Article
 ChingKan Lo, HsingChung Chen, PeiYuan Lee, MingChou Ku, Lidia Ogiela, ChengHung Chuang
 Smart Dynamic Resource Allocation Model for PatientDriven Mobile Medical Information System Using C4.5 Algorithm
 Journal of Electronic Science and Technology, 2019, 17(3): 231241
 http://dx.doi.org/10.11989/JEST.1674862X.71018117

Article History
 Manuscript received January. 13, 2017
 revised May. 31, 2017
H.C. Chen and C.H. Chuang are with the Department of Computer Science and Information Engineering and the Department of Bioinformatics and Medical Engineering, Asia University, Taichung 41354 (email: cdma2000@asia.edu.tw; chchuang@asia.edu.tw);
P.Y. Lee and M.C. Ku are with the Department of Orthopedics, ShowChwan Health Care System, Changhua 50008 (email: b1208@ms26.hinet.net; showjoeku@gmail.com);
L. Ogiela is with the Department of Applied Informatics, Akademia GórniczoHutnicza University of Science and Technology, Krakow 30059 (email: logiela@agh.edu.pl)
With increasing number of autonomous heterogeneous devices in the mobile networks, an efficient resource allocation scheme is required to maximize network throughput^{[1]}, memory, and energy optimization^{[2]}, so as to achieve higher efficient and better performance. Mobile health (mHealth), which is the usages of mobile computing together with communications technologies in health care and public health, is a rapidly expanding ehealth^{[3]} area, such as the usages of applications (apps) for postoperative followup in orthopedic surgery patients^{[4],[5]}. There are huge potential demands for mHealth interventions to obtain the beneficial effects of health, the delivery processes of health service as well as the improvement of the working time and satisfaction of the nurses^{[6]}. In a wireless sensor network (WSN), the usages of resources are usually highly related to the execution of tasks which consume a certain amount of computing and communications bandwidth^{[7]}. Due to the limitations of resource availability and communications medium, these existing algorithms cannot be directly addressed for the requirements of the mHealth system.
In addition, decision tree methodology has become more popular in medical researches. Some examples may include a predictive computerassisted decisionmaking system for traumatic injury using machine learning algorithms^{[8]} and diagnosis of coronary artery stenosis. The decision tree method is a powerful statistical tool for classification, prediction, interpretation, and data manipulation, which has several potential applications in medical research^{[9],[10]}. The Waikato Environment for Knowledge Analysis (Weka) is a popular suite of machine learning software written in Java, developed at the University of Waikato, New Zealand^{[11]}. It is a workbench that contains a collection of visualization tools and algorithms for data analyses and predictive modeling together with graphical user interfaces for easily accessing to these functions^{[11]}.
C4.5 is an algorithm used to generate a decision tree developed by Quinlan^{[12]}. C4.5 is an extension of Quinlan’s earlier ID3 algorithm. The decision trees generated by C4.5 can be used for classification, and for this reason, C4.5 is often referred to as a statistical classifier. Weka supports several standard data mining tasks, more specifically, data preprocessing, clustering, classification, regression, visualization, and feature selection. All of Weka’s techniques are predicated on the assumption that the data are available as a single flat file or a relation, where each data point is described by a fixed number of attributes (normally, numeric or nominal attributes, but some other attribute types are also supported)^{[11]}.
The rest of this paper is organized as follows: In Section 2, the related works are reviewed. Section 3 introduces the proposed algorithm model in detail. In Section 4, the simulation results of the proposed are presented and the algorithm’s performance is discussed. Finally, in Section 5, conclusions are made.
The goal of this paper is to create a smart and dynamic resourcemapping management table for the medical management information systems (MMISs) to optimize the weights of each server.
2. Related WorksIn general, network resources, storages, and energy allocation are a fundamental challenge in the mHealth system due to their unique features. Most of those traditional solutions do not consider resource consumption during communications and task execution. Therefore, they cannot be implemented efficiently. Furthermore, the allocation becomes a topic that remains largely unexplored for the mHealth system.
Several algorithms^{[13][15]} have been proposed for the task allocation and scheduling problem. Giannecchini et al.^{[13]} proposed an online task scheduling mechanism called CoRAl to allocate the network resources, between the tasks of periodic applications in WSNs^{[16],[17]}. Xie and Qin^{[14]} proposed another allocation strategy called balanced energyaware task allocation (BEATA) for collaborative applications running on heterogeneous networked embedded systems. Their strategies aimed at making the best tradeoffs between energy savings and schedule lengths^{[14],[16],[17]}. Lee and Jeong^{[15]} proposed a fuzzy relevancebased cluster head selection algorithm (FRCA) to solve problems, such as energy consumption, transmission rate reduction, decrease in throughput, and incorrect cluster head election.
