JMIR Publications

ABSTRACT

Background:

Artificial Intelligence (AI) based assistive diagnosis systems mimic the deductive reasoning process of a human physician in biomedical disease diagnosis and treatment decision making. Impressive progress in this area has been reported. But most of the reported successes are the applications of AI in Western medicine. The application of AI in Traditional Chinese Medicine (TCM) is relative lagging behind mainly because TCM practitioners need to perform syndrome differentiation (Bian Zhen) as well as biomedical disease diagnosis before a treatment decision can be made. Syndrome, a unique concept to TCM, is an abstraction of a variety of signs and symptoms. The fact that the relationship between diseases and syndromes is not one-to-one, but rather many-to-many makes it very challenging for a machine to perform syndrome differentiation. So far, only a handful AI-based assistive TCM diagnosis models have been reported, and they are limited to a single type of TCM disease.

Objective:

To develop an AI-based assistive diagnosis system that can handle multiple types of TCM diseases. Given a patient’s electronic health record (EHR) notes, the system can simultaneously predict TCM disease and differentiate the corresponding TCM syndromes.

Methods:

Unstructured free style EHR notes are processed by a natural language processing (NLP) technique to extract clinic information such as signs and symptoms, as represented by named entities. The NLP technique employees a recurrent neural network named bidirectional long short-term memory network-conditional random field (BiLSTM-CRF) model. A convolutional neural network (CNN) is then used to predict the disease type out of 187 known TCM diseases. A novel TCM syndrome prediction method named integrated learning model is used to produce a list of syndromes. By following a majority-rule voting method, the proposed syndrome prediction method can take advantage of four existing methods while avoiding their respective weaknesses, resulting in a consistent high prediction accuracy.

Results:

A dataset with a total of 22,984 copies of EHR notes from Guanganmen Hospital of China Academy of Chinese Medical Sciences collected between January 1, 2017 and September 7, 2018 was used for this project. The dataset contained a total of 187 common TCM diseases. The proposed diagnosis system can detect any one of the said 187 TCM diseases. The dataset was partitioned into a training set, a validation set and a testing set following the customary partition ratio of 8:1:1. The test results suggested that the proposed system had a good diagnosis accuracy and strong generalization capability. Specifically, for the disease type prediction, accuracies of the top one (T1), top three (T3) and top five (T5) were 80.5%, 91.6% and 94.2%, respectively.

Conclusions:

An AI-based TCM assistive diagnosis system is proposed. Two main contributions of this work are the proposal of a TCM disease diagnosis system that can diagnose 187 known common TCM diseases and the proposal of a novel TCM syndrome differentiation method named integrated learning model. The proposed syndrome differentiation method outperforms all the four existing methods according to our preliminary experimental results. It is expected that more TCM disease types can be diagnosed and even better diagnosis accuracies can be achieved with further improvement of the proposed algorithms and the availability of more quality EHR data.