1 Introduction

Diabetes mellitus (DM) is a chronic metabolic disorder characterized by persistent hyperglycemia and disruptions in carbohydrate, fat, and protein metabolism. It arises due to insufficient insulin secretion, impaired insulin action (insulin resistance), or a combination of both (Singh et al., 2022). A precursor to DM, prediabetes, is a critical health condition marked by elevated blood glucose levels that fall short of meeting the diagnostic criteria for type 2 diabetes (CDC, 2022). Prediabetes is typically diagnosed based on impaired fasting glucose (IFG) levels ranging from 5.7 to 6.3 mmol/L, impaired glucose tolerance (IGT) levels between 7.8 and 11.0 mmol/L, or glycated hemoglobin (HbA1c) levels of 6.4 % to 6.9 % (Barr et al., 2007; Echouffo-Tcheugui et al., 2023). Prediabetes affects approximately 96 million American adults, significantly increasing their risk of developing diabetes, cardiovascular disease, kidney disease, and other associated complications (Schlesinger et al., 2022; Zheng et al., 2018). Globally, an estimated 7.5 % of the population was diagnosed with prediabetes in 2019, a figure projected to rise to 8.0 % by 2030 (Saeedi et al., 2019). The asymptomatic nature of prediabetes often leads to low patient awareness and detection rates, exacerbating poor adherence to treatment.

Evidence suggests that early detection and intervention in prediabetes can substantially reduce the progression to type 2 diabetes. Identifying individuals with prediabetes early facilitates the implementation of preventive measures, such as lifestyle modifications and pharmacotherapy, to delay or prevent the onset of diabetes (Thipsawat, 2023; Tabák et al., 2012). However, current screening methods for prediabetes are often invasive, time-consuming, and costly, limiting their feasibility for large-scale application, particularly among low-income populations.

Machine learning models present a promising and efficient approach for prediabetes screening. By training these models on large-scale population data, they can accurately identify individuals at high risk of developing prediabetes. These models utilize diverse demographic, socioeconomic, and health-related variables to predict prediabetes risk, thereby facilitating targeted screening and early intervention strategies. Numerous studies have successfully developed machine learning models for prediabetes prediction, incorporating a wide range of features, including age, gender, body mass index (BMI), blood pressure, dietary habits, physical activity, family history, laboratory findings, and socioeconomic factors such as income, education, and marital status (Abbas et al., 2021; Li et al., 2021; Rajput et al., 2019; Štiglic et al., 2018; Walker et al., 2016; Wang et al., 2020; Wang et al., 2021). Various algorithms can be employed for predicting prediabetes using a diverse set of features. Logistic regression estimates the likelihood of prediabetes by modeling it as a linear combination of predictors (Kleinbaum et al., 2002). Support vector machines delineate the hyperplane that optimally separates individuals with low risk from those at high risk (Cortes and Vapnik, 1995). Neural networks are capable of capturing complex nonlinear interactions between variables through the use of hidden layers and weight optimization (LeCun et al., 2015). Ensemble methods, such as random forests, enhance generalizability by aggregating the outputs of multiple decision trees, each trained on different subsets of features and samples (Breiman, 2001).

The aim of this study is to develop a machine learning model leveraging extensive data from the Centers for Disease Control and Prevention (CDC) to predict prediabetes. Advanced methodologies, including adjusted feature importance analysis and multivariate logistic regression, were employed to identify the most significant factors associated with prediabetes prediction.

2 Materials and methods

3 Results

3.3

3.3 Features importance

In the analysis of the association between various features and the target variable “PreDiabetes,” five features emerged as the most influential in predicting the occurrence of PreDiabetes. These features, ranked in descending order of importance based on their AMI values, cholesterol status, hypertension status, BMI category, income category, and age group. Cholesterol status demonstrated the highest importance, with approximately 4.03 % higher importance than hypertension status. Moreover, when comparing cholesterol status to BMI category, the third-ranked feature, it displayed a substantial importance that was around 1.5 times higher (141.55 %). Similarly, in comparison to income category and age group, cholesterol status exhibited approximately 2.7 times (272.30 %) and 3.1 times (310.00 %) higher importance, respectively. These findings underscore the significance of cholesterol status as the most influential feature in predicting PreDiabetes compared to the other top-ranking features. Hypertension status, the second-ranked feature, displayed notable importance compared to BMI category, the subsequent feature in the ranking, with a significance approximately 1.4 times higher (136.27 %). Similarly, when compared to income category and age group, hypertension status demonstrated an approximately 2.6 times (261.49 %) and 3 times (297.69 %) higher importance, respectively. BMI category, the third-ranked feature, displayed significance approximately 1.9 times (191.89 %) higher than income category and roughly 2.2 times (218.46 %) higher than age group. The income category is the fourth-ranked feature. When compared to age group, the subsequent feature in the ranking, it displayed an importance approximately 1.1 times (113.85 %) higher. Age group, the fifth-ranked feature, exhibited an importance approximately 1.5 times (146.07 %) higher than education level, which did not make it to the top five in terms of importance. Other features had lower importance compared to the top five features, Fig. 1.

