Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Forecasting Drug Demand for Optimal Medical

Inventory Management: A Data-Driven Approach
with Advanced Machine Learning Techniques
Vibin Ravi Kumar 1 Pallavi Waghmare2 Sampath Bukya3
Research Associate, Mentor, Team Leader,
Innodatatics, Hyderabad, India Research and Development, Research and Development,
Innodatatics, Hyderabad, India Innodatatics, Hyderabad, India

*
Bharani Kumar Depuru4 Dr. Ilankumaran Kaliamoorthy5
Director, Innodatatics, Hyderabad, India CEO, Rela Institute and Medical Centre,
OCR ID: 0009-0003-4338-8914 Chennai, Tamil Nadu, India

Corresponding Author: *Bharani Kumar Depuru4

Abstract:- A hospital's capacity to allocate resources significance of using suitable prediction models, such as
efficiently and guarantee drug supply depends on collaborative predictions based on end-user consumption
effective medical inventory management. This study data, economic order quantity, or the Min/Max formula,
paper offers a thorough data-driven strategy for drug to ascertain the necessary dosages of critical medications
demand forecasting that makes use of cutting-edge while taking into account available resources, supply
machine learning methods, intending to improve medical chain information, and inventory levels. Healthcare
inventory management procedures. A range of machine organisations can considerably reduce prediction errors
learning algorithms were used to precisely model and and improve the efficiency of medical inventory
anticipate drug demand trends using historical data, management by utilising the results of this extensive
including Deep Learning-based models, time series research.
forecasting techniques, and ensemble learning methods.
To determine the best strategy for predicting drug Keywords: - Drug Demand Forecasting, Machine Learning
demand, the study compares the performance of various in Medicine, Medical Inventory Management, Healthcare
algorithms. Supply Chain, Predictive Analysis, Hospital Management

Healthcare facilities can improve patient care, I. INTRODUCTION

reduce waste, and achieve optimal supply chain
performance by minimising stockouts, lowering surplus The key objective of this study is to improve drug
inventory, and optimising the supply chain. The findings demand prediction accuracy through the use of data-driven
of this study increase medical inventory management strategies and cutting-edge machine-learning techniques.
procedures by offering insightful information on the use The research attempts to analyse historical medication sales
of cutting-edge machine learning methods for precise data and identify significant patterns and trends using RNN
drug demand forecasting. In turn, this promotes the use (Recurrent Neural Network), BI RNN (Bidirectional
of evidence-based decision-making and medical Recurrent Neural Network), LSTM (Long Short-Term
resources. Machine learning for forecasting has Memory), BiLSTM (Bidirectional Long Short-Term
enormous potential for revealing previously unknown Memory), GRU (Gated Recurrent Unit), BiGRU
patterns in disease, treatment, and care as the healthcare (Bidirectional Gated Recurrent Unit), and Ensemble models
sector experiences a data revolution with the growing (Gradient Boost). These models have the potential to
use of Artificial Intelligence (AI), Predictive Analytics, increase the accuracy and dependability of medication
and Business Intelligence. The research intends to demand forecasting since they are built to handle sequential
enhance people's health outcomes, socioeconomic status, data and capture dependencies over time. By assessing the
and day-to-day activities by resolving the difficulties efficiency of these cutting-edge machine-learning
caused by the complexity of pharmaceuticals and algorithms in predicting drug demand and improving
ensuring the supply of vital medications. medical inventory management.

The supply of essential drugs and life-saving India's pharmaceutical industry, which mostly depends
supplies can be less uncertain with accurate demand on China for the supply of some key chemicals and
estimates, which helps to create a well-organised and intermediates, has been constantly watched to assess the
effective health supply chain. The study highlights the impact of Covid-19 on its supply chain. At the time, Covid-

