New Framework Uses Inference-Time Augmentation to Improve Atrial Fibrillation Detection from Wearable Sensors
Researchers developed a comprehensive framework that applies data augmentation techniques during inference—rather than retraining—to improve the accuracy of atrial fibrillation detection from photoplethysmography (PPG) signals collected by wearable devices. The framework tested 13 different augmentation methods optimized via Bayesian optimization across five datasets with over 400 patients. The approach achieved improvements of up to 8.5% in detection accuracy for transformer-based models and reduced false positive rates, offering a practical solution for deploying AF detection systems without model retraining.
Researchers at arXiv have proposed a unified inference-time augmentation (ITA) framework designed to improve the robustness of atrial fibrillation detection from photoplethysmography (PPG) signals—a non-invasive measurement technique commonly used in wearable devices. The framework addresses a key challenge in real-world medical device deployment: sensor noise, motion artifacts, and distribution shifts between training data and actual deployment conditions. The researchers incorporated 13 augmentation methods spanning time-domain, amplitude-domain, frequency-domain, and artifact-injection transformations, with hyperparameters optimized using Bayesian optimization. Evaluation across five datasets comprising more than 400 patients and approximately 9,800 hours of recording showed that standard ITA consistently improved detection metrics, with area under the receiver operating characteristic curve (AUROC) improvements up to 8.5% for transformer-based models and up to 0.7% for ResNet architectures. The selective ITA approach further reduced false positive rates by up to 4.4% on non-AF datasets, establishing ITA as a practical, model-agnostic method for improving physiological signal classification in deployment settings where model retraining is not feasible.
Limitations & open questions
The study's limitations and open questions are not detailed in the abstract provided. Specific limitations regarding the generalizability of the framework to other physiological signals, potential computational overhead of inference-time augmentation in real-time deployment scenarios, and whether the optimized hyperparameters transfer across different patient populations or device manufacturers would strengthen understanding of practical applicability.
What different sources said
- arXiv cs.LGCenter
A Comprehensive Inference-Time Augmentation Framework in Physiological Signals: Application to PPG-Based AF Detection
Related
Study suggests asexual reproduction slowed early animal evolution during Ediacaran period
Researchers from the University of Cambridge found that early animals during the Ediacaran period (635-539 million years ago) reproduced asexually through runners, which limited competition and slowed evolutionary diversity. The study used fossil analysis, spatial modeling, and artificial intelligence to examine ancient ecosystems at Mistaken Point in Newfoundland. The findings help explain why animal diversity remained limited for millions of years before a dramatic burst of innovation in the Cambrian period.
UK Science Facilities Face £162m Funding Crisis With Potential Closures
Britain's world-leading science facilities, including the Diamond Light Source and ISIS Neutron and Muon Source, face potential closure or significant cuts due to a £162m funding crisis at the Science and Technology Facilities Council caused by rising electricity costs, staff expenses, and foreign exchange pressures. These facilities serve hundreds of companies and thousands of scientists domestically and internationally, with Diamond producing light 10 billion times brighter than the sun for materials research. Scientists and research leaders warn that short-term funding cuts could cause decades-long damage to the UK's scientific capability and international competitiveness.
Mitochondrial ROS Signaling Drives Avoidance Learning in C. elegans
Researchers discovered that reactive oxygen species (ROS) produced by mitochondria in postsynaptic neurons are necessary and sufficient to drive avoidance learning in C. elegans, using optogenetic stimulation of nociceptive neurons. The study demonstrates that activity-dependent mitochondrial ROS production increases glutamate receptors at synapses and strengthens neural circuits controlling avoidance behavior. This finding reveals a novel molecular mechanism linking neuronal activity to synaptic plasticity and behavioral learning.