Lead Time Forecasting with Machine Learning Techniques for a

Pharmaceutical Supply Chain

Maiza Biazon de Oliveira1 a

, Giorgio Zucchi2,3 b , Marco Lippi4 c , Douglas Farias Cordeiro5 d
,
Núbia Rosa da Silva1,6 e and Manuel Iori4 f
1 Special Academic Engineering Unit, Department of Production Engineering, Federal University of Goiás,
St. Dr. Lamartine Pinto de Avelar, 1120, 75704020, Catalão, Goiás, Brazil
2 Fondazione Marco Biagi, University of Modena and Reggio Emilia, Largo Marco Biagi 10, 41121 Modena, Italy
3 R&D Department, Coopservice S.Coop.p.A, Via Rochdale 5, 42122 Reggio Emilia, Italy
4 Dipartimento di Scienze e Metodi dell’Ingegneria, University of Modena and Reggio Emilia,

Via Amendola 2, Pad. Morselli, 42122 Reggio Emilia, Italy

5 Faculty of Information and Communication, Federal University of Goiás, Campus Samambaia,

74690900, Goiânia, Goiás, Brazil

6 Institute of Biotechnology, Federal University of Goiás, St. Dr. Lamartine Pinto de Avelar,

1120, 75704020, Catalão, Goiás, Brazil

Keywords: Lead Time Forecasting, Machine Learning, Pharmaceutical Supply Chain.

Abstract: Purchasing lead time is the time elapsed between the moment in which an order for a good is sent to a supplier
and the moment in which the order is delivered to the company that requested it. Forecasting of purchasing
lead time is an essential task in the planning, management and control of industrial processes. It is of particular
importance in the context of pharmaceutical supply chain, where avoiding long waiting times is essential to
provide efficient healthcare services. The forecasting of lead times is, however, a very difficult task, due to
the complexity of the production processes and the significant heterogeneity in the data. In this paper, we use
machine learning regression algorithms to forecast purchasing lead times in a pharmaceutical supply chain,
using a real-world industrial database. We compare five algorithms, namely k-nearest neighbors, support
vector machines, random forests, linear regression and multilayer perceptrons. The support vector machines
approach obtained the best performance overall, with an average error lower than two days. The dataset used
in our experiments is made publicly available for future research.

1 INTRODUCTION gravate pathologies, given the possibility of deterio-

ration of health conditions over time. Longer wait-
Long waiting times for service interventions are a ing times for medical intervention can increase read-
recurring feature in the health sector, especially for mission rates as well (Moscelli et al., 2016). Nowa-
public services. Clearly, timely treatments and drug days, this is even more crucial because of the recent
administrations are crucial factors for improving the COVID-19 pandemic, which is causing an increase
quality of healthcare services, and often also for in the number of pharmaceutical products urgently
saving the lives of patients, mainly in emergen- required by the many patients affected by the dis-
cies (Brown et al., 2016; Tetteh, 2019). The delay ease (Harapan et al., 2020).
for medical interventions, whether through medica- Among other factors, long waiting times for re-
tion, diagnosis or surgical procedures, can indeed ag- ceiving medicines can be associated with delay in
the administrative packaging, logistic problems with
a https://orcid.org/0000-0002-8981-1314 tracking and delivery (Haugh, 2014) and several other
b https://orcid.org/0000-0002-5459-7290 factors that could be outside the control of patients
c https://orcid.org/0000-0002-9663-1071 or healthcare professionals. Within this scenario, the
d https://orcid.org/0000-0002-5187-0036 analysis and proposition of measures to reduce wait-
e https://orcid.org/0000-0003-1982-5144 ing times for all possible related factors is important
f https://orcid.org/0000-0003-2097-6572 in healthcare policy guidelines (Moscelli et al., 2016).

634
Biazon de Oliveira, M., Zucchi, G., Lippi, M., Cordeiro, D., Rosa da Silva, N. and Iori, M.
Lead Time Forecasting with Machine Learning Techniques for a Pharmaceutical Supply Chain.
DOI: 10.5220/0010434406340641
In Proceedings of the 23rd International Conference on Enterprise Information Systems (ICEIS 2021) - Volume 1, pages 634-641
ISBN: 978-989-758-509-8
Copyright c 2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
 Lead Time Forecasting with Machine Learning Techniques for a Pharmaceutical Supply Chain