IJISRT23SE159 www.ijisrt.com 221

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
19 was in charge of numerous ground-breaking innovations important medications by comprehending and forecasting
in pharmacy material management. Lack of information future drug demand. Accurate demand forecasting also aids
regarding the spread of Covid-19 and lack of visibility over in cost control, waste reduction, and supply chain logistics
the delivery of some bulk medications causes speculative optimization.
price hikes [8].
The outcomes of this study could be advantageous to
Also, drugs have always been a vital component of several stakeholders in the healthcare ecosystem. The
both curative and preventive healthcare. A country's entire findings can be used by hospitals, pharmacies, and
public health budget, which ranges from 40 to 60 percent, is healthcare organisations to streamline their drug
spent on purchasing medications. Building a capable procurement processes, cut costs, and guarantee the
procurement system is the only option to increase the availability of critical pharmaceuticals. The insights can be
majority of the population's access to medications while used by pharmaceutical makers and companies to streamline
staying within the constraints of the available budget [9]. inventory management, distribution, and production. In the
end, greater drug demand forecasting will help patients and
Forecasting drug demand accurately is crucial for the healthcare system as a whole by enhancing patient care,
managing healthcare systems. It makes it possible for reducing pharmaceutical shortages, and better-allocating
healthcare providers and organisations to maximise their resources. The optimization of medical inventory
inventory, lessen stockouts, and prevent drug overstock. management has grown in relevance as computer systems
Healthcare institutions may streamline operations, improve have advanced, emphasising the necessity of our research in
patient outcomes, and ensure the timely availability of this area.

Fig 1 This Figure Depicts the CRISP-ML(Q) Architecture that We have followed for this Research Study.
(Source: Mind Map - 360DigiTMG)

To progress with this study in a structured manner we To better understand the factors impacting drug
have utilised and followed the CRISP-ML(Q) (CRoss- demand, we have gathered and analysed pertinent data sets
Industry Standard Process model for the development of during the “Data Understanding” [Fig.1] phase. Data from
Machine Learning applications with Quality assurance the past on patient demographics, disease prevalence, and
methodology) Mindmap (ak.1) [Fig.1]. Understanding the other contextual factors are also included. To build the
goals and needs of the healthcare sector concerning the groundwork for further modelling stages, exploratory data
forecasting of drug demand and the management of medical analysis techniques are used to find patterns, correlations,
inventory is part of the first step, referred to as "Business and outliers within the data.
Understanding" [Fig.1]. We aimed to deal with issues like
stockouts, excess inventory, and inefficient resource Pre-processing the gathered data to verify its integrity
allocation that healthcare providers encounter. Healthcare and usefulness for modelling is known as “Data
providers may optimise their inventory levels, cut expenses, preparation” [Fig.1]. The data must be cleaned, missing
and guarantee that patients have timely access to values must be handled, and any necessary variable
pharmaceuticals by correctly forecasting drug demand. transformations must be made. Techniques for feature

IJISRT23SE159 www.ijisrt.com 222

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
engineering can be used to extract useful predictors and Deployment” [Fig.1], allowing for real-time forecasting and
improve the models' capacity for prediction. optimization of drug inventory levels. To enhance their
supply chain management and resource allocation strategies,
“Data mining” [Fig.1] is the study of gathering, healthcare practitioners, policymakers, and pharmaceutical
cleaning, processing, analysing, and deriving practical corporations can use the research findings as practical
insights from data. To put it another way, data mining is the guidance.
practice of looking for patterns in datasets that contain a lot
of data (big data) to uncover undiscovered information or The CRISP-ML(Q) is one of the standards used in
knowledge. This is done by extracting and examining data mining. Because CRISP-ML(Q) is most frequently
significant or interesting patterns from data stored in used in data mining development, business problem
databases [11]. analysis, and data mining projects, Mariscal, Marba, and
Fernandez [12] declared it to be the de facto standard for the
The power of cutting-edge machine-learning creation of data mining and knowledge discovery projects
approaches is revealed during the “Model Building” [Fig.1] [3].
stage. To create a reliable and precise forecasting model, we
use a variety of algorithms, such as ensemble approaches  Architecture Diagram
and deep learning Models. These methods are excellent at Before going deeper into possible issues, we would
identifying intricate connections and trends within the data, like to have an analogy to an English idiom that says "A
allowing us to produce accurate and trustworthy predictions picture is worth a thousand words". As per this wiki
of the demand for drugs. explanation, "it refers to the notion that a complex idea can
be conveyed with just a single still image or that an image of
The generated model is then linked to the medical a subject conveys its meaning or essence more effectively
inventory management system during the “Model than a description does"

Fig 2 Architecture Diagram Representing a Drug Demand for Optimal Medical Inventory Management Incorporating Forecasting
Models

