Because of human healthcare, the pharmaceutical industry can be considered as one of the most significant industrial sector. For this reason, demand ...
Proceedings of 7th International Symposium on Intelligent and Manufacturing Systems (IMS 2010) , International University of Sarajevo, Sakarya University Department of Industrial Engineering, Sarajevo, Bosnia Herzegovina, September 15-17, 2010: 52-59
Demand Forecasting In Pharmaceutical Industry by Using Neuro-Fuzzy Approach Gökçe CANDAN1, Mehmet Fatih TAŞKIN2, Harun Reşit YAZGAN3 {gcandan1, mftaskin2, yazgan3}@sakarya.edu.tr *Sakarya University, Department of Industrial Engineering, Esentepe Campus 54187 Sakarya/ TURKIYE
Abstract Because of human healthcare, the pharmaceutical industry can be considered as one of the most significant industrial sector. For this reason, demand forecasting in pharmaceutical industry is more complex structure than other sectors. Human factors, seasonal diseases, epidemic diseases, market shares of the competitive products and marketing conditions can be considered as main external factors for forecasting pharmaceutical product. In addition that active ingredients rate is also important factor for forecasting process. The main objective of the research is to predict future periods demand from previous sales quantity with considering effects of the external factors by employing a neuro-fuzzy approach. Because of the biases of external effects in Artificial Neural Network (ANN) topology, an ANFIS as neuro fuzzy approach is applied. An example is given to illustrate effectiveness of the approach. Keywords: Demand forecasting, Neuro-Fuzzy, Pharmaceutical Industry
1.Introduction Today, health and treatment services are leading issues about prosperity and democracy concept. Therefore the pharmaceutical industry can be defined as a strategic sector. Pharmaceutical sector is quite moving and different from other sectors because of the drug, raw material inputs, production process, many factors such as customers and industry are different in terms. The sector as a distinctive structural reasons can be summarized as follows [1]: to be consumed by the disease and the related uncertainty about the availability of drugs, Tripartite structure of demand (patients, doctors and health insurance) and so different consumption structure patent protection and brand loyalty, by manufacturer is highlighting the market power highlighting the market power Industrial policy and politics in health policy where practitioners have to make choices.
52
In this study, the actual sales amounts of a pharmaceutical company for seven years period were considered and it was aimed to forecast of next period’s demands. A product which has three active raw materials called zinc oxide, lidocain HCI, diphenhydramine HCI was chosen for prediction. It is a dermatologic lotion and it is used to eliminate the pain temporarily in insect bites, simple skin irritation and itching bond cases. A year was divided as four periods by marketing departments. Because of the selling quantity is affected from seasonal demands, next period demand was predicted with employing previous periods’ demands. 2.Demand Structure of Pharmaceutical Industry Demand forecasting is one of the main the process of planning and main objective is to find out which products will be purchased, where, when, and in what quantities. Pharmaceutical companies and vaccine manufacturers are well aware of the political and economic dangers of poor forecasting. Developed markets are also characterized by relatively good information and market research, and by purchasers and suppliers with established relationships and balanced market power. In other industries, the current conditions call for forecasting methods that encourage dialogue among a diverse set of players through systematically gathering and sharing available information, creating scenarios independent of political pressure, and combining forecasts from various sources for greatest accuracy. Demand in the pharmaceutical market is a complex combination of push and pull driven demand and regulatory and pricing pressures (Figure 1). Third parties decide on formularies that influence prescription behavior, while pharmaceutical companies use lawyers, managed care, detailing, and direct-to-consumer campaigns to influence the Food and Drug Administration (FDA), pharmacy benefit providers, doctors and patients to prolong the patent, promote on formulary, prescribe and request their products [2]. There are, in fact many who could provide valuable input to forecasting for health products, but no focused forum exists to bring these groups together. The Forecasting Working Group will serve as such a forum to bring together a broad range of global health stakeholders with experts from other industries and disciplines in order to generate critical thinking on the subject of forecasting and develop recommendations for a broad set of actors on ways to improve forecasting and better match supply to demand [3].
53
Figure 1: Complex Structure of Pharmaceutical Market
3.The proposed methodology Improved forecasting methodologies exist, but the methods most common in pharmacy are fairly straightforward, relying on direct human judgments with implicit rather than explicit assumptions and limited quantitative data. The two most commonly used methods are “Consumpion Method” and “Morbidity Method” [3]. The consumpion was employed method in the study. This method causes historical data of past consumption to predict future requirements. This is obviously the most reliable method when good consumption data are available, for existing products, and for products where use patterns are well established. In this study, a neuro-fuzzy approach was employed for determining next period’s demands. 3.1.Neuro-fuzzy Approach Artificial neural networks (ANN) are among the newest signal-processing technologies in the engineer's toolbox. When creating a functional model of the biological neuron, there are three basic components of importance. First, the synapses of the neuron are modeled as weights. The strength of the connection between an input and a neuron is noted by the value of the weight. Negative weight values reflect inhibitory connections, while positive values designate excitatory connections [Haykin]. The next two components model the actual activity within the neuron cell. An adder sums up all the inputs modified by their respective weights. This activity is referred to as linear combination. Finally, an activation function controls the amplitude of the output of the neuron. An acceptable range of output is usually between 0 and 1, or -1 and 1 [4].
54
A general structure and mathematically structure of this process is described in Figure 2 and Figure 3 respectively.
Figure 2: General structure of ANN [5] Fixed input x0= ±1
wk0=bk (bias)
x0
wk0
x1
wk0
wk0
x2 ● ● xp Input Signals
● ●
Activation Function vk
₰(●)
∑
Output yk
Summing Function
wk0
Өk
Synaptic Weights
Threshold
Figure 3: Mathematical structure of ANN
The interval activity of the neuron can be defined as follows:
The output of the neuron, yk, would therefore be the outcome of some activation function on the value of vk [4]. 3.2.ANFIS ANFIS is the implementation of fuzzy inference system (FIS) to adaptive networks for developing fuzzy rules with suitable membership functions to have required inputs and outputs. In ANFIS (Adaptive Network Based Fuzzy Inference System) structure, both artificial neural networks and fuzzy logic are employed [6]. ANFIS has a five-layer feedforward neural network structure. These structures are as follows [7];
55
1. In first layer; the node number is equal to the number of input variables membership functions of these variables are used as node functions. Membership functions of these parameters, are called "the premise parameters” 2. In second layer; the nodes are constant and node number is equal to the number of rules 3. Normalizing weights of the rules is completed in the third layer. 4. The nodes in the fourth layer are adaptive. The node function is from the Sugeno method 5. The output of fifth layer consists of a single node and crisp character is the output of the model. The layers of the ANFIS is illustrated in Figure 4. X Y
A1 W1
X
A2
T T
1
N
1f1
f B1 Y
B2
T T
W2
N
2
2f2
X Y Layer 1
Layer 2
Layer 3
Layer 4
Layer 5
Figure 4: Layers of ANFIS [7]
Learning type in adaptive ANFIS is chosen generally as hybrid learning and there is a type of a hybrid learning system defined by Jang is shown in Figure 5.
Figure 5: Learning System of Anfis [8]
To design a fuzzy controller interface with ANFISedit, the following steps can be applied [8,9]; 1. To collect the system data and introducing the system data.
ANFISedit interface will be employed for
56
2. To build an appropriate “fismat” structure 3. At this stage, the proposed neural network is trained. Ideally, training error is expected as zero value. 4. Fismat structure caused by using performance data to determine and control the user is presented graphically. While designing a fuzzy logic controller with ANFIS, the following arrangements should be met “Sugeno” type must necessarily be established for controller. There should be input and output data pairs that identify the system better. Anfis is employed as predefined Matlab membership function In the system, each rule must be work on a single output member 4.Implementation of ANFIS According to the demand structure, the whole sales in a year were divided as four periods. Each of the periods effect on the output variable with a bias constant. The membership functions were employed in order to describe the sales volumes of the periods. The Membership functions consist of three variables such as low, normal and high. Next period’s demand quantity was forecasted with employing defined membership functions. The general “fis” structure was illustrated in Figure 6a The model consisted of four periods as four inputs, next period’s demand as one output and Sugeno method was chosen. The classification and description of membership functions in inputs is illustrated as follows in Figure 6b.
a) a)
b)
Figure 6: a) General FIS Structure of Model
b) Membership Functions of Input2
According to demand structure of pharmaceutical industry, the rules for evaluating were defined with using the Rule Viewer in Figure 7.The numerical volumes of inputs and next period’s amount were calculated by Anfis model. It was realized that periodical variations of inputs directly affected to the output demand amount.
57
Figure 7: Rule Viewer
Figure 8: Anfis Model Structure
The inputs, membership functions of inputs and output, rules, etc. in a whole structure were described in the Anfis. The proposed Anfis model structure is shown in Figure 8. Actual and predicted data are shown at Table 1. Table 1. Actual and predicted demand value of the product Periods 1 2 3 4 Year Actual Predicted Actual Predicted Actual Predicted Actual Predicted 92000 46172 50000 2003 153814 165000 187958 192000 92000 27500 2004 106574 105000 131295 135000 252000 252000 26508 92000 196490 198000 223000 223000 92731 90250 2005 88009 52500 2006 137126 138000 314252 300000 175000 175000 51109 72500 2007 138045 142000 337552 325000 155000 155000 71400 2008 162821 180000 149113 150000 203000 203000 170245 173000 2009 99878 105000 493709 500000 242500 242500 261800 265000 After 200 epochs, average testing error was found as 0,18169. Error between the training data and the output is shown in Figure 9.
Figure 9: Error Graphic of Model
Figure 10: Surface Period2 and Period3
58
Analysis
Between
Because of the product’s dermatological effect, in period2 and period3, there are large amounts of the product. The same effect results can be seen in Figure 10. 5.Conclusion In the company, an amount of product for each period is required to predict by the marketing department. In the production planning department, the predicted values are considered as production quantity to achieve with taking into consideration of technical production constraints. Actual demands quantity of the product for six years period, were taken into consideration. It is seen that the proposed approach predicts demands value within reasonable error ranges.
6.References [1] Giuffrida, A. “Learning from the Experience: The Inter-American Development Bank and Pharmaceuticals”, Inter-American Development Bank, Washington, (2001), www.iadb.org/sds/soc [2] Prest, R.“Real Demand Forecasting” http://www.pharmamanufacturing.com/articles/2007/178.html [3] Sekhri, N. Forecasting for Global Health: New Money, New Products & New Markets Center for Global Development February 2006 [4] http://www.learnartificialneuralnetworks.com/ [5] Watson, G.
http://glennwatson.net/
July 2010
[6] Kosko, B.. Neural networks and fuzzy systems, A Dynamical Systems Approach, Englewood Ciffs., NJ: Prentice Hall. 1991 [7] Caner, M. and Akarslan, E. “ Estimation of Specific Energy Factor in Marble Cutting Process Using ANFIS and ANN”, March 2009 [8] Jang J.S., ANFIS: Adaptive-Network-based Fuzzy Inference System, IEEE Trans. On System, Man and Cybernetics. Vol.23, No 3, 665-685, May/June 1993. [9] Jang, J. S., Gulley N., Fuzzy Logic Toolbox User’s Guide, The Mathworks Inc., 1995
59