The availability of medicines in healthcare service real and/or simulated. In this research, we exploit in-
networks, pharmacies and hospitals is directly related telligent methods, leaving conventional methods to an
to the lead time of the supply chain (Tetteh, 2019). analysis of the literature.
Our work is motivated by the activity of a logistic Recently, there have been significant advances in
company, Coopservice group, that receives the phar- this research field using artificial intelligence (Ioan-
maceutical products from the suppliers and then or- nou and Dimitriou, 2012; Gyulai et al., 2018). This
ganizes the shipping, when needed, to the healthcare process is mainly due to the growing availability of
facilities. To organize the service in the best possi- large data collections in different fields of manufac-
ble way, it is crucial for the company to correctly turing, that can enable data-driven technologies such
estimate the purchasing lead time, that is, the time as machine learning, data mining, knowledge discov-
that is elapsed between the moment in which an or- ery in databases, and big data analytical tools (Fayyad
der for a good is sent to a supplier and the moment et al., 1996; Tsai et al., 2016; Frank et al., 2019;
in which the good is delivered to the company. Cor- Kabugo et al., 2020)). Nevertheless, most of the in-
rectly forecasting this purchasing lead time (lead time telligent techniques used in recent research do not
for short, in the following) in the supply chain of the make use of real data (Öztürk et al., 2006), while us-
pharmaceutical sector is a crucial task, as it largely ing computer simulations to generate data and consid-
affects the whole industrial process of the healthcare ering many simplifying assumptions for the internal
services. In addition, proper estimation of lead time manufacturing process.
is a critical parameter in the relationship between the Given the limitations of the methods mentioned
management process and the customer (Noori-Daryan so far, in this paper we aim to use intelligent meth-
et al., 2019), being lead time one of the most impor- ods to predict the delivery times of suppliers who
tant performance indicators for the management of have to deliver the goods to a company that man-
manufacturing and service production processes (Kim ages the pharmaceutical supply-chain of hospitals. To
et al., 2014). Furthermore, accurate forecasting of this aim, we compared five different machine learn-
lead times can assist in optimizing the production pro- ing regression approaches, namely: k-nearest neigh-
cesses, by more accurately selecting the needed quan- bors (KNN), support vector machines (SVM), ran-
tities and thus shortening the overall production times dom forests (RF), linear regression (LR) and multi-
(Gyulai et al., 2018). layer perceptrons (MLP).
Besides, lead time prediction is a crucial aspect The use of accurate lead time forecast can be
to keep under control in the pharmaceutical supply highly beneficial in the planning of both production
chain, because sometimes having the medicine avail- and logistic services in the pharmaceutical field. We
able at the right time can save lives. Lead time fore- mention, to this regard, the work by (Gatica et al.,
casting could allow the pharmaceutical companies to 2001), who studied stochastic aspects related to prod-
predict and to avoid possible out of stock, caused by uct development and capacity planning in the phar-
a supplier. Besides, based on the lead time, it is pos- maceutical sector, by proposing a multistage stochas-
sible to evaluate the different suppliers and select the tic programming approach, and that of (Kramer et al.,
best ones. In addition, with a good prediction it is 2019), who proposed a metaheuristic algorithm for
possible for the pharmaceutical companies to define the delivery of pharmaceutical products in the region
different level of security stock of the goods for each of Tuscany (operated by the Coopservice group). In
month, making the procurement process leaner and the former work, accurate prediction of the lead time
more cost effective. for purchasing the products could be used within the
However, lead time forecasting is an extremely what-if analysis, while in the latter work, accurate
challenging task. In general, the estimation of lead predictions could be used to define the starting points
times from historical data has been a recurrent issue in of the deliveries, as multiple depots are available, and
the literature since the 1960s, and even in recent years the possible use of temporary depots at the hospitals,
some traditional systems simply obtain lead time by so as to reduce transportation costs and times.
computing average values based on historical data, The reminder of the paper is organized as follows.
with the result of deficiencies in production planning In Section 2 we present the related works and com-
and control (Lingitz et al., 2018). The proposed ap- pare our work with the literature. In Section 3 we
proaches in this research field can be divided into con- briefly present the classic techniques that we used to
ventional methods and intelligent methods, with the predict the lead time. In Section 4 we describe the
former not using artificial intelligence and the latter dataset used in the experiments, which are illustrated
exploiting data mining and machine learning. In both in Section 5. Finally, Section 6 concludes.
cases, data used for experimental evaluation can be

