This paper introduces an innovative AI-powered platform designed to enhance comprehensive diabetes management. The platform leverages advanced machine learning (ML) and deep learning (DL) algorithms to significantly improve the processes of diagnosis, continuous monitoring, and overall patient care. By utilizing a substantial dataset obtained from a Taipei Municipal medical center, the platform integrates a range of AI techniques, such as Logistic Regression, Decision Trees, Random Forest, K-Nearest Neighbor (KNN), Support Vector Machines (SVM), and Long Short-Term Memory (LSTM) networks. These algorithms work in tandem to provide accurate predictions and personalized insights into patient health. Key pre-processing steps ensure high data quality, including handling missing values, assessing the relevance of attributes, and balancing the dataset using the Synthetic Minority Over-sampling Technique (SMOTE). These measures enhance the robustness of the models, resulting in improved prediction accuracy and model performance. Notably, the Random Forest model emerged as a standout performer, achieving an impressive accuracy rate of 92.78%, significantly advancing the accuracy, sensitivity, and specificity of diabetes prediction. The platform is built with a scalable software architecture, complemented by an intuitive user interface that caters to a variety of clinical applications, making it a valuable tool for healthcare providers. This study highlights the transformative potential of AI in revolutionizing diabetes care, empowering clinicians to make informed decisions, and creating personalized treatment plans. Future research aims to expand the diversity of datasets, further refine the AI models, and incorporate real-time patient feedback to optimize the platform's efectiveness.
ing (ML) technologies[
The management of diabetes involves multiple compo- AI in diabetes diagnosis and monitoring. For instance, ML
nents, including early diagnosis, continuous monitoring algorithms have been used to analyze patient data and
of blood glucose levels, lifestyle modifications, and per- predict the onset of diabetes with high accuracy. DL
modsonalized treatment regimens[
cation. Navya Pratyusha Miriyala et al. [18] suggested a di4. Evaluation: To analyze the empirical findings in agnostic analysis of diabetes mellitus using a machine terms of accuracy, sensitivity, and specificity, and learning approach. The study utilized the Pima Indians to discuss the implications for diabetes care and Diabetes Dataset (PIDD) to train six diferent machine future research directions. learning (ML) algorithms, including Naïve Bayes, KNN,
Random Forest, Logistic Regression, Decision Tree, and
The subsequent sections of this paper will provide eXtreme Gradient Boosting (XGBoost). According to the
a detailed literature review, describe the methodology observed experimental data, the Decision Tree algorithm
used in developing the platform, present the results and delivered an accuracy of 85.3%, while XGBoost provided
ifndings, explore the technical aspects of the software the best accuracy at 88.2%. The study suggests that future
architecture, discuss the implications and future research work could focus on handling the sampling strategy to
directions, and conclude with the key takeaways from balance the data, as there is a slight imbalance present.
the study. Jobeda JK et al. [
The study reported varying degrees of accuracy for including K-Nearest Neighbor (KNN), Decision Tree (DT), each algorithm. The Artificial Neural Network (ANN) Random Forest (RF), AdaBoost, Naive Bayes, and XGachieved an accuracy of 72%, Support Vector Machine Boost. The results showed that the Decision Tree Clas(SVM) attained 79%, K-Means Clustering and K-Nearest sifier achieved 85.2% accuracy, the XGBoost Classifier Neighbors (KNN) both reached 77%, Random Forest (RF) achieved 88.8% accuracy, the KNN Classifier achieved showed 80%, and Naive Bayes (NB) achieved the highest 86.2% accuracy, the Random Forest Classifier achieved accuracy at 82%. This research highlights the significant 88.1% accuracy, the AdaBoost Classifier achieved 87.7% potential of machine learning techniques in improving accuracy, and the Naive Bayes Classifier achieved 80.7% the prediction and early detection of gestational diabetes, accuracy. Consequently, the study concluded that the ofering valuable insights for developing more eficient XGBoost Classifier is the best among all the classifiers diagnostic tools. mentioned.
