PanLUNA: An Efficient and Robust Query-Unified Multimodal Model for Edge Biosignal Intelligence

Marija Zelic; Anna Tegon; Yawei Li; Thorir Mar Ingolfsson; Luca Benini

cs.AIApr 5, 2026

PanLUNA: An Efficient and Robust Query-Unified Multimodal Model for Edge Biosignal Intelligence

Authors: Marija Zelic, Anna Tegon, Yawei Li, Thorir Mar Ingolfsson, Luca Benini

Organizations: Integrated Systems Laboratory, ETH Zürich, Zürich, Switzerland · DEI, University of Bologna, Bologna, Italy

Abstract

Physiological foundation models (FMs) have shown promise for biosignal representation learning, yet most remain confined to a single modality such as EEG, ECG, or PPG, largely because paired multimodal datasets are scarce. In this paper, we present PanLUNA, a compact 5.4M-parameter pan-modal FM that jointly processes EEG, ECG, and PPG within a single shared encoder. Extending LUNA's channel-unification module, PanLUNA treats multimodal channels as entries in a unified query set augmented with sensor-type embeddings, enabling efficient cross-modal early fusion while remaining inherently robust to missing modalities at inference time. Despite its small footprint, PanLUNA matches or exceeds models up to 57 $\times$ larger: 81.21% balanced accuracy on TUAB abnormal EEG detection and state-of-the-art 0.7416 balanced accuracy on HMC multimodal sleep staging. Quantization-aware training with INT8 weights recovers $\geq$ 96% of full-precision performance, and deployment on the GAP9 ultra-low-power RISC-V microcontroller for wearables achieves 325.6 ms latency and 18.8 mJ per 10-second, 12-lead ECG inference, and 1.206 s latency at 68.65 mJ for multimodal 5-channel sleep staging over 30-second epochs.

Figures & tables

Dataset	Modality	# Subjects	Ch.	FS (Hz)	T (s)
Pretraining (self-supervised)
TUEG [ 30 ]	EEG	14,987	20/22	250	5
Siena [ 7 , 8 , 15 ]	EEG	14	29	512	5
MIMIC-IV [ 16 , 15 ]	ECG	161,352	12	500	5
CODE-15% [ 32 , 33 ]	ECG	233,700	12	400	5
PulseDB [ 42 ]	ECG,PPG	5,361	2	125	5

TABLE I : Summary of pretraining and downstream datasets. Ch: number of channels. FS: sampling frequency. T: window length.

Fig. 1 : Per-patch channel-query cross-attention scores overlaid on raw signals. Each query specializes to a distinct modality: Query 3 activates on ECG QRS complexes, Query 1 on PPG waveforms, and Query 2 on EEG.

Model	Size	Bal. Acc. (%) $\uparrow$	AUC-PR $\uparrow$	AUROC $\uparrow$
Self-supervised Models
BIOT [ 43 ]	3.2M	79.59 $\pm$ 0.57	0.8692 $\pm$ 0.0023	0.8815 $\pm$ 0.0043
FEMBA-Huge [ 36 ]	386M	81.82 $\pm$ 0.16	0.9005 $\pm$ 0.0017	0.8921 $\pm$ 0.0042
CEReBrO-Huge [ 9 ]	85.15M	81.67 $\pm$ 0.23	0.9049 $\pm$ 0.0026	0.8916 $\pm$ 0.0038
LaBraM-Base [ 24 ]	5.9M	81.40 $\pm$ 0.19	0.8965 $\pm$ 0.0016	0.9022 $\pm$ 0.0009
LaBraM-Huge [ 24 ]	369.8M	82.58 $\pm$ 0.11	0.9204 $\pm$ 0.0011	0.9162 $\pm$ 0.0016

TABLE II : Performance comparison on TUAB abnormal EEG detection. Best results are bolded .

Method	Training Modality	Test Modality	Balanced Accuracy	Cohen’s Kappa	Weighted F1
EEG-Conformer [ 34 ]	EEG	EEG	$0.6767\pm 0.0200$	$0.5886\pm 0.0397$	$0.6550\pm 0.0463$
BIOT [ 43 ]	EEG	EEG	$0.6862\pm 0.0041$	$0.6295\pm 0.0113$	$0.7091\pm 0.0147$
LaBraM-Base [ 24 ]	EEG	EEG	0.7286 $\pm$ 0.0101	$0.6812\pm 0.0073$	$0.7554\pm 0.0024$
CBraMod [ 41 ]	EEG	EEG	$0.7177\pm 0.0072$	$0.6653\pm 0.0057$	$0.7388\pm 0.0052$
PhysioOmni [ 23 ]	EEG+EOG+EMG	EEG	0.7289 $\pm$ 0.0010	0.6880 $\pm$ 0.0097	0.7635 $\pm$ 0.0053
PanLUNA (FF)	EEG+ECG+PPG	EEG	0.7416 $\pm$ 0.0068	0.6946 $\pm$ 0.0110	0.7659 $\pm$ 0.0086

TABLE III : The results of different methods on HMC (sleep stage detection). Best results are bolded ; second-best results are underlined . Quantization results are obtained for INT8/INT8 combination in weights and activations.

