Enhanced diabetes prediction using pre-trained CNNs, LSTM, and conditional GAN on transformed numerical data.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Computational Biology

Background:

• Diabetes is a significant public health issue with severe complications.
• Accurate prediction from structured data is challenging due to limited samples and feature diversity.
• Early detection is crucial for effective diabetes management and intervention.

Purpose of the Study:

• To develop and evaluate a deep learning framework for enhanced diabetes prediction.
• To investigate the efficacy of tabular-to-image transformation and generative models in improving prediction accuracy.
• To assess the generalizability of the proposed model on independent datasets.

Main Methods:

• A deep learning pipeline combining tabular-to-image transformation, pre-trained Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks.
• Utilized Pima Indians Diabetes Dataset, transforming numerical features into 2D images and employing Conditional Generative Adversarial Networks (CGANs) for data augmentation.
• Feature extraction using DenseNet201, ResNet152, Xception, and EfficientNetB4, followed by LSTM classification optimized via Bayesian search.

Main Results:

• Achieved 94% accuracy and 98% AUC on the augmented PIMA dataset via five-fold cross-validation, outperforming existing benchmarks.
• Demonstrated comparable performance on the Frankfurt Diabetes Dataset, though generalizability requires further investigation due to sample limitations.
• Results suggest synthetic data may have influenced performance on the PIMA dataset.

Conclusions:

• The proposed deep learning framework shows promise for diabetes prediction using structured biomedical data.
• The methodology may be applicable to other biomedical classification tasks.
• Further validation on large, diverse, multi-institutional datasets is necessary for clinical translation.