A novel approach for personalized response model: deep learning with individual dropout feature ranking

Area of Science:

• Artificial Intelligence
• Machine Learning
• Bioinformatics

Background:

• Deep learning (DL) offers transformative innovations but lacks interpretability in biology and healthcare.
• Interpretability is crucial for hypothesis-driven research and clinical applications.

Purpose of the Study:

• Propose a novel DL model with individual feature ranking for enhanced interpretability.
• Compare its performance against existing models using simulated and clinical data.
• Enable identification of impactful features at the individual patient level.

Main Methods:

• Developed a novel deep learning model incorporating individual level dropout feature ranking.
• Utilized simulated datasets with correlated and buried features to test the model.
• Applied the model to a publicly available clinical dataset.
• Compared performance against Artificial Neural Network, Random Forest, and population level dropout models.

Main Results:

• The individual level dropout feature ranking model demonstrated reasonable prediction accuracy.
• It successfully identified impactful features at the individual level, unlike population-level models.
• This approach allows for patient subgroup identification based on feature impact.

Conclusions:

• The proposed model enhances DL interpretability in complex datasets.
• It facilitates personalized medicine by identifying individual-specific impactful features.
• This offers a new tool for patient stratification in clinical drug development and personalized therapies.