Data Collection from the medical inventory system errors, and Data Transformation [Fig.2, 3] is used to
[Fig.2, 3] is the first step in the process of acquiring details guarantee the consistency and quality of the data.
on inventory, sales, and other pertinent information. A SQL
(Structured Query Language) is then used to store and After pre-processing, exploratory data analysis (EDA)
retrieve the obtained data effectively. Next, Data Pre- is carried out to discover patterns, comprehend the data's
processing [Fig.2, 3] is applied, which includes data characteristics, and obtain new insights. To analyse
integration to combine data from different sources, Data sequential patterns and identify temporal correlations,
Reduction [Fig.2, 3] is used to pick out key features, Data machine learning models like RNN, BI RNN, LSTM, and
Cleaning [Fig.2, 3] is used to deal with missing values and Ensemble model (Gradient Boost) [Fig. 2, 3] are trained on

IJISRT23SE159 www.ijisrt.com 223

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
the pre-processed data after EDA. We decide on the making it simple for other systems or users to access and use
ensemble model because it provides the highest accuracy the model artefacts.
among these models.
The deployed models and the entire system are always
Using Streamlit [Fig.2, 3], a web application being watched over to guarantee continued performance and
framework that enables the development of interactive correctness. This entails monitoring model metrics, doing
interfaces for visualising and engaging with the models, the quality assurance checks on the data, and releasing timely
trained models are then deployed for use in real-time. The updates as soon as new data becomes accessible or as
deployed models are kept in an AWS S3 bucket [Fig. 2, 3], business requirements alter.

Fig 3 This Machine Learning Architecture Diagram Illustrates the Stages Involved in Achieving Optimal
Medical Inventory Management by Forecasting Drug Demand.
(Source: Open-Source ML Workflow Tool- 360DigiTMG) (ak.2)

II. METHODOLOGY AND TECHNIQUES patterns over a longer period and allowed for a thorough
investigation of the pharmaceutical inventory.
A. Data Collection
Using the hospital's Health Information system (HIS), The initial shape of the dataset was (5643260, 23),
a 4-year dataset for this Research was gathered. The dataset reflecting the considerable number of records and features
included drug sales information gathered over four years, included in the analysis.
giving a complete picture of the hospital's medicine supply.
The study provides a summary of the number of drugs
Data has been utilised and suitably processed for the consumed for both the train and test sets. The train set
prediction process. Predicting future data that can be used in included 39 months of data and the test set included 12
the healthcare supply chain is, in this regard, one of the most months of observations [1].
important objectives of time series analysis.
B. Data Pre-processing
For this work, health supply chain management data For this report, information on medical inventory
were employed. They featured information on consumption, management from 4+ years was gathered. They featured
inventory, orders and reorders, and purchasing prices [1]. information on consumption, inventory, orders and reorders,
and purchasing prices.
The dataset covered a wide range of medications and
the sales records for them, gathering details such as drug Before analysis, the dataset underwent Data Cleaning
category, volumes, and sales dates. The medicine categories, [Fig.2, 3] and Exploratory Data Analysis (EDA) [Fig.1] to
such as injections, tablets, fluids, and more, were explained uncover insights relevant to forecasting. Specifically, the
by the subcategory feature. While the date of the sale data considered for the analysis were limited to the date of
provided the precise day the drug was sold, the quantity drug purchase and the subcategory of drugs. The
feature stated how much of each drug was purchased. subcategory feature includes a total of 17 categories
encompassing 2,470 individual drugs.
The historical records from the HIS that were used in
this study's dataset covered a sizable 4-year time range. This We don't go with total counts, consequently, the reason
larger time frame made it easier to explore trends and why we use the method below.

IJISRT23SE159 www.ijisrt.com 224

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
To select the drugs for forecasting, a weighted direct future resource allocation and decision-making
approach was employed. This technique involved processes.
calculating the weightage of each drug based on its cost and
quantity. The cost weightage was computed as the product Utilising the aforementioned approach provided
of the medicine cost and the quantity purchased, divided by valuable insights and enabled a deeper understanding of the
the total cost of all drugs in the inventory. Similarly, the subject matter;
quantity weightage was calculated as the ratio of the
quantity purchased for each drug to the total quantity of all Below are the insights we have drawn from the Pareto
drugs. The total weightage of each drug was obtained by Chart of Medicines in terms of Total Cost, Total Quantity,
combining the cost weightage and the quantity weightage. and Profit