635
ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

2 RELATED WORKS Hosoda and Disney, 2018; Lingitz et al., 2018; Golt-
sos et al., 2019). Additionally, databases generated
by simulation often consider a perfect production sys-
In an Industry 4.0 scenario, big data analytics can be
tem, without introducing machine breakdowns, main-
divided into five different categories: predictive, de-
tenance downtime and raw material delays (Lingitz
scriptive, inquisitive, preventive and prescriptive an-
et al., 2018). When performing lead time analysis
alytics. Predictive analytics aims to anticipate what
and forecasting, it is important to consider external
will happen in the future: descriptive analytics in-
factors too, such as relationships and interactions be-
stead provides information and explanations about
tween different supply chains (Hosoda and Disney,
what has happened; inquisitive analytics tries to an-
2018; Ponte et al., 2018; Goltsos et al., 2019; Noori-
swer why it has happened, and preventive analytics
Daryan et al., 2019). (Chung et al., 2018) showed
provides insight to understand what is necessary to be
that lead time prediction is a key factor because the
done. Finally, prescriptive analytics provides infor-
lead time uncertainties can affect service level and or-
mation for decision-making (Sivarajah et al., 2017;
der lead time performance. Understanding these dy-
Cabrera-Sánchez and Villarejo-Ramos, 2020). Big
namics allows companies to reduce their exposure to
data analytics is very often associated with artificial
different types of delivery risk and to better manage
intelligence, data mining, and machine learning in-
their supply chain.
struments (Dean, 2014), with the aim to develop sys-
Despite the large amount of works in this area,
tems that can automatically extract information and
we could not find comprehensive studies on machine
discovery patterns in large data collections (Lu et al.,
learning algorithms for lead time forecasting in the
2015; Kuo et al., 2018), so as to provide beneficial
field of pharmaceutical distributions. Related works
insights to decision makers (Chamikara et al., 2020).
are limited to the use of Monte Carlo simulation to
By mid-1980s, many studies on operating and lead predict the production lead time (Eberle et al., 2014),
time estimation through mathematical formulations, and to the proposal of cyclic production plans com-
as well as statistical methods with analysis of variance bined with outsourcing in the packaging of medicines
(ANOVA) were proposed (Chang, 1997; Tatsiopou- in the Netherlands (Strijbosch et al., 2002). With this
los and Kingsman, 1983). Forecasting through math- paper we aim at filling this research gap.
ematical modeling approaches has also been recently
proposed for a custom system disregarding the current
system workload (Vandaele et al., 2002). In a more
complex product development scenario, a heuristic 3 METHODOLOGY
approach was proposed, by explicitly modeling net-
works of operating system activities (Jun et al., 2006). As already stated in Section 1, we employ a machine
Other research has proposed the use of queuing net- learning approach for purchasing lead time forecast-
works for lead time analysis and prediction (Ioannou ing of pharmaceutical services. We formulate the task
and Dimitriou, 2012; Berling and Farvid, 2014) with as a regression problem, where the aim is to predict
the use of discrete event simulation through mathe- a single real number y ∈ R as a function of a set of
matical expressions, assuming a continuous demand features x ∈ Rd . Supervised machine learning ap-
and studying the variance of the lead time. Con- proaches are able to learn a function f that computes
versely, a case-based reasoning approach was pro- a value ŷ from a given input vector x̂. Such a function
posed in (Mourtzis et al., 2014) to predict the lead is learned from a dataset D , which consists of a col-
time of complex engineered-to-order products. (Pfeif- lection of N pairs (xi , yi ) where each input example
fer et al., 2016) made use of multivariate regression xi is associated with the corresponding target yi , that
statistical methods using simulated data to obtain the is the target of the forecasting system. In this work,
production lead time of a flow-shop system. we compare several simple, classic regression algo-
Mathematical and statistical formulas were refor- rithms, largely used in statistics and machine learn-
mulated and proposed for production lead time esti- ing applications, with the aim of finding the one that
mates in chemical sector modular production plants performs the best on our real-world data set, without
(Sievers et al., 2017). However, the main disadvan- resorting to more sophisticated approaches. We com-
tage of all the methods cited so far is that they con- pare two efficient linear methods, namely linear re-
sider that past trends could possibly be repeated in the gression and linear support vector machines, against
future (Öztürk et al., 2006; Ioannou and Dimitriou, three simple non-linear ones, namely random forests,
2012). Moreover, there are few researches evaluating k-nearest neighbors, and multi-layer perceptron. We
the interactions of supply chain elements such as lead leave the use of more advanced machine learning ap-
times and forecasting procedures (Sievers et al., 2017; proaches for future research.

