Artificial intelligence’s impact on drug delivery in healthcare supply chain management: data, techniques, analysis, and managerial implications

Abstract

Healthcare supply chain management’s (HSCM) significance to economic and societal development is huge. In today’s very competitive market, supply chains have seen significant changes in the last several years. There is a need for technology that can handle the increasing complexity of today’s dynamic supply chain activities. Both machine learning (ML) and the quick dissemination of information have the potential to revolutionize the supply chain. ML has spawned a slew of useful supply chain applications in recent years, HSCM has received comparatively less attention. In this study, we applied three ML algorithms such as gradient boosting (GB), histogram gradient boosting (HGB), and cat boosting (CB) with data preprocessing tools to predict whether the medicines are delivered on time or not in the HSCM. The data preprocessing tools are used to manage datasets and increase the performance of ML algorithms. There are three methods of feature selection that are applied in this study such as Pearson correlation, chi-square test, and principal component analysis to select the best features to push in the ML algorithms. The main results show the CB is the best algorithm with the highest accuracy, precision, recall, and f1 score with values respectively. The three ML algorithms are compared with other ML algorithms to show the robustness of the applied ML algorithms. We made a sensitivity analysis to show the chaining in learning rate (LR) and compute the accuracy of the ML algorithms. We show the CB is not sensitive to values between 0.1 and 1.

Introduction

Increased knowledge and digital technology have a profound impact on healthcare, one of the most essential service industries. Overcrowding, lack of availability, and inaccessibility are all problems that have disrupted patient treatment in healthcare systems. Unpredictable calamities, such as earthquakes, floods, fires, and pandemic breakouts, amplify these interruptions and cause permanent harm to the healthcare system and many other essential sectors [1, 2]. These unforeseen occurrences throw healthcare systems off-kilter and cause bottlenecks. Different sections of healthcare supply chains (HSCs) should perform their crucial responsibilities successfully to address these difficulties, especially during emergencies [3, 4].

HSCs are made up of a network of suppliers, manufacturers, hospitals, blood banks, and pharmacies all working together to meet patient needs and provide optimal care. HSCs often have vendors of medical equipment and gadgets at the top of the food chain, and patients at the bottom [5,6,7]. Any error or disturbance at any level of HSCs may endanger lives and have far-reaching consequences. Complexity, diversity, service types, unpredictability, and objectives are what set HSCs apart from other supply chains (SCs). Big data analytics, AI, Blockchain, and cloud computing are just a few of the digital technologies that have greatly benefited HSCs during the last decade. Many new business possibilities arise as a result of the use and development of these innovations in healthcare systems, including the development of new business models for enhancing performance and generating anticipated values [8, 9]. Several studies applied blockchain and big data for drug SCM such as hospital waste management systems [10], pharmaceutical cold chains [11], evaluating service supply chain performance [12], and reduction maps for COVID-19 vaccine supply chain [13].

Machine learning (ML) algorithms can be applied in HSCM in various case studies. In this study, we apply the ML algorithms to predict the delivered medicines on time or not. The ML is a subset of the AI [14, 15]. To enable a system to automatically learn and develop from experience without having to be explicitly coded, ML is a vital component of artificial intelligence (AI) [16, 17]. ML methods are data-driven, meaning they actively look for patterns and adapt as they learn more. Opportunities for evaluating, categorizing, and forecasting on-time or delayed medication delivery are greatly enhanced by ML [18, 19].

ML uses the historical data in HSCM for patients in organizations to analyze some features such as ecological factors, and disease states to forecast the best delivery method for organizations and ensure the patient receives the drug as fast as possible. ML can be connected with sensors in the smart system to monitor the real-time delivery of products and can change the delivery in real-time to reduce costs and time delivery. In the traffic ways, ML can change the ways of delivering drugs to reach patients quickly. In the long ways, ML can process large historical data to predict the drug delivery is on time or not, so it can change the delivery system to reduce cost and faster reach to patients and organizations.

