Before modeling, the data were standardized using a Z-score transformation to harmonize the scale of the variables and facilitate the convergence of machine learning algorithms [2] . The transformation is given by:

$x_i^{\left(standard\right)}=\frac{x_i-{\mu }_i}{\sigma }$

where $x_i$ represents a value of variable $i$, ${\mu }_i$ is the mean, and ${\sigma }_i$ the standard deviation of this variable. This step ensures that each variable contributes equally to the model.

The main model is based on a linear kernel Support Vector Machine [3] . The principle is to find the hyperplane that maximizes the margin between the two classes. This hyperplane is defined by the decision function:

$f\left(x\right)=w^{\top }x+b$

where:

The optimization criterion is to maximize the margin while correctly separating the classes. This corresponds to solving the following primal problem: $\underset{w,b}{\min }\frac{1}{2}{\Vert w\Vert }^2$

Subject to $y_i\left(w^{\top }{x}_i+b\right)\ge 1$, $\forall i$

where $y_i\in \left\{-1,+1\right\}$ is the target class.

For visualization, a two-dimensional projection via Principal Component Analysis (PCA) [4] was applied: $z=P^{\top }x$.

With PPP being the matrix of the first two eigenvectors (principal components). This projection allows graphical illustration of the hyperplane, defined in

2D by: $z_2=-\frac{w_1}{w_2}{z}_1-\frac{b}{w_2}$

The margins are given by: $z_2=-\frac{w_1}{w_2}{z}_1-\frac{b\pm 1}{w_2}$

For comparison purposes, a multilayer perceptron (MLP) was also implemented [5] . The network consists of one or more fully connected hidden layers, each neuron being defined by:

$a^{\left(l\right)}=\sigma \left(W^{\left(l\right)}{a}^{\left(l-1\right)}+b^{\left(l\right)}\right)$

where:

For the output layer, a sigmoid function was used to model the probability of belonging to class 1 (diabetic):

$\hat{y}=\sigma \left(z\right)=\frac{1}{1+{\text{e}}^{-z}}$

with $z=a^{\left(l\right)}=\sigma \left(W^{\left(l\right)}{a}^{\left(l-1\right)}+b^{\left(l\right)}\right)$ where $l$ is the last layer.

The network is trained via gradient descent by minimizing the binary cross-entropy loss function:

$L\left(y,\hat{y}\right)=-\left[y\log \left(\hat{y}\right)+\left(1-y\right)\log \left(1-\hat{y}\right)\right]$

Both models were evaluated on the same data split, with 70% used for training and 30% for testing, stratified according to the target class [6] . The selected metrics are:

$\text{Accuracy}=\frac{\text{VP}+\text{VN}}{\text{Total}}$

A new patient, characterized by their medical and behavioral data, was inserted into both models [8] . The classification produced by the SVM was projected in 2D using PCA to visualize class membership. In the case of the neural network, the result was interpreted through the sigmoid output probability. These approaches facilitate interpretation and allow for intuitive integration into a computer-assisted medical diagnosis process.

The performance of the SVM and Neural Network models for diabetes prediction is presented in the following tables. Table 1 details the metrics by class, including precision, recall, F1-score, and the number of observations per category. Table 2 provides an overview of the overall performance of the models, with indicators

Caption: This table presents the performance of the SVM and Neural Network models by class, based on precision, recall, and F1-score. Support refers to the number of observations in each class.

Caption: This table presents the overall performance metrics for the SVM and Neural Network models. Accuracy measures the proportion of correct predictions, while the F1-scores summarize the trade-off between precision and recall.

Legend: This table presents the confusion matrices of the SVM and Neural Network models. Each cell indicates the number of correct or incorrect predictions for each actual and predicted class. These results allow for the evaluation of classification errors specific to each model.

such as accuracy and average F1-score. Finally, Table 3 shows the confusion matrices, allowing us to identify correct predictions and errors made by each model according to the classes.

The following illustrations show the results obtained with the SVM model applied to our dataset. Figure 1 displays the data projected onto two principal components

(PCA) with the SVM decision boundary. Figure 2 presents the confusion matrix associated with the model, evaluated on the full set of explanatory variables.

