Primal--Dual Alternating Neural Learning for Timely Classification with Performance Guarantees

TL;DR - This paper presents a recurrent neural network method for classifying clinical risks early while explicitly controlling sensitivity, specificity, and monitoring cost. A primal–dual optimization scheme provides sequential decision rules that satisfy prespecified performance constraints.

• Frames timely classification as a multi-objective sequential optimization problem balancing immediate decisions against collecting more observations.
• Derives a value recursion that determines whether to classify at each time point or continue monitoring.
• Trains an RNN to approximate evolving value processes while primal–dual updates enforce sensitivity and monitoring-cost constraints.
• Demonstrates the approach through simulations and hypoglycemia prediction using continuous glucose-monitoring data.