To predict the long-term viability of HSCs, Azadi et al. [20] developed a network data envelopment analysis (NDEA) framework and a deep learning technique. To foretell HSCs’ long-term viability, they used an NDEA model and a deep learning strategy. Bounded connection values may be optimized with their help. The DEA abilities are used to determine the threshold for every of these constrained connections to improve the performance of decision making units (DMUs). Each DMU’s dual-role connections function is also specified. Their primary findings indicate that the highest-scoring HSCs are those that consume the fewest resources while producing the highest quality goods and the fewest problems.

An ML-based methodology for choosing vendors’ incoterms (contracts) for direct drop-shipping in a worldwide omnichannel pharmaceutical supply chain was suggested by Detwal et al. [21]. They brought to light the most important considerations when deciding on an incoterm to use with a pharmaceutical vendor. Their results demonstrate that, for a wide range of input characteristics, the suggested model can reliably forecast a vendor incoterm (contract).

Critical success factors (CSFs) for artificial intelligence adoption in healthcare service delivery were the subject of Kumar et al.‘s research [22]. They utilize a rough version of Step-wise Assessment Ratio Analysis (SWARA) to rank HSC CSFs for the application of AI. According to their findings, technical, institutional/environmental, human, and organizational elements are the most significant in determining whether or not AI is used in HSC in developing nations.

Anil Kumar et al. [23]investigated what makes HSCs resilient and how to build resilient omnichannel HSCs (OHSCRs) of the future. A thorough literature analysis and in-depth interviews with subject-matter experts yielded the components that contribute to HSCs’ resilience. In the second stage, the cutting-edge building blocks of OHSCR were created using a machine-learning technique called K-means clustering. In the last stage, we spoke about what this all means and what directions we think future research should go in. The researchers concluded that the healthcare industry should evaluate OHSCR with a focus on six key building blocks. Table 1 shows the related works.

Table 1 Related works by ML and HSCM

The main contributions of this study are in summary:

• I. We gathered the data in HSCM to predict whether the medicine was delivered on time or not under different ML models.

• II. We apply feature selection methods such as Pearson correlation, chi-square test, and principal component test and show the ML model’s performance under different feature selection methods.

• III. The data preprocessing methods are applied in the dataset such as dealing with missing values and data normalization. Normalization is used to manage the outliers in the data and put all data in a specific range. The encoding data is used to encode the text categorical data. We compared the three models before and after applying the normalization method, we show higher accuracy after applying the normalization method.

• IV. We applied the three ML algorithms such as gradient boosting, histogram gradient boosting, and cat boosting in HSCM.

• V. We compare the applied ML with other ML algorithms. Our models show higher accuracy and performance compared with other models. The cat boosting obtained a higher accuracy equal to 98.8%.

• VI. We conducted a sensitivity analysis to show the changing of parameters and then computed the performance measurement. We showed the cat-boosting algorithm is not sensitive under different learning rates.

The rest of this paper is organized as: in Sect. 2 we introduce the experimental dataset. In Sect. 3, we introduce the three ML algorithms. In Sect. 4, we introduce the experimental setup. The results of the applied three ML algorithms are presented in Sect. 5 with the comparative and sensitivity analysis. In Sect. 6 we introduce the managerial implication from this study. In Sect. 7, we introduce the conclusions of this study.

