Deep learning for survival outcomes.

Area of Science:

• Machine Learning
• Biostatistics
• Survival Analysis

Background:

• Deep learning models are widely used for risk prediction.
• Standard deep learning methods are unsuitable for censored data, where outcomes are partially observed.
• Censoring is common in medical research, complicating outcome prediction.

Purpose of the Study:

• To develop novel deep learning algorithms capable of handling censored data.
• To enable accurate estimation of survival probabilities and restricted mean survival with censored outcomes.
• To provide a practical implementation of these algorithms using existing software.

Main Methods:

• Developed a new class of deep learning algorithms specifically designed for censored data.
• Replaced the standard loss function with a censoring-unbiased transformation to account for partial observations.
• Adapted existing software for uncensored data through response transformation techniques.
• Validated the algorithms using simulated datasets and real-world breast cancer patient data.

Main Results:

• The proposed deep learning algorithms effectively handle censored observations.
• The new methods accurately estimate survival probabilities and restricted mean survival.
• Comparisons demonstrate superior or competitive performance against existing risk prediction algorithms.
• Successful application to breast cancer patient data highlights clinical relevance.

Conclusions:

• Novel deep learning algorithms offer a robust solution for risk prediction with censored data.
• These algorithms advance the application of deep learning in survival analysis.
• The methods provide valuable tools for medical research and clinical decision-making.