Innovative Applications of O.R.

Accepted Manuscript

The 2Bin System for Controlling Medical Supplies at Point-of-Use

Claudia R. Rosales, Michael Magazine, Uday Rao
PII:
DOI:
Reference:

S0377-2217(14)00863-7
10.1016/j.ejor.2014.10.041
EOR 12594

To appear in:

European Journal of Operational Research

Received date:
Revised date:
Accepted date:

2 July 2013
10 October 2014
17 October 2014

Please cite this article as: Claudia R. Rosales, Michael Magazine, Uday Rao, The 2Bin System for
Controlling Medical Supplies at Point-of-Use, European Journal of Operational Research (2014), doi:
10.1016/j.ejor.2014.10.041

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Highlights
We model the 2Bin system under periodic and continuous review inventory policies.
We test the benefits of parameter optimization, optimizing the periodic review interval, versus policy
improvement, changing to a continuous review policy.

CR
IP
T

Policy improvement outperforms parameter optimization in most instances tested.

We identify especial cases when the benefit of parameter optimization is higher.

AC

CE

PT

ED

AN
US

We provide insights to hospital management on the efficient use of 2Bin systems.

ACCEPTED MANUSCRIPT

The 2Bin System for Controlling Medical Supplies at Point-of-Use

CR
IP
T

Abstract
The increase in cost of supplies and services is outpacing the increase in revenues at many hospitals. To address
this cost increase hospitals are seeking more efficient ways to store and manage vast inventories of medical
supplies. A parsimonious and efficient inventory system which we call 2Bin is becoming increasingly popular
in North American hospitals. Under the 2Bin system inventory is stored in two equal-sized bins. 2Bin systems
are reviewed periodically and empty bins are replenished. In recent years the adoption of RFID technology

AN
US

for 2Bin systems is allowing continuous-time tracking of empty bins, increasing inventory visibility. In this
paper we model the 2Bin inventory system under periodic and continuous review. For periodic review we show
that the long-run average cost per unit time is quasi-convex, enabling a simple search for the optimal review
cycle. For continuous review, we present a semi-Markov decision model, characterize the optimal replenishment
policy, and provide a solution approach to obtain the long-run average cost per unit time. Using data obtained

from hospitals currently using RFID-enabled 2Bin systems, we estimate the economic benefits of using the
best periodic review length (i.e., parameter optimization), and of using a continuous review inventory policy

ED

(i.e., policy improvement). We characterize system conditions such as the number of medical supplies used,
replenishment costs, stock-out costs, etc. that favor each option, and provide insights to hospital management

PT

on system design considerations that favor the use of periodic or continuous review.
Key words: OR in Health Services, Healthcare Operations Management, Inventory Management, Markov
Processes, Stochastic Methods

CE

Dispensing and Controlling Medical Supplies at the

AC

Point-of-Use: the RFID Enabled Two-Bin System

Claudia R. Rosales

Department of Supply Chain Management, The Eli Broad College of Business

Michigan State University
email: Rosales@bus.msu.edu
Michael Magazine
Department of Operations, Business Analytics, and Information Systems
Carl H. Lindner College of Business, University of Cincinnati

Preprint submitted to Elsevier

February 7, 2014

ACCEPTED MANUSCRIPT

email: magazim@ucmail.uc.edu
Uday Rao
Department of Operations, Business Analytics, and Information Systems
Carl H. Lindner College of Business, University of Cincinnati
email: raous@ucmail.uc.edu

CR
IP
T

1. Introduction

Hospitals are part of complex supply chains including purchase, storage, distribution, and inventory control
of drugs and medical supplies. Several factors increase the complexity of hospital supply chains, including
the large variety of items used by clinicians. Hospitals often manage hundreds of different items. Inventory
availability is critical for patient care. However, high inventory levels increase costs and create a significant

AN
US

economic impact for the hospital. In North America, more than 40% of hospital expenses are related to supply
chain activities (AHRMM [1]). While excess inventory is undesirable, enough inventory must be available to
avoid disruption to nursing activities that may hinder patient care. Per the report by Meyer and Meyer [24],
one of the future research areas in health care delivery is the study of inventory management practices within
hospitals.

In response to the need for storing and managing a large number of low-cost but critically important
medical supplies, hospitals have relied on inventory systems such as exchange carts and automated storage

ED

cabinets. The use of automated storage cabinets has increased in recent years (Pedersen et al. [30]), but given
the capital investment these systems require and the limited storage space available, many hospitals are using

PT

this technology mostly to store and dispense drugs or expensive supplies. In contrast, the inventory system
studied in this paper provides an efficient yet simple way to control and dispense hundreds of high-volume,
low-cost items without investing in expensive storage units (Denton [11]).

CE

We use the term 2Bin system to denote an inventory-control method similar to a kanban system in which
an items inventory is divided and stored in two bins (Hall [15]). Under a 2Bin inventory system only empty

AC

bins are tracked in lieu of unit-level item consumption. This reduces the data collection costs and the inventory
control parameters are on bin level as well, such as the bin replenishment frequency or the number of empty
bins that trigger a replenishment. More traditional inventory replenishment policies found in the literature
are on unit level, which require unit-level demand tracking and inventory control parameters per item, such
as reorder points and order-up-to levels (Rosales et al. [33] and Khouja and Goyal [18]).
In our experience with 2Bin systems, bins are usually denoted as primary and secondary. Although the
number of units contained in each bin will differ by item, the primary and secondary bin of the same item have
equal capacity. For each item the amount of inventory stored in a bin is typically determined based on the
3

ACCEPTED MANUSCRIPT

number of desired inventory turns per year. Bins are usually located back-to-back in racks with the secondary
bin placed behind the primary bin. When the primary bin is exhausted, it is removed from the racks, and
the inventory in the secondary bin becomes available. When the empty primary bin is replenished the bins
are exchanged, i.e., the secondary bin becomes the new primary bin, supporting product rotation, and the
recently replenished bin becomes the secondary bin. Traditionally 2Bin inventory systems used in hospitals

at the beginning of every review interval, usually once daily.

CR
IP
T

are replenished using periodic reviews (Landry and Beaulieu [19]), with empty bin replenishments carried out

Although the simplicity of 2Bin systems allows hospitals to efficiently replenish a large number of items,
policy parameters such as replenishment frequency are often based on values historically used and are typically
suboptimal. Performing the replenishment of empty bins requires time from materials management personnel,
incurring a fixed cost for each replenishment carried out. In addition, materials management personnel has a

AN
US

limited availability to perform the replenishments. On the other hand, replenishments avoid stock-outs, which
can be costly as nurses typically waste valuable time obtaining any items that are short, potentially disrupting
patient care. To balance fixed order costs and stock-out costs can be challenging. One of the goals in this
paper is to find optimal values for the periodic review interval to improve the inventory systems of hospitals
that use a periodic review 2Bin system..

