• interpretability
• recurrent neural networks
• deep-learning
• biomedical signals
• gated recurrent units

A large research effort is spent in the field of biomedical signal analysis. Great progresses have been made recently in the field by applying novel Machine Learning techniques, especially in classification tasks.
The adoption of new models is however slowed down by their lack of interpretability, a key requirement in the medical domain.
While some interpretability approaches have been proposed for models and architectures dealing with images data and classification tasks, little to none work exists for biomedical signal regression.
This work present results obtained by adapting well-established interpretability techniques to deep learning models dealing with arterial blood pressure estimation.
Different models have been trained on Electrocardiogram (ECG) and Photoplethysmogram (PPG) data to predict arterial blood pressure (ABP).
First we describe our analysis on filters learned by models using convolutional layers.
Then we focus on gated recurrent models to analyze the effects of input signals occlusion on both output and gates activations.
Experiments with different types of occlusion have been performed to explore the behaviour of different architectures.
Both a qualitative and a quantitative evaluation is presented.