Experimental dataset

This work uses the dataset to predict whether the medicine is delivered on time or not in HSCM. This study uses this dataset from the GitHub website and the details of this dataset are shown in Table 2. The dataset can be downloaded from [36]. This dataset includes 10,324 samples and 33 columns of information about drugs from pharmaceutical companies. The samples were collected from 2006 to 2016 delivering medicines to different countries. This study tends to predict whether the medicine is delivered on time or not. So we created a column named target by subtracting the delivered to client date and scheduled delivery date. If the delivered data is before the scheduled date, we put 1 in the target column. If the delivery date is after the scheduled date, we put 0 in the target column. So, we have a target column consisting of 0 (not delivered on time) and 1 (delivered on time). The column data includes day, month, and year. Figure 1 shows the number of orders by the target column, where x refers to the target column (0 not delivered on time) and (1 delivered on time), and y refers to the number of orders. The number of orders delivered on time is 4000 orders and not delivered on time is 6324. We analyze the dataset by computing the number of orders delivered on time and not in each country as shown in Fig. 2. We show that South Africa has the largest number of orders with 1013 orders not delivered on time followed by Vietnam with 666 orders and Cot d’Ivoire, and Afghanistan has the least orders not delivered on time followed by Liberia. Nigeria and Cot d’Ivoire are the largest countries that have orders delivered on time with 699 and 483 orders, and Libya and Benin have the least orders that are delivered on time. Figure 3 shows the number of dosage forms in the dataset. We show that Tablet has 34.2% of the dataset (3532 orders), followed by Tablet FDC 26.6% (2749 orders) and test kit 15.3% (1572 orders). So the Tablet is important to be delivered on time in the HSCM due to has largest number of orders in the dataset.

Table 2 The description of the dataset with the information of features

Fig. 1

Visualizing the target column with the number of orders

Fig. 2

The number of orders delivered and not delivered on time in each country

Fig. 3

The size of dosage form from the dataset

Fig. 4

The relation between the delivered date to the client and the scheduled date

Figure 4 shows the relation between the date delivered to the client and the date scheduled. We observe the difference between the two dates. So, we proposed this study to predict the medicines are delivered on time or not to aid organizations in healthcare. Figure 5 shows the number of orders in each shipment mode. This study has four types of shipment modes such as air, truck, air charter, and ocean. The air shipment mode has the largest number of orders (6113 orders), followed by truck (2830 orders), air charter (650 orders), and ocean (371). So, the transportation cost is increased in the air shipment mode due to having the largest number of orders. Figure 6 shows the top 20 order names in our dataset. We observed that Efavirenz has the largest number of orders (755 orders). Table 3 shows the ten transactions of the dataset with 33 columns. Table 4 shows the summary statistics of the dataset. From Table 4, there are inconsistent values in line item value, pick price, weight, fright cost, and line item insurance. So, we drop all the missing values for our dataset.

Fig. 5

The number of orders with shipment mode

Fig. 6

The top 20 orders in the dataset

Table 3 The sample of a dataset with ten transactions

Table 4 Statistics summary of dataset

Figure 7 shows the correlation map between features of our dataset. There are three values in correlation such as positive relationships (strong, medium, weak), no relationships, and negative relationships. There is a correlation between nine features and the target class. The line measurement, pack price, and unit price have a negative relationship with the target class. ID feature has a medium relationship with the target class. The line item quality, line item value, weight, freight cost, and line item insurance have positive relationships with the target class but weak relationships.

Fig. 7

The heat map between different features of the dataset

Figure 8 shows the outlier in the country, freight cost, line item insurance, and line item value. To mitigate the impact of the outliers, normalizing the data is recommended. Since the dataset is small, it’s preferable not to eliminate any rows unless necessary.

Fig. 8

The outliers of features in the dataset

The dataset has many missing values, so first, we drop all missing values. Then, we perform data preprocessing. Then, we perform the feature selection to select the best feature. Then, we applied the ML algorithms.