However, there is little literature discussed a decision tree algorithm model to allocate multiple, heterogeneous resources in the mHealth system. This paper proposes to construct a decision tree more efficiently by reducing the incorrectness and ambiguity in the selection of service’s resources.
3. Resource Allocation Using Modified C4.5 AlgorithmIn this section, there are three subsections consisting of the machine learning analysis (MLA) using the C4.5 algorithm, the definition of symbols and formula, and the smart allocation algorithm using the modified C4.5 algorithm.
3.1 MLA Using Modified C4.5 AlgorithmThe C4.5 algorithm is applied to MLA in our algorithm as shown in subsection 3.1, which is one of the best decision tree algorithms^{[12]}. It builds decision trees using the concept of information entropy from a set of training data in the same way as iterative Dichotomiser 3 (ID3). The training data are a set of already classified samples. At each node of the tree, C4.5^{[12]} chooses the attribute of the data that most effectively splits its set of samples into subsets enriched in one class or the other. The splitting criterion is the normalized information gain (difference in entropy). The attribute with the highest normalized information gain (νγ) is chosen to make the decision. The algorithm then recurs on a smaller sub list.
Weka freeware with the J48 classifier^{[12]} selected for the C4.5 algorithm is used in this study. It does not require the discretization of numeric attributes, in contrast to the ID3 algorithm from which C4.5^{[12]} has evolved Microsoft Excel 2010 for statistical computation of the formula in comparing with the Weka. The entropy, split entropy, normalized information gain (NIG), and visualization of tree structure will be evaluated and shown.
3.2 Definition of Symbols and FormulasSome symbols and formulas applied in this paper are defined below.
1) g is the generation of algorithm, where g
2) n is the collection of resources or attributes to allocate.
3) m is the collection of the node, where m
4) i is the attribute index, and its value is from 1 to n.
5) j is the attribute value index from 1 to its final value that depends on its attribute (refer to Table 1).
Attribute  Attribute value 
1. Picture archiving and communications system (PACS)  1. Magnetic resonance image (MR) 
2. Computed tomography (CT)  
3. Xray (XR)  
4. Echography (Ec)  
2. Laboratory information system (LIS)  1. Culture (Cu) 
2. Biochemistry/serology (Bc)  
3. Hemogram (Hm)  
3. Nursing information system (NIS)  1. Vital sign chart (VS) 
2. Nursing education (NE)  
3. Nursing record (NR)  
4. Pharmaceutical management information system (PMIS)  1. Inpatient stat order (IS) 
2. Inpatient regular order (IR)  
3. Outpatient stat order (OS)  
5. Report information system (RIS)  1. Consultation sheet (CS) 
2. Inpatient note (IN)  
3. Image report (IR)  
6)
7) X^{g}^{,}^{m} is the total instance in the gth generation at the node m, where m
8) A is a number represented as the total counts of optimal performance for X^{g}^{,}^{m}, similarly, A_{i} is that for X_{i}, and A_{i,j} is that for X_{i,j}.
9) B is a number represented as the total counts of suboptimal performance for X^{g}^{,}^{m}, similarly, B_{i} is that for X_{i}, and B_{i,j} is that for X_{i,j}.
10) α^{g}
11) β^{g}
12) Set ệ^{g} and ệ
13) Set ệ
14) Set ệ
15) Set
16) Set
17) Set
18) κ is the node or relevant attribute with
19) Set
20) Set
21)
22)
23)
The purpose of this study is to tune the weight for resource allocation in accordance to β^{g}, θ_{κ}, and the cumulative of B_{i} from the bottommost node resource to its parent node and substantially to the uppermost root node. The updated weight is the sum of
Using the new weight for resource allocation, another iterative of generation g+1 will restart the above algorithm to tune the weight, so as to minimize β^{g}. And β^{g}<0.05 is set as a significant optimization or stop criteria of the algorithm in this paper.
4. Experimental Results and AnalysesIn MMIS, the five major resources of HIS will be the attributes of the modified C4.5 algorithm in a cooperation communications system, including PACS, LIS, NIS, PMIS, and RIS, where i equals to 1 to 5 as its ID, individually. The class of concepts is scored as A if the performance of MMIS is optimal or B if the performance is suboptimal. The details of the attributes and attribute values are shown in Table 1.
Moreover, the modified C4.5 algorithm is processed by three phases consisting of the root generation phase, child nodes and decision tree generation phase, and weight updating for resource allocation phase. The details are described below.
4.1 Root Generation PhaseThe follows are the processed steps by using the modified C4.5 algorithm. The details are listed below.
1. γ
2.
3. X^{0,0}
4. A
5. Entropy of parent node:
ệ^{0}
6. Entropy of MR:
ệ
7. Entropy of CT:
ệ
8. Entropy of XR:
ệ
9. Entropy of Ec:
ệ
10. Total entropy of PACS:
ệ
11. Information gain of PACS:
12. Split entropy of PACS:
13. Normalized information gain of PACS:
14. The ệ
Attribute  Value  A  B  ệ