Thotad et al. [
The review also examined how continuous glucose pected continuous glucose monitoring (CGM) values. To monitoring and artificial pancreas devices could aid in evaluate the model, their study utilized the OhioT1DM diabetes management. However, it highlighted the chal- dataset, which includes eight weeks of data from five T1D lenges of dealing with significant fluctuations in blood patients. The proposed method outperformed previous sugar levels and maintaining them within target ranges. approaches, achieving root mean squared errors (RMSE) The authors discussed various deep learning architec- of 6.45 mg/dL and 17.24 mg/dL for the 30-minute and tures, such as Deep Multilayer Perceptrons (DMLPs), 60-minute prediction ranges, respectively. Convolutional Neural Networks (CNNs)[20, 21? ], and Hatice Vildan Dudukcu et al. [25] proposed a method Recurrent Neural Networks (RNNs), that have been used for blood glucose prediction using deep neural networks in diabetes research. They noted that these models excel with weighted decision level fusion, leveraging patients’ at handling complex data but face issues such as limited past BG data to address the challenge of accurately foredata for training and the interpretability of their predic- casting BG levels for diabetic patients. The authors emtions[? ]. ployed three neural network architectures: Long Short
The authors concluded that future advancements in Term Memory (LSTM), WaveNet, and Gated Recurrent deep learning have the potential to significantly improve Units (GRU). They combined these models to enhance diabetes management strategies. prediction accuracy by fusing the outputs of these net
Rahman et al. [
The researchers utilized the Pima Indians Diabetes and 60-minute prediction horizons. The results
demonDatabase (PIDD) to test their Conv-LSTM model against strated that the fusion of the three models yielded the
three other well-known models: CNN-LSTM, Traditional best results for short-term blood glucose prediction, with
LSTM (T-LSTM), and Convolutional Neural Network RMSE values of 21.90 mg/dL for 30 minutes, 29.12 mg/dL
(CNN)[22]. They employed the Boruta method to iden- for 45 minutes, and 35.10 mg/dL for 60 minutes.
tify the most significant features in the data, such as Martinsson et al. [
The study underscores the importance of using ad- engineering or complex data preprocessing. The study vanced methods and feature selection techniques for dia- demonstrated that the model performs comparably to betes prediction. The Conv-LSTM model addressed sev- state-of-the-art methods on the OhioT1DM dataset, useral issues inherent in other LSTM models, such as the ing metrics such as root-mean-squared error (RMSE) and vanishing gradient problem and challenges related to the blood glucose-specific surveillance error grid (SEG) temporal data changes. to evaluate performance. Furthermore, by incorporating
Swapna G. et al. [23] present a methodology for the a variance estimation method, the model generates a conclassification of diabetic and normal HRV signals using ifdence measure in the form of a univariate Gaussian disdeep learning architectures. They employed a combina- tribution for every prediction. This feature enhances the tion of convolutional neural networks (CNN) and long interpretability and reliability of the forecasts, allowing short-term memory (LSTM) networks applied to HRV users to know when to exercise caution based on predata, achieving an accuracy of 95.1%. The study fur- dicted accuracy. Because this method is computationally ther improves upon this methodology by incorporating eficient, it can be used on devices with low computing a support vector machine (SVM) for classification, which capacity, such as cell phones and CGM devices. increased the accuracy to 95.7%.
3.1.6. Artificial neural networks
In the diagnosis of diabetes we used several Machine ANNs mimic real neural networks by connecting their
learning and deep learning algorithms as follows: artificial neurons in a manner akin to that of the brain
network. The brain, or neural network, is made up of
3.1.1. Logistic Regression (LR) connections between these cells, also known as neurons
[
jected data. The data label represents the constants as 1 and 0. 3.1.2. Decision Trees (DTs)
3.1.3. Random Forest
consists of many decision trees for diferent subjects from
the provided dataset. To boost forecast accuracy, the
algorithm takes the average of subsets from each tree.
Instead of depending on a single decision tree, RF uses
the majority vote from all trees to forecast the result.
Each node in the decision tree answers a query about the
data [26].
3.1.4. K-Nearest Neighbor (K-NN)
KNN is a popular machine learning algorithm that uses
the Supervised Learning approach. According to
Brownlee (2016b), K-NN is commonly used for regression and
classification. The K-NN method compares the
similarities between new and current cases/data. The new case is
allocated to the most similar category from the available
possibilities [28].
3.1.5. Support Vector Machines (SVM)
SVM is non-parametric algorithms that solve regression
and classification problems with linear and non-linear
functions. These functions assign input feature vectors
to an n-dimensional space known as the feature space
[27].
3.1.7. Convolutional Neural Networks
Convolutional Neural Networks (CNN) are a powerful
class of deep learning models that are widely applied to
various tasks, including object detection, speech
recognition, computer vision [? ], classification imaging [ ? ?
