Deep Neural Network for Heart Disease Medical Prescription Expert System

ABSTRACT


INTRODUCTION
Ischaemic heart disease (IHD) is related to various cardiovascular circumstances or risk factors and consequently, treatment can be complex.Thus, "treatment deserves a comprehensive management approach, including pharmacotherapeutic and invasive or surgical therapies, professional lifestyle interventions based on behavioural models of change, with different strategies from more basic, family-based to more structured and complex modalities, depending on the cardiovascular risk assessment and on concomitant diseases" [1].Risk factor management concentrating on monitoring related cardiovascular risk factors, covering psychosocial support, physical activity advice and appropriate prescription of an adherence to cardioprotective drugs are integral to serving patients recover as normal a life as possible and advance their quality of life.
IHD is one of the most common causes of death in the world.This is the reason why the development of a broadly accepted expert system for accurately performing diagnostic and therapeutic  ISSN: 2089-3272 IJEEI, Vol. 6, No. 2, June 2018 : 217 -224 218 decision model of ischaemic heart disease constitutes a general need.In addition, insufficient IHD specialist in developing countries especially in rural area requires a support of an expert system to help non-specialist to accurately diagnose and treat IHD patients.
Physicians need a systemic approach to actual clinical problems.Such a strategy ensures the maximum diagnostic accuracy at minimum risk and expense to the patient.Most cardiologists pursue a line of questioning to establish whether the actual chest pain is typically cardiac, a typical or non-cardiac in origin.It is therefore important to determine the sequence of the steps in the diagnosis of IHD.Clinical decisions are often made based on doctors' intuition and experience rather than on the knowledge-rich data hidden in the database.This practice leads to unwanted biases, errors and excessive medical costs which affects the quality of service provided to patients.
Many works (such as [2][3][4][5][6]) have already been carried out on the development of expert systems for IHD diagnosis and treatement with the use of intelligent systems such as fuzzy logic, genetic algorithm and artificial neural network.The existing expert systems using conventional neural network face problem with accuracy.This work attempts to develop an intelligent engine to improve the accuracy of such expert system by leveraging the deep learning concept.The expert system is expected to help cardiologists as well as general physicians (in case the absence of the cardiologists in rural area) to accurately treat the IHD patients.

RESEARCH METHOD
This work uses the deep neural networks due to its advantace in processing inter-related string/text input data.The details of the patient data/information are shown in Table 1 to Table 10.The medicines as treatments for the IHD along with their codes (MIDs) are shown in Table 11.While Table 12 and Table 13 show the twenty of clinical data examples on symptoms examinations, and the medicines prescribed by the physician, respectively.

Deep neural network
Deep Neural Learning or Deep Neural Network (DNN) is a subset of machine learning in Artificial Intelligence (AI) that has networks which are capable of learning unsupervised from data that is unstructured or unlabeled.The deep neural network consists of more than 1 layer in its hidden layer [7][8][9][10].The DNN in this work, has 152 neurons in the input layer, 52 neurons in the output layer, and 4 layers in the hidden layer.Each attribute of patient is represented by binary numbers.The number of digit representing the attribute depends on the number of fields shown in Table 1 to Table 10.The number of digit of P1=18, P2=29, P3=4, P4=25, CVS=8, RS=5, PA=6, CNS=5, ECG=21, BI=24.Thus, the total is 152 as the number of neurons for the input.Digit 1 represents that the symptom exists, otherwise is 0. For example, the input for the fourth patient record in Table 12 is as follows.0110111 010000000000000000 11001000000010000000000000000 0001.0000001001000000000000000 00000001 00010 000010 00010 101000000000000000000 000000000000010000000000.The output is medicine prescriptions as shown in Table 11.So, there are 52 neurons in the output layer.As illustration, the fourth patient's medicines prescription in Table 13 is represented as:

