Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpretable Neural Temporal Point Processes for Modelling Electronic Health Records

Bingqing Liu

TL;DR

This work addresses the lack of interpretability in neural temporal point processes for electronic health records by introducing inf2vec, an interpretable framework that learns per-type local embeddings and type-wise encoders/decoders to capture type-type influences end-to-end. It decouples history information across event types and uses local decoding to compute type-specific conditional intensities, enabling direct interpretation of inter-type influences. Experiments on real EHR data and Hawkes-simulated datasets demonstrate improved event prediction (weighted F1 and MAE) and qualitative alignment of learned influences with ground-truth dependencies. The approach is model-agnostic with respect to the encoder/decoder topology, offering practical insights into interactions among event types in continuous time.

Abstract

Electronic Health Records (EHR) can be represented as temporal sequences that record the events (medical visits) from patients. Neural temporal point process (NTPP) has achieved great success in modeling event sequences that occur in continuous time space. However, due to the black-box nature of neural networks, existing NTPP models fall short in explaining the dependencies between different event types. In this paper, inspired by word2vec and Hawkes process, we propose an interpretable framework inf2vec for event sequence modelling, where the event influences are directly parameterized and can be learned end-to-end. In the experiment, we demonstrate the superiority of our model on event prediction as well as type-type influences learning.

Interpretable Neural Temporal Point Processes for Modelling Electronic Health Records

TL;DR

This work addresses the lack of interpretability in neural temporal point processes for electronic health records by introducing inf2vec, an interpretable framework that learns per-type local embeddings and type-wise encoders/decoders to capture type-type influences end-to-end. It decouples history information across event types and uses local decoding to compute type-specific conditional intensities, enabling direct interpretation of inter-type influences. Experiments on real EHR data and Hawkes-simulated datasets demonstrate improved event prediction (weighted F1 and MAE) and qualitative alignment of learned influences with ground-truth dependencies. The approach is model-agnostic with respect to the encoder/decoder topology, offering practical insights into interactions among event types in continuous time.

Abstract

Electronic Health Records (EHR) can be represented as temporal sequences that record the events (medical visits) from patients. Neural temporal point process (NTPP) has achieved great success in modeling event sequences that occur in continuous time space. However, due to the black-box nature of neural networks, existing NTPP models fall short in explaining the dependencies between different event types. In this paper, inspired by word2vec and Hawkes process, we propose an interpretable framework inf2vec for event sequence modelling, where the event influences are directly parameterized and can be learned end-to-end. In the experiment, we demonstrate the superiority of our model on event prediction as well as type-type influences learning.
Paper Structure (11 sections, 6 equations, 3 figures, 1 table)

This paper contains 11 sections, 6 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Embedding Layer
  4. Sequence Encoder
  5. Event Decoder
  6. Experiment
  7. Datasets
  8. Compared Models and Implemented inf2vec
  9. Prediction Results
  10. Type-Type Influences Learning
  11. Conclusion

Figures (3)

  • Figure 1: An overview of our proposed type-type influences learning framework. The framework creates separate embedding, encoding and decoding space for each event type. In the illustrated example, brighter color in the embedding layer means stronger influence and we see event of type 2 is more likely to occur.
  • Figure 2: The illustration of our learned local embeddings (after dimension reduction) over dataset Haw5, Haw9 and HawC9 (the first three plots), coordinates $(x,k)$ denotes the embedding of event type $x$ in the context of event type $k$. The last three plots are the ground truth influences and coordinates $(x,k)$ denotes the influence of $x$ on $k$, brighter color means stronger influence.
  • Figure 3: The illustration of our learned local embeddings over dataset SynEHR1. The first and second row show how other events influence "wellness" and "ambulatory", respectively. Brighter color indicates stronger influence.