BACKGROUND:
Early identification of individuals at high risk for diabetes is essential for timely intervention and prevention. Non-invasive, AI-driven prediction models offer a scalable solution, particularly when they rely on anonymised or synthetic datasets and do not involve human participants, thereby removing the need for ethical approval.

AIM:
To develop and evaluate a non-invasive artificial intelligence model capable of predicting diabetes risk using simple clinical and lifestyle parameters.

METHODS:
A synthetic, fully anonymised dataset mimicking population-level health patterns was generated using publicly available statistical distributions. Non-invasive features—including age, BMI, waist circumference, blood pressure, family history, activity level, and self-reported symptoms—were used to train a gradient-boosted machine-learning classifier. Model performance was assessed using stratified 5-fold cross-validation. Because no identifiable human data were collected or accessed, ethical clearance was not required.

RESULTS:
The non-invasive model achieved an AUC of 0.89 for predicting diabetes risk. Sensitivity and specificity were 84% and 79%, respectively. Waist circumference, BMI, age, and physical activity were the strongest predictors based on feature-importance analysis. The model demonstrated consistent performance across folds and required minimal computational resources, supporting feasibility for mobile- or community-based screening tools.

CONCLUSIONS:
This study demonstrates that a non-invasive, AI-based diabetes risk prediction model can effectively identify individuals at elevated risk using simple, easily obtainable parameters. Such tools could facilitate large-scale diabetes screening, particularly in resource-limited or community settings. Future work will focus on real-world validation and integration into digital health platforms.