ệ

ệ^{0} 

γ  νγ 
PACS (i=1)  MR  0  25  0  0.60  1.00  2.00  0.40  0.20 
CT  10  15  0.24            
XR  21  4  0.16            
Ec  19  6  0.20            
LIS (i=2)  Cu  10  23  0.29  0.94    1.58  0.06  0.04 
Bc  19  16  0.35            
Hm  21  11  0.30            
NIS (i=3)  VS  14  31  0.40  0.91    1.54  0.09  0.06 
NE  21  10  0.28            
NR  15  9  0.23            
PMIS (i=4)  IS  14  13  0.27  1.00    1.57  0  0 
IR  16  17  0.33            
OS  20  20  0.40            
RIS (i=5)  CS  19  15  0.34           
IN  17  13  0.30  0.98    1.58  0.02  0.01  
IR  14  22  0.35            
15. The maximum of
The algorithm then recurred in the smaller sub lists of PACS. As all samples of MR were B, the other 3 attribute values of CT, XR, and Ec were used separately to find the maximum
In the attribute value of CT, there were 25 instances of training samples,
1. X^{0.1}
2. ệ
3. Entropy of Cu of LIS is ệ
4. The maximum
5. The rest of the child node is shown in Fig. 1.
4.3 Weight Updating for Resource Allocation PhaseThe cumulative suboptimal performance B from the bottommost child nodes of LIS and NIS to the uppermost root node of PACS is 13, 25, and 50. So,
$ \begin{array}{l} {\small\sum}{\text{θ}} = 14 + 26 + 51 = 91.\\ {{\text{α}} ^0} = A/{X_0} = 49/100 = 49\text{%}. \end{array} $ 
The suboptimal performance rate is
$ {{\text{β}} ^0} = 1  {{\text{α}} ^0} = 51\text{%} . $ 
The base weight of all resources is
$ {\textit{Ƃ}}_i^0 = {W_i}\left( {{{\text{α}} ^0}} \right) = 10\text{%}\quad \left( {i = 1,2, \cdots\!,5} \right) $ 
There are three resources on the nodes of the decision tree. The weight gain of the root node PACS is
$ {\text{λ}} _1^0 = {{\text{β}} ^0}({{\text{θ}} _1}/{\small\sum}{\text{θ}} ) = 50\text{%} \times 51/91 = 28.6\text{%}. $ 
The weight gain of the child node NIS is
$ {\text{λ}} _3^0 = 50\text{%} \times 26/91 = 14.6\text{%}. $ 
The weight gain of the second child node LIS is
$ {\text{λ}} _2^0 = 50\text{%} \times 14/91 = 7.8\text{%} . $ 
The weight gain of both RIS and PMIS, that is, the resources that are not on the decision tree, is
$ {\text{λ}} _4^0 = 50\text{%} \times 0/91 = 0. $ 
The updated weight for the root generation is shown in Table 3.
Resources  PACS  NIS  LIS  PMIS  RIS 
θ_{κ}  51  26  14  0  0 
θ_{κ/Σ}_{θ}  56.0%  28.6%  15.4%  0  0 