Let us provide a comprehensive analysis of the  The top 10 medicines constitute 49.61% of the total cost.
Exploratory Data Analysis (EDA) [Fig.2, 3] research The top 60 medicines constitute 80% of the total cost.
findings:  The top 10 medicines constitute 35.46% of the total
Quantity. The top 64 medicines constitute 80% of the
EDA is an evaluation method that uses data to uncover total Quantity.
designs, solve issues, and test hypotheses using analytical  The top 10 medicines constitute 55.47% of the profit.
and visual representations. EDA is carried out using data The top 35 medicines constitute 80 % of the profit.
sets [9]. EDA on data collection allows us to continuously
learn new perceptions. In this work, we used specific E. Model Approach
analytical techniques on the HIS pre-processed data set [14]. Our drug forecasting model serves as an example of
realistic and evidence-based predictive model’s ability to
C. Variables and Measures revolutionise the management of vital medicine stock levels
The variables used in this study include 23 among and store replenishment. We can maximise access to
those important variables are subcategory, quantity, and essential medications while lowering safety inventory and
sales date. The subcategory represents the category to which cutting waste by utilising precise and data-driven
each drug belongs. The quantity denotes the number of predictions. Our forecasting model, which is specially
drugs purchased. The sales date indicates the date when the designed for predicting drug demand, takes into
drug was sold. consideration both the complexity of the pharmaceutical
supply chain and the peculiar properties of healthcare
D. Statistical Analysis products. We strive to reduce forecasting errors and improve
The statistical analysis primarily involved descriptive the precision of our predictions through ongoing
statistics to summarise the dataset and Pareto analysis to improvement and refinement. We can produce forecasts that
identify the drugs contributing to a significant portion of the are closer to the demand point by combining our model with
sales, cost, and profit. supply chain activities and using real-time data, leading to
more accurate and trustworthy insights.
The examination of the distribution of drug demand
benefits greatly from the addition of a Pareto chart. Based The main goal of the research was to increase the
on the Pareto principle or 80/20 rule, the Pareto chart predictability of drug demand. We built several different
graphically illustrates the total percentage of drug demand forecasting models, including Recurrent Neural Network
attributable to various categories or causes. The figure (RNN), Bidirectional RNN (BI RNN), Long Short-Term
effectively illustrates the most important drivers of demand Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated
by grouping the categories or elements in descending order Recurrent Unit (GRU), and Bidirectional GRU (BiGRU),
of their influence on drug demand [7]. [Fig.2, 3] to do this. Each of these models was created
independently as a part of our research strategy.
The Pareto chart is a useful tool for pinpointing the
"vital few" classes or elements that significantly influence  Recurrent Neural Network (RNN) [Fig.2, 3]: RNN is
drug demand. It enables scholars and decision-makers to essential in anticipating medication demand because of
concentrate their attention and resources on these significant its capacity to handle sequential data and capture
areas first. Stakeholders can create focused interventions, dependencies between previous and present inputs. RNN
improve inventory management methods, and more wisely can find hidden patterns and trends in past sales data that
deploy resources by knowing which categories or factors might not be visible using conventional statistical
most influence demand. techniques. The RNN's memory cells allow the model to
keep track of data from earlier time steps, enabling it to
The Pareto graphic also makes it easier to have data- factor previous observations into its forecasts. Since
driven debates and make decisions. Through the visual historical sales data significantly affects future demand,
portrayal of complex information, stakeholders may better this skill is particularly useful in predicting drug
understand the distribution of drug demand and pinpoint demand. RNN helps in the correct prediction of drug
areas for improvement. The Pareto chart can be used in the demand, assisting in the best management of medical
study to gain a more thorough understanding of the main inventory, by using its grasp of temporal dependencies
factors influencing the demand for drugs, which will help [5].

