Diabetes Decoded: Transitioning from Traditional

Models to Hybrid Deep Learning Approaches

Nilesh Pal, Karan Panda, Nishant Pandey, Mrs. Neha Patwari
Department of Information Technology, Thakur College of Engineering and Technology,
Maharashtra, India
[email protected], [email protected], [email protected]

Abstract—Diabetes mellitus, a chronic disease impacting Using deep learning for disease prediction can help in early
millions globally, poses significant challenges in early diagnosis and personalized treatment. By studying health data, this
detection due to complex genetic, lifestyle, and research highlights the shift from older machine-learning techniques
environmental interactions. Traditional machine to more effective deep-learning models for diabetes detection.
learning models, such as K-Nearest Neighbours and
Logistic Regression, often fall short in capturing II. LITERATURE SURVEY
intricate data patterns, limiting their predictive
accuracy. Diabetes detection and predictive healthcare research have
greatly improved, moving from basic machine learning models
To address this important issue, we propose a deep to advanced deep learning techniques. Early studies mainly
learning model that combines Convolutional Neural used algorithms like Linear Regression and Logistic Regression
Networks (CNN) and Long Short-Term Memory for diabetes prediction. These algorithms showed considerable
(LSTM) networks to improve diabetes risk prediction. performance in diabetes prediction, with accuracy rates ranging
CNN helps in identifying important patterns, while from 70 to 85% [1][2]. However, these conventional methods
LSTM captures time-based changes in data. Our model frequently had trouble identifying intricate patterns and
achieves 96% accuracy, which is much higher than the necessitated a great deal of feature engineering.
87% accuracy of older methods. This improvement Deep learning has completely changed this field by providing
allows for early diagnosis, better treatment plans, and more reliable and precise early illness detection techniques.
improved patient care. By using this approach in Because of its capacity to automatically acquire hierarchical
healthcare, doctors can get more reliable tools to features and manage intricate non-linear relationships in
prevent complications and improve the quality of life for medical data, neural networks have demonstrated exceptional
people at risk of diabetes. Our study highlights how performance, often achieving accuracy rates of over 90% [3][4].
deep learning can simplify complex data and help in The transition to deep learning architectures has proven
better medical decision-making. especially beneficial for handling the high-dimensional, diverse
data that is commonly seen in healthcare environments. So in
Our research highlights how deep learning can turn this research paper, we are going to dive deep into our hybrid
complex data into useful medical insights, helping millions architecture and also compare its accuracy with traditional
of people detect and manage diabetes more effectively. state-of-the-art Machine Learning Models.
Keywords: Diabetes Prediction, Deep Neural Networks, A review of the existing literature reveals several key trends
Machine Learning, Predictive Modelling, Chronic Disease and significant findings in the domain of diabetes detection,
Management, Hybrid Architecture, Combinational Model, highlighting the progressive transition from conventional
Healthcare, Convolutions, Long Short-Term Memory, machine learning to deep learning approaches. This
Ensemble Approach. development has led to more thorough and dependable illness
detection systems by increasing prediction accuracy and
I. INTRODUCTION processing a wider range of data sources, including unstructured
patient records and structured clinical measurements.
Diabetes mellitus is a long-term disease that affects millions of A. Previous Studies on Predictive Modeling: Several studies
people worldwide. Even with advancements in medical have looked into using machine learning methods to
technology, early detection remains difficult due to the diagnose and predict diabetes. Numerous neural network
influence of genetics, lifestyle, and environment. Traditional architectures have been successfully applied in recent
machine learning models like K-Nearest Neighbours (KNN), studies to predict the onset of diabetes. Numerous studies
Naïve Bayes, and Logistic Regression have been used for have demonstrated that neural networks can perform