Close

Fig. 1

Feature importance using Adjusted Mutual Information.

Export to PPT

4 Discussion

The descriptive analysis of the study population provides valuable insights into the distribution of key demographic and health-related characteristics in the sample. The relatively equal gender distribution enables an unbiased assessment of prediabetes risk factors across males and females. The heterogeneity across age groups suggests that the data sufficiently captures prediabetes risk across the adult age spectrum, allowing for age-stratified analysis. The fact that married and unmarried people are represented in the data in roughly equal proportions means that we can be confident that the results of the study are not biased toward either group. The educational diversity of the sample population is a strength, as it enables quantifying differential prediabetes susceptibility across education levels, which is a known socioeconomic determinant (Walker et al., 2014). The income variation also permits elucidating income-related prediabetes disparities. A substantial proportion of overweight and obese individuals is expected, given the rising obesity prevalence (Hales et al., 2020). The high rate of physical inactivity aligns with literature indicating that around 43 % of adults are physically inactive (Silveira et al., 2022). The sizable subgroup with hypertension mirrors nearly half of U.S. adults with hypertension (Virani et al., 2021) and enables quantification of its prediabetes association. The elevated cholesterol percentage conforms to research estimating a prevalence of around 36 % (Fryar et al., 2012). The relatively low rate of prediabetes is consistent with the latest global study of prevalence that ranges from 3 % to 13 % (Rooney et al., 2023).

The logistic regression analysis yielded valuable insights into the impact of key demographic, clinical, and socioeconomic factors on prediabetes odds. In line with previous findings, dyslipidemia characterized by elevated cholesterol emerged as the most influential risk factor, increasing the likelihood of prediabetes by 73 % (Bianchi et al., 2008). This underscores the importance of cholesterol screening for prediabetes risk assessment. Higher BMI categories also substantially elevated the odds, emphasizing the well-established dose–response relationship between excess adiposity and prediabetes (Kahn et al., 2005). Hypertension was associated with a 28 % increase in prediabetes odds, consistent with its connections to insulin resistance and metabolic dysfunction (Berbari et al., 2023). Advancing age significantly heightened the odds of prediabetes, reflecting the rising prevalence observed nationally among middle-aged individuals (CDC, 2022). Higher levels of education and income conferred protective effects, reducing the odds by 17.8 % and 19 %, respectively. These reductions likely stem from healthier behaviors and increased access to healthcare among higher socioeconomic groups (Walker et al., 2016). Greater physical activity and being married were also linked to a lower likelihood of prediabetes, suggesting the influence of social and lifestyle factors in mitigating risk. Notably, males had a 40 % reduction in odds compared to females. While premenopausal women are relatively protected from diabetes, males experience higher rates of impaired fasting glucose, contributing to their prediabetes risk (Kautzky-Willer et al., 2016). The higher prediabetes odds among females warrant further investigation into gender-specific trajectories. In summary, the analysis quantitatively demonstrated significant associations between prediabetes and a range of clinical and demographic factors.

The feature importance analysis using AMI scores provided critical insights into the relative predictive capacity of different demographic, clinical, and socioeconomic factors for prediabetes. In alignment with the logistic regression findings, elevated cholesterol was identified as the most influential prediabetes predictor, with importance scores exceeding all other features. This reinforces the value of dyslipidemia screening and monitoring for prediabetes risk assessment (Bianchi et al., 2008). Hypertension ranked second, again validating its significant contribution to risk, likely via mechanisms involving inflammation and insulin signaling dysfunction (Berbari et al., 2023). Higher BMI and older age categories also ranked highly, reflecting the well-established dose–response relationship between excess weight, advancing age, and prediabetes odds (Kahn et al., 2006; CDC, 2022). Lower income status emerged as the fourth most significant feature. This aligns with the socioeconomic prediabetes disparities seen nationally and the heightened susceptibility among disadvantaged groups (Walker et al., 2016).

