INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

Predictive Machine Learning Framework for Medical Relief Supply

Decision-Making in Disaster Response Operations

Roman B. Villones, Rachel J. Vergara

Graduate School, La Consolacion University Philippines, Philippines

DOI: https://doi.org/10.51583/IJLTEMAS.2025.1411000016

Received: 10 November 2025; Accepted: 20 November 2025; Published: 02 December 2025

ABSTRACT

This study aims to strengthen decision-making in medical relief supply management through computational

modeling and machine learning. Specifically, it seeks to develop a predictive model that supports effective

medical supply allocation during disaster response operations. By experimenting with multiple algorithms and

optimization techniques, the study endeavors to identify the most suitable machine learning approach for

accurate and real-time prediction of supply reorder needs. The study adopts the SEMMA methodology as the

framework for developing decision-making models that focus on optimizing medical supply distribution during

natural calamities. Both ensemble and classical algorithms were implemented to compare predictive

performance. Ensemble methods included Random Forest, Gradient Boosting, HistGradientBoosting, AdaBoost,

XGBoost, LightGBM, and CatBoost, while classical algorithms comprised of Decision Tree, Logistic

Regression, K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). Model evaluation utilized the

F1-score and ROC-AUC metrics, with hyperparameter tuning conducted via Randomized Search and 5-fold

cross-validation to ensure robust generalization. Results revealed that CatBoost achieved the highest F1-score

(0.9316), outperforming other models such as XGBoost (0.9282), Gradient Boosting (0.9279), and LightGBM

(0.9277). These findings underscore the superiority of gradient boosting ensemble methods in capturing complex

relationships and improving prediction accuracy. Traditional models, including AdaBoost, Decision Tree,

Logistic Regression, KNN, and SVM are yielded with lower F1-scores ranging from 0.8051 to 0.8738 only and

indicating its reduced predictive capability after optimization. The CatBoost algorithm demonstrated the most

reliable and accurate predictive performance by highlighting its potential as a robust decision- support tool for

optimizing medical supply allocation during disaster response operations.

Keywords — Predictive Modeling, Machine Learning, Decision Support Systems, Disaster Response, Medical

Relief Supply Management.

INTRODUCTION

In recent years, the growing complexity of disaster environments has encouraged the adoption of data-driven

and machine learning–based decision-support systems in humanitarian logistics (Rodríguez-Espíndola et al.,

2023). The predictive modeling and artificial intelligence (AI) approaches are increasingly used to forecast

medical supply demand and improve real-time decision-making during emergencies (Bastani et al., 2023).

Furthermore, the integration of big data analytics and optimization algorithms has enabled more adaptive and

responsive medical relief operations where critical resources can be dynamically deployed to high-impact

areas (Bag et al., 2021).

Despite this advancement, the real-world disaster response operations remain constrained by data fragmentation

and uncertain demand mechanisms. Inconsistencies in data collection and the absence of unified decision-

support frameworks often result in inefficiencies such as overstocking and shortages of essential medicines and

equipment (Ma, K. et al., 2022). Additionally, many existing models lack validation through real-world

experiments by limiting their reliability and adoption by organizations.

This study aims to strengthen decision-making in medical relief supply management through computational

modeling and machine learning. Specifically, it seeks to develop a predictive machine learning model to support

www.ijltemas.in

Page 169

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

decision-making for medical supply allocation during disaster response operations. Experimenting with multiple

machine learning algorithms to evaluate predictive performance and suitability in disaster contexts. And

implementing optimization techniques and cross-validation methods to fine- tune model accuracy for real-time

deployment.

The study contributes to the field of logistics and AI- driven decision systems in several ways. In methodological

contribution, it introduces a predictive framework that integrates supervised machine learning, optimization, and

cross-validation by addressing the gap between theoretical modeling and applied decision-making. For empirical

contribution, experiments on disaster-response datasets, this study will benchmark multiple algorithms and

identify which are most effective for high-uncertainty and time-sensitive environments. By practical

contribution, the resulting decision-support model enhances situational awareness and enables data-informed

allocation of medical supplies by helping relief organizations respond faster and more efficiently to disasters.

REVIEW OF RELATED LITERATURE

The SEMMA methodology was developed by the SAS Institute. It provides a systematic framework for

developing and evaluating data mining and machine learning models. It focuses on transforming raw data into

reliable insights through an iterative and structured process (Al Alawi et al., 2022). In emergency logistics

research, constructing composite indices to quantify localized demand pressure between supply and need, in such

composite metrics combine observed demand, casualty/impact estimates, and infrastructure accessibility to

guide prioritization when raw demand signals are noisy (Ma, Z. et al., 2023).

