A Deep Bidirectional LSTM Model Enhanced by Transfer-Learning for the Classification of Peripheral Arterial Blood...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Physiology

Background:

• Arterial blood pressure waveform (BPW) morphology offers critical insights into cardiovascular status.
• BPW alterations can be early markers of pathological changes, especially in critically ill patients.

Purpose of the Study:

• To develop a deep learning (DL) framework for classifying BPWs into distinct morphologies (Type A vs. B/C).
• To assess the clinical relevance and potential translation of DL models for cardiovascular monitoring.

Main Methods:

• A bidirectional long short-term memory (BiLSTM) model was trained on invasive BPWs from animal models.
• Deep transfer learning (DTL) was applied to human BPWs from the MIMIC-III database, adapting animal-trained models.
• Waveform Separation Analysis (WSA) principles were extended to inform model development and application.

Main Results:

• The BiLSTM model achieved 73% accuracy on aortic and 77% on femoral BPWs in animal test sets.
• DTL enabled classification of human femoral BPWs with 77% accuracy, comparable to animal model performance.
• The study demonstrated effective adaptation of animal-trained models to human data via DTL.

Conclusions:

• Deep transfer learning shows potential for adapting DL models trained on animal data to human clinical data.
• DL-based BPW classification may enhance cardiovascular monitoring in critically ill patients through early detection of vascular alterations.
• Challenges remain regarding the need for large, labeled clinical datasets for robust validation of DL models.