] and bioinformatics. CNN is a feed forward neural
network that leverages convolutional structures to extract
features from data. Unlike traditional methods, CNN
automatically learns and recognizes features in data without
the need for manual feature extraction by humans. The
design of CNN is inspired by visual perception. The main
components of a CNN include a convolutional layer, a
pooling layer, and a fully connected layer [
Taipei Municipal medical center, with 15,000 women aged 20-80 as samples. These ladies were hospitalized between 2018 and 2020, as well as 2021 and 2022, with or without a diabetes diagnosis. The study looked at eight patient parameters, including number of pregnancies, plasma glucose level, diastolic blood pressure, sebum thickness, insulin level, BMI, diabetes pedigree function, and age where the patients with diabetic are 5000 and the healthy patients count is 10000. Initial inspection shows an imbalance in the dataset, with more non-diabetic instances than diabetic ones (Figure 1). 3.1.9. Data Pre-processing
substantial correlation, while a value close to zero
indicates no correlation [
addressed any missing values in the dataset by
either eliminating rows/columns with missing 3.2.1. AI Background
data or imputing them using statistical methods. An improved version of recurrent neural networks (RNN)
In our dataset, there were no missing values, as called long short-term memory (LSTM) deals with the
shown in Figure 2. problem of storing long-term dependencies. The LSTM
2. Determining Attribute Relevance: The rele- was first presented by in 1997.The current input in an
vance of each attribute was assessed using Pear- LSTM network at a given moment in time step and the
son’s correlation coeficient, illustrated in Figure output from the preceding time step are supplied to the
3. This method calculates a correlation coeficient Long Short-Term Memory (LSTM) unit, which produces
between − 1 and 1 to quantify the relationship an output that is forwarded to the subsequent time step.
between input and output properties. A coefi- For categorization purposes, the last hidden layer of the
cient value above 0.5 or below − 0.5 indicates a last time step—and occasionally all hidden layers—are
frequently used [
where the input and output gate vectors are denoted by the letters f, i, and o, respectively. W , w, b and ⊗
By including three gated units—a forget gate, input represent weights of input,weights of recurrent output,
gate, and output gate—that allow for efective control bias and element-wise multiplication respectively.
over the memory of previous states, LSTM circumvents
the vanishing gradient problem in RNN .Based on the 3.2.2. Dataset
current input and the prior internal state, the input gate
determines how to update the internal state. How much
of the prior internal state should be lost is decided by the
forget gate. Lastly, the output gate controls how much
the internal state afects the system [
• Study 1: Extracting glucose level. Figure 6: preprocessing steps . • Study 2: Extracting glucose level and carbohy
drates to analyze their efect on glucose levels. • Study 3: Extracting glucose level, carbohydrates, 4. Results and Findings and steps. • Study 4: Extracting glucose level, carbohydrates, 4.1. Diagnosis of diabetes and quality of sleep (1 for Poor, 2 for Fair, 3 for
Good). 4.1.1. Performance criteria
• Study 5: Extracting glucose level, carbohydrates, Model Evaluation Metrics The following measures
and intensity of exercise (on a scale of 1 to 10, were utilized to assess the suggested model. When
makwith 10 being the most physically active). ing predictions about occurrences, there will be four
cat• Study 6: Extracting glucose level, carbohydrates, egories of outcomes[
quality of sleep, and intensity of exercise.
Handling Missing Values Missing values for extracted data expected glucose level are filled with -1, indicating nothing was recorded at those times. Rows with missing glucose levels are dropped to maintain data integrity.
Loading and Splitting Data The integrated data for each patient is loaded, with timestamps converted to datetime objects. The data is split into training (80%) and testing (20]%) sets and then consolidated into two comprehensive data frames: integrated_train_data and integrated_test_data.
Acc =
TP + TN TP + TN + FP + FN
Sensitivity (Recall) = Specificity (Sp.) = TP TP + FN
TN TN + FP
Data Scaling The data is normalized using MinMaxS- by the classifier and is expressed as: caler to ensure all features are on a similar scale. (7) (8) (9)
Precision (Pr.) =
TP TP + FP
The F1-Score (F1.,) represents the precision and recall
harmonic mean, with a range of [
1 1 1 + 4.1.2. Machine learning results (10) (11)
the performance of the five machine learning models.
