A Novel Deep Learning Framework for Diabetes Detection: DiabetesNet

Due to the steadily rising prevalence of diabetes, early detection is essential for both prevention and effective management. Pima Indians Diabetes dataset has been employed in this study for the verification of proposed approach based on machine learning. Eight clinical characteristics, including age, BMI, insulin concentration, and glucose levels, are included in the dataset and are used to determine if a person has diabetes (1) or not (0). In order to avoid overfitting, we suggest a neural network-based model with three hidden layers that uses Swish activation functions and a Dropout layer for regularization. Dataset is divided into two balanced group called training and testing set. Features are getting normalized using the Min-Max scaling. Early Stopping is used to ensure that the best model parameters are maintained by stopping training as soon as the validation loss stops improving. This maximizes the model's performance. We obtain a classification accuracy of 76% when testing the model on the test set, which offers a strong basis for future model enhancements. F1-score, precision, and recall are used as a performance parameter to evaluate the efficacy of proposed model in classifying patients into two classes viz diabetic or non-diabetic. Although issues like class imbalance and model interpretability still need to be addressed, the results show the promise of deep learning techniques in medical diagnostics.