrain generalizability. This underscores the importance of robust datasets for reliable diabetes prediction and classification. In [92], the aforementioned five standard ML classifiers were applied, and HbA1c and fasting plasma glucose (FPG) were considered as input features. The findings highlight the potential of ML algorithms in facilitating the early detection of T2D and mitigating the disease’s impact.

Meganathan [93] examined the two existing prediction models—the Hybrid Prediction Model for T2D and the T2D Mellitus Prediction Model—by introducing a novel ML-based method called the Diabetes Pattern Detection Technique using the Tree Ensemble Clustering Classifier (DDTEC). DDTEC is specifically designed to improve the classification of T1D and T2D. The experimental results highlight the challenges that existing models face in accurately predicting diabetes patterns. In comparison, DDTEC demonstrates superior performance across multiple metrics, including accuracy, recall, specificity, precision, and F-measure. By enhancing early detection and classification, DDTEC establishes itself as a valuable tool for diabetes screening, diagnosis, treatment planning, and holistic patient management. Using supervised ML techniques and key predictors, including physical traits (waist size, leg length, and gender), dietary factors (water, protein, and sodium intake), and demographic variables, valuable insights into diabetes risk in children and adolescents are obtained [94]. Using algorithms such as LR, SVM, RF, XGB tree, and a WVC, the model shows promise in early detection. The limitations include a small sample size and a lack of data distinguishing between diabetes types. Further research was expected to expand datasets, incorporate additional features, and utilize advanced neural networks to improve prediction accuracy and prevention strategies.

5.3. Multilayer ML Algorithm

To eliminate the lack of missing values for predicting diabetes, the Multilayer Perceptron—a neural network composed of multiple layers with a weighted ensemble technique—is introduced [95,96]. This ML algorithm, with a robust prediction framework, integrates outlier rejection, missing value imputation, data standardization, feature selection, and cross-validation. Using this technique, by analyzing large datasets that include demographic, clinical, and lifestyle factors, complex relationships between risk factors for diabetes are revealed [95]. The multilayer ML algorithm is introduced as a novel no-prop [96] to improve the classification of the two types of diabetes mellitus: T1D (caused by insufficient insulin production by beta cells) and T2D (arising from the body’s inability to effectively utilize insulin). Using a multilayer neural network, the algorithm trains each layer separately, employing an attribute-selection process to identify the most relevant features for each patient. Experimental analysis, using a confusion matrix, revealed high classification performance.

5.4. ML Based Ensemble Model

As a promising approach for predicting diabetes in various medical settings [97], the ML-based Ensemble Model, with preprocessing as a key component, is developed. The preprocessing techniques improved dataset quality through feature selection and imputation of missing values, with their effectiveness rigorously evaluated through ablative analysis. The study achieved superior predictive accuracy compared to prior research, even with the use of just four key features: BMI, age, average systolic, and diastolic blood pressure. These features were chosen for their clinical relevance and interpretability. The framework employs a weighted ensemble of ML classifiers, assigning probabilities to outcomes generated by individual models to enhance classification performance.

5.5. Deep Learning Approach

Deep learning (DL) is a modern technology that expands the machine learning (ML) technique, which is a subdomain of artificial intelligence. DL uses supervised deep neural networks for data processing, classification, and analysis of large datasets. It allows for the input of raw data and requires minimal feature engineering in the preprocessing phase. This capability helps in detecting various diseases in their early stages, including diabetes in children and adolescents [87]. Recently, it has been suggested that the DL approach, which is interested in the prediction of diabetes in adults, should be applied in children and adolescents [98]. In [99,100], the utilization of the DL algorithms for diabetes detection was discussed. Since that time, a significant number of research studies have been conducted on DL techniques in diabetes prediction and diagnostics (see [101,102]). The recent research of Naseem et al. [103] gives a highly sophisticated Internet of Things-based approach toward diabetes prediction by using DL models.

5.6. Deep Neural Network

A deep neural network (DNN) is an artificial neural network (ANN) with deep layers. Deep layers means that the network consists of multiple layers: an input layer, a hidden layer, and an output layer. The number of hidden layers is two or more, and they are connected to enable data processing and learning. Based on this, an enhanced DNN diabetes prediction model is developed [104]. The main advantage of the DNN is found in the possibility of it being based on a classification model, which is already in application, and for it to be used to predict the risk [105]. In [106], a case study of DNN leveraging the Pima dataset for diabetes prediction is considered. The supervised DNN for processing, classifying, and analyzing large volumes of data allows for the input of raw data that are specific to the disease. In study [107], the new DL method that utilizes a multilayer perception (MLP) algorithm based on DNN is introduced. The proposed model is built by implementing ten hidden layers and a large number of epochs, taking into account 18 significant features. Various validation metrics were used to ensure the reliability of the results, including cross-validation methods and statistical measures, such as accuracy, F-score, sensitivity, specificity, and the Dice similarity coefficient. The method achieved a very high diabetes prediction accuracy, reaching up to 99.8% [107].