Hospitals are investing in technology such as RFID or barcodes in order to avoid stock-outs, to reduce
inventory cost of supplies and to increase inventory tracking. Item-level RFID tags can be prohibitively

ED

expensive for low-cost items but not for bins (Johnson and Lee [16]). The second goal in this paper is to
study the use of RFID tags on hospital bins. This research is motivated by our interactions with consulting
company Logi-D (www.logi-d.net/en/) which developed such a system, called the 2BIN-iD system. In the

PT

2BIN-iD system a passive radio frequency identification (RFID) transponder is attached to each bin. When
a bin becomes empty, the passive RFID tag is removed and placed on an RFID-enabled replenishment board

CE

that has an RFID reader capable of reading all tags on the board. The replenishment board is connected
to the IT network alerting the system that a product is due to be replenished, which increases inventory

AC

visibility. The use of this technology provides hospital management the opportunity to achieve additional
economic benefits by using a continuous-review 2Bin system.
Hospitals currently using traditional 2Bin systems have two alternatives to improve their inventory man-

agement practices, (1) parameter optimization - starting with a typical hospital periodic review system and
optimizing parameters such as the periodic review interval, or (2) policy improvement - changing to a continuous review policy with the use of technologies such as bin-level RFID or barcodes. Therefore the two goals
in this paper are referred to as parameter optimization and policy improvement, respectively. We model the
2Bin system under both periodic and continuous review. The periodic review model provides a performance

ACCEPTED MANUSCRIPT

benchmark representing current hospital practice and it allows us to estimate the cost benefit of optimizing the
periodic review interval. To achieve this, we show that the long-run average cost per unit time is quasi-convex
in the review interval, enabling a simple search to obtain the optimal review cycle length. Conversely, the
continuous review model allows us to estimate the benefits of using a new inventory policy enabled by the
presence of new technology, we present a semi-Markov model to characterize the optimal policy (i.e., find the
optimal number of empty bins over all items that should trigger a replenishment), and to provide an approach

CR
IP
T

to estimate the optimal long-run average cost per unit time. We use data from hospitals currently using
2BIN-iD systems, to conduct a performance comparison between both types of review systems as well as a
sensitivity analysis.

In the sensitivity analysis, the following questions are answered: When is the benefit of using a continuousreview policy sufficiently greater than the benefit of using an optimal periodic review policy to warrant

AN
US

investments in new continuous-review technology? How are the relative benefits of using new technology
affected by system conditions such as the number of stocked medical supplies, the cost of performing a
replenishment, or the cost of material stock-outs? In addition our study allows hospitals to evaluate the cost
impact of having several storage locations with few items or fewer storage locations with more items, and how
these costs can be affected by using optimal periodic or continuous review policies.

This paper is organized as follows. In 2 we review the literature, in 3 we describe the 2Bin system and
introduce basic notation and assumptions common to both periodic and continuous-review models. In 4 we

ED

present closed form expressions for the average cost per unit time under periodic review as well as a Golden
Section search approach to estimate the optimal periodic review interval. In 5 we present a semi-Markov
decision model for the continuous-review 2Bin system, as well as a proof of the optimal replenishment policy.

PT

We also present a linear programming approach to obtain the optimal threshold value for the continuousreview 2Bin system. In 6 we provide a numerical comparison of both models, and finally in 7 we summarize

CE

our findings and indicate future research directions.

AC

2. Literature Review

2Bin systems have been in use at hospitals since the 1980s, and Kanban-type inventory systems are widely

used particularly in lean manufacturing environments. Inventory models in the literature that are closest to
the periodic review 2Bin systems are (nQ, r, T ) replenishment policies. In such policies, every T time periods
an order is placed whenever the inventory position falls to or below reorder level r. The order size is always
and integer multiple of Q such that the inventory position exceeds r after ordering. This policy can be mapped
to the 2Bin formulation by setting r = Q equal to the bin size capacity. Inventory systems with the (nQ, r, T )
policy are first studied by [14]. However, the authors require a fixed order cost per item and it is based on

ACCEPTED MANUSCRIPT

unit-level demand (assumed to be Poisson distributed). Larsen and Kiesm

uller [21] study the same single-item
inventory policy using a generalized Erlang distribution.
In the hospital setting, a fixed order cost is incurred per replenishment order regardless of the number
of items or number of units ordered. Therefore, our work is related to the joint inventory replenishment
literature. Research on multi-item replenishment policies with stochastic demand include the work of Balintfy
[4], Renberg and Planche [32], Atkins and Iyogun [3], Pantumsinchai [29], Viswanathan [39], Rao [31], and

CR
IP
T

Ozkaya
et al. [28] among others. Khouja and Goyal [18] provide a detailed discussion of these papers. More
recently Mustafa Tanrikulu et al. [25] consider a multi-item inventory system in which the setup costs for orders
have a step-wise structure. The authors propose a new (s, Q) policy, where s = {s1 , s2 , ..., sN } representing
a vector of reorder levels for each item in the system. Under this policy a combined replenishment order of
size Q is triggered whenever the inventory of one item falls to its corresponding si value. Under the (Q, S)

AN
US

policy proposed by Renberg and Planche [32], whenever the combined demand of all items reaches a value Q,
an order of size Q is placed to raise the inventory position of all items up to the vector S = {S1 , S2 , ..., SN }.
The (s, Q) policy outperforms the (Q, S) when back-order costs are high and lead times are short. None of
this multi-item research deals with kanban-type 2Bin inventory systems that we study in this paper.
Research that focuses on inventory replenishment policies in hospital settings includes the work of Lapierre

and Ruiz [20] who present a periodic review multi-item model to coordinate supply operations under deterministic demand. Little and Coughlan [23] present a periodic-review constraint-based model considering

ED

capacity and service level constraints. Bijvank and Vis [7] present two periodic-review lost sales models, either maximizing service level subject to capacity constraint, or minimizing capacity subject to service level
constraint. Opolon [27] studies the impact of inventory inaccuracies with the use of automated dispensing

PT

machines (ADMs) at hospitals and demonstrates that frequent periodic inventory counts can mitigate the
impact of imperfect demand recording. Rosales et al. [34] address the impact of ADMs on inventory replen-

CE

ishment practices by studying a new hybrid inventory policy and propose closed-form heuristic expressions to
set parameter values for hospital settings. The impact of multi-item replenishment policies with the use of

AC

ADMs under unit-level demand is addressed by Rosales et al. [33]. They test four new replenishment policies
that are found to offer significant benefits such as reductions in the number of replenishments performed and
inventory carried. With the exception of Rosales et al. [34] and Rosales et al. [33] all of the work mentioned
follows a periodic review policy while we consider both periodic and continuous review policies.
The impact of the adoption of new technologies such as RFID on supply chain performance has received
some attention. Delen et al. [10] propose and evaluate performance metrics to assess the value of RFID, and

address RFID data quality challenges. Lee and Ozer

[22] perform an extensive review of the RFID literature in operations management and emphasize the importance of analytical modeling tools linking underlying

ACCEPTED MANUSCRIPT

operating characteristics to control decisions and performance measures. Gaukler et al. [13] analyze the impact of increased supply chain visibility from the use of RFID allowing placement of emergency orders as
order progress information becomes available. Zhou [40] quantifies the benefits of RFID adoption through
uncertainty reduction from a manufacturers perspective. A simulation study by Sari [37] explores different
operational and system environmental conditions under which RFID is more beneficial. The author considers
a four-echelon supply chain under different levels of supply chain collaboration, finding that RFID is most

CR
IP
T

beneficial under conditions of high supply chain collaboration, low demand uncertainty, and long lead times.
The work of Sarac et al. [36] provides a detailed discussion of supply chain research concerning RFID technology. To our knowledge, there are only two papers that study RFID technology for the management of supplies
in a hospital setting. Cakici et al. [8] study the operational and economical differences between a single-item
inventory system using barcode technology and one using RFID technology within a hospital setting. Landry

AN
US

and Beaulieu [19] describe the use of 2Bin systems at hospitals in Europe, Asia, Canada and the United States,
with several hospitals implementing RFID with 2Bin systems in Canada. Our work differs from these as we
focus attention on a multi-item hospital inventory system, and provide insights into the economic benefits
of the use of RFID technology to track bin consumption and replenish in continuous time. In summary, we
consider the unique elements of inventory systems observed at hospitals: bin-level demand data rather than

unit-level demand, batch size replenishments equal to the bin capacity and multi-item settings.

ED

3. The 2Bin System

In this section we first describe the hospital environment and data collected from a facility currently using
RFID-enabled 2Bin systems. Then we introduce general model assumptions that will be common to both

PT

periodic and continuous-review models that will be introduced in sections 4 and 5 respectively.
We obtained data from a hospital currently using RFID technology in conjunction with 2Bin systems.

