TLDA: A transfer learning based dual-augmentation strategy for traditional Chinese Medicine syndrome differentiation in rare disease

Abstract

The Traditional Chinese Medicine (TCM) has demonstrated its significant medical value over the decades, particularly during the COVID-19 pandemic. TCM-AI interdisciplinary models have been proposed to model TCM knowledge, diagnosis, and treatment experiments in clinical practice. Among them, numerous models have been developed to simulate the syndrome differentiation process of human TCM doctors for automatic syndrome diagnosis. However, these models are designed for normal scenarios and trained using a supervised learning paradigm which needs tens of thousands of training samples. They fail to effectively differentiate syndromes in rare disease scenarios where the available TCM electronic medical records (EMRs) are very limited for each unique syndrome. To address the challenge of rare diseases, this study proposes a simple yet effective method called Transfer Learning based Dual-Augmentation (TLDA). TLDA aims to augment the limited EMRs at both the sample-level and feature-level, enriching the pathological and medical information during training. Extended experiments involving 11 comparison models, including the state-of-the-art model, demonstrate the effectiveness of TLDA. TLDA outperforms all comparison models by a significant margin. Furthermore, TLDA can also be extended to other medical tasks when the EMRs for diagnosis are limited in samples.