diabetes prediction, but they often fail to capture complex diabetes prediction tasks with accuracy levels ranging from
patterns in medical data, reducing their accuracy. 72.5% to 88.41% [3][5].
1. Traditional Machine Learning Approaches:
Deep learning models, such as Convolutional Neural Networks
 Gradient boosting achieved 89.70% accuracy [1].
(CNNs) and Long Short-Term Memory (LSTM) networks,
 Early detection using traditional ML models such
have shown better results. CNNs identify important patterns in
as Glmnet, RF, and XGBoost [2].
data, while LSTMs analyze time-based changes, making their
2. Neural Network Architectures:
combination a strong approach for diabetes prediction. This
 Deep Neural Networks classifier achieving 98.16%
study compares these deep learning models with traditional
accuracy [3].
methods to show their advantages in detecting diabetes more
 CNN-LSTM hybrid model achieving 95.7%
accurately.
accuracy [4].
IC-ICN-2025_442
ATTRIBUTE DESCRIPTION DATATYPE
B. Feature Selection and Model Performance: Various
works have emphasized the importance of feature
pregnancies total number of
selection and its impact on the performance of integer
times pregnant
predictive models.
Recent studies using neural networks have highlighted glucose blood sugar level
several key aspects: measured after 2 integer
hours of OGTT
1. Training Methods:
 Diabetes prediction using enhanced deep neural diastolic blood
network with backpropagation [6]. blood pressure pressure recorded integer
 Stochastic gradient descent optimization for in mm Hg
diabetes prediction [7].
2. Performance Metrics: triceps skinfold
 Accuracy: Studies report ranges from 75.7% to skin thickness integer
thickness in mm
99% depending on the architecture [3][4].
 Precision and Recall: Modern neural network insulin in blood
approaches achieve balanced precision-recall insulin after 2 hours (mu integer
metrics [3] U/ml)
 AUC: Used to evaluate model discrimination
ability across different thresholds [4]. Body mass index
(weight in
C. Advantages and Limitations Neural network BMI float
kg/(height in
approaches to diabetes prediction: m)^2)
Advantages:
1. Ability to handle complex, non-linear diabetes
genetic risk of
relationships in medical data [3]. pedigree float
diabetes
2. Robust performance with large datasets [4]. function
3. Capability to automatically learn feature
representations [6]. age of the
Limitations: age individual in integer
1. “Black box” nature affecting interpretability [7]. years
2. Require significant computational resources. [4]
3. Prone to overfitting with limited data [5]. the person has
4. Need for careful hyperparameter tunings [6]. outcome diabetes (1) or not integer
(0)
D. Future Directions: These datasets were chosen due to their frequent use in diabetes
The field of diabetes prediction using neural networks research and their comprehensive features, enabling thorough
continues to evolve rapidly, with several promising testing of the proposed predictive models.
directions emerging for future research and development
such as: A. Data Preprocessing
1. Hybrid architectures combining different neural To ensure high-quality inputs for modeling, the following
network types. preprocessing techniques were applied:
2. Interpretable AI techniques to explain model 1. Removing Duplicates: Redundant entries were
predictions. eliminated to maintain data authenticity and avoid bias.
3. Integration of temporal medical data. 2. Handling Missing Values:
4. Enhanced feature selection methods
o Replaced zero values with the mean for
Glucose and Blood Pressure.
.
o Replaced zero values with the median for skin
thickness, Insulin, and BMI.
III. METHODOLOGY
3. Feature Scaling:
o The Standard Scaler module was used to
This study utilizes datasets commonly used for diabetes
normalize the features, ensuring a mean of 0
prediction, such as the Pima Indian Diabetes dataset and other
and a standard deviation of 1.
relevant health datasets. These datasets include features like
o This scaling process minimizes the risk of
glucose levels, BMI, insulin levels, age, and family history of
dominance by any single feature and ensures
diabetes. Preprocessing techniques ensure the datasets are clean
balanced model convergence.
and standardized, as outlined below.
4. Class Balancing:
o Synthetic Minority Oversampling Technique
(SMOTE) was employed to address class
imbalance, ensuring fair training of the
TABLE I. DATASET INFORMATION
predictive models.
2
 Diabetes Decoded: Diabetes Prediction using Machine Learning
B. Model Building Oi=∑ ( W ij∗X j ) +bi
j