CE

From the data obtained the complexity of hospital supplies management is readily apparent. The hospital
has many different departments, each having several storerooms. Whenever items are replenished, one order

AC

is placed for all items in one storeroom. The number of items per storeroom can vary greatly but typically
ranges between 100 to 500 different items. Given the number of items and storerooms in each facility, it is
critical to propose replenishment policies that are both simple and efficient.
The data obtained includes bin consumption and replenishment data for all medical supplies used through-

out the hospital in different nursing units. An example of the data is provided in Table 1 which shows that
in the emergency department, Item 89017 with 36 units/bin had a primary bin appear on the replenishment
board on June 27 at 8:02pm. The replenishment of this empty bin was triggered June 28 at 7:20am and was
completed June 28 at 11:24am. Note that the hospital data obtained does not include unit-level demand; the

ACCEPTED MANUSCRIPT

demand captured by the hospital is only at the bin level. Therefore, our unit of analysis is on bin level as well.
For scenarios where unit-level demand is available, the online supplement provides extensions to our models
that consider unit-level demand.
Another important characteristic of the system under study is the fact that hospitals using RFID-enabled
2Bin systems will typically trigger a replenishment order whenever a certain number of primary bins is empty

possible.

CR
IP
T

or whenever both the primary and secondary bin of an item are empty to disrupt patient care as shortly as

Table 1: Format of Hospital Data Provided

Item
Code
89017

Units/bin
36

Bin
Type
P

Bin on
Board
6/27 8:02PM

Replenishment
Ordered
6/28 7:20AM

Replenishment
Completed
6/28 11:24AM

AN
US

Hospital
Department
Emergency Dpt.

Next, we describe the general characteristics with the basic assumptions of the 2Bin inventory system
and how the data of Table 1 can be used to set the input parameter values for the inventory system. These
characteristics will be common to both periodic and continuous review models, which will be introduced in

the following sections. We use N to denote the number of medical supplies or items stored in a particular
storeroom using the 2Bin system. Inventory for each medical supply is divided into two equal-sized bins

ED

referred to as primary and secondary. For the remainder of the paper we use the following assumptions:
Assumption 1: Bins become empty following a Poisson process, having probability mass function p(), cumulative distribution function P (), and upper tail probability P () with equal bin consumption rates, , for

PT

each item.

Although demand for medical supplies at the unit level will differ by item, we assume that the consumption

CE

rate will be the same at the bin level. This is based on the fact that in our experience hospitals often determine
the number of units stored in each bin to achieve a target number of inventory turns. Therefore bins of different
items may contain different numbers of units, but bins will have on average similar consumption rates. Based

AC

on the bin replenishment data records, we computed the average consumption rates of bins. Table 2 shows
the average consumption rates for a subset of items tested in the hospital data set. It becomes apparent that
bins usually have enough capacity to satisfy the demand for two weeks on average (0.003 24 14 1 bin

per 2 weeks)

Table 2: Interarrival rates for a subset of items

Item number
bins/hr

89157
0.003

89444
0.0029

89445
0.0028

89659
0.003

89745
0.003

ACCEPTED MANUSCRIPT

In addition Assumption 1 also states that bin consumption follows a Poisson process. Our analysis of
hospital data identified several items for which interarrival times for bin demand is exponential, and therefore
a Poisson process can be a good fit for these items. In Table 3 we show the results of the Kolmogorov-Smirnoff
test for a subset of items tested, where large p-values indicate that the exponential distribution is a good fit
for the data (Kelton and Law [17]). When the bin consumption process is not Poisson we believe that the
assumption of exponential interarrival times could be replaced by a phase-type distribution (see Asmussen

CR
IP
T

et al. [2], Neuts [26]) that permits an extension of the analysis in this paper.

Table 3: Kolmogorov-Smirnoff test results for the interarrival times of bin orders to follow an exponential distribution.

Item number
Test Statistic
p-value

89207
0.082
> 0.15

89542
0.081
> 0.15

89609
0.054
> 0.15

89649
0.085
> 0.15

89695
0.092
> 0.15

AN
US

Assumption 2: The probability that demand for a particular item during the replenishment lead time L
exceeds the quantity stored in both primary and secondary bins is negligible.

Replenishment lead times are on average a few hours. Since the bin demand is on average 1 bin per 2 weeks,
the probability that both primary and secondary bins of the same item are consumed during the replenishment
lead time is negligible. This assumption implies that at most one bin of each item may become empty during

the replenishment lead time. This means that after an order is delivered at most one bin of each item can be
empty (the bin used during the lead time).

ED

We obtain estimates for the fixed replenishment cost and the stock-out cost based on a cost analysis recently
performed at the hospital under study. In order to obtain the replenishment cost, the hospital estimated the

PT

average amount of time materials management personnel spent on processing, refilling, and delivering full bins
to the storage locations (i.e., storerooms). The hospital considered that most of the replenishment cost is
associated with the time and effort of these activities during the replenishment process. The cost associated

CE

with retrieving specific items from the warehouse is considerably smaller and relatively constant. Therefore,
it is assumed that the fixed cost of a replenishment is independent of the number of items or the number

AC

of units ordered. This means that every time a replenishment order is triggered at the storeroom, a fixed
replenishment cost K is incurred.
The stock-out cost was obtained by estimating the amount of time nurses spent requesting and obtaining

stock-out items. The unit-level stock-out cost must be converted in our model to a bin level stock-out cost. To
obtain the stock-out cost per bin, we multiply the unit-level stock-out cost by a conversion factor. Consistent
with the way we selected other parameter values, we use the median number of items stored per bin as a
conversion factor. Whenever there is a bin demand while the secondary bin is empty, there is a stock-out cost
. Furthermore, we consider a time-weighted shortage cost for each unit time that both bins (primary and
9

ACCEPTED MANUSCRIPT

secondary bin) are empty (see Ozkaya

et al. [28]). This cost allows us to penalize long-standing shortages that
tend to undermine nurse satisfaction and confidence in the stock replenishment process. Cost parameter values
are presented in 6. As previously mentioned, we capture demand and costs at the bin level. An extension
of shortage cost calculations to the case where unit level demand data is available is presented in the online
supplement.
The hospital did not monitor inventory holding costs at point-of-use because inventory costs were already

CR
IP
T

incurred once items were purchased by the hospital and the incremental cost of holding stock at point-of-use
as opposed to the central storage location was considered negligible. Nevertheless our model can be extended
to consider inventory holding costs when these costs are considered relevant. Further, hospital management
controlled inventory levels at point-of-use by restricting the number of units per bin; this bin-size parameter
was computed exogenously based on target inventory turns and was considered a pre-specified constant in our

AN
US

model; optimizing bin sizes can be an area of future research.

4. Periodic Review

In this section we develop a model to estimate the long-run average cost for the periodic review 2Bin
system. The periodic review case provides the performance benchmark corresponding to current hospital

practice. Unless otherwise noted, proofs for all expressions included in this section are available in a separate
supplemental file.

ED

We formulate the expected stock-out costs for one representative item and then scale this cost using the
number of items in the system (based on Assumption 1, all bins have equal consumption rates). In order to
obtain the stock-out costs for one representative item, we first need to understand how the inventory profile

PT

for this representative item changes with time. Recall that inventory status is reviewed periodically every T
time units, and empty bins of all items are replenished at the review epoch. For the hospital supplying the

CE

data for our research, T = 24 hrs. Figure 1 illustrates an inventory profile for a representative item with one

bins

AC

full and one empty bin at time 0. After a delivery lead time L, the order of one bin is received, and on-hand

time

On hand inventory

T+L

Inventory position

Figure 1: Illustrative Inventory Profile for Periodic Review 2Bin system

10

ACCEPTED MANUSCRIPT

inventory and inventory position are equalized to two bins. Prior to the next periodic review epoch, one more
bin is demanded, consequently when the system reaches time T another order for one bin is triggered and a
renewal cycle occurs with inventory position equal to two full bins.
Based on renewal theory (Ross [35]) we estimate the long-run average cost per unit time for the periodic
2Bin system by the average cost per unit time for one renewal cycle. Total stock-out costs include the stockout cost per occurrence at rate , and the time-weighted stock-out cost at rate . For a representative item,

CR
IP
T

stock-out costs during interval t depend on the number of non-empty bins at the beginning of the time interval,
denoted by m {0, 1, 2} as there are only two bins per item. When m = 0, products may be in back-order,
with incurring shortage costs at rate ( + t). When m = 1 or 2, the total stock-out cost over interval t is
denoted as C(m, t).

