Explicit-Blurred Memory Network for Analyzing Patient Electronic Health Records

Prithwish Chakraborty, Fei Wang, Jianying Hu, Daby SowarXiv 2019 · arXiv

Introduction

Predicting patient outcomes from longitudinal Electronic Health Records (EHR) is a critical task for modern healthcare. While deep learning models like Recurrent Neural Networks (RNNs) have shown promise, they often function as "black boxes," mixing information in complex, non-linear ways that lack clinical interpretability. Furthermore, models that attempt to use "explicit" memory (storing exact event copies) are often difficult to train due to non-differentiable operations.

To address these challenges, researchers have developed the Explicit-Blurred Memory Network (EBmRNN). This novel architecture mimics human cognitive processes by partitioning memory into an "explicit" block for exact event recall and a "blurred" block for high-level abstractions. This approach achieves state-of-the-art predictive accuracy while providing clinicians with a traceable path to understand the specific past events triggering a medical alert.

Methodology

The EBmRNN architecture is inspired by the Atkinson-Shiffrin model of human psychology and consists of three primary components:

  1. The Controller: A standard RNN (e.g., LSTM or GRU) that processes incoming patient data and manages memory operations.
  2. Explicit Memory Block: Stores exact replicas of recent clinical events. To ensure differentiability during training, the authors employ the Gumbel-Softmax trick, allowing the model to make discrete choices within a gradient-based optimization framework.
  3. Blurred Memory Block: As explicit memory reaches capacity, older events are "compressed" into this block as higher-level abstractions. This provides long-term context without the overhead of storing every raw data point.
  4. Read Gate: A mechanism that dynamically balances the model's reliance on raw recent data (explicit) versus long-term trends (blurred).
Figure 1: Architecture of the Explicit-Blurred Memory Network (EBmRNN), showing the interaction between the controller, explicit memory, and blurred memory blocks.

Datasets and Evaluation

The model was evaluated using the MIMIC-III database across three critical clinical tasks:

  • In-Hospital Mortality: Predicting survival based on the first 48 hours of ICU data.
  • Decompensation: Real-time prediction of health deterioration within the next 24 hours.
  • Phenotyping: Multi-label classification of 25 common ICU conditions (e.g., sepsis, diabetes).

Key Insights and Results

The EBmRNN performed on par with or better than standard baselines (Logistic Regression, LSTM, GRU) across all tasks. For mortality prediction, it achieved an AUC-ROC of 0.8612, and for decompensation, it reached 0.8989, the highest among tested models.

A significant advantage of this architecture is its traceability. By analyzing "slot utilization" maps, researchers can identify exactly which time steps the model prioritized for a given prediction. This transparency allows for case-based reasoning in a clinical setting.

Figure 2: Memory slot utilization maps for In-Hospital Mortality and Decompensation tasks, illustrating how the model tracks specific clinical events over time.

The authors also found that the "blurred" memory bank stabilizes the training of the explicit memory by providing a tractable path for gradients, preventing the mathematical instabilities common in simpler explicit-only models.

Conclusion and Future Work

The EBmRNN offers a high-performing and clinically interpretable alternative to traditional "black box" models. While it provides the data necessary for interpretation, future work should focus on developing user-facing interfaces to present these insights to clinicians effectively. Additionally, the sensitivity of the read gate suggests that task-specific tuning remains an important consideration for deployment.

Summary updated

Abstract

In recent years, we have witnessed an increased interest in temporal modeling of patient records from large scale Electronic Health Records (EHR). While simpler RNN models have been used for such problems, memory networks, which in other domains were found to generalize well, are underutilized. Traditional memory networks involve diffused and non-linear operations where influence of past events on outputs are not readily quantifiable. We posit that this lack of interpretability makes such networks not applicable for EHR analysis. While networks with explicit memory have been proposed recently, the discontinuities imposed by the discrete operations make such networks harder to train and require more supervision. The problem is further exacerbated in the limited data setting of EHR studies. In this paper, we propose a novel memory architecture that is more interpretable than traditional memory networks while being easier to train than explicit memory banks. Inspired by well-known models of human cognition, we propose partitioning the external memory space into (a) a primary explicit memory block to store exact replicas of recent events to support interpretations, followed by (b) a secondary blurred memory block that accumulates salient aspects of past events dropped from the explicit block as higher level abstractions and allow training with less supervision by stabilize the gradients. We apply the model for 3 learning problems on ICU records from the MIMIC III database spanning millions of data points. Our model performs comparably to the state-of the art while also, crucially, enabling ready interpretation of the results.
BibTeX
@misc{chakraborty2019explicit,
  author = {Prithwish Chakraborty and Fei Wang and Jianying Hu and Daby Sow},
  title = {{Explicit-Blurred Memory Network for Analyzing Patient Electronic Health Records}},
  year = {2019},
  eprint = {1911.06472},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.06472},
}
Figure