Balancing Interpretability and Flexibility in Modeling Diagnostic Trajectories with an Embedded Neural Hawkes Process...

Area of Science:

• Computational statistics
• Machine learning in healthcare

Background:

• Hawkes processes (HP) model event sequences with self-reinforcing dynamics, common in electronic health records (EHRs).
• Traditional HPs offer interpretability but lack flexibility, while neural network HPs are flexible but less interpretable.
• A critical need exists for interpretable yet flexible models in healthcare event sequence analysis.

Purpose of the Study:

• To develop a novel Hawkes process (HP) formulation that enhances interpretability without sacrificing performance for event sequence modeling.
• To address the tradeoff between flexibility and interpretability in neural network-based HPs.
• To enable the modeling of large-scale event sequences with numerous event types.

Main Methods:

• Proposed a flexible impact kernel instantiated as a neural network in event embedding space.
• Modeled impact functions within the event embedding space for enhanced flexibility and interpretability.
• Incorporated optional transformer encoder layers to further contextualize event embeddings, allowing a tunable tradeoff between flexibility and interpretability.

Main Results:

• The proposed method accurately recovered impact functions in simulation studies.
• Achieved competitive performance on the MIMIC-IV procedure dataset.
• Demonstrated clinically meaningful interpretations on the Duke-EHR children diagnosis dataset, even without transformer layers.

Conclusions:

• The flexible impact kernel effectively captures self-reinforcing dynamics in EHRs and other event sequence data.
• Interpretability can be maintained in advanced HP models without compromising predictive performance.
• This approach offers a promising solution for interpretable event sequence modeling in critical healthcare applications.