Diabetes is a major global health burden that is rapidly expanding and needs to be detected early and prevention

strategies to be effective. Identifying the risk of diabetes early is very important to prevent complications such

as cardiovascular diseases, kidney failure and nerve damage. But, conventional predictive methods aren't always

able to find subtle and complex patterns in patient data, which makes them less effective when it comes to early

diagnosis. Research’s objective is to seek innovative, accurate and strong strategies for early detection of high

risk individuals. The aim of this study is to create an anomaly based machine learning model for diabetes risk

prediction based on physiologic and lifestyle data. Parameters measured and included in the data are key

physiological parameters such as blood glucose, BMI, blood pressure, insulin level, age and cholesterol, as well

as lifestyle parameters such as physical activity, smoking status, alcohol consumption, sleep length, and dietary

habits. The target variable is the outcome of diabetes (positive or negative). For anomaly detection, the study

risk prediction and aids in the detection of anomalies, which may be beneficial for diabetes prevention services

and clinical actions.

Keywords: Diabetes Risk Prediction, Anomaly Detection, Machine Learning, Physiological and Lifestyle Data.

Diabetes is one of the most serious health emergencies affecting life of numerous and millions of people

throughout the world. Chronic metabolic condition with high blood glucose levels due to a lack of insulin or

ineffective use of it (Sapra & Bhandari, 2026). The prevalence of diabetes has rapidly risen over the recent

decades worldwide, attributed to unbalanced diet, inactivity, obesity, aging and increased genetic risk. The

disease has been linked with severe health complications like heart disease, kidney failure, stroke, nerve damage

and vision loss, and is a significant cause of morbidity and mortality in both developed and developing countries.

may suggest future risk. This has made it more imperative to develop intelligent predictive systems that can

effectively process vast amounts of healthcare data efficiently and accurately. Physiological and lifestyle data

are also now considered as important information for disease prediction. The physiological measures that are

used to indicate patient health include blood glucose level, body mass index (BMI), blood pressure, insulin level,

and cholesterol (Mathew & Zubair, 2026). Lifestyle changes, such as physical activity, diet, smoking, alcohol

and sleep are also significant factors in the development of diabetes. These data sources can be combined to

enhance the predictive models of healthcare. The healthcare sector has also seen the rise of new machine learning

and AI technologies that aid in the automated prediction of disease and pattern recognition (Li et al., 2025). One

of these methods, anomaly detection, has proven to be a potential solution for detecting abnormal or hidden

patterns in medical data. Thus, the present work proposes to build the anomaly-based machine learning model

learning classification models are developed for balanced and fully labeled data sets and are not as effective with

real-world health care data. But medical data often has anomalies, missing information, noisy data, and

imbalanced class distributions that can impair prediction accuracy and reliability. This can make it more difficult

for the people who have an increased risk of developing diabetes to be identified at an early stage by the existing

prediction systems (Guo et al., 2025). This restriction impacts on prompt medical interventions and preventive

healthcare efforts. So, there is a need for an abnormal pattern detection approach for physiological and lifestyle

data that can measure abnormal patterns and enhance the early prediction of diabetes risk.

Diabetes is a chronic metabolic disorder that occurs when the body is unable to properly regulate blood glucose

levels due to inadequate insulin production or ineffective use of insulin. There are several different types of

Signs and symptoms include thirst, frequent urination, fatigue, visual changes, weight loss, and failure of wounds

to heal. Diabetes can cause serious problems if it's not controlled, including heart disease, kidney failure, nerve

damage, stroke, vision problems and foot ulcers. There are several physiological and lifestyle risk factors that

play a significant part in diabetes risk. One of the most common signs of insulin regulation problems is elevation

in blood glucose levels. Another significant factor is Body Mass Index (BMI), which is higher in obese persons,

which leads to insulin resistance (Młynarska et al., 2025). Cardiovascular complications that are common among

diabetics include elevated blood pressure and high blood cholesterol. Age is also a factor, older people are at

risk of developing Type 2 diabetes. Other lifestyle factors also affect the risk of developing diabetes. Physical

activity decreases the body's ability to control blood sugar levels well and smoking increases insulin resistance

and poor circulation (Petrie et al., 2018). Heavy drinking can interfere with your body's ability to regulate blood

In the healthcare sector, machine learning has emerged as a revolutionary technology that can be used to create

smart systems that assist in diagnosis, prediction, and decision-making. It is also used in medicine for predictive

analytics, the algorithms are able to process massive amounts of patient information and predict outcomes based

on patterns (Fahim et al., 2025). Predictive models can use historical and real-time data to predict the risk of

diseases like diabetes, heart disease, and cancer, which can help healthcare providers take proactive action and

enhance patient care (Nelson & Chalotte, 2025).