20.0%  20.0%  20.0%  20.0%  20.0% 

9.8%  9.8%  9.8%  9.8%  9.8% 

28.6%  14.6%  7.8%  0  0 

38.4%  24.4%  17.6%  9.8%  9.8% 
After the updated weight allocation table is assigned to the resources, a new testing sample of 100 instances is recurred and shows a different class of concept. The optimal performance rate is improved to 78%, α^{1}=A/X_{1}=78/100=78%. Owing to β^{1}=1–α^{1}>0.05, the loop is going on to do the algorithm above.
The new decision trees of the first and second generations are drawn and shown in Fig. 2 and Fig. 3, respectively. In Fig. 2, the root node is RIS, and PMIS, NIS, and LIS are as child roots, respectively. After applying the updated resource allocation table, the second generation decision tree is shown with a root node of LIS, and NIS and RIS as its child nodes.
The resource allocation tables of the first and second generations are shown in Tables 4 and 5. The new resource allocation table in the first generation improved the optimal performance rate A and α^{2}
Resources  RIS  PMIS  LIS  NIS  PACS 
θ_{κ}  22  22  2  7  0 
θ_{κ}_{/Σ}_{θ}  41.5%  41.5%  3.8%  13.2%  0 

9.8%  9.8%  17.6%  24.4%  38.4% 
Ƃ^{0}_{κ}  7.6%  7.6%  13.8%  19.0%  29.9% 

9.1%  9.1%  0.8%  2.9%  0 

16.8%  16.8%  14.6%  21.9%  29.9% 
Resources  LIS  NIS  RIS  PMIS  PACS 
θ_{κ}  9  9  9  0  0 
θ_{κ}_{/Σ}_{θ}  33.3%  33.3%  33.3%  0  0 

