Efficient Normalized Conformal Prediction and Uncertainty Quantification for Anti-Cancer Drug Sensitivity Prediction...

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Computational Biology

Background:

• Deep learning models are crucial for medical tasks but lack confidence estimates.
• Uncertainty quantification is vital for clinical trust in AI.
• Conformal Prediction offers prediction intervals but struggles with heteroskedasticity.

Purpose of the Study:

• To develop a method for accurate uncertainty estimation in deep learning models.
• To improve the efficiency and coverage of conformal prediction intervals.
• To enhance deep learning model trustworthiness in medical applications.

Main Methods:

• Utilized a Deep Regression Forest to estimate sample-wise uncertainty.
• Calculated variance from the Deep Regression Forest for uncertainty quantification.
• Applied normalized inductive conformal prediction with the proposed uncertainty estimation.

Main Results:

• The Deep Regression Forest variance significantly improved conformal prediction efficiency.
• Enhanced coverage of prediction intervals was observed.
• The method demonstrated effectiveness in anti-cancer drug sensitivity prediction.

Conclusions:

• Deep Regression Forest variance is a powerful tool for uncertainty estimation.
• This method enhances the reliability of deep learning in critical medical tasks.
• Improved uncertainty quantification can increase medical practitioners' trust in AI models.