The Algorithm
This work adopts the deep learning method for Retristic Boltzman Machine (RBM) from the works in [6] and [7].250 patient's data are used for training and the other 55 data are for testing.
The main aim of the RBM training algorithm is to maximize the product of probabilities assigned to some training set V (a matrix, each row of which is treated as a visible vector v): The algorithm optimizes the weight matrix W [6].The algorithm performs Gibbs sampling and is used inside a gradient descent procedure (similar to the way backpropagation is used inside such a procedure when training feedforward neural networks) to compute weight update.The basic, single-step contrastive divergence procedure for a single sample is summarized as follows: a.Take a training sample v, compute the probabilities of the hidden units and sample a hidden activation vector h from the probability distribution.b.Compute the outer product of v and h and call this the positive gradient.c.From h, sample a reconstruction v' of the visible units, then resample the hidden activations h' from this.
(Gibbs sampling step) d.Compute the outer product of v' and h' and call this the negative gradient.e.Let the update to the weight matrix W be the positive gradient minus the negative gradient, times some learning rate: f. Update the biases a and b analogously: (4)

Implementation
The algorithm is implemented using a high-end PC machine with the following specifications: Processor name: Intel Core i7-7700, Processor speed 3.8 GHz, RAM 16 GB, Storage 2.5 TB and Windows 10 O/S.Tensorflow utility/library is used on the Phyton platform.

RESULTS AND ANALYSIS 3.1. Training Results
Five (5) series of training were conducted with the same number of epochs.Figure 2 shows the training results.The average error is 0.009895.The training time was also recorded and in average the proposed system achieved a rational training time compared to the standard/conventional neural network.It was 15% slower.

Testing Results
Fifty-five (55) data testing were used and some data that used during the testing.We compared the accuracy of the proposed expert system (using DNN) with standard/conventional Neural Network (NN).Five experiments are carried out and the results are shown in Figure 3.The average accuracy of the proposed expert system with DNN is: (99.60 +99.75+99.801+99.801+99.851)/5=99.7874%,whereas the average accuracy of the expert system with NN is: (99.16+99.002+99.1+99.002+99.012)=99.0552.Thus, the proposed expert system with DNN improves the accuracy with 0.7322%.

CONCLUSION
A deep neural network has been used to increase the accuracy of the expert system for IHD treatment.The proposed expert system that uses deep neural network has improved the accuracy by 0.7322% compared to the expert system that uses conventional neural network. in some cases, the system suggested new prescriptions for some patients, however, those prescriptions do not affect the cardiatic problems, according the advices from the physician.On the other hand, the medicines that were prescribed by the physician but were not prescribed by the system do not significantly affect the IHD.In other word, it increases efficiency in term of cost for IHD treatment.

2. 1 .
Dataset 305 Patient clinical data related to IHD are collected from a hospital in Jakarta, Indonesia.The data is collected as symtomps and medicine prescriptions.Each patient information has ten attributes: -

Figure 1 .
Figure 1.Architecture of the DNN for Heart Disease Medical Prescription Expert System (Majzoob K. Omer) 223

Figure 3 .
Figure 3. Accuracy of the proposed system

Table 5 .
Cardio Vascular System

Table 6 .
Respiratory System

Table 11 .
The Medicines and Their Codes Deep Neural Network for Heart Disease Medical Prescription Expert System (Majzoob K. Omer) 221

Table 12 .
Examples of Symptoms Examinations of 20 Patients

Table 13 .
The Corresponding Medicine Prescriptions for the 20 Patients

Table 14
shows the sample results of the testing that have discrepancies compared with the actual medicine prescriptions given by the physician.Analysis from the physician are given for some discrepancies of the results.For example, for patient number 4: -Medicine no.16 is suggested by the system as additional prescription.It is beneficial as it reduces the heart rate and thereby reduces workload and improves outcome.-Medicine no.19 is not given by the system.It is an antipyretic drug (to reduce fever) or an analgesic that if it is given, will not affect the cardiac outcome.
-Medicine No. 28 is injectable form of medicine no.14 which the system has already prescribed.-Medicine no.36 is not given by the system.It is a laxative (Stool Softener) is given to patients who complain passing hard stools which cannot be judged by the system.