636
 Lead Time Forecasting with Machine Learning Techniques for a Pharmaceutical Supply Chain

3.1 Linear Regression the Euclidean distance, which we use in our experi-
mental evaluation. Once the k nearest neighbors are
Linear regression (LR) is a widely employed paramet- found, the algorithm computes the prediction as an
ric regression technique (Montgomery et al., 2012), average, or voting procedure, among them. In a re-
where function f is computed as a linear combination gression setting, the predicted target value ŷ is simply
of input features: f (x) = βT x + β0 . The vector of pa- computed as the weighted average of the targets y j of
rameters β is typically learned by minimizing the sum all k neighbors.
of squared errors on the training set. Clearly, this ap-
proach achieves good results when a linear function 3.5 Multi-Layer Perceptron
results to be a reasonable approximation of the de-
pendency relation holding between input and output A multi-layer perceptron (MLP) is a very simple ar-
variables, while suffering when such dependency is tificial neural network that can learn non-linear func-
strongly non-linear. tions between input and output variables (Rumelhart
and McClelland, 1987). An MLP consists in a stack
3.2 Linear Support Vector Machines of layers, each consisting of a certain number m of
neurons. The first layer consists of input variables.
In the second layer, named hidden layer, the output
Support vector machines (SVM) are a classic machine
of each neuron is computed as a non-linear combina-
learning approach that can be used both for classifica-
tion of input variables, whose weights are learned dur-
tion and for regression. In the regression setting, the
ing a training phase. Finally, the last layer computes
goal is to find a function f for which the forecasting
the output of the network as a non-linear combina-
error with respect to target y is at most equal to a pre-
tion of the output of the hidden neurons, again with
defined tolerance threshold ε for the elements in the
adjustable, learnable weights.
training set (Drucker et al., 1997). In its linear formu-
lation, which is the one we employ in this paper, the
function to be learned is still a linear combination of
the features. The optimal (or close to optimal) param- 4 DATASET
eters are found by heuristically solving a constrained
quadratic optimization problem (Albers et al., 2011). A crucial ingredient of any machine learning applica-
tion is the preparation of the dataset used for training
3.3 Random Forests and evaluation (Ristoski and Paulheim, 2016). The
database used in this research was made available
by an integrated service company, the Coopservice
A random forest (RF) is an ensemble classifier that
Group. Founded in 1992, the Coopservice Group pro-
consists in a collection of n different decision trees
vides specialised services to private companies and
(Breiman et al., 1984). A decision tree is an inter-
public entities. The Group operates worldwide, with
pretable classifier that inductively learns classification
its headquarters in Italy, and counts around 20,000
rules by testing the informativeness of the attributes
employees. It offers a variety of facility services, es-
(features) with respect to the category (in case of clas-
pecially the ones that are not part of the core busi-
sification) or the target value (in case of regression) to
nesses of the clients, including: industrial, commer-
be predicted. Several different decision trees can be
cial and healthcare cleaning; management and main-
obtained either considering different sets of features,
tenance of buildings and systems; management of en-
or by subsampling different sets of training examples.
ergy supplies; security and surveillance; transport and
In the regression setting, the output prediction of the
handling of goods; industrial and commercial mov-
RF is computed as the average of the predictions of
ing; collection and transport of special waste. With
individual trees.
18 logistic warehouses and a storage area of over
150,000 squared meters, Coopservice Group is the
3.4 k-Nearest Neighbors leader in healthcare and pharmaceutical logistics in
Italy, and a key provider of management and distri-
Based on the concept of distance (or similarity) be- bution services for pharmaceuticals, medical-surgical
tween examples, k-nearest neighbors (KNN) is not devices and non-medical consumables. The key as-
properly a learning algorithm. Given a test exam- pects for the services are relying on a large workforce,
ple x, the KNN algorithm looks for the k examples working at client-sites, maintaining consistent quality
in the training set that are the most similar to x, i.e., and monitoring performance.
the nearest ones according to a given metric, such as Forecasting lead times is a crucial task for

