QualityFM: a Multimodal Physiological Signal Foundation Model with Self-Distillation for Signal Quality Challenges in Critically Ill Patients (2509.06516v1)

Abstract: Photoplethysmogram (PPG) and electrocardiogram (ECG) are commonly recorded in intesive care unit (ICU) and operating room (OR). However, the high incidence of poor, incomplete, and inconsistent signal quality, can lead to false alarms or diagnostic inaccuracies. The methods explored so far suffer from limited generalizability, reliance on extensive labeled data, and poor cross-task transferability. To overcome these challenges, we introduce QualityFM, a novel multimodal foundation model for these physiological signals, designed to acquire a general-purpose understanding of signal quality. Our model is pre-trained on an large-scale dataset comprising over 21 million 30-second waveforms and 179,757 hours of data. Our approach involves a dual-track architecture that processes paired physiological signals of differing quality, leveraging a self-distillation strategy where an encoder for high-quality signals is used to guide the training of an encoder for low-quality signals. To efficiently handle long sequential signals and capture essential local quasi-periodic patterns, we integrate a windowed sparse attention mechanism within our Transformer-based model. Furthermore, a composite loss function, which combines direct distillation loss on encoder outputs with indirect reconstruction loss based on power and phase spectra, ensures the preservation of frequency-domain characteristics of the signals. We pre-train three models with varying parameter counts (9.6 M to 319 M) and demonstrate their efficacy and practical value through transfer learning on three distinct clinical tasks: false alarm of ventricular tachycardia detection, the identification of atrial fibrillation and the estimation of arterial blood pressure (ABP) from PPG and ECG signals.

• The paper introduces QualityFM, a multimodal foundation model that uses dual-track self-distillation and self-supervised pretraining to robustly analyze physiological signals.
• It employs a windowed sparse attention Transformer combined with frequency-domain loss functions to preserve waveform characteristics in noisy PPG and ECG data.
• The model outperforms state-of-the-art baselines on tasks like arrhythmia detection, AF identification, and blood pressure estimation, demonstrating solid cross-task generalization.

QualityFM: A Multimodal Foundation Model for Robust Physiological Signal Analysis in Critical Care

Introduction

The paper presents QualityFM, a multimodal foundation model designed to address persistent signal quality challenges in physiological monitoring of critically ill patients. The model targets photoplethysmogram (PPG) and electrocardiogram (ECG) signals, which are highly susceptible to noise, artifacts, and incomplete recordings in clinical environments such as ICUs and operating rooms. QualityFM leverages large-scale self-supervised pretraining, a dual-track self-distillation architecture, and frequency-domain loss functions to build generalizable representations that transfer effectively across diverse downstream tasks, including false arrhythmia alarm detection, atrial fibrillation (AF) identification, and arterial blood pressure (BP) estimation.

Figure 1: Pipeline of QualityFM pretraining and downstream tasks. Stage 1: self-distillation with quality-divergent pairs; Stage 2: adaptation for clinical downstream tasks.

Model Architecture and Pretraining Strategy

QualityFM employs a dual-track encoder architecture, processing paired segments of high- and low-quality signals. The teacher encoder, updated via Exponential Moving Average (EMA), extracts robust features from high-quality signals and guides the student encoder to distill meaningful representations from low-quality counterparts. This self-distillation paradigm is inspired by DINO and iBOT, but is specifically tailored for physiological time series with quality-divergent pairs.

The backbone is a windowed sparse attention Transformer, which mitigates the quadratic complexity of standard self-attention for long sequences. Windowed attention enables efficient modeling of local waveform morphology and quasi-periodic patterns, with receptive fields expanding in deeper layers. Layer normalization is applied to Q, K, and V tensors to stabilize training and prevent attention logit explosion.

A composite loss function integrates direct distillation (cross-entropy between teacher and student outputs) and indirect spectral reconstruction losses. The latter enforces preservation of amplitude and phase spectra via DFT-based reconstruction, ensuring that frequency-domain characteristics critical for clinical interpretation are retained.