Methods

This section offers ML models to predict the medicines delivered on time or not in HSCM. This study uses boosting ML models. Boosting is a subset of Ensemble Learning that combines many low-performing predictors to produce a single high-performing model. This is effective because successive models learn from the errors of their predecessors. We use the three types of GB, CB, and HGB. The following detailed of using models:

Gradient boosting classifier (GB)

Decision trees (categorization trees for problematizing patterns and regression trees for approximating functions) are the backbone of GB, an ensemble of machine-learning algorithms. Each succeeding tree in an ensemble is reliant on the trees that came before it. That’s because the second tree works to minimize the former’s mistakes as shown in Fig. 9. During model training as shown in Fig. 10, this technique reduces the loss function that represents the ensemble’s misclassification rate. Importantly, by going in the opposite direction as the negative gradient, GB reaches global convergence. A strong committee may be established after training, even if the basic classifier was rather poor. The last step is to compile the predictions made from all of the decision trees to acquire the outcomes of new data samples [37, 38].

Fig. 9

The structure of the GB model

Fig. 10

The training process of the GB model

Cat boosting classifier (CB)

The Train Using Auto ML program implements CB, a supervised machine-learning technique that uses decision trees for regression and categorization. CB’s two primary characteristics are its ability to handle categorical data and its reliance on CB. Iteratively building multiple decision trees is the core of the GB procedure. As more trees are added, the quality of the final product increases. CB is a speedier alternative to the traditional gradient boost approach.

With CB, you may skip the tedious pre-processing step of converting category text variables to numbers, one-hot-encodings, and so on, a common requirement of other decision tree-based approaches. Without the need for any preprocessing, this approach may be fed a mixture of category and non-categorical explanatory factors. As part of the method, it does preprocessing. CB employs an approach to encoding categorical characteristics known as ordered encoding. To find a number to stand in for the categorical characteristic, ordered encoding takes into account the desired statistics from all rows before the data point in question [39, 40].

CB is distinct in its usage of symmetric trees, for example. This implies that the same split condition is used by all decision nodes across all depths.

Histogram Gradient boosting classifier (HGB)

Furthermore, the histogram-based technique is a powerful tool for training models using GB. This approach uses tiny bins to categorize the spectrum of continuous information utilized by decision trees. Hossain and Deb proposed a value of 255 for the total amount of bins. Histograms showing the distribution of characteristic values are constructed using these bins. The amount of data examples and the total gradients for every bin may be calculated, among other elementary statistics. Using these measures, we may zero in on the sweet spots for splitting apart the base-level learners throughout their instruction. As the histogram-based technique does not need to scan all ranges of characteristics for split point evaluation during the training stage of a decision tree, it may greatly minimize the computing cost. Although the learning phase is more resistant to noise, the histogram-based technique may also aid in greater generalization. This research employs a histogram-based technique to build the categorization tree structure because of its benefits in computing efficiency and learning effectiveness [41, 42].

Experimental setting

In this section, we offer the feature selection methods to select the best features in our dataset to obtain the highest accuracy, the performance measurement, and the implementation settings of our experiments.

Feature selection

We applied three feature selection methods in this study and obtained the ML model accuracy under three feature selection methods.

Pearson Correlation (PC)

Correlation-based Feature Selection (PC) calculates inter-correlation values to compute the correlation between features. This method can obtain links and relationships between the features and target features [43, 44]. A linear relationship between an input X and an output Y may be measured using a statistic called PC. It may be positive or negative depending on the strength of the relationship being measured. Therefore, a value of 0 indicates the absence of a linear relationship between features and target features. The PC is computed by dividing the covariance of input X and output Y at the input by the standard deviation of input X times and Y at the output.

$$\:PC=\frac{cov\left(x,y\right)}{std\left(X\right)std\left(Y\right)}$$

(1)

Cov refers to the covariance and std refers to the standard deviation.

Chi-Square Test

The chi-square test is useful for resolving the feature-selection issue since it examines the correlation between features [45, 46]. To determine whether or not two occurrence features are independent, researchers applied the chi-square test. Chi-Square calculates the deviation between the predicted output (E) and the actual output (O).

$$\:{X}_{c}^{2}=\sum\:\frac{{\left({O}_{i}-{E}_{i}\right)}^{2}}{{E}_{i}}$$

(2)

$$\:\left\{\begin{array}{c}c:degrees\:of\:freedom\\\:O:actual\:value\\\:E:predicted\:value\end{array}\right.$$

The number of possible, arbitrary values, or “degrees of freedom,” is defined as the largest possible set of independent variables. It may be written as the number of observations minus the number of constraints that are not based on the observations themselves.