IJISRT23SE159 www.ijisrt.com 225

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Bidirectional RNN (BI RNN) [Fig.2, 3]: By processing inventory management, the forecasting model uses the
sequential data in both forward and backward Gated Recurrent Unit (GRU), in conjunction with other
orientations, BI RNN improves prediction skills. The cutting-edge machine learning approaches. The GRU is a
model may acquire a more complete understanding of type of recurrent neural network that is made to handle
the sequential patterns and dependencies contained in the sequential data and identify temporal relationships. GRU
drug demand data thanks to this bidirectional processing. effectively captures both short-term oscillations and
BI RNN can find subtle trends and patterns by taking long-term patterns in drug demand by employing gating
information from both past and future inputs, which may mechanisms, which let it retain and update knowledge
not be visible with unidirectional techniques. The from past inputs in a selective manner. By adding GRU
forecasting algorithm can now predict drug demand to the forecasting model, its predictive powers are
more precisely thanks to this bidirectional analysis, improved, enabling precise projections and assisting in
which also helps to manage medical inventory the effective management of medical inventory [6].
effectively.  Bidirectional GRU (BiGRU) [Fig.2, 3]: The forecasting
 Long Short-Term Memory (LSTM) [Fig.2, 3]: Hospitals model must include the Bidirectional Gated Recurrent
can use Long Term Short Memory (LSTM) to anticipate Unit (BiGRU), which collaborates with other cutting-
when patients will require medications. This approach edge machine learning methods like RNN, BI RNN,
was selected since it is well recognized to be highly LSTM, BiLSTM, and GRU. To correctly forecast drug
accurate at foretelling stationary data. demand and improve medical inventory management,
 By overcoming the vanishing gradient issue and BiGRU is essential. The BiGRU is made to handle
identifying long-term dependencies in sequential data, sequential input and capture temporal dependencies,
LSTM improves predictive skills. Similar to the RNN, much like other recurrent neural networks. BiGRU
LSTM plays a critical role in medication demand obtains a thorough grasp of the underlying patterns and
forecasting by successfully simulating complicated trends by analysing data in both forward and backward
patterns and trends over time. This helps to estimate orientations.
future demand accurately. The LSTM can record both  Ensemble Model (Gradient Boost) [Fig.2, 3]: A
short-term fluctuations and long-term dependence predictive modelling technique called an ensemble
because it has memory cells and gating mechanisms that model combines the results of various independent
allow it to selectively store and use information from models to create a single, more reliable prediction. An
prior inputs. Similar to the RNN, this ability of the ensemble model was used to increase prediction
LSTM aids in improving forecasting accuracy and accuracy when projecting drug demand for the best
optimising the medical inventory management system management of medical inventories.
[4].  The forecasts from different individual forecasting
 Bidirectional LSTM (BiLSTM) [Fig.2, 3]: A vital part of models, including RNN, BI RNN, LSTM, BiLSTM,
the forecasting model, Bidirectional Long Short-Term GRU, and BiGRU, were integrated with the ensemble
Memory (BiLSTM) expands on the strengths of LSTM model. Each of these models has particular advantages
and takes on similar problems. By successfully capturing and traits. The ensemble model sought to improve
long-term dependencies and intricate patterns in prediction performance by utilizing the variety of these
sequential data, BiLSTM, which is similar to LSTM and models and their capacity to capture various facets of
RNN, is essential in accurately forecasting future drug demand trends.
medication demand. BiLSTM combines information
from past and future inputs by processing the sequential F. Research Findings
data in both forward and backward directions, giving a In our research, we conducted model building and
more thorough knowledge of the underlying patterns. evaluation using different resampling frequencies, including
Similar to LSTM and RNN, the use of BiLSTM daily, weekly, and monthly, to assess their impact on
improves forecasting accuracy while optimising medical forecasting accuracy.
inventory management procedures.
 Gated Recurrent Unit (GRU) [Fig.2, 3]: For precisely The performance of each model was evaluated, and the
forecasting drug demand and improving medical results are presented below.

Table 1 MAPE Values of Each Model

MAPE Meropenem Noradrenaline Pantoprazole Paracetamol
Train Test Train Test Train Test Train Test
RNN 41.7 21.8 50.5 56.3 9.52 10.7 53.6 75.8
BI-RNN 43 69 43 118.8 21.6 7.9 33.3 6.4
LSTM 44.7 81.8 30.7 48.4 21 9.69 30.3 6.7
BI-LSTM 35.7 76.8 32.2 76.1 22.8 14.6 22.1 13.9
GRU 64.7 78.8 50.5 56.3 65.7 82.7 53.6 75.8
BI-GRU 36.7 67.8 32 67.2 20.7 10.9 22.3 11.7
Gradient Boost 5.1 1.98 2.82 4.91 0.78 2.78 1.34 2.5

