CycleTrans: a transformer-based clinical foundation model for safer prescription
Track: Traditional track
Keywords: Cycle Transformer; drug set recommendation; EMR; clinic foundation models
TL;DR: CycleTrans
Abstract: Deep learning techniques are extensively utilized in prescribing drug combinations, drawing on extensive electronic medical records (EMRs). A prescription assistant may be able to provide immediate guidance on drug combinations for some urgent clinical situations. A well-controlled drug-drug interaction (DDI) rate and high recommendation precision are of great importance for a safe prescription. A lower DDI often implies the set of drug combinations should be as small as possible, which is challenging because EMR prescriptions for certain symptom(s) are often highly noised due to the diversity side symptoms of individuals. We propose a model comprised of cycle transformers (CycleTrans) to handle these challenges. CycleTrans employs cross-attention and transformers, integrates patients' longitudinal EMRs, enhances knowledge representations through the so-called cycle-embedding module, and thus predicts safer and better essential drug combinations for new-coming cases. The new model achieves the state-of-the-art in three dimensions: high precision (89%), low DDI rate (0.34%), and small drug set size (3.02) on the MIMIC-III benchmark dataset, surpassing previous bests of 73%, 5%, and 17 in each dimension, respectively. Such a significant advancement makes a much safer clinic prescription possible. The idea of the cycle transformer we proposed has considerable potential for other domains besides clinics, such as set recommendations, translation, and unsupervised representation learning in knowledge graphs.
Presentation And Attendance Policy: I have read and agree with the symposium's policy on behalf of myself and my co-authors.
Ethics Board Approval: No, our research does not involve datasets that need IRB approval or its equivalent.
Data And Code Availability: Yes, we will make data and code available upon acceptance.
Primary Area: Clinical foundation models
Student First Author: Yes, the primary author of the manuscript is a student.
Submission Number: 37
Loading