The traditional approach of Artificial Intelligence Disease Prediction systems use machine learning techniques

such as decision trees, SVM, NN, and Ensemble models for disease classification and prediction. These systems

handle large volumes of intricate data, such as patient health records, lab results, imaging data, and lifestyle

details. For diabetes prediction, AI systems can analyze various physical indicators such as blood glucose, BMI

can be trained more and more with the introduction of more data and thus get more efficient over time. However,

there are certain hindrances in using it (Faiyazuddin et al., 2025).

These include problems of data privacy and security, training set bias, difficulty of interpretability of more

advanced models, and a need for significant, high quality medical data. In addition, there could be regulatory

and ethical concerns with the implementation of AI systems in clinical workflows. Despite these challenges,

machine learning continues to play a crucial role in advancing modern healthcare systems.

In this study, the data set is comprised of publicly available healthcare data sets commonly used in diabetes

prediction research. These are datasets from the National Institutes of Health (NIH), the PIMA Indians Diabetes

an individual.

Furthermore, lifestyle risk factors like physical activity, smoking, alcohol consumption, sleep duration and

dietary habits are incorporated because these are important lifestyle factors that can affect the development of

diabetes. The data set contains the target variable which is usually a binary class, indicating whether the person

is a diabetic (positive class) or not (negative class).

This physiological and lifestyle data can be used for more comprehensive analysis and to create an anomaly-

based machine learning model with better accuracy for predicting diabetes risk.

The model performance evaluation plays an important role in the effectiveness evaluation of the diabetes risk

prediction system based on the anomalies. A combination of classification, anomaly detection and computational

metrics are used to ensure comprehensive evaluation in this study. Some key metrics are accuracy (what percent

of the time the model gets predictions right), precision (what percent of the time the model predicts a positive

class, it is correct), recall (what percent of the time the model predicts a negative class, it is correct) and F1 score

(a balance between precision and recall). For anomaly detection performance, ROC-AUC score is used to

of 89.6%, demonstrating its ability to learn complex patterns. Isolation Forest achieved moderate accuracy while

One-Class SVM and LOF achieved lower accuracy. Overall, it is generally found that the hybridization results

in significant improvement in reliability and robustness of the predictions.

means the model had the greatest accuracy in discriminating between diabetic and non-diabetic patients for all

thresholds. The Autoencoder model also had a good performance of 0.91, which indicated good performance of

anomaly detection. The scores of the other methods, such as “Isolation Forest”, “One-Class SVM”, and “LOF”

ranged from 0.83 to 0.88. Overall, the results validate hybrid and deep learning methods being superior in terms

of classification sensitivity and discrimination ability.

Table 9 shows the overall performance of the Hybrid Model in most of the evaluation metrics. It achieved the

best accuracy (92.3%), recall (92.1%), and F1 score (92.2%) which shows that it performed well in identifying

diabetic patients, while maintaining a balanced performance. The Hybrid Model also had the best ROC-AUC

(0.95) score indicating good discrimination between the diabetic and non-diabetic groups. Overall, the Hybrid

Model gave the best performance in predicting diabetes risk while maintaining a high level of reliability and

comprehensiveness, although the Autoencoder had the best precision (90.8%).

indicating the potential of deep learning in uncovering patterns in medical data. The findings also showed that

the physiological parameters like blood glucose, BMI, insulin level, and blood pressure level have significant

influences in diabetes risk prediction. The blood glucose level was determined as the most important predictor,

and BMI as a proxy of metabolic health, were found to be closely related to the occurrence of diabetes. Bad

lifestyle habits also have a significant impact, in that individuals who are inactive, smoke, drink alcohol, and eat

a poor diet develop insulin resistance and poor health in general. The results are in line with recent studies

emphasising the need for clinical and lifestyle data to be correlated for better prediction of diabetes. Previous

studies indicated that traditional machine learning algorithms are less effective in imbalanced data, whereas the

high-risk anomaly-based and hybrid algorithms are more effective in detecting high-risk scenarios. The findings

in this study agree with and replicate the conclusions of the previous study, further indicating that using both the

could lead to an increase in the likelihood of diabetes in a person. For this purpose, a number of anomaly

detection methods, including Isolation Forest, One-Class SVM, Local Outlier Factor, Autoencoders, and Hybrid

Model were implemented and tested on healthcare datasets. The methods used included data pre-processing,

feature selection, model training, and model evaluation based on different performance measures such as

which is higher the more accurate a model can predict. Furthermore, certain important predictive features were

identified in the study: blood glucose level was the most important predictor, followed by BMI, then insulin

level and age and finally blood pressure. Further, it was discovered that lifestyle factors poor diet, smoking,

alcohol, and physical inactivity – also have a significant impact on the risk of diabetes. The contributions of this

research are important in a number of aspects. It adds a more potent anomaly-based approach to the early

diagnosis of disease and improves prediction of healthcare. It also has the potential to support preventive health

the field of medical analytics and artificial intelligence since it demonstrates how the hybrid anomaly detection

models can be applied in medical fields. Finally, the outcome of this research would be helpful for developing

and improving the intelligent diabetes prediction systems in future.

https://doi.org/10.1007/s12652-021-03612-