Evaluating the Performance of Fully Connected, Convolutional, and Recurrent Neural Networks in Diabetes Risk Prediction

Diabetes prediction is a new task in the healthcare system that uses machine learning to find out who is at high risk of getting diabetes. The present study investigates the performance of Fully Connected Neural Networks (FCNN), Convolutional Neural Networks (CNN), and Recurrent Neural Networks in recognition predictive modeling for diabetes using the Pima Indians Diabetes dataset. The dataset contains various attributes such as Glucose levels, Blood pressure, BMI, and Age. A normalization methodology was adopted for the data, followed by splitting it into training and validation sets. A fully connected layer with ReLU activation functions consists of the FCNN model. On the other hand, the CNN model combines convolutional and pooling layers with multiple dense layers. The RNN model contains LSTM layers: Long short-term memory (LSTM), which captures the temporal dependencies. The training process for each model employs the binary cross-entropy loss function, optimized using the Adam algorithm. Based on the results of the ROC AUC (Receiver Operating Characteristic Area Under the Curve), the novel model based on gfDF (Global Feature Descriptors Fusion) achieved the highest performance in determining whether a patient has diabetes, with an FCNN (Fully Convolutional Neural Network) score of 0.87903. Sensible performance of CNN and RNN models with ROC AUC scores of 0.87158 and 0.86116, respectively. To evaluate the prediction capability of each model in a global view, precision, recall, and F1-score are adopted as evaluation criteria. Those results demonstrate the predictive power of neural networks, but importantly also illustrate that different network topologies are better suited for specific regions in prediction space.