637
ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

January
February
0.10 March
April
May
0.08 June
July
August
0.06

Density
September
October
November
0.04 December

0.02

0.00
10 0 10 20 30 40
Days
Figure 2: Lead time distribution, as a function of the month.
Figure 1: Distribution of the number of samples in the
dataset, for each different category. • ordered quantity (pills);
• distance between supplier and the pharmacy ware-
Coopservice, because with an accurate prediction it is house (km).
possible to optimize and manage the scheduling of the
A standard pre-processing phase was applied to the
truck deliveries, as well as predict the unloading pro-
database, including explorative data visualization,
cess schedule for the inbound area. Thanks to this, it
cleaning and removal of duplicates and corrupted
is possible to better organize the shifts of the employ-
data, outlier detection, manipulation of missing val-
ees in the warehouse. In addition, lead time prediction
ues (Ristoski and Paulheim, 2016). In particular, we
allows the company to have a better knowledge of the
used boxplots to identify outliers and extreme values
supplier and to evaluate its performance. In order to
(Hu et al., 2018; Sagaert et al., 2019) to remove cor-
do this, a supplier rating system can be created, con-
rupted data. Figure 2 shows the distribution of the
sidering the historical data and the prediction. Finally,
lead time for each month. It can be noticed that the
with an accurate forecasting of lead times, the man-
trend is quite similar for all the months, with a peak
agement of safety stock in the warehouse can be safer,
between 3 and 7 days, and very few values exceed-
avoiding negative events like overstock and stockout.
ing 32 days. After a detailed analysis of the cases
In the pharmaceutical database provided by
with such a large lead time, we noticed they were
Coopservice, the total number of samples was 42,753
due to insertion errors in the original database, and
collected during 2018.
hence we discarded them. Overall, around 5% of data
All pharmaceutical products in the database are
were removed following the whole pre-processing
associated with some specific categories, namely: tu-
and cleaning procedure. The resulting dataset is avail-
moral, diagnostic, medicine, nutritional, prostatic,
able for research at https://github.com/regor-unimore/
sanitary, dialysis, heavy items, toxic, narcotic, and
Pharmaceutical-Lead-Time-Forecasting.git.
economale (that are all the non-medical items like
pens, papers...). All these categories were used in our
study, although most of the data belong to economale,
medicine, or sanitary categories, as shown in Figure 1. 5 EXPERIMENTAL RESULTS
For each sample in the database, eight indepen-
dent variables were considered as the input vector x To compare the machine learning systems employed
for our machine learning systems used to forecast lead in our analysis, we performed two different experi-
times: ments, splitting the whole corpus by category, as well
• day of the month of the customer order (1 to 31); as by month.
Initially, in order to select the best hyper-
• weekday of the customer order (1 to 5, from Mon- parameters of each algorithm, we employed a stan-
day to Friday); dard 10-fold cross-validation procedure, where the
• month (1 to 12) of the customer order; whole dataset is partitioned into 10 different groups,
• supplier code identifier; named folds. In turn, each fold is considered as test
set, whereas the remaining folds were split into 2/3
• product name identifier; for the training set, and 1/3 for the validation set.
• pharmaceutical product type category; The training set is the set of examples used during

