Title: ADAPTOOD: Uncertainty-Aware Fine-Tuning for Out-of-Distribution ECG Time Series Models

Abstract:

There is often a discrepancy between the data utilized for initial training and the samples encountered during fine-tuning and subsequent deployment. While machine learning models hold significant potential, their efficacy is constrained when limited annotated datasets are available. Performance typically deteriorates in the face of distribution shifts stemming from varied sensors, distinct populations, and differing application contexts. Although pre-training offers a foundation, models frequently encounter out-of-distribution (OOD) data in practical scenarios, which undermines their robustness. Current adaptation strategies generally presuppose static distribution shifts and falter when confronted with multiple types or varying degrees of severity. Specifically, these methods often neglect the intensity of the shift; for instance, they may treat adaptation to a large, familiar dataset equivalently to adaptation to a small dataset involving a novel task, thereby hindering generalization.

To overcome these limitations, we introduce ADAPTOOD, a novel framework that employs data uncertainty to measure the severity of distribution shifts and direct the fine-tuning process for time series data. This uncertainty metric quantifies the extent to which samples from the target deployment distribution diverge from the pre-training distribution, serving as a direct indicator of OOD severity. By integrating this uncertainty measure with low-rank model updates and adaptive hyperparameter optimization, our framework enhances the adaptation process. Our results demonstrate that ADAPTOOD surpasses existing methods in OOD tasks, achieving up to 7% greater accuracy and 12.9% higher precision, while sustaining strong performance even as the severity of the distribution shift increases.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC