Data-Driven Simulator for Mechanical Circulatory Support with Domain Adversarial Neural Process

TL;DR

The model Domain Adversarial Neural Process (DANP) employs a neural process architecture, allowing it to capture the probabilistic relationship between MCS pump levels and aortic pressure measurements with uncertainty, resulting in a more realistic and diverse representation of potential outcomes.

Abstract

Mechanical Circulatory Support (MCS) devices, implemented as a probabilistic deep sequence model. Existing mechanical simulators for MCS rely on oversimplifying assumptions and are insensitive to patient-specific behavior, limiting their applicability to real-world treatment scenarios. To address these shortcomings, our model Domain Adversarial Neural Process (DANP) employs a neural process architecture, allowing it to capture the probabilistic relationship between MCS pump levels and aortic pressure measurements with uncertainty. We use domain adversarial training to combine simulation data with real-world observations, resulting in a more realistic and diverse representation of potential outcomes. Empirical results with an improvement of 19% in non-stationary trend prediction establish DANP as an effective tool for clinicians to understand and make informed decisions regarding MCS patient treatment.

Figures (7)

• Figure 1: Left: Cardiovascular System overview (adapted from Pearson's "Essentials of Human Anatomy and Physiology" Ch. 11 "The Cardiovascular System". Middle: Translation to Lumped Parameter circuitry model, the source of our simulated data. Right: Visualizations of a patient data sample. Note that as the clinician increases the pump level, the mean aortic pressure (MAP) increases as a result.
• Figure 2: Architecture of DANP with the neural process encoder, decoder, and a domain classifier. Input ${\mathbf{x}}$ (along with P-level $\mathbf{pl}$, combined for conciseness) is passed into the encoder. The output (feature $z$) is then passed separately into the decoder and domain classifier to obtain regression ${{\mathbf{y}}}$ prediction (MAP) and domain prediction ${d}$ (0 or 1), respectively. The Gradient Reversal Layer (GRL) for the domain classifier acts as an identity transform. The loss of $y$ prediction ($L_{y}$) and domain prediction ($L_{d}$) will be calculated separately but backpropagated together.
• Figure 3: Left: Neural Process Architecture Details and Decoder Details. The input data $\boldsymbol x$ is initially processed by an R Encoder to obtain its latent representation. This latent representation is then further processed by a Z Encoder, which parametrizes it into the latent space $z$. When concatenated with the target $\boldsymbol x$, the sampled values from this latent space result in a combined feature ${\mathbf{x}}_T \oplus \mathbf{pl}_T \oplus z$. This is passed through a sequential Decoder, where it undergoes a series of transformations to ultimately generate predictions for MAP. Right: Graphical model of DANP. A grey background indicates that the variable is observed.
• Figure 4: Sample Prediction using Model Trained and Tested on HR-PCI Cohort. In scenarios where both direct transfer and CLMU can only produce flat forecasts, DANP is able to produce sensible MAP that corresponds to p-level shifts.
• Figure 5: Sample Pump Performance (25 Hz) Data