We use expression C(m, t) to capture expected stock-out costs for a single representative item during time

AN
US

interval t = L and t = T L. To obtain an expression for C(m, t) we use x(t) to denote bin demand for a single
representative item during time interval t. Note that stock-out costs will be incurred only if the secondary
bin becomes empty. Therefore, any demand x(t) m will imply an empty secondary bin and potentially a
stock-out. The expression for C(m, t) equals:

or
C(m, t) =

C(m, t) = E[x(t) (m 1)]+ +

ED

=m1

( (m 1))p(, t) +

y=0

E[x(y) (m 1)]+ dy

( (m 1))p(, y)dy

(1)

y=0 =m1

Using standard properties of Poisson demand, we obtain (see supplemental file for details):
=

CE

PT

C(m, t)

[tp(m 1, t) + (t m + 1)P (m, t)]

 2
t
m(m 1)
+

P (m 1, t)
P (m + 1, t)
2
2
i
m
+(1 m)[tP (m, t) P (m + 1, t)]

(2)

To scale the stock-out costs for one representative item to all N items we define the following: L denotes

AC

the expected number of primary bins empty after replenishment lead time L. 1 denotes the expected number
of items with at least the primary bin consumed during the periodic review interval. 2 denotes the total
number of secondary bins consumed plus any additional unmet bin demand across all items, representing total

11

ACCEPTED MANUSCRIPT

stock-outs during the periodic review interval. Expressions for L , 1 , and 2 are provided below.
L

N L

(3)

N (1 p(0, T ))

(4)

N (T p(1, T ))

(5)

CR
IP
T

Note that by Assumption 2, all the primary empty bins immediately after delivery lead time L must have
been demanded during this lead time. Hence N L represents the expected number of bins depleted during
the delivery lead time.

Based on expressions (2) - (5) we provide an expression for the long run average cost per unit time for the
whole system (that is, for the storeroom). This is denoted by J P (T ) and includes fixed order and stock-out
costs for all N items. The fixed order cost is incurred when at least one item in the storeroom has an empty bin

AN
US

during the review interval T . Hence, the expected order costs per replenishment cycle are K(1 p(0, N T )).
To obtain the expected shortage costs we divide T in two time intervals, one of length T L (the length of
the review period after order delivery) and one of length L (the delivery lead time). During interval L, only
items with at most one empty bin at the beginning of the review interval (m = 0, 1) can incur stock-outs, since
x(L) is 0 or 1 by Assumption 2. When m = 1, expected stock-out costs during L for a single item are given

by C(1, L). Multiplying this by 1 yields the expected stock-out cost during lead time L for all items with
m = 1. For items with m = 0, there are no non-empty bins, so stock-out costs incurred over interval L are

ED

given simply by ( + L). Multiplying this by 2 yields the expected stock-out cost for all items with m = 0.
Shortage costs over time interval T L are obtained analogously by setting t = T L. Note that m = 1, or 2

PT

because the T L interval starts immediately after the delivery time L (see also Assumption 2). Furthermore,
note that the expected number of items where m = 1 (m = 2) after lead time L equals L (N L ). As a


result, the total shortage cost over T L is 1 C(1, L) + 2 ( + L) + L C(1, T L) + N L C(2, T L).

AC

CE

Putting all these expressions together yields:

J p (T )


1
K(1 p(0, N T )) + 1 C(1, L)
T

+2 ( + L) + L C(1, T L)



+ N L C(2, T L)

(6)

)), the long-run

Proposition 1: When T 2 , with = ((KN e(N 1)T )/( + L
) + L( + L2 )/( + L
average cost per unit time J p (T ) for the periodic review 2Bin policy is quasi-convex in the cycle length T , and
therefore unimodal.

12

ACCEPTED MANUSCRIPT

Proof:
We only outline the proof here; detailed calculations can be found in the online supplement. We start by
expressing (6) as
J p (T ) =

F (T )
T

where
=




L C(1, T L) + N L C(2, T L)

CR
IP
T

F (T )

+K(1 p(0, N T )) + 1 C(1, L)


+2 ( + L)

AN
US

The proof is based on two key concepts: first, convexity of expression F (T ), which we prove holds whenever
T 2 , with = ((KN e(N 1)T )/( + L
) + L( +
expression J p (T ) =

F (T )
T .

L
2 )/(

+ L
)); and second, the quasi-convexity of

It is worth noting that the hospital data provided resulted in all scenarios tested

satisfying the condition in Proposition 1, with 0.07, and T 0.82.

Based on Proposition 1, the optimal review cycle T can be found readily by solving:

T = arg min J p (T ).
T

(7)

We code expressions (2) - (6) in C++ and by virtue of unimodality, we find T as well as the optimal cost for

PT

5. Continuous Review

ED

the periodic 2Bin system, J p (T ), using a Golden section search (Bazaraa et al. [5]).

Under periodic review each empty bin is replenished at the beginning of a replenishment period. By

CE

contrast, under continuous review empty bins are replenished when a certain number of items have empty
primary bins or when an item has an empty primary and secondary bin, since hospital management can
observe the total number of empty bins at all times. In this section, we develop a semi-Markov decision

AC

process (SMDP) model to analyse such a continuous-review 2Bin system. We first define the possible SMDP
states in the system; we then derive basic mathematical expressions for transition probabilities, expected
transition times, and expected costs for each possible state in the system. We then use these expressions to
obtain the Bellman optimality equations and propose a solution approach. All proofs for expressions found in
this section are included in the online supplement unless noted otherwise.
We define the state of the system by the pair (n1 , n2 ) where n1 and n2 denote the number of empty primary
and secondary bins respectively; n1 can take any value between 0 and N , while n2 = 0, 1 as a secondary empty

13

ACCEPTED MANUSCRIPT

bin will always trigger a replenishment under continuous review. Note that triggering a replenishment order
when a secondary bin becomes empty may not be the most economical strategy. Without this restriction,
when stock-out cost is low, the optimal cost solution may allow more than one secondary bin to become empty
before a replenishment order is triggered. Forcing an order to be placed with one empty secondary bin may
increase costs, particularly at high values of fixed order cost K. Nevertheless we incorporate this restriction

practice at hospitals using RFID-enabled 2Bin or 2BIN-iD systems.

CR
IP
T

in our model given the importance of having stock available for patient care and also because this is common

This means that a new replenishment order can only be triggered after a bin demand. Each time a
replenishment order is placed, all empty bins at the order placement epoch will be replenished after delivery
lead time L. If no replenishment order is placed, the next system transition time is random and marked
by the consumption of one more bin. When a replenishment order is placed, the next system transition

AN
US

time is fixed and marked by the replenishment lead time L. Note that during the replenishment lead time
additional secondary bins may become empty; nevertheless, following Assumption 2, the number of secondary
bins empty after delivery lead time will always be zero and therefore there is no need to track the number of
empty secondary bins during the replenishment lead time.