Task	AUROC FP32 (FE)	W: INT8 A: INT8		W: INT4 A: INT8		W: INT2 A: INT8		W: INT4 A: INT4
		PTQ	QAT	PTQ	QAT	PTQ	QAT	PTQ	QAT
PTBXL-Super	0.9083	0.8448	0.9012	0.5892	0.9011	0.4976	0.8675	0.4993	0.8514
PTBXL-Sub	0.8880	0.7877	0.8784	0.6718	0.8797	0.4877	0.8272	0.5310	0.8170
PTBXL-Form	0.8331	0.7215	0.8009	0.5528	0.8114	0.4422	0.6710	0.4663	0.7378
PTBXL-Rhythm	0.9641	0.8897	0.9623	0.6526	0.9508	0.5443	0.9034	0.5340	0.9099
CSN	0.9505	0.8743	0.9507	0.6256	0.9447	0.5297	0.8821	0.5384	0.9109

TABLE IV : Quantization Results on Cardiac Experiments. Bold results are closest to the FP32 performance.

Model	Platform	Params	Signal	Cycles	Latency	Energy	Power	MACs/cyc
Task-specific models
AF-LSTM [ 1 ]	STM32L4	$<$ 50k	ECG (1)	–	143 ms	3.5 mJ	–	–
CLTC [ 20 ]	STM32F7	$\sim$ 240k	ECG (12)	–	$\sim$ 1 s	–	137.4 mW	–
Foundation models
TinyMyo [ 12 ]	GAP9	3.56M	EMG	291.3 M	785 ms	44.9 mJ	57.2 mW	6.65
FEMBA [ 36 ]	GAP9	7.8M	EEG (22)	629.4 M	1.70 s	75 mJ	44.1 mW	1.13

TABLE V : On-device biosignal inference on ultra-low-power MCUs. All GAP9 models operate at 370 MHz.

Explore similar work

Sense Less, Infer More: Agentic Multimodal Transformers for Edge Medical Intelligence
Apr 12, 2026Chengwei Zhou, Zhaoyan Jia, Haotian Yu +6Efficient Multimodal Inference Wearable Health Monitoring
Case Western Reserve University, USA · University of Southern California, USA
SignalMC-MED: A Multimodal Benchmark for Evaluating Biosignal Foundation Models on Single-Lead ECG and PPG
Mar 10, 2026Fredrik K. Gustafsson, Xiao Gu, Mattia Carletti +3Electrocardiogram Foundation Models Multimodal Model Evaluation
Department of Engineering Science, University of Oxford · NIHR Health Protection Research Unit in Healthcare Associated Infections and Antimicrobial Resistance · Big Data Institute, Nuffield Department of Population Health, University of Oxford +2
Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model
May 10, 2026Zhangdaihong Liu, Chang Liu, Fenglin Liu +4Cross-Modal Representation Learning Masked Autoencoders
Department of Engineering Science, University of Oxford, Oxford OX3 7DQ, UK · Oxford Suzhou Centre for Advanced Research, Suzhou, 215123, China · School of Public Health, Shanghai Jiao Tong University, Shanghai, China

Apr 12, 2026J/K move · Enter open · S save

Sense Less, Infer More: Agentic Multimodal Transformers for Edge Medical Intelligence

Chengwei Zhou, Zhaoyan Jia, Haotian Yu, Xuming Chen+5

Case Western Reserve University, USA · University of Southern California, USA

Efficient Multimodal Inference Wearable Health Monitoring Edge ML Adaptive Sampling Multimodal Sensor Fusion

Open paper AskPDF arXiv

Edge-based multimodal medical monitoring requires models that balance diagnostic accuracy with severe energy constraints. Continuous acquisition of ECG, PPG, EMG, and IMU streams rapidly drains wearable batteries, often limiting operation to under 10 hours, while existing systems overlook the high temporal redundancy present in physiological signals. We introduce Adaptive Multimodal Intelligence (AMI), an end-to-end framework that jointly learns when to sense and how to infer. AMI integrates three components: (1) a lightweight Agentic Modality Controller that uses differentiable Gumbel-Sigmoid gating to dynamically select active sensors based on model confidence and task relevance; (2) a Learned Sigma-Delta Sensing module that applies patch-wise Delta-Sigma operations with learnable thresholds to skip temporally redundant samples; and (3) a Foundation-backed Multimodal Prediction Model built on unimodal foundation encoders and a cross-modal transformer with temporal context, enabling robust fusion even under gated or missing inputs. These components are trained jointly via a multi-objective loss combining classification accuracy, sparsity regularization, cross-modal alignment, and predictive coding. AMI is hardware-aware, supporting dynamic computation graphs and masked operations, leading to real energy and latency savings. Across MHEALTH, HMC Sleep, and WESAD datasets, it reduces sensor usage by 48.8% while improving state-of-the-art accuracy by 1.9% on average.

Abstract

Explore similar work

Sense Less, Infer More: Agentic Multimodal Transformers for Edge Medical Intelligence

SignalMC-MED: A Multimodal Benchmark for Evaluating Biosignal Foundation Models on Single-Lead ECG and PPG

Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model