Ankush Jindal, MD, a student in the Health Science Informatics Research Masters’ program, recently completed his masters’ thesis project entitled, “AI-Driven Triage of Patient Messages in Telemedicine Systems for Optimized Clinician Workload.” Simply put, Ankush and colleagues developed a tool for patient message triaging using AI.
Since the COVID-19 pandemic, the use of patient messaging portals for medical advice (called patient medical advice requests, or PMARs) has increased by 157%, which created a challenge for healthcare providers. About 30% of these messages are administrative and require no clinical input, but they are nonetheless routed to physicians. Dr. Jindal’s study addressed the need for an efficient digital tool for PMAR message triaging to reduce physician burden, optimize workflow, and improve patient-provider communication. The goal was to develop robust tools for patient portal systems like MyChart which facilitate message triaging by interpreting semantically meaningful clinical information. The scope for this analysis was to accurately classify between administrative and clinical messages that can potentially help in triaging.
To accomplish this goal, the team developed an inbox classifier called OPTIC (Optimizing Patient-Provider Triaging and Improving Communications in Clinical Operations) which uses the GPT-4 model for data labeling and the open-source bidirectional encoder representations from transformers (BERT) model for distillation.
In their study, the team curated approximately 100K messages from all medical specialties and grouped them based on semantic similarity, which reduced redundancy. The most representative messages from each cluster were selected and manually labeled (approximately 347 messages). These messages were used to train and test OPTIC.
OPTIC was deployed through Nebula as part of Epic’s SaaS offering, which ensures accessibility and scalability across healthcare systems. This research underscores the potential of pairing Large Language Models like GPT-4 for large scale data labelling and with small language models like BERT for deployment on edge devices ensuring cost-effectiveness for systems similar in scale to Johns Hopkins Medicine.
The team concluded that OPTIC offers a scalable and efficient tool for triaging PMARs in health care systems. Leveraging advanced natural language processing techniques can reduce the administrative burden on healthcare providers and streamline patient care coordination.
The team’s future work will focus on enhancing interpretability and expanding the model’s capabilities to handle additional nuances in patient communication. This exciting research highlights the transformative potential of leveraging advanced natural language processing techniques to address healthcare challenges and improve patient outcomes.