Recent benchmarking literature categorizes machine learning models for classification tasks, for example, a

comparative study of gradient-boosting algorithms under hyperparameter tuning on multiple datasets then noting

a trade-off between accuracy and computational cost (Florek & Zagdański, 2023). Well-tuned boosting

implementations typically achieve top performance on complex tabular tasks, while ensemble strategies provide

robustness across noisy and imbalanced settings (Ayodele, 2023). At the same time, applied evaluations continue

to include classical models as interpretable baselines to quantify practical gains from more complex learners (El

Guabassi et al., 2021).

Recent analyses, threshold-sensitive and threshold- independent metrics together like F1-score for balancing

precision/recall in imbalanced and ROC-AUC for overall discrimination across thresholds (Li, 2024). In

addition, empirical work highlights AUC’s relative stability across prevalence scenarios including the prediction

accuracy for determining the model’s performance in healthcare system (Owusu-Adjei et al., 2023). For

hyperparameter tuning, modern applied studies increasingly use randomized search strategies as efficient in

reproducible alternatives to exhaustive grids (A Ilemobayo et al., 2024). In addition, hyperparameter

optimization documented that randomized sampling as a practical approach for large model families and limited

compute budgets (Ali et al., 2023).

METHODOLOGY OF THE STUDY

In this study, SEMMA serves as the methodological foundation for building decision-making models aimed at

optimizing medical supply allocation during natural calamity relief operations.

Fig. 1. SEMMA Methodology according to Al Alawi et al. (2022)

www.ijltemas.in

Page 170

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

Sample

The dataset used in this study consists of 76,000 records and 10 attributes, representing critical variables that

influence medical supply allocation during natural calamity relief operations. The dataset includes features such

as Item ID and District, which identify the supply items and their geographic distribution. Operational variables

such as Inventory Level, Relief Aid, and Inventory Re-Ordered capture logistical and replenishment patterns.

Contextual variables like Calamities provide disaster-related information affecting supply demand. Derived

indicators such as Simple Ratio Method, Weighted Demand Index, Normalized Score, and Overall Demand

Score were computed to quantify supply prioritization and regional demand intensity.

Explore

This phase focuses on identifying patterns, correlations, and data quality issues that may influence model

accuracy.

Fig. 2. Distribution of Medical Relief across Districts during the Calamities

Figure 2 presents the distribution of data entries by district in the NCR, Philippines. The 2nd District recorded

the highest count, followed by the 4th and 3rd Districts, while the 1st District showed the lowest frequency. This

indicates a significant variation in data concentration across districts, suggesting that the 2nd and 4th Districts

may have higher levels of calamities encountered compared to the others.

Fig. 3. Distribution of Medical Relief by Calamities

Figure 3 presents the frequency distribution of various calamities. Among the four categories, Typhoons

recorded the highest count, followed by Floods and Earthquakes, while Volcano incidents showed the lowest

occurrence. This indicates that typhoons are the most frequently experienced calamity in the NCR, Philippines,

suggesting a strong climatic influence in the observed area whereas volcanic activities are relatively rare.

www.ijltemas.in

Page 171

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

Fig. 4. Distribution of Inventory Level and Relief Aid

Figure 4 illustrates the distribution patterns of Inventory Level and Relief Aid. The histogram for inventory

levels (left) shows a right-skewed distribution, indicating that most inventory counts are relatively low

with fewer instances of high inventory accumulation. Similarly, the relief aid distribution (right) also

demonstrates a right-skewed pattern, where most of the relief aid quantities are concentrated at lower values.

This suggests that both inventory and relief aid are generally maintained at moderate levels with occasional

peaks possibly during times of high demand or major calamities.

Fig. 5. Distribution of Inventory Re-Ordered by Type of Calamity

Figure 5 illustrates the distribution of inventory re- ordered across different types of calamities. The boxplot

shows that all calamity categories exhibit a wide range of inventory re-orders with numerous outliers that

indicates an instance of exceptionally high demand. Among the calamities, typhoons appear to have slightly

higher median and upper-range values, suggesting that they tend to generate greater inventory demand compared

to other disaster types. This trend implies that typhoon-related events may exert a more substantial impact on

inventory replenishment activities and reflecting their frequency and severity in affected regions.