The Area Under the Curve (AUC) measures the model’s ability to discriminate between classes, with a higher AUC indicating better performance. As shown in Figure 4.1.3. Deep learning results 7 , the logistic regression model (blue curve) has a mod- Artificial Neural Network The following table 3 proerate AUC, indicating adequate but not optimal perfor- vides a detailed overview of the architecture and training mance. The decision tree model (orange curve) performs details of the Artificial Neural Network (ANN) employed slightly better with an AUC of 0.89. The random forest in this study. model (green curve) demonstrates the best performance with an AUC of 0.98, indicating excellent classification • Results Before Enhancement Strategies The power. The support vector machine (SVM) model (red initial performance of the ANN was evaluated curve) also performs well with an AUC of 0.94. The K- using the original architecture and training setup. Nearest Neighbor (KNN) model (purple curve) shows The plots in Figures 9 and 10 indicates that the good performance with an AUC of 0.92, although it is model performs well and learns eficiently. Both slightly less eficient than the random forest and SVM the training and validation accuracy curves indimodels cate a consistent rise, beginning at 0.75 and attain
Based on the evaluation metrics in Figure 8, the Ran- ing 0.93 after 50 epochs. The validation accuracy dom Forest classifier stands out as the best-performing closely tracks the training accuracy, indicating model for diabetes prediction in this study achieved the high generalization with minimal overfitting. The highest accuracy (92.78%) . Building the Neural Network Model Compilation Callbacks Enhancement Strategies red dashed line at 0.93 represents the best validation accuracy attained, and the model’s accuracy plateaus around this value after about 20 epochs, suggesting convergence. The loss curves for both training and validation data fall significantly in the early epochs before stabilizing at low values, showing efective learning and minimal overfitting. Overall, the closely aligned training and validation curves for both accuracy and loss show that the model generalizes efectively to unknown data. • Results After Enhancement Strategies Following the implementation of enhancement strategies, the performance of the ANN was reevaluated.
Based on the plot in Figure 11 The validation accuracy curve, although plateauing after a certain epoch (around 30), still remains high throughout.
This suggests the model has achieved a good level of performance on unseen data. Even though it might not be significantly improving after that point, it’s maintaining a strong performance overall. From the classification report in Figure 12 we observe that the Precision is high for both classes, at 0.94 for class 0 and 0.92 for class 1.Recall (how many of the actual positive cases did the model predict correctly) is also high for both classes, at 0.96 for class 0 and 0.88 for class 1. This means that the model is good at not missing actual positive cases.finally the Accuracy is 0.94, which is also high. This means that the model is performing well overall.
around 0.93, which is a positive sign. The training loss and validation loss generally decrease over time, which is what you expect as the model learns. which suggests the model is generalizing well to unseen data. Based on confusion matrix and classification report the model performs well in classifying occurrences, as evidenced by its 93% overall accuracy. It shows good recall (0.94) and precision (0.95) for class 0, indicating that true negatives can be identified efectively, however there are some misclassifications as class 1. Class 1 precision (0.89) and recall (0.92) are marginally poorer, suggesting that some cases were incorrectly classified as class 0. Overall, the model works well in both classes; however, it could be even more accurate if it could be optimized to accurately categorize instances of class 1.
The plot in Figure 15 shows the accuracy of the model over 100 epochs for both training and validation datasets.The validation accuracy is consistently higher than the training accuracy, indicating that the model is performing well on unseen data. In the right plot, both training and validation losses decrease rapidly, indicating efective learning. The close alignment of training and validation losses suggests the model generalizes well and does not sufer from significant overfitting.The confusion matrix (Figure 16 ) demonstrates that the model has a large number of right predictions for both the negative and positive classes, indicating good performance. 4.2.2. Lstm model
Compiling the Model Training the Model Evaluating the Model
Details - Optimizer: Adam - Learning Rate: 0.001 - Loss Function: Mean Squared Error (MSE) - Epochs: 100 - Batch Size: 32 - Callbacks: - Early Stopping: Monitors validation loss, patience=10 - Model Checkpoint: Saves best model based on validation loss to best_model_g.keras - Metrics: - Root Mean Squared Error (RMSE) - Mean Absolute Error (MAE)
based in figure 24 in all studeis the validation loss generally remains lower than the training loss after the initial epochs, indicating good generalization performance. The consistent patterns across diferent studies suggest robustness in the model training process.