To model the decisions at the decision epochs (i.e., after bin demand), an action set is specified. Two

ED

actions a {do not order, order}={0,1} are possible. The set of possible actions in state (n1 , n2 ) are:

{0, 1},
if n2 = 0
A(n1 ,n2 ) =

{1},
if n2 = 1

Each time an order is placed, a = 1, a fixed order cost K is incurred. In addition to fixed order costs,

PT

we compute stock-out costs per occurrence at rate , and the time-weighted stock-out cost at rate as
in 4. When a secondary bin is empty at the order placement epoch, n2 = 1, we incur stock-out costs

CE

( + L) during the replenishment lead time. In addition other secondary bins may get consumed during the
replenishment lead time causing additional stock-outs. Based on Assumption 2 only secondary bins of items
with an empty primary bin at the order placement epoch have the potential for stock-out by depleting during

AC

the replenishment lead time. Following the same notation as in 4, we use x(L) to denote bin demand for
one item during replenishment lead time, and express the expected additional number of secondary bins that
become empty during L as n1 E[x(L)]. We can then express the total stock-out costs during L as:
(n1 , n2 ) = n1 E[x(L)] + n1

E[x(t)]+ dt + n2 ( + L).

t=0

The above expression can be simplified to:

(n1 , n2 ) = n1 (L + L2 /2) + n2 ( + L).
14

(8)

ACCEPTED MANUSCRIPT

Using expression (8), we define the expected cost for each possible state-action triplet (n1 , n2 , a) as:

K + (n1 , n2 ),
if a = 1
G(n1 ,n2 ) (a) =

0,
if a = 0
Note that no costs are incurred if no order is placed.

Transition probabilities from state (n1 , n2 ) to state (i, j) for given action a are denoted by P(n1 ,n2 )(i,j) (a).

and transition probabilities are defined as:

N n1

N
n1
P(n1 ,n2 )(i,j) (0) =
N ,

0,

CR
IP
T

When no order is placed (a = 0), the next transition depends solely on the number of empty primary bins

for i = n1 + 1, j = n2 = 0

for i = n1 1, j = 1, n2 = 0
otherwise

AN
US

In the event that a = 1 and an order is placed, all empty bins at the order placement epoch will be replenished

after L time periods. After stock is replenished the system may not necessarily start in state (n1 , n2 ) = (0, 0)
because of the possibility of additional bins becoming empty during the replenishment lead time. By Assumption 2, if the secondary bin of an item becomes empty during replenishment lead time, the corresponding
primary bin will be filled when stock is replenished and only new primary bins may be empty after replenish-

transition probabilities are given by:

ED

ment lead time. After stock is replenished, n1 may take any value in 0, ..., N , while n2 = 0. For a = 1, the

PT

P(n1 ,n2 )(i,j) (1) =

((N L)i eN L )/i!

N
X

m N L

(N L) e

m=0

0,

/m!

for j = 0.

for j = 1.

So far in this section we have described the possible states and actions for the 2Bin system under continuous

CE

review, the expected cost for each possible state-action triplet, and the transition probabilities. In order to
define the Bellman optimality equations, the only element that needs to be defined is the expected transition

AC

times. For the state-action triplet (n1 , n2 , a), the expected transition times are defined by:

2 ,
if a = 0
N
(n1 ,n2 ) (a) =

L,
if a = 1

We now seek to find the average cost per unit time for the continuous 2Bin system, J c . We use h(i,j) to denote
the relative or differential cost for each state (i, j). This h(i,j) provides the minimum over all policies of the
difference between the expected cost to reach state (0, 0) starting from state (i, j) for the first time and the
cost that would be incurred if the cost per stage at all states were equal to the average J c , (see Bertsekas [6]).

15

ACCEPTED MANUSCRIPT

The average cost per unit time, J c and the replenishment policy can be obtained by solving the following
optimality equations. Using standard dynamic programming notation we have:
"
#
N X
1
X
c
h(n1 ,0) =
min G(n1 ,0) (a) (n1 ,0) (a)J +
P(n1 ,0)(i,j) (a)h(i,j) , n1 {1, ..., N }
a{1,0}

h(n1 ,1)

= G(0,0) (0) (0,0) (0)J c +

=

"

N X
1
X
i=0 j=0

G(n1 ,1) (1) (n1 ,1) (1)J +

P(0,0)(i,j) (0)h(i,j)

N X
1
X
i=0 j=0

(9)

(10)
#

CR
IP
T

h(0,0)

i=0 j=0

P(n1 ,1)(i,j) (1)h(i,j) , n1 {1, ..., N }

(11)

For each state (n1 , n2 ) we use h(n1 ,0) to denote what is commonly referred to in SMDPs as the relative or
differential cost (Bertsekas [6]). As is customary in dynamic programming formulations, the first and second
terms in expressions (9), (10), and (11) represent the differential cost and the cost to go for each possible

AN
US

decision a respectively. Equation (9) provides the best decision, whether to order or not, when there are no
secondary empty bins; i.e., the state is characterized by (n1 , 0). Equation (10) states that no replenishment
order is placed in the absence of empty bins; i.e., the state then is (0, 0). Finally equation (11) states that
whenever a secondary bin becomes empty an order is triggered; i.e., the state is characterized by (n1 , 1). The
system of equations (9)-(11), has a unique solution that is independent of the initial state if all stationary

policies are unichain, i.e., the underlying Markov chain contains only one recurrent class. Furthermore, this
unique solution provides the optimal average cost per unit time for the policy structure used in this paper and

ED

it characterizes the optimal actions for each state (n1 , n2 ).(see Bertsekas [6], Feinberg [12]).
For the remainder of this subsection we develop various properties of the continuous 2Bin model.
Lemma 1: All stationary 2Bin replenishment policies are unichain.

PT

By Lemma 1, and results presented by Feinberg [12] and Bertsekas [6] we have the following result:
Proposition 2: The system of equations (9)-(11) has a unique solution.

CE

Having established the existence of a unique optimal average cost per unit time for the 2Bin policy, now we
focus on characterizing the optimal replenishment policy for the continuous-review 2Bin system, based on

AC

equations (9)-(11).

Theorem 1: Under system conditions specified below in expression (13), the optimal 2Bin replenishment

policy is a threshold policy.

A threshold policy is defined as a policy that triggers a replenishment whenever the number of empty bins

reaches or exceeds a threshold value n1 . In order to prove the optimality of a threshold policy for the 2Bin
inventory replenishment system, we prove that the following relationship holds:
h(n1 ,0) (0) h(n1 1,0) (0) h(n1 ,0) (1) h(n1 1,0) (1), for all n1 > 0

16

(12)

ACCEPTED MANUSCRIPT

This expression states that the rate of increase of the relative value function is higher when no replenishment
order is placed (a = 0) than when an order is placed (a = 1); therefore if it becomes economical to place an
order at a certain value n1 , then it will be economical to replenish at any value greater than n1 . We show the
validity of expression (12) using induction. We show that the optimal replenishment policy for the continuous
review 2Bin model is a threshold policy whenever (see online supplement for details):
(13)

CR
IP
T

(n1 ,0) (0)J c (1, 0)

The right hand side of expression (13) denotes the expected penalty cost for a secondary bin to become
empty during replenishment lead time when we place an order with only one empty primary bin. This
expected penalty cost will tend to be very low for system parameters commonly found at hospitals since
the replenishment lead time is relatively low compared to the bin consumption rates. We tested expression

AN
US

(n1 ,0) (0)J c (1, 0) using hospital data and found that the condition holds in every tested instance. When
the condition doesnt hold we expect bins to be replenished only when a secondary bin becomes empty.
Solution methodologies such as value iteration, policy iteration, or linear programming can be used to find
the optimal average cost per unit time J c as well as the optimal threshold value, n , for the 2Bin system.
The next subsection introduces a solution approach for the 2Bin system using linear programming.

