Temporal Label Smoothing for Early Event Prediction

TL;DR

This paper introduces Temporal Label Smoothing (TLS), a novel method that leverages temporal dependencies to improve early event prediction accuracy and reduce missed events, especially in medical decision support systems.

Contribution

The paper proposes TLS, a new approach that unifies survival analysis and early event prediction, enhancing performance by focusing on predictive signal regions and maintaining monotonicity.

Findings

TLS outperforms baselines on benchmark tasks

Significant reduction in missed events, up to 50%

Improved event recall at low false-alarm rates

Abstract

Models that can predict the occurrence of events ahead of time with low false-alarm rates are critical to the acceptance of decision support systems in the medical community. This challenging task is typically treated as a simple binary classification, ignoring temporal dependencies between samples, whereas we propose to exploit this structure. We first introduce a common theoretical framework unifying dynamic survival analysis and early event prediction. Following an analysis of objectives from both fields, we propose Temporal Label Smoothing (TLS), a simpler, yet best-performing method that preserves prediction monotonicity over time. By focusing the objective on areas with a stronger predictive signal, TLS improves performance over all baselines on two large-scale benchmark tasks. Gains are particularly notable along clinically relevant measures, such as event recall at low false-alarm rates. TLS reduces the number of missed events by up to a factor of two over previously used approaches in early event prediction.