Developing machine learning algorithms for dynamic estimation of progression during active surveillance for prostate...

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Active Surveillance (AS) is a prostate cancer management strategy involving continuous monitoring for disease progression.
• Current AS lacks robust methods for real-time updates on individual progression risk.
• This gap necessitates advanced predictive modeling for personalized patient care.

Purpose of the Study:

• To develop and validate a novel deep learning model for predicting prostate cancer progression during AS.
• To enhance the accuracy of risk prediction by incorporating longitudinal follow-up data.
• To identify distinct patient subgroups with similar progression trajectories.

Main Methods:

• Development of a Dynamic-DeepHit-Lite (DDHL) model, a deep learning-based longitudinal survival model.
• Application of Actor-Critic temporal predictive clustering (AC-TPC) to identify outcome-related patient clusters.
• Validation using data from 585 men on AS with a median follow-up of 4.4 years.
• Comparison of DDHL model performance against Cox regression and landmarking methods using C-indices and Brier scores.

Main Results:

• The DDHL model demonstrated improved predictive performance with additional follow-up data compared to baseline models.
• With 3 years of data and follow-up, DDHL achieved a C-index of 0.79, outperforming landmarking Cox (0.70) and baseline Cox (0.67).
• AC-TPC identified 4 distinct temporal clusters, with progression risks ranging from negligible to 50% by 5 years in the highest-risk group.

Conclusions:

• A novel machine learning approach, DDHL combined with AC-TPC, significantly enhances personalized follow-up for prostate cancer AS.
• The model's predictive power increases with more longitudinal data, offering dynamic risk assessment.
• This approach supports tailored clinical decisions by identifying patients with varying progression risks.