Signal Quality Index (SQI) and Data Curation

Signal segments are assigned quality labels based on a composite SQI, aggregating metrics such as power ratio, perfusion index, skewness, relative power, and entropy for PPG, and energy, sample entropy, heart rate range, and rhythm stability for ECG. Only pairs from the same subject, within a five-minute interval, and with differing quality labels are used for pretraining.

Figure 2: Example of ECG and PPG signals with different SQI values.

The pretraining dataset comprises over 21 million 30-second waveform segments (179,757 hours) from VitalDB and MIMIC-III, representing the largest scale for this domain to date.

Downstream Task Performance and Transferability

QualityFM is evaluated on three clinical tasks: false ventricular tachycardia alarm detection (VTaC), AF identification (MIMIC PERform AF), and BP estimation (UCI). Three model variants (Base: 9.6M, Large: 70M, Huge: 319M parameters) are pretrained and fine-tuned.

QualityFM consistently outperforms state-of-the-art baselines (ResNet, Transformer, BeatGAN, TAnoGAN, CNN-CL, Diffusion+CL, SiamQuality) across all metrics. For VTaC, QualityFM-Huge achieves an accuracy of 0.8551, TPR of 0.8220, TNR of 0.8679, F1-score of 0.7607, and AUC of 0.9047. For AF detection, QualityFM-Huge attains an accuracy of 0.8764 and AUC of 0.9411. BP estimation yields MAE of 19.32 (SBP) and 7.61 (DBP), outperforming all compared models.

Ablation Studies

Impact of Pretraining

Ablation demonstrates that pretraining is essential: models trained from scratch show marked drops in classification and regression performance, especially for false alarm detection and AF identification.

Figure 3: Comparison of the model performances with and without pre-training.

Windowed Sparse Attention

Varying the window size in sparse attention reveals that a size of 8 is optimal; disabling the mechanism (window size 0) yields the worst results, confirming the necessity of local context modeling.

Figure 4: Comparison of the model performances with different window sizes of sparse attention.

Backbone Comparison

QualityFM-Base outperforms ResNet101 and standard Transformer backbones, achieving superior comprehensive performance profiles on both VTaC and AF tasks.

Figure 5: Comparison of QualityFM with different backbones architectures.

Indirect Loss Contributions

Ablation of amplitude and phase reconstruction losses shows that both contribute positively to performance, with the combined loss yielding the best results for both classification (AUC) and regression (MAE) tasks.

Practical and Theoretical Implications

QualityFM demonstrates that large-scale self-supervised pretraining on multimodal physiological signals can yield robust, transferable representations that generalize across tasks and mitigate the impact of poor signal quality. The dual-track self-distillation approach, combined with frequency-domain supervision, enables effective learning from noisy and artifact-laden data without reliance on extensive manual labeling.

The model's scalability is validated by consistent performance improvements with increased parameter count. However, the current architecture is limited to dual-channel input and requires substantial computational resources for full fine-tuning. Future work should address adaptive input mechanisms for variable signal modalities and explore parameter-efficient fine-tuning (e.g., LoRA) to facilitate practical deployment.

Future Directions

Potential avenues for further research include:

• Scaling pretraining to billion-parameter models and larger, more diverse datasets.
• Developing adaptive architectures for flexible multimodal input.
• Integrating PEFT techniques to reduce resource requirements for downstream adaptation.
• Extending the foundation model paradigm to other biosignal domains (EEG, EMG, etc.).
• Investigating real-time deployment and edge inference for continuous patient monitoring.

Conclusion

QualityFM establishes a robust foundation for physiological signal analysis in critical care, overcoming longstanding challenges of poor signal quality, label scarcity, and limited cross-task transferability. Its architecture and training paradigm set a precedent for future foundation models in biomedical signal processing, with strong empirical results and clear pathways for further scaling and generalization.