5.1. Solution Methodology

A linear program can be formulated to optimize J c subject to equations (9)-(11). The dual of this linear

ED

program provides a useful interpretation and is usually the model of choice in the literature for solving SMDP
models (Feinberg [12], Tijms [38]). This dual program provides the value of J c , the optimal long-run average

PT

cost per unit time, as well as the optimal replenishment policy. Let x(n1 ,n2 ) (a) denote the long-run fraction of
decision epochs, or steady-state probabilities, at which the system is in state (n1 , n2 ), and decision a is made.

min

N
1 X
1
X
X

G(n1 ,n2 ) (a)u(n1 ,n2 ) (a)

n1 =0 n2 =0 a=0

AC

CE

Define u(n1 ,n2 ) (a) = x(n1 ,n2 ) (a)/

(n1 ,n2 ) (a). Then the average-cost optimal policy is given by the solution to

17

ACCEPTED MANUSCRIPT

subject to:

N X
1
X
u(0,0) (1) + u(0,0) (0)
P(i,j)(0,0) (1)u(i,j) (1)

(14)

i=0 j=0

u(n1 ,0) (1) + u(n1 ,0) (0) P(n1 1,0)(n1 ,0) (0)u(n1 1,0) (0)

N X
1
X

P(i,j)(n1 ,0) (1)u(i,j) (1) = 0, n1 1, .., N

(15)

i=0 j=0

0, n1 1, .., N

(16)

(n1 ,n2 ) (a)u(n1 ,n2 ) (a)

(17)

0 i, j, a

(18)

N
X

1
X

1
X

n1 =0 n2 =0 a=0

CR
IP
T

u(n1 ,1) (1) P(n1 ,0)(n1 ,1) (0)u(n1 ,0) (0)

u(n1 ,n2 ) (a)

The objective function minimizes the steady-state average cost per unit time. Constraints (14) - (16)

AN
US

represent the balance equations stating that the long-run average number of transitions to any state (n1 , n2 )
is equal to the long-run average number of transitions from state (n1 , n2 ). Constraint (17) requires that
the sum of the long-run fractions of decision epochs at which the system is in state (n1 , n2 ) be equal to
one. Non-negativity constraints are included in expression (18). The objective function of this linear program
provides the value of J c , while the optimal threshold value n is obtained from the solution to the steady state

probabilities, x(n1 ,n2 ) (a). With 4N variables and (11N + 2) constraints, excluding non-negativity constraints,
the size of this linear program increases linearly in N . Therefore, this linear program can be solved for values

ED

of N found in real settings. We solve this model using AMPL/CPLEX.

6. Numerical Results and Sensitivity Analysis

PT

In this section we test the benefits of optimizing the periodic review interval and compare them to the
additional benefits that can be obtained by implementing a continuous review policy. In addition we perform a

CE

sensitivity analysis using data provided by hospitals currently using RFID-enabled 2Bin systems in all nursing
wards throughout their facilities. For the periodic review 2Bin model we implement a C++ algorithm to find
the optimal review interval T and the corresponding long-run average cost per unit time J p (T ) as described

AC

in 4. We also obtain the long-run average cost per unit time J p (T ) for review interval T = 24 hrs as a
benchmark, since this is typically used at hospitals. For the continuous review 2Bin model we implement the
linear program described in 5.1 using AMPL/CPLEX to obtain the long-run average cost per unit time J c

and the optimal threshold value n .

We select a base scenario choosing parameter values corresponding to median values that represent system
conditions commonly found at the hospital from which data was obtained. Based on the data provided as
described in 3, we set N = 200, = 0.002778 bins/hr, L = 4 hrs, T = 24 hrs, K = $100 and = $55.

18

ACCEPTED MANUSCRIPT

The only cost parameter the hospital was unable to provide was the penalty cost per unit time . In the
absence of hospital data to obtain this parameter, we compare the value of to the median cost of capital
traditionally used in industry (Damodaran [9]). The cost of capital is usually related to the cost of holding
inventory. Stock-out cost is typically considered to be greater than inventory holding costs, therefore the
median industry cost of capital provides us with a conservative estimate for = 0.04 $/bin/hr.
In Table 4 we present the results for the above-mentioned median values of system parameters obtained

CR
IP
T

from hospital data. In this table we also include additional scenarios with one parameter varied at a time to
represent typically the 25th and 75th percentiles for each parameter. In the case of we tested additional
values 0.0 and 0.08. For the base scenario, we observe that the hospital is performing at conditions close
to optimal under periodic review triggering replenishment orders every T = 24 hrs, compared to T = 28.8
hrs; with ratio J(T )/J(T ) = 1.02. So using the optimal periodic review interval will not provide significant

from a periodic to a continuous review policy.

AN
US

benefits. By contrast ratio J(T )/J c = 2.83 indicates that significant benefits can be obtained by changing

In general we observe that the continuous review 2Bin system significantly dominates the periodic review
system for the conditions typically found at the hospital that supplied the data. Only when the per unit
penalty cost () was low the cost of the periodic review policy approaches the cost of the system under

continuous review. In most instances tested the cost of the optimal periodic review policy was at least two
and a half times the cost under continuous review. This is significant, especially when we consider that costs

ED

provided in Table 4 represent costs per hour for one storage location only. Hospitals usually have several
storage locations or supply rooms throughout their facilities. The costs obtained when scaled to a yearly basis
at potentially dozens of supply rooms throughout the hospital can become significant. Based on our results

PT

the hospital can realize significant savings using the information available and allowing system conditions to

CE

trigger replenishment orders in continuous time.

From Table 4 we observe that certain parameters such as , , and N can increase or decrease the benefits of

AC

continuous review. This is good news for hospital management, as these parameters, particularly and N ,
can be directly controlled by hospital management. It may be possible for hospital management to vary the
value of by changing the amount of units stored per bin, while N may be changed by effectively allocating
items to different supply rooms used throughout the hospital. In order to better understand the effect of these
parameters on the models considered, we run additional experiments using the base scenario and testing,
ceteris paribus, 30 different values of each parameter at a time, allowing us to observe the behavior of each
model under different values of system parameters.

19

ACCEPTED MANUSCRIPT

Table 4: Illustrative examples - optimal solution 2Bin system under continuous and periodic review. Under periodic review J p (T )
denotes the long-run average cost per unit time with T =24 hr.
Denotes order triggered only when secondary bin becomes empty. Base Scenario corresponds to median parameter values for
hospital data N = 200, = 0.002778 bins/hr, L = 4 hrs, T = 24 hrs, K =$100, and =$55. In addition for the Base Scenario
we set = 0.04. Numbers in the Value column correspond to the 25th and 75th percentile of data, except for which hospital
data was unavailable.

Ratios
J p (T )
J p (T )
1.02
1.05
1.00
1.09
1.60
1.02
1.01
1.00
1.05
1.92
1.02
1.02
1.02

J p (T )
J c
2.79
2.82
2.73
2.63
2.92
2.89
2.69
2.88
2.67
1.38
2.97
2.78
2.79

J p (T )
J c
2.83
2.97
2.73
2.85
4.67
2.94
2.73
2.89
2.79
2.66
3.02
2.83
2.83

CR
IP
T

Periodic
Review
J p (T ) T
J p (T )
($/hr) (hr) ($/hr)
7.33
28.8
7.45
6.30
33.5
6.63
9.08
23.4
9.09
14.17
15.3 15.41
2.95
70.4
4.71
7.18
29.0
7.30
7.47
28.8
7.58
6.40
25.0
6.41
8.30
33.1
8.70
2.36
87.8
4.53
10.83
19.7 11.01
7.32
29.0
7.44
7.35
28.8
7.46

AN
US

Parameter
Value
Modified
Base Scenario
N
150
300

0.00521
0.00114
L
2
6
K
75
130

6
115

0
0.08

Continuous
Review
J c
n Avg. Cycle
($/hr)
Length (hr)
2.63
33
119
2.23
28
134
3.33
41
98
5.40
35
67
1.01
32
280
2.48
32
115
2.77
34
122
2.22
28
101
3.11
39
140
1.71

212
3.64
21
76
2.63
33
119
2.63
33
119

Figure 2 illustrates the behavior of J p (T ), J p (T ) and J c for different values of and N . We see that at

low values of we obtain lower economic benefits from implementing a continuous review process, and using
an optimal periodic review may provide higher benefits. Low values of imply more units stored per bin and
lower inventory turns. If the hospital chooses to operate under periodic review and reduce costs, then more

ED

inventory should be held in each bin, reducing the number of replenishments. Nevertheless factors such as
supplies expiration dates and inventory rotation need to be considered. By contrast our research indicates

PT

that hospitals focusing on reducing the amount of stock held in nursing wards (corresponding to moderate
values) can obtain greater benefits by implementing continuous time tracking inventory control.