Finally, in review [108], it is concluded that the potential of recent ML and DL models has demonstrated significant promise in predicting diabetes effectively. However, further validation using larger datasets is essential to confirm their reliability and scalability. A key advantage of these models is their potential integration into mobile applications and diabetes management devices, offering improved health monitoring for patients while reducing the risk of complications. Additionally, these models can help identify critical risk factors, supporting early intervention in healthy individuals to prevent the onset of diabetes. Published studies highlight their ability to enhance accuracy and efficiency in managing blood glucose levels, but refinement and testing remain crucial. As ML techniques advance, they are expected to play a central role in disease prognosis, personalized treatment strategies, and preventative care, transforming the landscape of diabetes management and healthcare.

Paper [86] provides an overview of patient and healthcare staff perceptions regarding ML-based predictive models. Generally, perceptions were positive, although there is some uncertainty and skepticism about the accuracy of the models, likely stemming from the lack of sufficiently large datasets for system modeling or the insufficient comprehensiveness of factors. For this reason, further research is necessary.

6. Future Research

In spite of the large amount of research on diabetes prediction in children and adolescents, many challenges remain. Let us mention some of them:

-

The natural history of early-onset diabetes remains poorly understood and explained.

-

Further determining risk factors for diabetes that are specific in children and adolescents and are related to growing up and maturing is necessary.

-

Some of the risk factors are not sufficiently researched yet. This is the case, for example, with the influence of psychological conditions or disorders in adolescence on the occurrence of diabetes; however, there is an opposite effect between the influence of early-onset diabetes on psychological conditions.

-

The effect of the mutual influence of different risk factors during growing up on the occurrence of diabetes in adolescents.

The complexity of early-onset diabetes requires a multidisciplinary approach and collecting a significant amount of data, which represent the risk factors. Modern techniques, along with ML and DL, enable the processing of a large number of data, and the DNN analysis of large quantities that are intertwined with each other. In order to ensure the highest possible prediction accuracy, these AI tools will need to be adapted and adjusted for problem-solving. Thus, future research focused on predicting diabetes in children and adolescents should encompass the following phases:

Identification of Key Variables and Relationships: Incorporate tools that identify the most relevant variables (family history, BMI, age, average systolic and diastolic blood pressure, fasting glucose, and glycol hemoglobin level), and improve understandings of the relationships between different parameters. This step is crucial to enable better model interpretation and improve prediction accuracy.

Development of Interpretable Models: The focus has to be on creating interpretable models that provide clinical insights that are useful for decision-making. Existing models, such as RF, are efficient, but their complexity limits transparency, making it necessary to further refine them for improved usability.

Validation Across Diverse Populations: For achieving the generality of the model, there is a need for the validation of large and diverse datasets across multiple populations (various ethnic groups, different genders, and rural and urban population; with various SES, nutrition, and activity levels). This step is vital to enhance model reliability and ensure the generalizability of different patient groups.

It can be concluded that these new prediction models would need tools for identification, interpretation and validation based on large dataset. A substantial dataset can be obtained from Electronic Health Records (EHRs). However, these records must be preprocessed to address their inherent incompleteness.

On the other hand, for the incomplete dataset with some missing values, another prediction procedure has to be developed in the future. Moving forward, it is essential to find a way to incorporate the missing values into the dataset and the existing models. It is recommended to integrate AI tools and techniques into the preprocessing stage to enhance efficiency and accuracy. To overcome the problem of incompleteness of data or of a small dataset the following recommendations are suggested:

Synthetic data generation: A model of the DNN type for generating completely new synthetic data which could expand the original dataset has to be developed. To generate more authentic new data from the training dataset, this has to be completed. The model must include the already missing values, which require preprocessing (psychology state or disorder indicators [109], specific features of children, and adolescents, etc.). For the appropriate DL architecture, the generative adversarial network is suggested.

Prediction DNN model: A predictive DNN model for diabetes based on the original and synthetic data has to be created. First, the model must be only tested on original data (without synthetic one) and, after receiving confirmation, it needs to be utilized for the complete dataset.

Due to the generalization of the procedure, the final goal of this research would be achieved, and that is found in the realization of an AI model that is specifically tailored for broader clinical applications, with the aim of predicting diabetes in children and adolescents within the context of clinical practices. In addition, future research is expected to integrate semantic methods, local health data, and sensor technologies to enhance real-time predictions, offer comprehensive health insights, and improve early-onset diabetes diagnosis. This system with AI is expected to be capable of diagnosis by itself, and a clinician should not be present anymore.