638
 Lead Time Forecasting with Machine Learning Techniques for a Pharmaceutical Supply Chain

the learning phase to find the optimal model parame- Table 1: Mean squared error obtained per each different cat-
ters, whereas the validation set is the set of examples egory (best results in bold).
that is employed to evaluate the performance of the KNN LR RF MLP SVM
learned model. In this way, we selected the following Tumors 3.37 2.23 2.39 3.87 1.94
hyper-parameters for our machine learning systems: Diagnosis 4.98 2.37 3.40 7.41 2.51
100 estimators (i.e., number of trees) for the RF, lin- General 4.59 2.22 3.48 8.12 2.30
ear kernel and a regularization term C = 1 for SVM, Medicine 4.10 2.22 2.71 5.51 2.02
Nutritional 2.90 2.21 2.28 4.60 2.28
a value of k=13 for the number of neighbors in KNN, Prostatic 3.11 1.75 3.07 3.15 3.38
and a single hidden layer with 3 neurons for MLP. Sanitary 3.11 2.22 2.49 6.98 2.30
Then, we performed two distinct experiments. As Dialysis 3.23 1.50 2.49 2.34 1.83
a first experiment, we partitioned the dataset by cat- Heavy Goods 2.66 1.79 2.40 5.40 1.86
egory, and we split each portion into 2/3 to be used Toxic 3.73 1.70 2.68 2.03 1.73
for training, and 1/3 to be used for test. As a sec- Narcotics 3.72 5.16 5.44 4.29 4.81
Average 3.59 2.31 2.99 4.88 2.45
ond experiment, we partitioned the dataset by month,
and again we split the data of each month into 2/3 for
Table 2: Mean squared error obtained per each different
training, and 1/3 for test. In both experiments, as a
month (best results in bold).
standard performance metric, we considered the mean
squared error (MSE) as the average of the squared dif- KNN LR RF MLP SVM
January 3.43 5.13 2.62 5.60 1.86
ference between true and predicted lead time: MSE = February 2.77 4.20 2.05 5.46 1.58
∑ni=1 (yi − ŷi )2 where yi is the true lead time, and ŷi is March 3.88 2.83 6.14 6.94 1.80
the forecast value. April 3.96 9.51 2.94 8.03 1.87
The two experiments have different goals. In the May 3.57 5.74 2.54 7.69 1.55
first case, one full year of data for each category is June 3.79 5.91 2.71 7.01 1.58
used both for training and for test, thus we can eval- July 3.84 2.69 3.00 8.75 2.09
uate the performance of a forecasting approach when August 4.01 2.43 3.15 13.47 2.02
September 3.49 5.47 2.55 6.44 1.55
a long period of data is available, for each single cat- October 3.87 2.36 2.95 7.36 1.76
egory. Conversely, in the second experiment, we take November 3.91 2.72 2.95 6.95 2.21
into account all the categories, by partitioning the data December 4.09 7.01 3.25 10.33 2.86
by month: in this way, we can evaluate whether data Average 3.72 4.67 3.07 7.84 1.89
from different categories can help in forecasting the
lead times of each product.
As for the first experiment, in Table 1 we report
the performance achieved by all the competitors on
6 CONCLUSIONS
each distinct category. The results show that LR is
the best performing method. A very similar perfor- This paper presented a methodology for lead time
mance is also obtained by the SVM approach, that forecasting in the pharmaceutical supply chain with
achieves the lowest error in two categories (Tumors machine learning techniques. In particular, we com-
and Medicine). Narcotics results to be the most diffi- pared support vector machines, random forests, multi-
cult category to forecast, which is not surprising, as it layer perceptron, linear regression, and k-nearest
contains very few examples. For that category, KNN neighbors on a very large collection of examples pro-
is the best-performing algorithm. vided by a large company with headquarters in Italy.
In our second setting, the samples of all the cate- Our experimental results are very encouraging, show-
gories are used within the training and test set of each ing how the purchasing lead time can be forecast with
month. As shown in Table 2, in this case SVM is high accuracy, especially for linear support vector re-
clearly the best performing algorithm, achieving the gression. In particular, the use of simple non-linear
lowest MSE in every month, with an average error approaches does not seem to yield significant im-
equal to 1.89 days, which is largely better than the provements in the forecasting.
second best approach, which is RF, that achieves an The research described in this paper aims to fill
MSE equal to 3.07 days only. Overall, the results a gap in the scientific literature regarding lead time
of both settings suggest that the use of non-linear ap- forecasting for the purchase of pharmaceutical prod-
proaches does not significantly lower the forecasting ucts. An accurate forecast of such lead time can
error. be crucial for decision making, optimization, and
planning in the overall pharmaceutical supply chain.
Waiting times for drugs and medicines could in fact be
reduced, and hospitals and pharmacies could choose

