DiseaseNet: a transfer learning approach to noncommunicable disease classification

Area of Science:

• Biomedical informatics
• Computational biology
• Genomics

Background:

• Noncommunicable diseases (NCDs) represent a major global health challenge.
• Epigenetic modifications, including DNA methylation, are promising biomarkers for NCDs.
• Neural networks can model complex biological data but face challenges with limited NCD datasets.

Purpose of the Study:

• To develop accurate diagnostic and predictive models for non-cancer NCDs.
• To explore the utility of transfer learning for NCD modeling.
• To investigate feature importance using concept-based explanations.

Main Methods:

• Leveraged a neural network pre-trained on cancer data for a transfer learning approach.
• Applied the model to predict arthritis, asthma, and schizophrenia from blood samples.
• Utilized Testing with Concept Activation Vectors (TCAV) for model interpretability.

Main Results:

• Achieved a high overall accuracy (f-measure) of 94.5% in predicting the selected NCDs.
• Demonstrated the effectiveness of transfer learning in NCD prediction.
• Identified key sample sources influencing model performance via TCAV.

Conclusions:

• Transfer learning is a viable and effective strategy for building robust NCD predictive models.
• Epigenetic markers combined with neural networks show significant potential for NCD diagnostics.
• Interpretability methods like TCAV can guide improvements in NCD model training datasets.