IJISRT23SE159 www.ijisrt.com 226

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
We evaluated the performance of various models for By calculating the MAPE for our ensemble forecasting
forecasting drug demand, namely RNN, BI-RNN, LSTM, model, we assessed its performance in accurately predicting
BI-LSTM, GRU, BI-GRU, and the Ensemble model drug demand. Additionally, we compared the MAPE scores
(Gradient Boost). We measured the Mean Absolute of the individual models (RNN, BI RNN, LSTM, BiLSTM,
Percentage Error (MAPE) for each model using different GRU, BiGRU) with that of the ensemble model to
drugs, including Meropenem, Noradrenaline, Pantoprazole, determine the improvement achieved through the ensemble
and Paracetamol. The MAPE values were calculated approach.
[Table.1] for both the training and testing datasets.
In addition to MAPE, we also considered other metrics
For the drug Meropenem, the Gradient Boost model such as Mean Absolute Error (MAE) and Root Mean
demonstrated the lowest MAPE values. While other models Squared Error (RMSE) to comprehensively evaluate the
such as RNN, BI-RNN, LSTM, BI-LSTM, GRU, and BI- performance of our ensemble model. These metrics
GRU showed higher MAPE values ranging from 21.8% to provided additional insights into the absolute differences
81.8% [Table.1], the Gradient Boost model achieved and dispersion of our predictions compared to the actual
significantly lower MAPE values of 1.98% for training and values.
5.1% for testing [Table.1]. This suggests that the Gradient
Boost model provided more accurate predictions, benefiting Through rigorous evaluation and comparison, we
from its ensemble nature and ability to capture complex aimed to validate the effectiveness of our ensemble
patterns in the data. forecasting model and demonstrate its superiority in
predicting drug demand.
Therefore, the Gradient Boost model was selected as
the preferred model for drug demand forecasting for  Reorder Point
Meropenem. Reorder point is the minimum inventory level at which
a new order should be placed to prevent stockouts. It is
G. Evaluation of the Forecasting Models required to ensure a continuous supply of products, handle
To evaluate the performance of the ensemble demand fluctuations, and optimise inventory efficiently
forecasting model, we employed various metrics such as
Mean Absolute Error (MAE), Root Mean Squared Error  Reorder Point Calculation:
(RMSE), and Mean Absolute Percentage Error (MAPE).
These metrics allowed us to assess the accuracy and Reorder point Formula: (Average Daily Use x Average
robustness of our predictions. Lead Time in Days) + Safety Stock

Mean Absolute Percentage Error (MAPE) is a To find Reorder point multiply the average daily usage
commonly used metric in forecasting to measure the by the average lead time in days, and then add the safety
accuracy of predictions relative to the actual values. It stock to determine the reorder point
calculates the percentage difference between the predicted
values and the corresponding actual values, averaging these  Method For Determining The Safety Stock
differences across the entire dataset. Safety stock refers to the additional stock held by a
company or organisation as a buffer to mitigate the risk of
MAPE is calculated using the following formula: stockouts or supply disruptions. It serves as a cushion to
account for uncertainties in demand, lead time, and supply
MAPE = (1/n) * Σ(|(Actual_i - Predicted_i)/Actual_i| * variability. The purpose of safety stock is to ensure that
100) there is sufficient inventory on hand to meet customer
demands during unexpected fluctuations or delays in the
Where: supply chain

 Actual_i represents the actual value of the target variable To calculate the Reorder point, we must first be aware
for the i-th observation. of our safety stock. Safety stock is calculated as follows:
 Predicted_i represents the predicted value of the target
variable for the i-th observation. Safety stock = (maximum daily usage x maximum lead
 n represents the total number of observations. time in days) – (average daily usage x average lead time in
days)
MAPE provides a relative measure of forecasting
accuracy, allowing us to understand the average percentage To calculate the safety stock, businesses take into
difference between our predictions and the actual values. A account the maximum daily usage, which represents the
lower MAPE indicates higher prediction accuracy, while a highest quantity of a product used or sold in a single day,
higher MAPE suggests a larger discrepancy between and the maximum lead time in days, which is the longest
predicted and actual values [4]. duration it takes for an order to be placed with a supplier
and for the inventory to be received. Subtracting the product
of the average daily usage and the average lead time in days
from this value results in the safety stock.

