A novel interpretable deep transfer learning combining diverse learnable parameters for improved T2D prediction based...

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Accurate deep learning (DL) models for type 2 diabetes (T2D) prediction require effective feature representation and interpretability.
• Existing DL tools often lack interpretability, hindering performance explanation and improvement.

Purpose of the Study:

• To develop an interpretable deep transfer learning (DTL) approach for predicting T2D using single-cell gene regulatory network (SCGRN) images.
• To enhance the performance and provide explainability for DL models in T2D detection.

Main Methods:

• Utilized pre-trained models (SEResNet152, SEResNeXT101) for knowledge transfer in DTL.
• Adapted classification layers with the Adam optimizer for T2D classification from SCGRN images.
• Implemented two DTL strategies: freezing convolutional base weights and fine-tuning consecutive layers.

Main Results:

• The TFeSEResNeXT101 model achieved a highest average balanced accuracy (BAC) of 0.97.
• Outperformed baseline models, which had an average BAC of 0.86.
• Simulation studies confirmed the superiority attributed to Adam optimizer's distributional conformance with pre-trained weights.

Conclusions:

• The proposed interpretable DTL approach significantly enhances T2D prediction accuracy.
• The model's interpretability aids in understanding and improving its predictive performance.
• This method offers a promising direction for developing reliable T2D diagnostic tools.