Principal component analysis (PCA)

Often used to decrease the dimensionality of huge datasets, PCA works by reducing a large number of parameters and dimensions by preserving the amount of information in the dataset. Accuracy suffers as a data set’s amount of variables is reduced, but the idea is decreasing dimensionality without loss in information. Because ML techniques can more quickly and easily analyze data points when working with smaller data sets there are fewer factors to consider when exploring and visualizing such sets [47, 48].

Implantation settings

We implemented this study by using Python 3.10 with the library Sklearn. We trained the three algorithms by various values in hyperparameters. Table 5 shows the hyperparameter values for three algorithms. We divide the dataset into training and testing. The training value is 85% of the dataset and the test value is 15% of the dataset.

Table 5 The hyperparameters setting of three ML algorithms

Performance measurement

In this study, we suggest four performance measurements such as accuracy, precision, recall, and f1 score to evaluate three applied ML algorithms to select the best one. The equations of four performance measurements are detailed:

$$\:Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

$$\:Precision=\frac{TP}{TP+FP}$$

$$\:Recall=\frac{TP}{TP+FN}$$

$$\:{F}_{1}Score=2\times\:\frac{\left(Precision\times\:Recall\right)}{\left(Precision+Recall\right)}$$

$$\:\left\{\begin{array}{c}TP:True\:Positive\\\:TN:True\:Negative\\\:FP:False\:Positive\\\:FN:False\:Negative\end{array}\right.$$

Results and analysis

This section offers the results of the applied three ML algorithms in the HSCM dataset to predict whether the medicines are delivered on time or not.

Data preprocessing results (missing values and feature selection)

In this part, we introduce the data preprocessing tools to obtain good data to push it into the ML models. The data preprocessing contains results of missing values, and results, of feature selection.

We split the missing values into serval processes. First, we dropped all missing values of the dataset, we obtained 1526 records from the 10,324 records. Then we applied the feature selection tools. Table 6 shows the results of ML algorithms after dropping all missing values and applying feature selection tools.

In PC, there are three numbers of features selected including 3,6,9. We applied the ML algorithms in the PC with three features were selected. Of the 3 features, the GB has the highest accuracy, precision, recall, and f1 score. In 6 features, the CB has the highest accuracy, precision, and f1 score, but the HGB has the highest recall score. Of the 9 features, the HGB and CB have the highest accuracy, the CB has the highest precision, and the GB has the highest recall and f1 score.

In the chi-square test, in 3 features, the CB has the highest accuracy, and precision and the GB has the highest recall score and f1 score. In 6 features, CB has the highest accuracy, precision, recall, and f1 score. Of the 9 features, the GB has the highest accuracy and recall, the HGB has the highest precision and f1 score, and CB has the highest accuracy with the GB.

In the PCA, in the 3 features, the CB has the highest accuracy, recall, and f1 score, and the HGB has the highest precision and f1 score. Of the 6 features, the HGB has the highest accuracy, precision, recall, and f1 score. In the 9 features, the CB has the highest accuracy, precision, recall, and f1 score.

Overall, the CB has the highest accuracy, precision, and score in 9 features with the PCA feature selection. We conclude the CB is the best algorithm in the applied HSCM dataset with the PCA feature selection.

