TinyML-Based Photoplethysmography (PPG) Analysis for Atrial Fibrillation Detection

This paper proposes an atrial fibrillation detection algorithm using PPG signals and presents a method to port it to edge devices such as the nRF52840. To classify the PPG signals, simple CNN-based, Autoencoder-based, and Transformer-based Deep Neural Networks (DNNs) were implemented and comparatively analyzed using the MIMIC PERform AF dataset. As a result, the Transformer-based algorithm exhibited the best performance with an average accuracy of 98.66%. While post-training quantization (PTQ) of the Transformer-based algorithm caused the average accuracy to degrade to 95.23%, applying Quantization-Aware Training (QAT) prevented this performance drop. Because some layers of the Transformer are not supported in TensorFlow Lite for Microcontrollers (TFLite Micro), they were modified, and the Multi-Head Attention was approximated using a CNN-based network for porting to the edge device, resulting in a final model size of 107 KB. The ported network demonstrated excellent performance, maintaining an average accuracy of 98.09%.

Authors

Institutions

Publication Details

Journal
Sensors
Published
2026-11-16
DOI
https://doi.org/10.3390/s26227229
Primary Topic
Non-Invasive Vital Sign Monitoring
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

TinyML-Based Photoplethysmography (PPG) Analysis for Atrial Fibrillation Detection

Kyoungtaek Choi
Sensors
Non-Invasive Vital Sign Monitoring
article

TinyML-Based Photoplethysmography (PPG) Analysis for Atrial Fibrillation Detection

Kyoungtaek Choi
article en

Abstract

This paper proposes an atrial fibrillation detection algorithm using PPG signals and presents a method to port it to edge devices such as the nRF52840. To classify the PPG signals, simple CNN-based, Autoencoder-based, and Transformer-based Deep Neural Networks (DNNs) were implemented and comparatively analyzed using the MIMIC PERform AF dataset. As a result, the Transformer-based algorithm exhibited the best performance with an average accuracy of 98.66%. While post-training quantization (PTQ) of the Transformer-based algorithm caused the average accuracy to degrade to 95.23%, applying Quantization-Aware Training (QAT) prevented this performance drop. Because some layers of the Transformer are not supported in TensorFlow Lite for Microcontrollers (TFLite Micro), they were modified, and the Multi-Head Attention was approximated using a CNN-based network for porting to the edge device, resulting in a final model size of 107 KB. The ported network demonstrated excellent performance, maintaining an average accuracy of 98.09%.

SensorsVol. 26(22)
Daegu Catholic University (KR)
Non-Invasive Vital Sign Monitoring
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon