Abstract—The healthcare sector has been confronted with a growing necessity to reduce operational cost. Many hospitals have been focusing their efforts in optimizing their inventory management procedures through the incorporation of technological solutions such as tracking devices and data mining to come up with an ideal inventory model. Demand forecasting is an integral part of inventory management and hospitals are no exception. Time series forecasting methods are widely used in traditional approaches. Limited studies integrated asset tracking technology and neural network analysis to facilitate demand forecast. This paper proves that neural network forecasting has a key edge over traditional time series forecasting methods. It also evaluates the improvements in the efficiency of the inventory management of infusion pumps at Tan Tock Seng Hospital (TTSH) due to the integration of radio frequency identification (RFID) tagging and neural network forecasting to the current work flow process to allow it to capture and manipulate the data relating to the movement and usage of the infusion pumps. Projected ward and the total in-patient usage data were compared using error analysis algorithms such as mean squared error (MSE), mean absolute deviation (MAD) and mean absolute percentage error (MAPE). The potential benefits of the proposed system, contribution of current study and recommendations for future research are also mentioned at the end of this paper.
Nowadays, the role of demand forecasting of medical assets has increased significantly due to the various innovative and effective concepts of forecasting science and inventory management which helps greatly to keep the hospital operations cost under control . Managing the inventory levels is important to the operations and the management of the hospital’s assets. Hospital operations have to take a look at their in-patient flow to make decisions on their resource capacity. In-patient care is one of the main drivers of demand for resources in hospitals . In-patient systems have very complex throughput systems that make the medical inventory planning much more complicated. Factors of the in-patient flow process such as non-stationary arrival and varying medical service processes make current static forecasting models rather obsolete  as they do not capture the compound behavior of the true inpatient system. The mismanagement of resources has considerably more impact on the lives and well-being of the patients being served. Forecasting plays a critical role in the medical inventory management. However the challenge that most hospital management faces is the lack of visibility and integration of already present data; data that is routinely collected but stored in differing information systems into useful demand forecasting that can help improve the medical inventory management . Current medical inventory management systems can be categorized into four main conceptual components which are physical infrastructure, inventory planning and control, information system as well as organizational embedding . However, due to a huge amount of medical items and human-intensive working processes, current systems cannot provide a timely and accurate inventory management and forecasting. To improve the situation, the future of inventory management is to build up an automated work flow system that requires minimal manual interaction. This represents a state where the medical amenities replenishment requirements are aggregated and an order is placed automatically. The usage data is also recorded to allow for the hospital management to use to predict for the future demand
- NEURAL NETWORK FOR INVENTORY MANAGEMENT
With the improvements of statistical models and forecasting techniques, the complex throughput can be studied and an ideal inventory which can process the data inputs to come up with an efficient state of inventory management can be modeled. Current time-series methodology attempts to first identity forecasting parameters such as trend cycle, seasonality and irregularity and then extrapolates these components to come up with the forecasts. However these trend-cycle and seasonal data components of a time forecast tends to evolve over time and needs to be continuously revised for higher accuracy in forecasting. In addition, a key assumption to the time-series forecasting model is that the activities responsible in influencing the past will continue to influence the future. This is often a valid assumption whilst forecasting for a short-term demand, but falls short when attempting to forecast for long-term analysis . A neural network forecast is proposed to handle the deficiency. It uses analytical methodologies that make use of the historic demand data as inputs and updates information over time as the number of training data sets provided is increased . The adaptive and learning abilities of this neural network improves the forecasting accuracy so that better decisions can be made. The key to achieve accurate demand forecasting is to have good pattern recognition. Back propagation algorithm of NN is a typical supervised learning algorithm, where the neural network is trained by setting the input vectors and the corresponding target vectors. After the neural network is changed, approximate function is used to recognize a pattern. Levenberg – Marquardt, which is the one of the most effective algorithm for function approximation problems, will be studied in this research. The advantage of Levenberg-Marrquart algorithm can approach second-order training speed without computing the Hessian matrix, which is the square matrix of second-order partial derivatives of errors with respect to weights.
- AN RFID BASED INVENTORY MANAGEMENT SYSTEM
An RFID based inventory management system (RFID-IMS) integrates RFID tracking services and the neural network model to assist in the tracking of the medical devices such as infusion pumps throughout the hospital and allows the storage of the real time data. Greater visibility on the infusion pump movement and the demand characteristics will allow the operations department to come up with more effective supply chain solutions to manage their infusion pumps. Data on the actual movement and usage of the infusion pumps are captured using the RFID technology and is feedback to the neural network platform for aggregated analysis of the inventory of the pumps. The proposed workflow has been modified to suit the infusion pump inventory management in TTSH from retail industry . The detailed framework is shown in Fig. 1. Firstly, the RFID-IMS uses RFID technology to capture of the usage data within a certain periods and this information is then used as input for the neural networking model to calculate the demand forecast. Then, neural network analysis is conducted to analyze the demand pattern and to predict the systematic and random component. Neural network forecasting is an enhancement of the time series and the casual forecasting templates. In this study, neural network toolbox from Matlab is used.