IJISRT23SE159 www.ijisrt.com 227

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Safety stock serves as an additional buffer to mitigate by looking at drug demand patterns at the regional level.
the risk of stockouts or supply disruptions caused by With the use of these insights, the health ministry is better
fluctuations in customer demand, lead time variability, or equipped to plan, use resources more effectively, and
other unforeseen circumstances. By having safety stock in enhance regional health outcomes.
place, companies can ensure that they have enough
inventory on hand to meet customer demands even during IV. CONCLUSION
unexpected situations, helping maintain a smooth and
reliable supply chain operation With a MAPE score of below 5%, our model for
forecasting drug demand and optimising medical inventory
Where: management demonstrates exceptional reliability. This level
of precision establishes a strong basis for its application in
 Average Daily Drug Consumption: numerous other healthcare initiatives, ensuring its
The average quantity of drugs consumed per day based effectiveness in addressing a wide range of challenges and
on historical data. opportunities within the field.

 Average Lead Time in Days: The proposed implementation of this model involves
The average time it takes for the drugs to be integrating real-time data streams from healthcare sources
replenished once an order is placed with the supplier. such as hospitals and public health agencies. The proactive
management of inventory and timely availability of
 Safety Stock: important pharmaceuticals made possible by this innovative
The buffer quantity of drugs kept on hand to account technique helps healthcare professionals improve patient
for unexpected fluctuations in demand or delays in the care while also maximising the management of medical
replenishment process. inventory.

By substituting the appropriate values into the formula, DECLARATIONS

we can calculate the reorder point for our drug forecasting
model. This value represents the inventory level at which a  Acknowledgments:
new order should be placed with the supplier to ensure a
continuous supply of drugs and prevent stockouts.  We acknowledge that with the consent from
360DigiTMG, we have used the CRISP-ML(Q)
III. RESULTS AND DISCUSSION methodology (ak.1) and the ML Workflow which are
available as open-source in the official website
The study aimed to enhance drug demand prediction of 360DigiTMG (ak.2).
through data-driven strategies and advanced machine
learning models. The models employed, including RNN, BI  Funding and Financial Declarations:
RNN, LSTM, BiLSTM, GRU, BiGRU, and the Ensemble
model (Gradient Boost), were assessed for their accuracy in  The authors declare that no funds, grants, or other
forecasting medication demand based on historical sales support were received during the research or the
data. Notably, the Gradient Boost model outperformed preparation of this manuscript.
others, achieving MAPE values as low as 1.98% for training  The authors declare that they have no relevant financial
and 5.1% for testing, compared to other models with values or non-financial interests to disclose.
ranging from 21.8% to 81.8%. This underscores the
effectiveness of the Ensemble approach, demonstrating its  Data Availability Statement:
ability to capture complex patterns and dependencies within The datasets used, generated and/or analysed during
the data. this study are not publicly available due to internal Data
Privacy Policy but are available from the corresponding
The success in achieving such low MAPE scores for author on reasonable request.
specific drugs demonstrates the effectiveness of our data-
driven approach and the power of advanced machine- COMPLIANCE WITH ETHICAL STANDARDS
learning techniques in forecasting drug demand. By
accurately predicting the demand for each drug, healthcare  Disclosure of Potential Conflicts of Interest:
organisations can optimise their inventory levels, reduce The authors declare no conflict of interest. The funders
wastage, and ensure the availability of critical medications had no role in the design of the study; in the collection,
for patients. analyses, or interpretation of data; in the writing of the
manuscript, or in the decision to publish the results.
The health ministry can identify regional health trends
by using the forecasting drugs inventory management model  Research involving Human Participants and/or Animals:
outlined in this study. The model may identify illness It is declared by all the authors that there was no
outbreaks, highlight healthcare inequities, personalise involvement of any human and/or animal trial or test in this
interventions, effectively allocate resources, assess program research.
performance, and encourage collaborative decision-making