639
ICEIS 2021 - 23rd International Conference on Enterprise Information Systems

the most convenient supplier at every moment on the stocks on logistical performance in a border-crossing
basis of accurate predictions. This can be very rele- jit supply chain. Computers & Industrial Engineering,
vant when treating patients with urgent needs, as well 118:440–450.
is fast-changing medical conditions, as the ones we Dean, J. (2014). Big data, data mining, and machine learn-
are currently facing in the COVID-19 pandemic. ing: value creation for business leaders and practi-
tioners. John Wiley & Sons.
Future research will incorporate forecasting of in-
ternal supply chain lead times of real service pro- Drucker, H., Burges, C. J., Kaufman, L., Smola, A. J.,
and Vapnik, V. (1997). Support vector regression ma-
cesses. In this way, the forecast of lead time for pur- chines. In Mozer, M. C., Jordan, M. I., and Petsche,
chasing products will be coupled with the forecast of T., editors, Advances in neural information processing
the entire supply chain lead time, providing decision systems 9, pages 155–161. MIT Press.
makers with a larger instrument of analysis. In addi- Eberle, L. G., Sugiyama, H., and Schmidt, R. (2014). Im-
tion, more sophisticated approaches to lead time fore- proving lead time of pharmaceutical production pro-
casting could be exploited, with simulation of non- cesses using monte carlo simulation. Computers &
linear systems to investigate how machine faults and Chemical Engineering, 68:255–263.
maintenance procedures can influence lead time. Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P.
(1996). From data mining to knowledge discovery in
databases. AI magazine, 17(3):37.
Frank, A. G., Dalenogare, L. S., and Ayala, N. F. (2019).
ACKNOWLEDGMENT Industry 4.0 technologies: Implementation patterns in
manufacturing companies. International Journal of
The research described in this paper was carried Production Economics, 210:15–26.
out with funding from the Brazilian State Funding Gatica, G., Shah, N., and Papageorgiou, L. G. (2001). Ca-
Agency of Goiás (FAPEG), Brazilian National Coun- pacity planning under clinical trials uncertainty for the
pharmaceutical industry. In Gani, R. and Jørgensen,
cil of State Funding Agencies (CONFAP-ITALY) and S. B., editors, European Symposium on Computer
Higher Education Personnel Improvement Coordina- Aided Process Engineering - 11, volume 9 of Com-
tion (CAPES). puter Aided Chemical Engineering, pages 865 – 870.
Elsevier.
Goltsos, T. E., Ponte, B., Wang, S., Liu, Y., Naim, M. M.,
and Syntetos, A. A. (2019). The boomerang returns?
REFERENCES accounting for the impact of uncertainties on the dy-
namics of remanufacturing systems. International
Albers, C. J., Critchley, F., and Gower, J. C. (2011). Journal of Production Research, 57(23):7361–7394.
Quadratic minimisation problems in statistics. Jour-
Gyulai, D., Pfeiffer, A., Nick, G., Gallina, V., Sihn, W., and
nal of Multivariate Analysis, 102(3):698–713.
Monostori, L. (2018). Lead time prediction in a flow-
Berling, P. and Farvid, M. (2014). Lead-time investiga- shop environment with analytical and machine learn-
tion and estimation in divergent supply chains. Inter- ing approaches. IFAC-PapersOnLine, 51(11):1029–
national Journal of Production Economics, 157:177– 1034.
189.
Harapan, H., Itoh, N., Yufika, A., Winardi, W., Keam, S.,
Breiman, L., Friedman, J., Stone, C. J., and Olshen, R. A. Te, H., Megawati, D., Hayati, Z., Wagner, A. L.,
(1984). Classification and regression trees. CRC and Mudatsir, M. (2020). Coronavirus disease 2019
press. (covid-19): A literature review. Journal of Infection
Brown, M. T., Bussell, J., Dutta, S., Davis, K., Strong, S., and Public Health, 13(5):667 – 673.
and Mathew, S. (2016). Medication adherence: truth
and consequences. The American journal of the med- Haugh, K. H. (2014). Medication adherence in older adults:
ical sciences, 351(4):387–399. The pillbox half full. Nursing Clinics of North Amer-
ica, 49(2):183–199.
Cabrera-Sánchez, J.-P. and Villarejo-Ramos, Á. F. (2020).
Acceptance and use of big data techniques in services Hosoda, T. and Disney, S. M. (2018). A unified theory of
companies. Journal of Retailing and Consumer Ser- the dynamics of closed-loop supply chains. European
vices, 52(C):101888. Journal of Operational Research, 269(1):313–326.
Chamikara, M., Bertok, P., Liu, D., Camtepe, S., and Khalil, Hu, X., Pedrycz, W., and Wang, X. (2018). Fuzzy classifiers
I. (2020). Efficient privacy preservation of big data for with information granules in feature space and logic-
accurate data mining. Information Sciences, 527:420– based computing. Pattern Recognition, 80:156–167.
443. Ioannou, G. and Dimitriou, S. (2012). Lead time estima-
Chang, F.-C. (1997). Heuristics for dynamic job shop tion in mrp/erp for make-to-order manufacturing sys-
scheduling with real-time updated queueing time esti- tems. International Journal of Production Economics,
mates. International Journal of Production Research, 139(2):551–563.
35(3):651–665. Jun, H.-B., Park, J.-Y., and Suh, H.-W. (2006). Lead time
Chung, W., Talluri, S., and Kovács, G. (2018). Investigat- estimation method for complex product development
ing the effects of lead-time uncertainties and safety process. Concurrent Engineering, 14(4):313–328.