The following two figures ( Figure 3 and Figure 4 ) present the results of the Neural Network (NN) model. Figure 3 , “NN Model on 2D PCA,” illustrates the data distribution in a space reduced by PCA. Figure 4 , “Confusion Matrix - MLP Classifier,” evaluates the model’s performance on the full set of variables.

The comparative analysis shows that the linear SVM model offers better overall performance than the multilayer perceptron (MLP) neural network on the diabete_custom.xlsx dataset. The SVM achieves an accuracy of 89%, with a precision of 93% for the non-diabetic class (0) and 81% for the diabetic class (1), while maintaining a high recall of 87% for positive cases. These results reflect a robust ability to effectively detect diabetic patients, which is essential in a clinical context [9] .

In comparison, the MLP model achieves an accuracy of 84%, with slightly lower performance in terms of F1-score and recall metrics. Although this type of model is well-suited for capturing complex nonlinear relationships between variables, it has a major drawback: its lack of interpretability, which is often perceived as a barrier to its adoption in medical settings [10] .

The SVM model presents significant advantages both educationally and medically. From a teaching perspective, it relies on a clear and structured mathematical framework: the concept of an optimal hyperplane, maximum margins, and regularization via the weight norm provide a concrete illustration of the foundations of supervised classification. These features make it a powerful educational tool, particularly for introducing students to explainable algorithms [11] .

From a medical standpoint, the SVM model promotes the interpretability of decisions: practitioners can visualize the data in a reduced space (via PCA) and observe the relative position of patients with respect to the separating hyperplane. This facilitates understanding of the decision-making process, thereby enabling practitioners to justify classifications to patients and healthcare professionals [12] . Such transparency is essential to meet the growing requirements for explainable artificial intelligence in healthcare (XAI - Explainable AI).

Thanks to the absence of prior dimensionality reduction, the SVM model utilized all 14 medical and behavioral variables. The PCA projection analysis highlighted several discriminative dimensions, including:

These results are consistent with the medical literature, which identifies these factors as major indicators in the detection of type 2 diabetes [13] .

Several limitations must be noted:

Several avenues for further development can be considered:

This study demonstrated the effectiveness of a Support Vector Machine (SVM) model applied to an enriched dataset, diabete_custom.xlsx, for diabetes diagnosis. Compared to a multilayer perceptron (MLP) neural network, the SVM exhibited better overall performance, achieving an accuracy of 89%, a macro-average F1-score of 0.88, and an enhanced ability to correctly identify diabetic patients. The SVM approach also stands out for its mathematical clarity, decision transparency (via 2D PCA projection), and alignment with explainability requirements in medical contexts.

In contrast, while the neural network performed reasonably well, it showed a slight shortfall in critical metrics such as recall and precision, highlighting the need for further adaptation to clinical settings and interpretability constraints.

The diabete_custom.xlsx file, derived from an enriched transformation of the Pima Indians Diabetes Dataset, provides a relevant foundation for future research in digital health. It includes not only standard medical parameters (glucose level, BMI, blood pressure, HbA1c) but also behavioral factors (smoking, alcohol consumption, physical activity) that are often overlooked in public datasets. This level of granularity makes it suitable for exploring other supervised or semi-supervised algorithms, as well as approaches such as feature selection, medical clustering, or interactive risk visualization.

To strengthen the results obtained and improve their clinical applicability, several avenues should be considered:

In conclusion, this work demonstrates that explainable machine learning methods, such as SVM, can effectively contribute to the early detection of diabetes, provided they are integrated into a rigorous, transparent, and human-centered approach.

The dataset used in this study, diabete_custom.xlsx, is a derived and enriched version of the Pima Indians Diabetes Dataset, which is available for download at the following address: https://archive.ics.uci.edu/ml/datasets/pima+indians+diabetes

Dua, D., & Graff, C. (2017). UCI Machine Learning Repository: Pima Indians Diabetes Dataset. University of California, Irvine.