Table 6 Results of ML algorithms after applying to drop missing values and feature selection

Enhancement model accuracy

The goal of this part enhance the accuracy of the three ML algorithms. We concluded when dropping the missing values, we obtained a few records of the dataset (14.7% of the dataset). So, there is a large amount of information lost. So, we dropped the irrelevant columns and then computed the correlation between the dataset as shown in Fig. 11. We dropped the missing values in the dataset after dropping all irrelevant columns. We obtained 2902 records (28% of the dataset). After this, we doubled the size of the selected dataset. After this, we selected the nine columns (Country, Fulfill Via, Vendor INCO Term, Shipment Mode, Scheduled Delivery Date, Delivered to Client Date, Line-Item Value, Freight Cost (USD), Line Item Insurance (USD)). The nine selected features have various factors such as the type of shipment to deliver the medicines to clients, data of scheduled and delivered date, cost of transportation, and country of delivered medicines.

These selected features have the text categorical data such as country. So, we encode these data via the Label Encoder method. Then there are outliers in the dataset. So, we normalize these data to ensure that all are in the specific range and prevent the outlier. We used the Standard Scaler method to normalize the dataset. Then we split the dataset into a training (2466 records) and a testing set (436 records). We applied the three ML algorithms to the selected feature. Figure 12 shows the performance measurement of the three ML algorithms.

We observed the CB has the highest accuracy (98.8%), recall (87.5%), and f1 score (93.3%), but all algorithms are equal in precision (100%). So, the CB is the best ML algorithm.

Table 7 shows the confusion matrix of the three ML algorithms. We found the CB has 396 records of true positive and five records are predicted false.

Fig. 11

The correlation between the columns and the target class

Fig. 12

The performance measurement of the selected features

Table 7 The confusion matrix of the selected features

ROC-AUC curve

The area under the receiver operating characteristics (ROC) curve, which represents the degree of class separation, is a crucial performance metric for categorization algorithms. The area under the curve (AUC) measures how likely it is that a random positive example would be ranked higher than a random negative instance by the model being depicted. The area under the ROC curve from 0 to 1 is represented by the AUC, which is a metric with a maximum value of 1. Figure 13 shows the ROC curve for the three ML algorithms. We found that CB is the best algorithm with an accuracy of 99.48% followed by GB with an accuracy (of 98.88), and the lowest accuracy is HGB with 98.79%.

Fig. 13

The ROC curve of three ML models

Analysis with other ML algorithms

We compare the three ML models (GB, HGB, and CB) with other ML models such as logistic regression, support vector machine, random forest, ada boosting, and k-nearest neighbors [49,50,51,52,53]. We used the ML in the previous studies and applied it to our dataset. We found the three applied ML models are the best algorithms. We found that ada boosting has the least accuracy, precision, recall, and f1 score followed by logistic regression, and decision tree. Table 8 shows the comparative analysis results.

Table 8 Comparative analysis with the ML models

Statistical test

We used the statistical test method to test the applied three models with the comparative models as shown in Table 9. We used McNemar’s test to test the significant statistical or not [54]. McNemar’s test determines whether two sets of conflicts are consistent with one another. In statistical parlance, this is referred to as marginal homogeneity of the contingency table. For this reason, we may classify McNemar’s test as a homogeneity test for contingency tables. The test provides feedback on whether or not two models disagree in the same manner (or not) when comparing binary classification algorithms. It does not evaluate the relative quantity of different models or the likelihood of their making errors. We compute the p-value to compare with the threshold value (0.05). If the p-value is less than the threshold value, there is a significant statistical. If not, we conclude there is no significant statistical. In this study, we conclude the applied three models have significant statistics with the comparative model.

Table 9 P-values of applied three models and comparative models

Ablation study

In this part, we discuss the impact of the normalization method and LR when applied in the ML algorithms.

Impact of normalization

In this part, we measure the impact of the normalization method on the ML algorithms. In this study, we applied the standard scaler method to manage outliers and normalize all data by putting it in a specific range. Table 10 shows the accuracy precision, recall, and f1 score after and before applying three ML models. Before applying the normalization method, we obtained the CB has the highest accuracy (97.2%), followed by HGB with accuracy (96.5%), and GB with accuracy (95.6%). After applying the normalization method, we found the accuracy is increased, the CB has the highest accuracy with (98.85%) followed by the HGB with accuracy (98.39), and GB with accuracy (98.17%). We conclude the normalization method has a big impact on the applied three ML algorithms.

