Reproduce “NLICE: Synthetic Medical Record Generation for Effective Primary Healthcare Differential Diagnosis” on OpenResearch

About this paper

Zaid Al-Ars, Obinna Agba, Zhuoran Guo, Christiaan Boerkamp, Ziyaad Jaber, Tareq Jaber

This research from Delft University of Technology and Medvice Digital Health introduces NLICE, a systematic method for generating medically-correct and highly-expressive synthetic patient records for differential diagnosis in primary healthcare. Training machine learning models on NLICE-augmented data improved Top-1 accuracy to 82.0% and Top-5 accuracy to 99.0% for 55 conditions, outperforming traditional binary symptom representation.

Key insights

• NLICE symptom modeling significantly enhances the expressiveness of synthetic data, allowing ML models to differentiate between conditions with superficially similar symptoms more effectively.
• Models trained on NLICE-augmented data achieve substantially higher diagnostic accuracy and confidence compared to those trained on data with binary symptom representations.
• The NLICE framework, combined with robust ML models like Random Forest, demonstrates strong resilience to varying symptom counts, perturbed probabilities, and the injection of additional, potentially misleading symptoms.

Summary by alphaXiv.