Fig. 6. Distribution of Inventory Re-Ordered by District

www.ijltemas.in

Page 172

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

Figure 6 presents the distribution of inventory re- ordered across the four districts. The boxplot indicates that all

districts exhibit a similar pattern of inventory demand with a concentration of lower re-order quantities and

several outliers representing high-demand instances. Among the districts, the 2nd District displays a slightly

higher median and broader range of re-orders, implying a relatively greater demand for inventory replenishment.

This suggests that the 2nd District may experience more frequent or severe calamity-related disruptions

compared to the other districts and necessitating increased inventory re-supply activities.

Fig. 7. Distribution of Reorder Occurrences

Figure 7 shows the frequency of reorder occurrences based on the Reorder_Flag variable, where 0 represents

“No” and 1 represents “Yes.” The chart reveals that items marked with No reorder have a higher count compared

to those that required reordering. However, a considerable number of items were flagged for reorder, signifying

certain supplies reached critical levels and required replenishment to meet ongoing demand. This suggests that

while inventory management is generally stable, there are instances where high consumption that creates a

pressing need for restocking to ensure continuous supply availability during crucial periods.

Modify

In this phase, a total of 114 duplicate rows were removed and reducing the dataset from 76,000 to 75,886 unique

records. To facilitate the application of classification algorithms, a binary target variable was created from the

Inventory Re- Ordered attribute. This transformation converted the quantitative reorder values into a categorical

indicator representing decision outcomes. A value of 1 denoted that an item was reordered, while 0 indicated no

reorder action. To enhance the dataset’s predictive capability, the inventory sufficiency and shortage indicators

were computed. These derived features assessed whether the current inventory levels could meet the expected

demand. The Inventory_Shortage variable represented the difference between available stock and projected

demand, where negative values indicated insufficient supply and a higher likelihood of reorder. Additionally, a

binary flag Is_Low_Inventory, was introduced to simplify classification by marking instances of low stock

conditions. To further enhance feature representation, a Demand Pressure Score was computed by combining

multiple demand-related metrics into a single standardized indicator. The score was calculated as a weighted

sum of the Simple Ratio Method (40%), Weighted Demand Index (30%), and Normalized Score (30%), resulting

in a unified measure of demand intensity for each item.

For model training and evaluation, the Reorder_Flag variable was designated as the target variable, representing

the binary classification outcome of whether an item required reordering. The dataset was then partitioned into

training (80%) and testing (20%) subsets to ensure proper model validation and performance assessment. This

split allowed the models to learn underlying patterns from the training data while maintaining an independent

dataset for unbiased evaluation of predictive accuracy and generalization capability.

Before proceeding to model training and evaluation, a data preprocessing procedure was implemented to

standardize and prepare the input variables. A ColumnTransformer was utilized to apply different

transformations to numerical and categorical features. Specifically, numerical attributes were standardized using

www.ijltemas.in

Page 173

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

the StandardScaler to normalize feature distributions and improve model convergence, while categorical

variables were passed through without alteration using the passthrough method to retain their original encoding.

This preprocessing step ensured that all input features were properly scaled and formatted, promoting

consistency and stability across the various machine learning algorithms applied in this study.

Model

For the Modeling phase, multiple machine learning algorithms were employed to develop predictive models for

medical supply reorder decisions. Both ensemble methods including Random Forest, Gradient Boosting,

HistGradientBoosting, AdaBoost, XGBoost, LightGBM, and CatBoost are implemented. Classical algorithms

such as Decision Tree, Logistic Regression, K-Nearest Neighbors (KNN), and Support Vector Machine (SVM)

were also applied. Each model was trained using the preprocessed training dataset and evaluated on the held-out

test set to determine its suitability for supporting data-driven decision- making in disaster relief operations.

The predictive performance of each machine learning model was evaluated using the F1-score and ROC-AUC

metrics. The F1-score measures the balance between precision and recall, it indicating how accurately the

models identify regions that require urgent medical supply reorders while accounting for both misses and false

alarms. The ROC- AUC metric evaluates the models’ ability to rank regions according to urgency by providing

insight into predictive discrimination independent of a specific classification threshold. To further enhance

model performance, Randomized Search with 5-fold cross-validation was applied for hyperparameter

optimization for systematically identifying the best parameter configurations for each algorithm. Together, these

metrics and optimization techniques offer a comprehensive assessment of model effectiveness and guiding the

selection of the most reliable algorithms for decision support in natural calamity relief operations.

