HerbMiners Informatics Limited is a clinical Traditional Chinese Medicine (TCM) intelligence software solutions company. HerbMiners Informatics Limited focuses on research in TCM data mining, which aims to reveal relationships between symptoms, illnesses, herbs and prescriptions. HerbMiners Informatics Limited also provides artificial intelligence software solutions which assist hospitals and clinics for TCM modernization and patient records digitization.

1. Artificial Neural Network based Chinese Medicine
Diagnosis in decision support manner and herbal
ingredient discoveries
Wilfred Lin, Jackei Wong
HerbMiners Informatics Limited
Unit 209A, Photonics Centre, Hong Kong Science Park, Hong Kong
wlin@herbminers.com, jwong@herbminers.com
Abstract. Artificial neural network (ANN) for fast and trusted herbal ingredient
discoveries is proposed. It is fast, because different ANN modules can be
executed in parallel, and the ANN results are trustworthy, because they can be
verified by TCM domain experts in real clinical environments in Hong Kong,
Nanning, GuangXi, China and New York, United States of America. The ANN
is able to learn the relationship between herbal ingredients and the set of
information given (e.g. symptoms and illnesses). The ANN output is called the
relevance index (RI), which conceptually associates two TCM entities.
Keywords: Chinese Medicine Diagnosis, Artificial Neural Network, HerbMiners Cloud
Computing Platform
1. Introduction
We propose a novel approach of applying the artificial neural network (ANN)
based on backpropagation for Chinese Medicine Diagnosis in decision support
manner and for fast, trusted herbal ingredient discoveries. The approach will be
verified in a real TCM (Traditional Chinese Medicine) clinical environment. Firstly,
the ANN modules will be trained with real patient cases, and secondly domain experts
will be enlisted to confirm the discoveries to make them trustworthy. If the relevance
between two TCM entities (e.g. an herbal ingredient and an illness) was never
explicitly defined/annotated but is revealed by the trained named ANN module (e.g.
named after an herbal ingredient or illness), it is a potential discovery in the context of
the proposed ANN approach.
The proposal of the novel ANN approach is inspired for the following reasons:
a) Our previous research and development experience in the area of clinical
TCM ontology [1] indicates that ontological constructs can be huge and
complex. Fast herbal discovery may therefore be quickened by parallel
processing [2]. In this light, many trained named ANN modules could be
invoked at the same time for parallelism and thus speedup.
b) The same basic ANN construct can be trained by different datasets to become
specialized ANN modules, named after the specific TCM entities (e.g. an
illness); for example the FluANN module is dedicated to Flu analysis.

2. c) The relevance between two TCM entities (e.g. an herbal ingredient U and an
illness (or a set of illness) V) can be computed to indicate the likelihood of a
discovery. The computed value is called the relevance index (RI) between U
and V, to explicitly show their association. The algorithm to compute the
specific RI should be already consensus-certified by domain experts.
2. Related Work
The HerbMiners Cloud Computing Platform, Chinese Medicine Clinical Data
Warehouse architecture is illustrated in Figure 1.
Figure 1. The HerbMiners Cloud Computing Platform, Chinese Medicine
Clinical Data Warehouse architecture
Our literature review of soft computing techniques [3, 4, 5, 6, 7] indicates that the
artificial neural network (ANN) based on backpropagation is suitable for fast, trusted
herbal ingredient discoveries. The reasons are: i) reusability – the same ANN
construct can be trained to become named ANN modules that assume different roles;
ii) simplicity – it is easy to program and less error-prone than the traditional
algorithmic programming approach; iii) data-orientation – the logical points inside an
ANN construct will converge to the required logical operation with respect to the

3. given training dataset; iv) versatility – an ANN construct can be combined with its
clones or other constructs to form larger, more complex ANN configurations; v)
adaptability – the neuron’s activation function can be replaced any time, and the input
parameters to a neuron can be weighted and normalized according to the needs; vi)
optimization – an ANN can be effectively optimized or pruned for a particular
operation [5]; vii) commodity – many ANN constructs in the form of freeware are
available in the public domain with rich user experience, viii) accuracy – as long as
the number of the hidden neurons is twice that of the input neurons the ANN output is
accurate [8, 9]; and ix) parallelism - many named ANN constructs can be invoked to
work in parallel for speedup. The ANN configuration by propagation has a 3-layer
architecture: i) a layer of input neurons; ii) a layer of hidden neurons interconnected
with the input neurons; and iii) one output neuron interconnected with the hidden
neurons. The behavior of every neuron is governed by its activation function (e.g.
Sigmoid) [3]
Figure 2. Modernized Chinese Medicine Clinic with Chinese Medicine Granules
for better quality control