MD-ViSCo: A Unified Model for Multi-Directional Vital Sign Waveform Conversion.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Signal Processing

Background:

• Deep learning models for vital sign waveform generation are typically specific to source-target pairs, limiting clinical usability.
• Developing distinct models for each vital sign conversion task requires significant resources and expertise.

Purpose of the Study:

• To propose a unified framework, Multi-Directional Vital-Sign Converter (MD-ViSCo), capable of generating any target vital sign waveform from any single input waveform using a single model.
• To overcome the limitations of existing specialized models in clinical settings.

Main Methods:

• MD-ViSCo utilizes a 1-Dimensional U-Net integrated with a Swin Transformer architecture.
• Adaptive Instance Normalization (AdaIN) is employed to capture distinct waveform styles.
• The framework was evaluated on multi-directional waveform generation using publicly available datasets.

Main Results:

• MD-ViSCo outperformed state-of-the-art baselines, reducing Mean Absolute Error (MAE) by 8.8% and improving Pearson Correlation (PC) by 4.9% on average across waveform types.
• Generated arterial blood pressure (ABP) waveforms met AAMI standards and achieved Grade B on the BHS standard.
• Generated electrocardiogram (ECG) waveforms showed superior performance in downstream Atrial Fibrillation (AF) classification tasks.

Conclusions:

• MD-ViSCo offers a unified framework for generating diverse vital sign waveforms, eliminating the need for task-specific models.
• The proposed method enhances clinical utility and simplifies healthcare monitoring by providing a versatile solution for vital sign waveform conversion.