640
 Lead Time Forecasting with Machine Learning Techniques for a Pharmaceutical Supply Chain

Kabugo, J. C., Jämsä-Jounela, S.-L., Schiemann, R., and mining and knowledge discovery: A comprehensive
Binder, C. (2020). Industry 4.0 based process data an- survey. Journal of Web Semantics, 36:1–22.
alytics platform: A waste-to-energy plant case study. Rumelhart, D. E. and McClelland, J. L. (1987). Learning
International Journal of Electrical Power & Energy Internal Representations by Error Propagation, pages
Systems, 115:105508. 318–362. MIT Press.
Kim, S. H., Kim, J. W., and Lee, Y. H. (2014). Simulation- Sagaert, Y. R., Kourentzes, N., De Vuyst, S., Aghezzaf, E.-
based optimal production planning model using dy- H., and Desmet, B. (2019). Incorporating macroe-
namic lead time estimation. The International Jour- conomic leading indicators in tactical capacity plan-
nal of Advanced Manufacturing Technology, 75(9- ning. International Journal of Production Economics,
12):1381–1391. 209:12–19.
Kramer, R., Cordeau, J.-F., and Iori, M. (2019). Rich vehi- Sievers, S., Seifert, T., Franzen, M., Schembecker, G., and
cle routing with auxiliary depots and anticipated deliv- Bramsiepe, C. (2017). Lead time estimation for modu-
eries: An application to pharmaceutical distribution. lar production plants. Chemical Engineering Research
Transportation Research Part E: Logistics and Trans- and Design, 128:96–106.
portation Review, 129:162–174. Sivarajah, U., Kamal, M. M., Irani, Z., and Weerakkody, V.
Kuo, C.-F. J., Lin, C.-H., and Lee, M.-H. (2018). Ana- (2017). Critical analysis of big data challenges and
lyze the energy consumption characteristics and af- analytical methods. Journal of Business Research,
fecting factors of taiwan’s convenience stores-using 70:263–286.
the big data mining approach. Energy and Buildings, Strijbosch, L., Heuts, R., and Luijten, M. (2002). Cyclical
168:120–136. packaging planning at a pharmaceutical company. In-
Lingitz, L., Gallina, V., Ansari, F., Gyulai, D., Pfeiffer, A., ternational Journal of Operations & Production Man-
and Monostori, L. (2018). Lead time prediction using agement, 22(5):549–564.
machine learning algorithms: A case study by a semi- Tatsiopoulos, I. and Kingsman, B. (1983). Lead time man-
conductor manufacturer. Procedia CIRP, 72:1051– agement. European Journal of Operational Research,
1056. 14(4):351–358.
Lu, W., Chen, X., Peng, Y., and Shen, L. (2015). Bench- Tetteh, E. K. (2019). Reducing avoidable medication-
marking construction waste management performance related harm: What will it take? Research in Social
using big data. Resources, Conservation and Recy- and Administrative Pharmacy, 15(7):827–840.
cling, 105(Part A):49–58. Tsai, C.-F., Lin, W.-C., and Ke, S.-W. (2016). Big data
Montgomery, D. C., Peck, E. A., and Vining, G. G. (2012). mining with parallel computing: A comparison of dis-
Introduction to linear regression analysis, volume 821 tributed and mapreduce methodologies. Journal of
of Wiley Series in Probability and Statistics. John Wi- Systems and Software, 122:83–92.
ley & Sons. Vandaele, N., Boeck, L. D., and Callewier, D. (2002). An
Moscelli, G., Siciliani, L., and Tonei, V. (2016). Do waiting open queueing network for lead time analysis. IIE
times affect health outcomes? evidence from coronary transactions, 34(1):1–9.
bypass. Social Science & Medicine, 161:151–159.
Mourtzis, D., Doukas, M., Fragou, K., Efthymiou, K., and
Matzorou, V. (2014). Knowledge-based estimation of
manufacturing lead time for complex engineered-to-
order products. Procedia CIRP, 17:499–504.
Noori-Daryan, M., Taleizadeh, A. A., and Jolai, F.
(2019). Analyzing pricing, promised delivery lead
time, supplier-selection, and ordering decisions of a
multi-national supply chain under uncertain environ-
ment. International Journal of Production Economics,
209:236–248.
Öztürk, A., Kayalıgil, S., and Özdemirel, N. E. (2006).
Manufacturing lead time estimation using data min-
ing. European Journal of Operational Research,
173(2):683–700.
Pfeiffer, A., Gyulai, D., Kádár, B., and Monostori, L.
(2016). Manufacturing lead time estimation with
the combination of simulation and statistical learning
methods. Procedia CIRP, 41:75–80.
Ponte, B., Costas, J., Puche, J., Pino, R., and de la
Fuente, D. (2018). The value of lead time reduction
and stabilization: A comparison between traditional
and collaborative supply chains. Transportation Re-
search Part E: Logistics and Transportation Review,
111:165–185.
Ristoski, P. and Paulheim, H. (2016). Semantic web in data

641