Based on Table 6 provided for Mean Absolute Error (mae) and Root Mean Squared Error (rmse) for each study, here’s a comparison of models performance. Generally, as more features are added to the model, the MAE tends Figure 17: Classification report. to decrease, indicating improved accuracy in predicting blood glucose levels. Models including sleep quality and exercise intensity show slightly lower MAE values com4.2. Monitoring of diabets pared to those with fewer features. Similarly, RMSE decreases as more features are incor4.2.1. Performance criteria porated, suggesting better overall predictive performance. We utilize two standard performance metrics: root-mean- Models with more features show lower RMSE values, insquare error (RMSE) and mean absolute error (MAE). Let dicating more accurate predictions. be the actual value, ˆ the predicted value, and the Including additional features such as carbs, steps, sleep sample size. quality, and exercise intensity consistently improves prediction accuracy (lower MAE and RMSE). However, the diferences between models with more features compared ⎯ to those with fewer are relatively small but generally conRoot-Mean-Square Error (RMSE) : RMSE = ⎷⎸⎸ 1 ∑︁(sis− te nˆt.)2 =1 (12)
This section discusses the prototype implementation of the AI-powered platform for comprehensive diabetes management, focusing on its architecture, key components, and functionalities that are important for ensuring the platform’s scalability, usability and efectiveness in real-world clinical settings.
The software architecture of the AI-powered diabetes management platform consists of several key components that collaborate to collect, process, store, and analyze data, providing valuable insights and supporting clinical decisions. The general architecture is illustrated in Figure 25 and includes the following steps: and deep learning algorithms, the proposed system effectively predicts diabetes onset, monitors glucose levels, and assists healthcare professionals in providing personalized care.
The performance evaluation indicates that the DNN achieved a validation accuracy of 94%, showcasing its robustness and generalization capabilities. Enhancement strategies, including increased model complexity and k-fold cross-validation, further improved the model’s performance, ensuring minimal overfitting and high precision. Similarly, the LSTM model demonstrated a strong ability to predict blood glucose levels, with validation losses indicating good generalization to unseen data.
Moreover, the inclusion of user-friendly interfaces for healthcare professionals and patients ensures that the platform is accessible and practical for everyday use. This fosters better communication between patients and doctors, streamlining the management of health records, prescriptions, and medical analyses.
Future work will focus on expanding dataset diversity, refining AI models, and incorporating real-time patient feedback to further optimize the platform, ultimately improving clinical decision-making and personalizing treatment plans for diabetes care. insights to assist with clinical decisions, which may include changes to prescriptions, lifestyle advice, or scheduling further appointments
The foremost aim of the UI design is to create a clean and easy-to-use enjoy for users. This method making the interface simple to navigate, ensuring all functions are smooth to discover, and displaying data in reality. By focusing on user-friendly layout, the platform objectives to improve consumer engagement and satisfaction. The following figures ofer an overview of the platform’s person interface:
This study assessed various machine learning methods for diabetes management. Using 5-fold cross-validation, the second DNN architecture achieved the highest accuracy of 94%, demonstrating the efectiveness of deep learning techniques for diabetes prediction. Compared to our results, other studies have shown lower accuracies for most algorithms, with their Random Forest model achieving only 80% accuracy, a notable diference from our 93%. This discrepancy can be attributed to diferences Declaration on Generative AI in feature selection, data preparation, or hyperparameter optimization methods. During the preparation of this work, the authors used
In monitoring results, models incorporating additional ChatGPT, Grammarly in order to: Grammar and spelling relevant features beyond glucose levels exhibited slightly check, Paraphrase and reword. After using this tool/serbetter predictive performance in terms of MAE and RMSE. vice, the authors reviewed and edited the content as However, the diferences between models were relatively needed and take full responsibility for the publication’s minor, suggesting diminishing returns as more features content. are added.
It is important to consider that diferences in datasets used across studies can significantly impact results. Vari- References ations in data characteristics, such as sample size, demographics, and data quality, may influence machine learning model performance. While our study shows promising results, future research should focus on reifning models, exploring advanced feature engineering methodologies, and validating these strategies across diverse datasets to ensure robustness and generalizability.
Future work may also involve developing new AI models for predicting diabetes-related complications and risk assessments and integrating wearable devices for real-time monitoring to enhance analytics capabilities and improve prediction accuracy.