IJISRT23SE159 www.ijisrt.com 228

Volume 8, Issue 9, September – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Informed Consent: [9]. Bhojwani, Divya & Bhojwani, Khushboo. (2019). A
As there were no human subject involved in this study to review drug inventory and pharmacy
research hence a informed consent is not applicable to the management with reference to I.V. & injectables at a
best of the authors’ understanding. tertiary municipal care hospital with 1800 bedded
hospital. 342-350.
 Conflict of Interest Statement: [10]. Binsar, Faisal. (2020). Drug Stock Optimization Based
The authors declare that there are no conflicts of on Consumption Patterns for Hospital Formulary
interest that could influence the results or interpretation of Using Deep Learning Approach. International Journal
this manuscript. This research was conducted in an impartial of Advanced Trends in Computer Science and
and unbiased manner, and there are no financial, personal, Engineering. 9. 2689-2697.
or professional relationships that might be perceived as https://doi.org/10.30534/ijatcse/2020/31932020.
having influenced the content or conclusions presented in [11]. Tuga, Mauritsius. (2019). Bank Marketing Data
this work. Mining using CRISP-DM Approach. International
Journal of Advanced Trends in Computer Science and
REFERENCES Engineering. 8. 2322-2329.
https://doi.org/10.30534/ijatcse/2019/71852019. .
[1]. Mbonyinshuti, Francois & Nkurunziza, Joseph & [12]. Mariscal, Gonzalo & Marbán, Oscar & Fernández,
Niyobuhungiro, Japhet & Kayitare, Egide. (2021). The Covadonga. (2010). A survey of data mining and
Prediction of Essential Medicines Demand: A Machine knowledge discovery process models and
Learning Approach Using Consumption Data in methodologies. Knowledge Eng. Review. 25. 137-166.
Rwanda. Processes. 10. 26. https://doi.org/10.1017/S0269888910000032.
https://doi.org/10.3390/pr10010026. . [13]. Zolnoori, Maryam & Fung, Kin & Patrick, Timothy &
[2]. Pall, Raman & Gauthier, Yvan & Auer, Sofia & Fontelo, Paul & Kharrazi, Hadi & Faiola, Anthony &
Mowaswes, Walid. (2023). Predicting drug shortages Wu, Shirley & Eldredge, Christina & Luo, Jake &
using pharmacy data and machine learning. Health Conway, Mike & Zhu, Jiaxi & Park, Soo & Xu, Kelly
Care Management Science. 1-17. & Moayyed, Hamideh. (2019). The PsyTAR Dataset:
https://doi.org/10.1007/s10729-022-09627-y. From Patients Generated Narratives to a Corpus of
[3]. M. J. Christ, R. Nikolaus, P. Tri, W. Chandra, and T. Adverse Drug Events and Effectiveness of Psychiatric
Mauritsius, Lending Club Default Prediction using Medications. Data in Brief. 24. 103838.
Naïve Bayes and Decision Tree, Int. J. Adv. Trends https://doi.org/10.1016/j.dib.2019.103838.
Comput. Sci. Eng., vol. 8, no. 5, pp. 2528–2534, [14]. Panigrahi, Chhabi & Panda, Bijayalaxmi & Pati,
2019. https://doi.org/10.1103/PhysRevC.46.1757. Bibudhendu. (2022). Exploratory Data Analysis and
[4]. Shao, Bilin & Song, Dan & Bian, Genqing & Zhao, Sentiment Analysis of Drug Reviews. Computación y
Yu. (2021). Wind Speed Forecast Based on the LSTM Sistemas. 26. https://doi.org/10.13053/cys-26-3-4093.
Neural Network Optimized by the Firework
Algorithm. Adv ]ances in Materials Science and
Engineering. 2021. 1-13.
https://doi.org/10.1155/2021/4874757.
[5]. Cheng, Lilin, Haixiang Zang, Tao Ding, Rong Sun,
Miaomiao Wang, Zhinong Wei, and Guoqiang Sun.
2018. "Ensemble Recurrent Neural Network Based
Probabilistic Wind Speed Forecasting Approach"
Energies 11, no. 8: 1958.
https://doi.org/10.3390/en11081958.
[6]. Qi, Chenyang & Ren, Jiaying & Su, Jin. (2023). GRU
Neural Network Based on CEEMDAN–Wavelet for
Stock Price Prediction. Applied Sciences. 13. 7104.
https://doi.org/10.3390/app13127104.
[7]. Alkiayat, Mohammad. (2021). A Practical Guide to
Creating a Pareto Chart as a Quality Improvement
Tool. Global Journal on Quality and Safety in
Healthcare. 4. https://doi.org/10.36401/JQSH-21-X1.
[8]. Basha, Mir Mohammed Junaid & V.S, Navya & Wani,
Sonali & Gogi, Vivekanand. (2020). Study of
Inventory Management in Pharmaceuticals: A Review
of COVID-19 Situation. International Journal of
Innovative Science and Research Technology. 5. 366-
371. https://doi.org/10.38124/IJISRT20AUG257.

IJISRT23SE159 www.ijisrt.com 229