Predicting Secure Messaging Traffic in Clinical Settings.

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Network Science

Background:

• Asynchronous text-based communication, or secure messaging, is prevalent among healthcare workers.
• This communication method can disrupt workflow and increase professional burden.
• Identifying communication peaks is crucial for interventions to reduce overload.

Purpose of the Study:

• To predict future secure messaging volume among healthcare professionals.
• To develop and evaluate a graph-based transformer model for this prediction task.

Main Methods:

• Utilized secure messaging metadata from 14 inpatient hospitals and 381 outpatient clinics (Feb 2023 - Apr 2024).
• Employed a multi-headed graph transformer for hourly communication structures.
• Integrated a transformer decoder architecture for future messaging volume prediction.
• Incorporated domain-specific, time-dependent positional encoding for improved accuracy.

Main Results:

• The graph-based transformer model successfully predicted secure messaging volume across test sets.
• Performance was enhanced by utilizing domain-specific time-dependent positional encoding.
• The model effectively integrates network topology, temporal patterns, and relational context.

Conclusions:

• Accurate forecasting of healthcare secure messaging traffic is achievable with advanced AI models.
• This approach offers a novel method for understanding and managing communication dynamics in healthcare settings.
• Findings support the development of targeted interventions to mitigate communication overload.