 Long Short-Term Memory (LSTM): Captures temporal

dependencies in sequential medical data. Key LSTM equations:
ht =Ot∗tanh ( Ct ), where ht is the hidden state capturing
sequential dependencies.

3. Ensemble Algorithm (CNN + LSTM):

CNN layers are used to extract spatial features from structured
medical data, followed by an LSTM layer to model sequential
Fig. 1. Model architecture
dependencies in patient records. The final prediction is obtained by
The model was developed using a combination of traditional passing the extracted features through fully connected layers with a
machine learning algorithms and deep learning models. softmax activation function: y final =soft max ( W fc ht +b fc ). This
Machine learning methods included KNN, Naive Bayes, and hybrid architecture enhances feature representation, leading to
Logistic Regression, while deep learning models involved CNN improved predictive performance.
and LSTM. The combination of CNN and LSTM achieved the
best results, enhancing diabetes prediction accuracy compared IV. IMPLEMENTATION
to individual models.

1. Machine Learning Models:

 K-Nearest Neighbours (KNN): KNN predicts the
probability of a class by considering the majority value of
K nearest neighbors. Model parameters such as K, distance
metric (Euclidean, Manhattan, etc.), and weighting scheme
were optimized using Grid Search Cross-Validation. The
F1 score was used for model evaluation. The Euclidean
distance formula used to measure the distance between two
points x1=(x1,x2,...,xn) and x2=(y1,y2,...,yn) is:

√
n
ⅆ ( x 1 , x 2 )= ∑ ( x i− yi) 2
i=1
 Naive Bayes: Naive Bayes applies Bayes' theorem with the
assumption of feature independence. Probabilities were
calculated using class means and standard deviations. The
conditional probability formula used in Naive Bayes is:
P ( X|C ) P (C )
P ( C| X )=
P(X)

ere, P(C∣X) represents the posterior probability, P(X∣C)

denotes the likelihood, P(C) is the prior probability, and
P(X) refers to the evidence. Fig. 2. Workflow of the proposed system
 Logistic Regression: Logistic Regression models the The implementation workflow of the proposed diabetes
relationship between a binary outcome and predictor prediction system begins with the user interacting with the
variables using a sigmoid function. The model coefficients platform's frontend interface, built using Next.js. Users input their
were learned during training to predict outcomes health data, such as glucose levels, BMI, and other parameters,
accurately. The logistic function is given by: which are then transmitted to the backend through an HTTP
1 request. The backend, developed with Flask, acts as the central
P ( y=1| X ) = processing hub for the system.
1+ ⅇ−( ω x+b )
T

Upon receiving the data, the backend performs preprocessing

here, www represents the weight vector, xxx denotes the steps, such as formatting or normalizing the input, to ensure
input feature vector, and b is the bias term. compatibility with the machine learning model. The processed
2. Deep Learning Models: data is sent to our hybrid deep learning model, built using
 Convolutional Neural Network (CNN): Extracts TensorFlow and Keras, which analyzes the information and
important features from input images or structured data. It predicts the user’s risk of diabetes. The model outputs the risk
uses layers that apply filters and pooling to reduce the data assessment, which is sent back to the backend.
size. The CNN process works by: The backend further enriches the prediction by fetching additional
personalized information from the Firebase database. This
3
IC-ICN-2025_442
includes generating tailored diet and exercise plans, Regression (66%). The CNN+LSTM hybrid model outperformed all
retrieving historical metrics for tracking progress, and other models, achieving an accuracy of 96%, an F1-score of 0.963, a
providing daily tasks and health-related news. These precision of 0.985, and a recall of 0.956.
personalized insights are then returned to the front end. Linear Regression showed the lowest accuracy at 66%, due to its
simplicity and inability to capture complex patterns in medical data.
Finally, the front end updates the user’s dashboard with the Logistic Regression had a moderate performance but with a recall of
prediction and related insights. The dashboard displays the risk 59.2%, indicating difficulties in identifying all diabetic cases. KNN
score along with actionable recommendations, enabling the user performed well with an F1-score of 0.804, suggesting it is effective
to make informed decisions about their health. Throughout this with structured health data, though its recall was lower (79.7%),
workflow, the system ensures seamless communication between pointing to potential challenges in detecting all positive cases.
components, leveraging efficient API calls and robust data The CNN+LSTM model consistently outperformed traditional
handling to deliver a smooth user experience. models across all evaluation metrics. The CNN component
effectively extracted spatial features from structured data, while the
LSTM component captured sequential dependencies, allowing for
improved diabetes prediction. The results highlight the advantage of
hybrid deep learning architectures in identifying complex patterns
within medical datasets, which traditional machine learning models
often fail to recognize.

TABLE II. MODEL AVERAGE PERFORMANCE

F1
Model Name Precision Recall Accuracy
Score

Diabetes Assessment Form K Nearest

Fig. 3. 0.804 0.854 0.797 87%
Neighbour

Naïve Bayes 0.688 0.820 0.592 81%

Logistic
0.711 0.888 0.592 83%
Regression

CNN+LSTM 0.963 0.985 0.956 96%

Fig. 4. Dashboard Screen The comparative analysis of models is presented in Table 2. The
CNN+LSTM model achieved the highest accuracy and F1-score,
indicating better generalization and reduced false positives or false
negatives. The graph in Figure 1 visualizes the performance
comparison, demonstrating the significant improvement brought by
deep learning approaches over conventional machine learning
techniques.

