A comprehensive inference-time augmentation framework in physiological signals: application to PPG-based AF detection

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Standard ITA improved AUROC by up to 8.5% (GPT-PPG) and 0.7% (ResNet); AUPRC improved by up to 10.6% (GPT-PPG) and 0.8% (ResNet). Selective ITA reduced average FPR by up to 4.4% (GPT-PPG) and 1.3% (ResNet) on non-AF datasets

• Primary Outcome: Inference-time augmentation (ITA) consistently improved AUROC and AUPRC for AF detection from PPG signals across all model-dataset combinations evaluated

• Nnt Or Sensitivity: Sensitivity, specificity, PPV, and NPV at defined clinical thresholds not reported; diagnostic performance expressed only as AUROC and AUPRC summary statistics

• Confidence Interval: Not reported

Clinical Application

The ITA framework is described as model-agnostic and requires no retraining, which is operationally attractive for deployment in existing clinical systems. However, computational overhead of 13 augmentation methods applied at inference time requires quantification before integration into real-time monitoring pipelines. Clinical adoption would require prospective validation, regulatory clearance, and integration with existing clinical workflows. AF affects approximately 500,000 Australians and is a leading cause of stroke, making accurate detection a national health priority. PPG-based AF detection via wearables (e.g., Apple Watch, Fitbit) is increasingly used in Australian community settings, though no TGA-cleared PPG-specific AF detection software based on this ITA framework currently exists. The RACGP supports opportunistic AF screening in primary care, and digital health tools are increasingly incorporated into Australian cardiovascular guidelines. PBS does not currently subsidise wearable-based AF monitoring, though MBS item numbers exist for Holter monitoring. This framework, if prospectively validated in diverse populations including Aboriginal and Torres Strait Islander communities (who have higher cardiovascular risk), could support scalable AF detection. TGA approval as a Software as a Medical Device (SaMD) under the Therapeutic Goods Act would be required before clinical deployment. Adults undergoing PPG-based cardiac monitoring where AF detection is clinically indicated, particularly in settings where wearable or remote monitoring devices are deployed and model retraining is not feasible (e.g., continuous hospital monitoring, community-based AF screening programmes)

Abstract

Objective.Accurate classification of physiological signals in real-world deployments is challenged by sensor noise, motion artifacts, and distribution shifts between training and deployment data. Inference-time augmentation (ITA), which applies augmentations during inference rather than retraining, offers a simple, model-agnostic mechanism to improve robustness. However, ITA application to physiological signals has remained narrow in scope, relying on limited augmentation methods with fixed, unoptimized parameters. This work proposes a unified ITA framework to address that gap.Approach.The framework incorporates 13 augmentation methods spanning time-domain, amplitude-domain, frequency-domain, and artifact-injection transformations, with hyperparameters systematically optimized via Bayesian optimization. We evaluate the framework on atrial fibrillation (AF) detection from 30 s photoplethysmography (PPG) signals using two deep learning architectures, generative pre-trained transformer (GPT)-PPG (in two sizes) and ResNet, across five datasets comprising more than 400 patients and ∼9800 h of PPG recording. Two evaluation strategies are assessed: standard ITA applied to all inputs, and selective ITA applied to initially positive predictions.Main results.Standard ITA consistently improved the area under the receiver operating characteristic curve (AUROC, up to 8.5% for GPT-PPG and 0.7% for ResNet) and the area under the precision-recall curve (AUPRC, up to 10.6% for GPT-PPG and 0.8% for ResNet) across all model-dataset combinations. Selective ITA further reduced average FPR by up to 4.4% (GPT-PPG) and 1.3% (ResNet) on non-AF PPG datasets.Significance.These findings establish ITA as a practical, model-agnostic approach for improving the reliability of PPG-based AF classification in deployment settings where retraining is not feasible, with broader applicability to physiological signal analysis.

References

1. 1.Fattahi, D., Yan, R., Kataria, S., Chen, Z., & Hu, X. (2026). A comprehensive inference-time augmentation framework in physiological signals: application to PPG-based AF detection. Physiological Measurement. https://doi.org/10.1088/1361-6579/ae7d82