CE

The effect of N on costs as shown in Figure 2(b) is useful as well from a hospital standpoint; we observe
that costs increase more dramatically with N when the system operates under periodic review. We also see
that it is less costly to run one supply room, for example with N = 200 items, than it is to run two separate

AC

supply rooms, with N = 100 items; moreover, the difference in cost of operating two small supply rooms vs.

one larger one is higher under periodic review than under continuous review. This is significant for hospitals
as it is very common to have several small supply rooms throughout their facilities. Operating several small or
medium-sized supply rooms can become very expensive under periodic review, even when the review interval
T is optimized. Therefore hospitals seeking to improve access to medical supplies by increasing the number

of supply rooms should consider using a continuous review policy. Note that in our problem setting the value
of K is not significantly affected by N . Nevertheless if K changes with N , then our model can be run with

20

ACCEPTED MANUSCRIPT

12

Average Cost ($/hr)

50

10

40

J p(T)

30

J p(T *)

20
10

J p(T *)

J p(T)

6
4

Jc

Jc

0
0

0.002

0.004

0.006

0.002778

0.008

0.01

0.012

100

CR
IP
T

Average Cost ($/hr)

60

200

300

400

500

(a) Impact of changes in

(b) Impact of changes in N

Figure 2: Impact of changes in Average Cost with changes in and N

Average Cost ($/hr)

12
10
8

J p(T)

J p(T *)

AN
US

additional values of K in order to obtain new insights.

2
0

20

ED

40

Jc
60

55

80

100

120

PT

Figure 3: Impact of changes in Average Cost with changes in

For hospital management it is easier to alter parameters N and , while cost parameters K and may be

CE

harder to change. It is therefore important to define how each inventory policy behaves at different values of
cost parameters so hospitals can choose the right policy for their system conditions. Figures 3 and 4 illustrate
the changes in average cost for the models considered with changes in and K. From Figure 3 we see that the

AC

benefits of optimizing the periodic review interval versus using a continuous review policy are highly dependent
on . At a very low value of (at < 6 in Figure 3) the benefit of using an optimal periodic review policy can
be higher than the benefit of switching to a continuous inventory policy. At higher values of ( > 55 in Figure
3), the benefit of using an optimal periodic review interval decreases relative to the benefit of implementing a
continuous review policy. Low values of may be expected when materials stored in nursing wards are mostly
low-cost, non-critical items. If hospitals can identify supply rooms where most items fall into this category,
then these supply rooms or items can be replenished periodically with low inventory costs. By contrast supply

21

ACCEPTED MANUSCRIPT

rooms where several critical supplies are stored can benefit considerably from continuous-time replenishments,
and potentially from more careful unit-level monitoring as is the case with the use of ADMs (see Rosales et al.

18

18

16

16

Average Cost ($/hr)

14

14

12

12
10

10

J p(T)
J p(T *)

8
6
4

Jc

J p(T *)

0
0

100

200

300

100

200

300

(b) Impact of changes in K with = 6

AN
US

(a) Impact of changes in K with = 55

J cc

J p(T)

CR
IP
T

Average Cost ($/hr)

[34]).

Figure 4: Impact of changes in Average Cost with changes in K

From Figure 4 it is worth noting that the benefits of using an optimal periodic review policy versus using
a continuous inventory policy depend not only on the changes in K, but also on the relationship between K
and . For higher values of continuous review always dominates periodic review, while at lower values of ,

periodic review provides more benefits than continuous review for high K. Forcing an order to be placed when
a secondary bin becomes empty makes this option less economical, particularly for high values of K/, making

ED

the periodic replenishment option more economically attractive. In addition replenishments under continuous
review can occur at any time, often disrupting personnels daily activities and therefore can be associated
with higher fixed order costs (Rosales et al. [34]). Higher fixed order costs for continuous replenishment can

PT

result in even higher K/ ratios, decreasing the attractiveness of the continuous replenishment option. In this
paper comparisons between periodic and continuous review policies are performed assuming equal values of K.

CE

When unequal values of K need to be considered, results illustrated in Table 4 and Figure 4 can still provide
an easy way to compare the benefits of using periodic or continuous review policies under different values of

AC

K.

Although hospitals currently operating periodic review 2Bin systems may obtain significant savings by

changing to a continuous review policy, many hospital managers are hesitant to operate under continuous
review. Periodic review provides an easy way to schedule the replenishment of several supply rooms sequentially throughout the day, simplifying the task of scheduling and assigning responsibilities among material
management workers. Allowing system conditions to trigger replenishment orders may imply the need for
more complicated workforce schedules, potentially requiring the assignment of additional workers during peak
demand hours. Our work provides hospital management with a quantitative tool to assess the potential bene-

22

ACCEPTED MANUSCRIPT

fits of using an optimal periodic review interval, or alternatively switching to a continuous review policy when
demand is observed at the bin level, so more informed decisions can be made.
7. Conclusions and Future Research
2Bin replenishment systems have been used at hospitals for many years because of their simplicity for
managing and replenishing medical supplies in a multi-item setting. In recent years technology such as RFID

CR
IP
T

has been used to enhance the 2Bin system by increasing inventory visibility. Based on our interaction with
consulting company Logi-D as well as with hospitals currently using RFID enabled 2Bin systems, we have
learned that although RFID technology has been implemented in many hospitals, the replenishment policy
to be used and the system conditions that should trigger a replenishment are not always completely clear to
hospital management. In addition many hospitals currently using RFID-enabled 2Bin systems continue to rely

AN
US

on traditional periodic replenishment policies when RFID technology provides the potential of replenishing
materials in continuous time. In this paper we analyze the 2Bin inventory replenishment system and provide
quantitative models to evaluate the cost benefit of improving system conditions by parameter optimization,
choosing the best review interval for a periodic review model, or via policy improvement, changing from a
periodic to a continuous review system. We present a semi-Markov decision model for continuous review and

closed form cost expressions for periodic review. We also prove that under conditions we specify (
(n1 ,0) (0)J c
(1, 0)), the optimal replenishment strategy for the 2Bin continuous review policy is a threshold value policy,

ED

characterized by a number of emptied primary bins in order to trigger a replenishment. We also provide a
solution methodology to obtain the optimal threshold value for the 2Bin policy using a linear programming
formulation.