The comparative feature importance scores provide a quantitative basis for emphasizing medical risks like dyslipidemia and hypertension over demographic factors in prediabetes screening frameworks. However, demographic features may serve as crucial indicators for targeted screening of high-risk subgroups that may have less frequent healthcare contact. For example, lower income groups could be proactively assessed given income category’s relatively high importance compared to education level. Machine learning models leveraging both clinical and sociodemographic inputs could balance sensitivity and precision to optimize predictive performance and cost-effectiveness. Overall, the feature importance analysis provides a quantitative basis for prioritizing medical and demographic screening indicators to enhance prediabetes prediction and early detection efforts.

The present study demonstrates the potential of machine learning algorithms in leveraging key demographic, clinical, and health-related data to predict prediabetes risk. Among the models tested, Random Forest emerged as the top performer, achieving high accuracy and strong discrimination with an AUC ROC exceeding 0.80. The high recall rates of this model highlight its effectiveness in correctly identifying individuals with prediabetes. The robust performance of ensemble tree-based methods like Random Forest for medical prediction has been noted previously (Christodoulou et al., 2019). Random Forest mitigates overfitting risks by aggregating outputs from diverse decision trees trained on subsets of data. KNN also generalizes well by basing predictions on similarity to multiple neighboring points. We identified elevated cholesterol, hypertension, higher BMI, lower income, and older age as the most influential prediabetes predictors. These findings corroborate established evidence on the significance of clinical and sociodemographic risk factors. Dyslipidemia marked by high cholesterol often arises before diabetes onset and indicates insulin resistance (Bianchi et al., 2008). Hypertension is also closely linked to prediabetes through overlapping pathophysiology related to insulin signaling, inflammation, and oxidative stress (Berbari et al., 2023; Usui, 2023). The dose–response relationship between excess weight and prediabetes risk is well-documented, with a 5-fold higher risk above BMI 30 kg/m2 compared to below 25 kg/m2 (Kahn et al., 2006). Lower socioeconomic status marked by income strongly predicts diabetes incidence, likely mediated through poor diet, inactivity, and limited healthcare access (Walker et al., 2016). Prevalence escalates with advancing age, peaking at 45–64 years (CDC, 2022). Our study provides quantitative validation of the relative importance of these known risk factors.

Machine learning models offer distinct advantages over traditional regression when predicting complex multifactorial conditions like prediabetes. While logistic regression evaluates the independent effects of single features, our machine learning approaches capture interactions among risk factors. For example, excesses in BMI and cholesterol likely confer a greater risk when considered together rather than individually. Machine learning models can automatically identify nonlinear relationships, such as the exponential growth of prediabetes with age, which are challenging to specify in advance using regression models. However, a trade-off of machine learning is that the models' decisions are less interpretable than the intuitive outputs of regression models, such as odds ratios from logistic regression. Therefore, in our study, we incorporated logistic regression analysis to provide crucial insights into the effects of key demographic, clinical, and socioeconomic factors on prediabetes odds. Additionally, we included feature importance analysis using AMI scores to gain critical insights into the relative predictive capacity of different demographic, clinical, and socioeconomic factors for prediabetes. To the best of our knowledge, this is the first study to combine these features to predict prediabetes and the first to use multivariate logistic regression in conjunction with AMI to identify the most significant risk factors for prediabetes, in comparison to previous studies (Fujiati et al., 2017; Koopman et al., 2008; Liu et al., 2011; Ouyang et al., 2016; Soo et al., 2017; Wang et al., 2015).

Our study presents several limitations, providing opportunities for future improvement. The models were developed using a single dataset, necessitating further external validation on diverse populations. The BRFSS data relied on self-reports, potentially introducing recall biases. Integrating objective clinical data could enhance prediction accuracy. Additionally, our models were cross-sectional, but incorporating longitudinal data would allow for an assessment of prediabetes trajectories.

5 Conclusion

This study provides valuable insights into prediabetes risk factors through a comprehensive analysis of the study population's demographics, socioeconomic and health-related characteristics. Further optimization in diverse populations over time is warranted to translate these tools towards targeted prevention efforts combating the prediabetes epidemic.

CRediT authorship contribution statement

Saeed Awad M Alqahtani: Writing – original draft, Validation, Software, Resources, Methodology, Investigation, Formal analysis, Data curation. Hussah M Alobaid: Writing – review & editing. Jamilah Alshammari: Writing – review & editing. Safa A Alqarzae: Writing – review & editing. Sheka Yagub Aloyouni: Writing – review & editing, Funding acquisition. Ahood A. Al-Eidan: Writing – review & editing. Salwa Alhamad: Writing – review & editing. Abeer Almiman: Writing – review & editing. Fadwa M Alkhulaifi: Writing – review & editing. Suliman Alomar: .

Funding

This project was supported by the Researchers Supporting Project number (RSP-2024/35), King Saud University, Riyadh, Saudi Arabia, and by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2024R713), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.