Table 10 The performance measurement before and after applying the normalization method

Sensitivity analysis

In this part, we applied the sensitivity analysis in this study to show the changes in the performance measures of the applied three ML algorithms. We change the values of the LR and then compute the performance measurement. We put three values of LR such as 1.0, 0.1, and 0.01, and then computed the performance measurements as shown in Table 11. We found the performance decreased from the original value of the LR. In GB, the highest accuracy in the LR is 0.1. The highest accuracy in the LR is 0.1. The highest accuracy in the CB is LR 0.1, 1. So, the CB is not sensitive to the LR between 0.1 and 1, and the other algorithms are sensitive to changing LRs.

We applied the k-fold cross-validation method for five folds for all models. We show a range of accuracy between 90 and 98.8% which means to accuracy is high and stable under different folds, the range of accuracy is small that show the consistent performance of the model and model has not the overfitting problem.

Table 11 The sensitivity analysis in the LR

Comparative analysis with previous studies

We compare the CB (which has the highest accuracy) with the previous studies as shown in Table 12. All these algorithms are applied in our dataset. From Table 12, we show that LSTM has 63.9% accuracy and CNN has 63.9% accuracy. The decision tree has two accuracies 88.98% and 80%. The KNN has a 98.09%. The SVM has an 83.3%. We show our applied algorithm is the best with 98.85% accuracy.

Table 12 Comparative with previous studies

Managerial implications

There are various managerial implications in the prediction of the delivery of medicines on time or not in HSCM.

The results of this study can aid managers and organizations in predicting the delivered medicines in the HSCM in inventory management to make considerations about the delay in the delivered medicine or not, so the inventory has an important role in storing more medicines if the delivered medicines are delayed.

The results of this study can aid organizations in demand for drugs. This can help the manager to order the most accurate number of medicines in the future based on the scheduled date of delivery.

Patient care and customer satisfaction can be increased by using this study model to predict whether medicines are delivered on time or not. This allows managers to enhance customer service by communicating with the service provider in the predicted delay.

This study can help managers in cost management in the HSCM by knowing the best shipment mode, inventory cost, and some shortcuts.

This model can aid in risk management by reducing the risk when predicting whether the medicines are delivered on time or not. So, the manager can develop a plan to manage the orders of medicines in the HSCM.

Conclusions

This study applied the data preprocessing methods and ML algorithms in the HSCM dataset to predict the delivered medicines on time or not. We applied the data preprocessing method to the dataset. We handle the missing values by dropping all of them. Then we use the normalization method to manage the outliers and put all data in a specific range. Then we encode the text categorical data. We applied the three ML algorithms in the HSCM dataset to show whether the medicines were delivered on time or not. We obtained the CB is the largest accuracy, followed by the HGB and GB. We used the ROC-AUC curve to show the accuracy of the three applied ML algorithms, we show the CB is the best algorithm. We made a comparative analysis to show the robustness of the applied three ML algorithms. We compare the applied ML with five ML algorithms. We show that the three applied ML algorithms have the highest accuracy, precision, recall, and f1 score. We made a sensitivity analysis to show the changes in the LR values and then computed the performance measurement. We put the leering rate with three values such as 0.1, 0.01, and 1. We conclude the CB is not sensitive between 0.1 and 1 LR and other ML algorithms are sensitive to changes in LR. There is a limitation in this study, a few number of samples in the dataset, in future studies this problem can be solved by collecting datasets from different sources to increase the volume of the dataset. Also, the deep learning models can be used in the future direction to show the accuracy, precision, recall, and f1 score to show the difference between deep learning and machine learning. Transfer learning can be used to train a large number of datasets to obtain higher accuracy and performance.