Neural network forecasting requires accurate analysis of smoothening parameters such as level, trend and seasonality which may not be acquired immediately without the RFID tracking and this will affect the neural network forecasting accuracy. Next, the forecasted values are then feed into the RFID-IMS to construct virtual aggregation of demand. It uses these forecasted values to aggregate demand for the individual wards. This simulation of the infusion pumps at each ward allows for better streamlining of process and improves on the current manual system. With the RFID tracking and neural network forecasting, RFID-IMS allows auto generation of the number of sets of the pumps to be issued from the central equipment base to the wards. This eliminates the need for end users to raise a request for the number of sets of infusion pumps to be issued and also physically count the number of sets returned. The workflow process is optimized with the automation of the infusion pump inventory. Healthcare workers now have more time to focus on patient care as there is no longer a need for physical stock-taking or to raise a request to receive the set. With the neural network forecasting showing high levels of accuracy in predicting the futuristic demand patterns, the wards would have the ideal number of infusion pumps that they require hence reducing the need to borrow the infusion pumps. This saves time for the healthcare workers who can concentrate their effects in taking care of the patients
- CASE STUDY ANDDISCUSSION
The proposed framework was trial implemented in a Tan Tock Seng Hospital (TTSH). It is the second largest hospital in Singapore and one of the nation’s biggest multi-disciplinary hospitals with more than 160 years of pioneering medical care and development. In 2012, TTSH had 36 clinical and allied health departments with 15 specialist centers and powered by more than 6000 healthcare staff. Tan Tock Seng Hospital currently held 827 infusion pumps that were manually tracked for usage and preventive maintenance by the healthcare workers in the wards. Due to the management of huge amount of infusion pumps, several problems regarding infusion pump are triggered and shown in A, B and C in section IV respectively. Table I shows the current time-line of the work flow process when shortage of infusion pumps.
A.Manual Search for the Pumps for Patient Usage
Healthcare workers had to perform a manual search within the ward for any available infusion pumps and then checked with other wards manually if there were no available infusion pumps within their wards. This caused an increase in the waiting time of the patients for the infusion pumps.
B.Manual Administrative and Paperwork
Due to a lack of a visibility over the infusion pumps a lot of administrative time was spent on locating the pumps for periodic maintenance, stock-taking or to find a ‘lost’ pump. When a certain ward had a shortage of a pump and there arose a need for loan or swapping, healthcare workers had to spend time doing manual searches for the pump. When they found the pump, they had to spend time doing manual paperwork and handover. This decreased the direct patient care time.
C.Fluctuation in Infusion Pump Supply
Pumps were periodically taken away for periodic maintenance with having replacement units. This batch maintenance approach created periodic variations to the supply of in-service pumps at the different wards as a shortage of pumps was triggered during the periodic maintenance wards. In this study, the RFID-IMS was developed to improve the situation. The RFID-IMS leverages the RFID technology based on the hospital’s existing wireless ‘N’ access points to track real-time location. It is able to track the location of tagged assets or individuals in real-time. It also allows for the provision of real time information system that offers more visibility on the infusion pump location and the utilization rate. In order to evaluate the forecasting performance of the RFID-IMS, the error difference between the forecasted values and the actual values for both the time-series forecasting and the nodal forecasting methodologies were measured and compared. The error analysis algorithms such as mean squared error (MSE), mean absolute deviation (MAD) and mean absolute percentage error (MAPE) were used for evaluation. Table II shows the evaluation of ward pump utilization data of Ward 3A and 3B. It shows that neural network model used in RFID-IMS holds a key edge over the rest.
Three potential benefits of using the RFID-IMS regarding workers’ productivity, patient safety and maintenance planning are shown as below.
1)The RFID-IMS should save time and allow healthcare workers to channel more of their time for more direct patient care by reducing the time spent on search for infusion pumps, periodic maintenance, stock-take or locating of the ‘lost’ pump and handling the paperwork when there is a pump loan/swap.
2)Automation of the current inventory management system should prevent and reduce the risks to improve the patient safety. The RFID-IMS ensures that there is an undisrupted infusion pump service available for the patients and also provide timely detection of defective infusion pumps during pump operations for enhanced patient safety. It collects the utilization and movement data of the infusion pumps and allows for the most efficient allocating of the infusion pumps giving patients with a greater need for the pump to be given a greater priority.
3)The RFID-IMS should provide better information to allow a maintenance schedule that is based on the pump utilization and availability, rather than on a batch basis. This allows a more proactive way to ensure that pump functionality and availability are still feasible during maintenance where there would be a shortage of infusion pumps.
The RFID-IMS for the inventory management of infusion pumps based on the RFID tracking system and the neural network was successfully deployed at TTSH. The selection of neural network and the tracking of the movement and the usage patterns of the infusion pump in the proposed medical inventory system are integrated into a process flow framework. This framework helps in the elimination of wastage in terms of manpower and administrative time and promotes lean and efficient inventory management in healthcare industry. The proposed integrated solution that combines both RFID tracking and neural network analysis provides TTSH a basic data flow framework that can be used as a blueprint for TTSH’s proposed Information Technology Unit (ITU) Management System with respect to their inventory management of their infusion pumps. However, all forecast based on the key assumption that for every five patients there is a demand for one pump as there is a limited knowledge of the actual number of pumps per patients. Also, the values for the smoothening parameter were based on trial and error and this compromises the accuracy of the forecasts. Hospital is one of the human-intensive working environments in healthcare industry. Most of the tasks are carried by healthcare workers manually. In future, studies regarding process resign and reengineering can be conducted to improve the productivity of the inventory management and reduce the operation cost. Also, medical assets managed in current study can be further expanded to a larger group of products with the use of RFID technology. A comprehensive inventory tracking and forecasting can be established to provide better medical services to patients