Adaptive risk prediction system with incremental and transfer learning.

Area of Science:

• Medical Statistics
• Genetics
• Machine Learning

Background:

• Prenatal risk assessment for fetal aneuploidies relies on static, multivariate probabilistic models.
• These models, developed from extensive research, are seldom updated to reflect evolving population characteristics in screening labs.

Purpose of the Study:

• To develop an adaptive risk prediction system (ARPS) that automatically updates risk models.
• To address the limitations of static models in capturing changing patient population characteristics for prenatal screening.

Main Methods:

• Developed a distribution shift detection method using 8 years of real-life Down syndrome screening data.
• Implemented a probabilistic risk modeling system that adapts to new data upon detecting population changes.
• Evaluated various transfer and incremental learning systems, including a novel Incremental-Learning-Population-to-Population-Transfer-Learning design.

Main Results:

• A windowed approach for distribution shift detection proved computationally efficient without performance loss, particularly with transfer learning.
• The best-performing ARPS design achieved prediction performance comparable to manually updated algorithms.
• Demonstrated that ARPS can operate effectively without human intervention.

Conclusions:

• Adaptive risk prediction systems (ARPS) can maintain high prediction performance for fetal aneuploidies without manual updates.
• ARPS offers a viable solution for dynamic adaptation to population changes in screening labs.
• The proposed methods have potential applications beyond fetal aneuploidy screening to other population screening challenges.