[Paper Review] Deep Transfer Learning for Physiological Signals.

Hugh Chen, Scott Lundberg+3 more|arXiv (Cornell University)|Feb 12, 2020

Phonocardiography and Auscultation Techniques61 references4 citations

TL;DR

PHASE is a deep transfer learning framework that learns shared, low-dimensional embeddings of physiological signals using signal-specific LSTMs, enabling improved prediction of adverse outcomes across hospitals and departments. It achieves superior performance and lower computational cost while successfully transferring knowledge across diverse clinical settings.

ABSTRACT

Deep learning is increasingly common in healthcare, yet transfer learning for physiological signals (e.g., temperature, heart rate, etc.) is under-explored. Here, we present a straightforward, yet performant framework for transferring knowledge about physiological signals. Our framework is called PHASE (PHysiologicAl Signal Embeddings). It i) learns deep embeddings of physiological signals and ii) predicts adverse outcomes based on the embeddings. PHASE is the first instance of deep transfer learning in a cross-hospital, cross-department setting for physiological signals. We show that PHASE's per-signal (one for each signal) LSTM embedding functions confer a number of benefits including improved performance, successful transference between hospitals, and lower computational cost.

Motivation & Objective

• Address the under-explored challenge of transfer learning in physiological signal analysis across hospitals and departments.
• Develop a framework that learns shared representations of physiological signals to improve prediction of adverse clinical outcomes.
• Enable knowledge transfer between hospitals with different data distributions and signal acquisition protocols.
• Reduce computational cost while maintaining or improving model performance on downstream prediction tasks.

Proposed method

• Train signal-specific Long Short-Term Memory (LSTM) networks to learn deep, low-dimensional embeddings from raw physiological signals such as heart rate and temperature.
• Use the learned embeddings as transferable representations for downstream prediction tasks, such as adverse outcome prediction.
• Apply the same embedding function across different hospitals and departments, enabling zero-shot or few-shot transfer learning.
• Train a final prediction head on top of the shared embeddings to predict clinical outcomes, with the embedding function fixed during fine-tuning.
• Optimize the framework end-to-end using cross-entropy loss for classification of adverse events.
• Evaluate transferability by testing the same embedding function on data from different hospitals and departments without retraining.

Experimental results

Research questions

• RQ1 Can signal-specific LSTM-based embeddings transfer knowledge effectively across different hospitals and clinical departments?
• RQ2 Does using shared embeddings from PHASE improve prediction performance for adverse outcomes compared to training models from scratch?
• RQ3 To what extent does PHASE reduce computational cost while maintaining or improving model accuracy?
• RQ4 How robust are the learned embeddings to variations in data distribution and signal acquisition protocols across hospitals?

Key findings

• PHASE achieves improved prediction performance for adverse outcomes compared to baseline models trained from scratch.
• The framework successfully transfers knowledge across hospitals and departments, even with differing data distributions and signal acquisition systems.
• PHASE reduces computational cost by reusing pre-trained signal-specific embedding functions across multiple prediction tasks.
• Signal-specific LSTM embeddings provide better representation learning than generic or shared embeddings across all signals.
• The framework demonstrates strong zero-shot transfer capability, where embeddings learned in one hospital generalize to another without fine-tuning.
• PHASE maintains high performance even when tested on data from departments with different monitoring protocols or signal types.