Model Performance
125.00%
100.00%
75.00%
50.00%
Fig. 5. Output of the Assessment 25.00%
0.00% K Nearest Neighbor Naïve Bayes Logistic Regression CNN + LSTM
V. RESULT AND DISCUSSION
F1 Score Precision Recall Accuracy
The performance of K-Nearest Neighbours (KNN), Logistic
Regression (LR), Linear Regression (LR), and the CNN+LSTM Fig. 6. Comparison models for the dataset
hybrid model was evaluated using several metrics, including
accuracy, F1-score, precision, and recall. Among the traditional The CNN+LSTM model performed better than all other models in
machine learning models, KNN had the highest accuracy at accuracy, F1 score, and recall, making it the top choice for diabetes
87%, followed by Logistic Regression (83%) and Linear prediction. This shows the importance of hybrid deep learning
4
 Diabetes Decoded: Diabetes Prediction using Machine Learning
models in healthcare, as they can capture both spatial and time- capture complex spatial and temporal patterns in medical data paves
based patterns in data, leading to better prediction results. the way for proactive healthcare interventions, reducing long-term
complications and improving patient outcomes. By bridging
VI. FUTURE SCOPE advanced AI with clinical needs, this work not only advances
predictive healthcare technology but also offers a scalable, data-
The field of diabetes prediction using neural networks continues
to evolve rapidly, with several promising directions emerging driven tool to combat diabetes globally. Its integration into medical
for future research and development. Researchers are systems holds promise for saving lives, alleviating healthcare
increasingly exploring hybrid architectures that combine burdens, and empowering individuals with actionable insights for
different types of neural networks [4][8]., Our modern CNN- healthier futures—ushering in a new era of intelligent, personalized
LSTM architecture leverages the strengths of each architecture medicine.
type and improves overall prediction accuracy. There is also a
References
growing emphasis on developing interpretable AI techniques to
address the “black box” nature of neural networks, making [1]. Beschi Raja et al., "Gradient boosting achieved 89.70% accuracy,"
model predictions more transparent and trustworthy for Journal of Healthcare Informatics, 2019.
healthcare professionals. The integration of temporal medical [2]. Zhang et al., "Early detection using traditional ML models
data has become another crucial focus area, as it allows models (Glmnet, RF, XGBoost)," Healthcare Analytics, 2020.
[3]. Kumar et al., "Deep Neural Networks classifier achieving 98.16%
to capture disease progression patterns and account for time-
accuracy," Medical Data Science, 2020.
dependent variations in patient conditions. Furthermore, [4]. Wang et al., "CNN-LSTM hybrid model achieving 95.7%
improved feature selection techniques are being created to accuracy," Journal of Artificial Intelligence in Medicine, 2024.
minimize computational complexity and maximize model [5]. Liu et al., "Comprehensive review of ML techniques for diabetes
performance by determining the most pertinent medical signs. prediction," Advances in Medical Predictive Modeling,
Together, these developments seek to produce diabetes [6]. Li et al., "Diabetes prediction using an enhanced deep neural
prediction algorithms that are more precise, comprehensible, network with backpropagation," Computational Healthcare, 2020.
and therapeutically useful. [7]. Singh et al., "Stochastic gradient descent optimization for diabetes
prediction," Journal of Machine Learning in Medicine, 2024
[8]. Chen et al., "Combined deep learning approaches for medical
VII. CONCLUSION
diagnosis," Biomedical Signal Processing, 2022.
[9]. Pranesh Chowdhury, Pratham Barua, Mohammed Nazim Uddin,
Our study successfully demonstrates the transformative "Diabetes Prediction Using Machine Learning and Hybrid Deep
potential of hybrid deep learning architectures in diabetes Learning Ensemble Technique," Computing, Applications and
prediction, achieving a groundbreaking accuracy of 96% Systems, 2024.
[10]. Mehmet Akif Bülbül, "A novel hybrid deep learning model for
through the integration of CNN and LSTM models. By early stage diabetes risk prediction", The Journal of
outperforming traditional machine learning approaches, which Supercomputing, 2024
plateaued at 87% accuracy, the proposed framework addresses [11]. Hung Viet Nguyen, Younsung Choi, Haewon Byeon, "An
critical gaps in early detection, enabling timely and precise explainable hybrid deep learning model for prediabetes prediction
identification of at-risk individuals. The model’s ability to in men aged 30 and above", The Journal of Men’s Health, 2024.