PT

We find that the benefits of using a continuous review policy can be substantial, typically greater than
50% reduction for typical hospital conditions, and discuss how hospitals should consider the type of inventory

CE

policy to be used when making system-design decisions such as the number and size of the storage areas to
maintain throughout the facility. We find that periodic review policies favor the use of fewer storage areas with

AC

a higher number of items. In addition periodic replenishment favors the use of larger bins as a way to reduce
fixed order costs, but this reduces stock rotation at point-of-use. Furthermore periodic review outperforms
continuous review only at low values of stock-out costs. Therefore periodic review is a suitable policy for
storage areas containing mainly low-cost, non-critical hospital supplies.
Future research will focus on assessing the impact of additional factors that may trigger inventory replenishment; for instance the length of time bins have remained empty may trigger a replenishment if bins are not
emptied following a Poisson process. For hospitals unable to invest in new technology or simply preferring to
replenish inventory periodically, alternative periodic-review policies, such as ones that place periodic orders

23

ACCEPTED MANUSCRIPT

only when a threshold number of bins are empty can be explored. Furthermore, questions such as how to
optimally determine the amount of inventory to be held in each bin, whether inventory should be equally
divided between both bins, and how to design a system with multiple storage locations in the facility remain
to be addressed.
Acknowledgements

CR
IP
T

We wish to thank Dr. Sylvain Landry, Professor in the Department of Logistics and Operations Management, HEC Montreal, as well as Richard Phillip, President of Logi-D, Quebec, CA, for sharing their valuable
knowledge and providing us access to relevant data. We also wish to thank the anonymous reviewers whose
suggestions helped improve our paper and provided us with promising avenues for future work.

AN
US

References

[1] AHRMM (2010). Healthcare supply chain, resources and logistics processes. Association for healthcare
resource and materials management, Chicago.

[2] Asmussen, S., O. Nerman, and M. Olsson (1996). Fitting phase-type distributions via the em algorithm.

Scandinavian Journal of Statistics, 419441.

[3] Atkins, D. and P. Iyogun (1988). Periodic versus can-order policies for coordinated multi-item inventory

ED

systems. Management Science 34, 791796.

[4] Balintfy, J. (1964). On a basis class of multi-item inventory problems. Management Science 10, 287297.

PT

[5] Bazaraa, M., H. Sherali, and C. Shetty (2006). Nonlinear Programing (2nd. Edition ed.). USA: Wiley.
[6] Bertsekas, D. (2007). Dynamic Programming and Optimal Control (4th ed.). Athena Scientific, Mas-

CE

sachusetts.

[7] Bijvank, M. and I. Vis (2012). Inventory control for point-of-use locations in hospitals. Journal of the

AC

Operational Research Socitey 63, 497510.

[8] Cakici, O., H. Groenevelt, and A. Seidmann (2010). Efficient inventory management by leveraging rfid in
service organizations. Proceedings if the 41st Hawaii International Conference on System Sciences.

[9] Damodaran, A. (2012). Cost of capital by sector. http://people.stern.nyu.edu/adamodar/New_Home_

Page/datafile/wacc.htm. [online access September 2012].
[10] Delen, D., B. Hardgrave, and R. Sharda (2007). A stochastic model for a visit to the doctors office.
Mathematical and Computer Modelling 45, 309323.
24

ACCEPTED MANUSCRIPT

[11] Denton, B. (2013). Handbook of Healthcare Operations Management Methods and Applications, Volume
184 of International Series in Operations Research and Management Science. USA: Springer.
[12] Feinberg, E. (1994). Constrained semi-markov decision processes with average rewards. Mathematical
Methods of Operations Research 39, 257288.
[13] Gaukler, G., O. Ozer, and W. Hausman (2008). Order progress information: Improved dynamic emergency

CR
IP
T

ordering policies. Production and Operations Management 17, 599613.

[14] Hadley, G. and T. Whitin (1963). Analysis of inventory systems. Prentice-Hall, Inc, Englewood Cliffs,
N.J..

[15] Hall, R. W. (2012). Handbook of Healthcare System Scheduling. Springer. New York, NY.

AN
US

[16] Johnson, M. E. and H. Lee (2002). Quad wants to be a Savi player in agribusiness: Case study 6-0015.
Tuck School of Business at Darmouth - Glassmeyer/McNamee Center for Digital Strategies.
[17] Kelton, W. D. and A. M. Law (2000). Simulation modeling and analysis (3rd. Edition ed.). McGraw Hill
Boston, MA.

[18] Khouja, M. and S. Goyal (2008). A review of the joint replenishment problem literature. European

Journal of Operational Research 186, 116.

[19] Landry, S. and M. Beaulieu (2010). Achieving lean healthcare by combining the two-bin kanban replen-

ED

ishment system with RFID technology. International Journal of Health Management and Information 1,
8598.

PT

[20] Lapierre, S. and A. Ruiz (2007). Scheduling logistic activities to improve hospital supply systems. Computers and Operations Research 34 (3), 624641.

CE

[21] Larsen, C. and G. P. Kiesm

uller (2007). Developing a closed-form cost expression for an (r, s, nq) policy
where the demand process is compound generalized erlang. Operations research letters 35 (5), 567572.

AC

Ozer

[22] Lee, H. and O.

(2007). Unlocking the value of RFID. Production and Operations Management 16,
4064.

[23] Little, J. and B. Coughlan (2008). Optimal inventory policy within hospital space constraints. Health
Care Management Science 11 (2), 177183.
[24] Meyer, A. and D. Meyer (2006).

Roundtable proceeding, transforming the health care supply

chain. http://ctl-test1.mit.edu/sites/default/files/library/public/proceedings_healthcare_
roundtable.pdf. [online access June 2012].
25

ACCEPTED MANUSCRIPT

[25] Mustafa Tanrikulu, M., A. Sen, and O. Alp (2010). A joint replenishment policy with individual control
and constant size orders. International Journal of Production Research 48, 42534271.
[26] Neuts, M. F. (1981). Matrix-geometric solutions in stochastic models: an algorithmic approach. Courier
Dover Publications.
[27] Opolon, D. (2010). Improving product availability in hospitals: the role of inventory inacurracies. Ph.D.

CR
IP
T

Thesis-Massachusetts Institute of technology, Engineering Systems Division.

[28] Ozkaya,
B., U. G
urler, and E. Berk (2006). The stochastic joint replenishment problem: a new policy,
analysis and insights. Naval Research Logistics 53, 525546.

[29] Pantumsinchai, P. (1992). a comparison of three joint ordering inventory policies. Decision Sciences 34,

AN
US

791796.

[30] Pedersen, C., P. Schneider, and D. Scheckelhoff (2009). ASHP national survey of pharmacy practice in
hospital settings:dispensing and administration-2008. Am J Health-Syst Pharm 66 (10).
[31] Rao, U. (2003). Properties of the periodic review (R,T) inventory control policy for stationary stochastic

demand. Manufacturing and Service Operations Management 5, 3753.

[32] Renberg, B. and R. Planche (1967). Un modle pour la gestion simultane des n articles dun stock. Revue

ED

Francaise dInformatique et de Recherche Operationnelle 6, 4759.

[33] Rosales, C., M. Magazine, and U. Rao (2013a). Hybrid-joint replenishment policies for medical supplies.

PT

Working Paper, Michigan State University, East Lansing, MI.

[34] Rosales, C., M. Magazine, and U. Rao (2013b). Point-of-use hybrid inventory policy for hospitals. Decision

CE

Sciences Journal, forthcoming.

[35] Ross, S. (2003). Introduction to Probability Models (8th ed.). Academic Press, An Imprint of Elsevier.

AC

[36] Sarac, A., N. Absi, and S. Dauzere-Peres (2010). A literature review on the impact of RFID technologies
on supply chain management. International Journal of Production Economics 128, 7795.

[37] Sari, K. (2010). Exploring the impacts of radio frequency identification RFID technology on supply chain
performance. European Journal of Operational Research 207, 174183.
[38] Tijms, H. C. (2003). A First Course in Stochastic Models (First ed.). Wiley, England.
[39] Viswanathan, S. (1997). Note. periodic review (s,S ) policies for joint replenishment inventory systems.
Management Science 43, 14471454.
26

ACCEPTED MANUSCRIPT

[40] Zhou, W. (2009). RFID and item-level information visibility. European Journal of Operational Re-

AC

CE

PT

ED

AN
US

CR
IP
T

search 198 (1), 252258.

27