14.6%  21.9%  16.8%  16.8%  29.9% 
Ƃ^{0}_{κ}  13.3%  19.9%  15.3%  15.3%  27.2% 

3.0%  3.0%  3.0%  0  0 

16.3%  22.9%  18.3%  15.3%  27.2% 
After the new weight allocation table of the second generation is assigned to the resources, the new different classes of concepts show that the optimal performance rate improved to 95%, α^{3}
In this study, a method of smart and dynamic resource allocation management algorithm is designed by the combination of the C4.5 algorithm and cumulative distribution function. The cumulative function accumulates the suboptimal classes of concepts in hierarchy tree nodes from the bottommost child nodes to the uppermost root nodes. The initial allocation is assigned with an evenly distributed weight. The optimal performance rate is improved from 49% to 78%, 91%, and 95% in the first, second, and third generations after the algorithm run. This algorithm could be applied in the mHealth system with different resources. Future work should be done for the services with more resources to allocate.
[1]  M. Hasan and E. Hossain, " Resource allocation for networkintegrated devicetodevice communications using smart relays,” in Proc. of IEEE Globecom Workshops, Atlanta, 2013, pp. 591596. 
[2]  X. Liu, K.M. Hou, C. D. de Vaulx,et al., " Memory and energy optimization strategies for multithreaded operating system on the resourceconstrained wireless sensor node,” Sensors (Basel), vol. 15, no. 1, pp. 2248, Jan. 2015. 
[3]  R. Whittaker, " Issues in mHealth: Findings from key informant interviews,” Journal of Medical Internet Researc, vol. 14, no. 5, pp. e129:19, Sept./Oct. 2012. 
[4]  J. L. Semple, S. Sharpe, M. L. Murnaghan, J. Theodoropoulos, and K. A. Metcalfe, " Using a mobile app for monitoring postoperative quality of recovery of patients at home: A feasibility study,” JMIR Mhealth and Uhealth, vol. 3, no. 1, pp. e18:111, Jan./Mar. 2012. 
[5]  C.K. Lo, Y.C. Wu, C.H. Chuang, and H.C. Chen, " A patientdriven mobile medical information system applying to physician clinical care,” in Proc. of the 9th Intl. Conf. on Innovative Mobile and Internet Services in Ubiquitous Computing, Blumenau, 2015, pp. 394398. 
[6]  C.K. Lo, S.S. Chang, C.H. Chuang, and H.C. Chen, " A mobile nursing app applying to the wound care and drug administration of patients,” in Proc. of the 9th Intl. Conf. on Innovative Mobile and Internet Services in Ubiquitous Computing, Blumenau, 2015, pp. 399402. 
[7]  W.Z. Guo, N.X. Xiong, H.C. Chao, S. Hussain, and G.L. Chen, " Design and analysis of selfadapted task scheduling strategies in wireless sensor networks,” Sensors (Basel), vol. 11, no. 7, pp. 65336554, Jul. 2011. 
[8]  S.Y. Ji, R. Smith, T. Huynh, and K. Najarian, " A comparative analysis of multilevel computerassisted decision making systems for traumatic injuries,” BMC Medical Informatics and Decision Makingvolume, vol. 9, pp. 2:117, Dec. 2009. 
[9]  Y.Y. Song and Y. Lu, " Decision tree methods: Applications for classification and prediction,” Shanghai Archives of Psychiatry, vol. 27, no. 2, pp. 130135, Apr. 2015. 
[10]  L. Ogiela and M. R. Ogiela, " Semantic analysis processes in advanced pattern understanding systems,” in Advanced Computer Science and Information Technology, T. H. Kim, H. Adeli, R. J. Robles, and M. Balitanas, Eds. Heidelberg: Springer, 2011, pp. 2630. 
[11]  G. Holmes, A. Donkin, and I. H. Witten, " WEKA: A machine learning workbench,” in Proc. of Australian New Zealnd Intelligent Information Systems Conf., Brisbane, 1994, pp. 357361. 
[12]  J. R. Quinlan, C4.5: Programs for Machine Learning, San Mateo California: Morgan Kaufmann Publishers, Inc., 1993. 
[13]  S. Giannecchini, M. Caccamo, and C. S. Shih, " Collaborative resource allocation in wireless sensor networks,” in Proc. of the 16th Euromicro Conf. on RealTime Systems, Catania, 2004, pp. 3544. 
[14]  T. Xie and X. Qin, " An energydelay tunable task allocation strategy for collaborative applications in networked embedded systems,” IEEE Trans. on Computers, vol. 57, no. 3, pp. 329343, Mar. 2008. 
[15]  C. Lee and T. Jeong, " FRCA: A fuzzy relevancebased cluster head selection algorithm for wireless mobile AdHoc sensor networks,” Sensors (Basel), vol. 11, no. 5, pp. 53835401, May 2011. 
[16]  N. Renugadevi and C. Mala, " Improved group key agreement for emergency cognitive radio mobile ad hoc networks,” Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, vol. 6, no. 3, pp. 7386, 2015. 
[17]  A. Skovoroda and D. Gamayunov, " Securing mobile devices: Malware mitigation methods,” Journal of Wireless Mobile Networks, Ubiquitous Computing, and Dependable Applications, vol. 6, no. 2, pp. 7897, 2015. 