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

Arterial blood pressure waveform (BPW) morphology provides critical insight into cardiovascular status and can serve as an early marker of pathological changes, particularly in critically ill patients where waveform alterations may occur over hours. This study aims to develop a deep learning (DL) framework to classify BPWs into type A and type B/C morphologies, reflecting distinct cardiovascular conditions.

A bidirectional long short-term memory (BiLSTM) model was trained on central (aortic) and peripheral (femoral) invasive BPWs from a large animal cohort. The framework extends the applicability of waveform separation analysis (WSA), as waveform type determines the choice and tuning of arterial blood flow estimation methods. To explore clinical relevance, deep transfer learning (DTL) was applied to femoral BPWs from the MIMIC-III database, retraining only the last 8 of 15 layers to balance accuracy and computational cost.

The BiLSTM achieved 73% accuracy on aortic and 77% on femoral BPWs in independent animal test sets. Using DTL, classification on human femoral BPWs reached 77% accuracy, comparable to animal model performance.

DTL effectively adapts animal-trained models to human data, supporting potential clinical translation. Results also highlight challenges due to limited large, labeled clinical datasets, essential for robust validation.

DL-based BPW classification may improve cardiovascular monitoring in critically ill patients by enabling early detection of vascular alterations.