Assess

Table I Performance Comparison of Machine Learning Algorithms Used

ROC-

AUC

#

Algorithms

F1-score

Rank

1.000000

0.999979

0.999970

0.999968

0.999913

0.999883

0.999882

0.994896

0.991523

1 Random Forest

8 CatBoost

0.999913

0.999740

0.999913

0.999827

0.999913

0.999827

0.946383

0.956899

1

9 Logistic Regression

6 XGBoost

3

4

3 Gradient Boosting

5 AdaBoost

5

6

2 Decision Tree

7 LightGBM

7

8

4 HistGradientBoosting

11 SVM

9

10

11

10 KNN

Table 1 presents the comparative performance of eleven machine learning algorithms in predicting medical

supply reorders during natural calamity relief operations. Random Forest and CatBoost achieved the highest

performance, both with an F1-score of 0.9999 and ROC-AUC of 1.000, ranking first. Logistic Regression

followed closely in third place, with slightly lower F1-score but perfect ROC-AUC. Other ensemble algorithms

such as XGBoost, Gradient Boosting, AdaBoost, Decision Tree, LightGBM, and HistGradientBoosting also

demonstrated high predictive capability, with F1-scores and ROC-AUC values near 1.000. Meanwhile, SVM

and KNN showed lower F1-scores of 0.9464 and 0.9569, despite with high ROC-AUC. Overall, the table

highlights that ensemble- based models, offer the most accurate and reliable predictions for decision-making in

medical supply allocation.

www.ijltemas.in

Page 174

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

Table III Optimized Performance of Machine Learning Algorithms Using Randomized Search CV (5-Fold)

#

8

Algorithms

CatBoost

Best F1- score

0.931610

0.928203

0.927939

0.927723

0.925507

0.919562

0.873800

0.868372

0.821085

0.814562

0.805065

Rank

1

2

6

XGBoost

3

Gradient Boosting

LightGBM

3

7

4

HistGradientBoosting

Random Forest

AdaBoost

5

1

6

5

7

2

Decision Tree

Logistic Regression

KNN

8

9

10

11

10

11

SVM

Table 2 summarizes the best F1-scores achieved by eleven machine learning algorithms after hyperparameter

optimization using Randomized Search CV with 5-fold cross-validation. CatBoost emerged as the top-

performing algorithm with an F1- score of 0.9316, followed closely by XGBoost (0.9282), Gradient Boosting

(0.9279), and LightGBM (0.9277), demonstrating the effectiveness of gradient boosting ensemble methods in

modeling reorder decisions. HistGradientBoosting and Random Forest also performed strongly by achieving F1-

scores above 0.92. Traditional and simpler models, including AdaBoost, Decision Tree, Logistic Regression,

KNN, and SVM, exhibited lower F1-scores ranging from 0.8051 to 0.8738 that indicating comparatively

reduced predictive capability after hyperparameter tuning. Overall, the results highlight that optimized

ensemble-based algorithms can provide a superior decision-making performance in predicting medical supply

reorders during natural calamity relief operations.

Table III Best Model and Its Hyperparameters

Best Parameters of CatBoost

learning_rate

l2_leaf_reg

Values

1

2

3

4

5

6

0.05

3

iterations

1000

6

depth

verbose

0

random_state

42

Table 3 shows an after hyperparameter optimization using Randomized Search CV and CatBoost was identified

as the top-performing algorithm. The optimal hyperparameters for this model included a learning rate of 0.05,

www.ijltemas.in

Page 175

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

L2 regularization (l2_leaf_reg) of 3, 1,000 iterations, and a tree depth of 6, with verbose set to 0 to suppress

training output and random_state set to 42 to ensure reproducibility. These parameters balance model complexity

and generalization by enabling CatBoost to effectively capture non-linear relationships and interactions in the

dataset while maintaining robust predictive performance in disaster relief operations.

CONCLUSION

The application of the SEMMA methodology proved highly effective in guiding the development of machine

learning models for medical supply decision-making during natural calamity relief operations. Its structured

approach ensured proper data sampling, preprocessing, feature engineering, modeling, and performance

assessment, resulting in robust and reliable predictive models. Among the algorithms tested, CatBoost emerged

as the top-performing model by achieving the highest F1-score and demonstrating superior capability in

accurately predicting reorder decisions.

For future studies, it is recommended to expand the experimentation to include additional machine learning

algorithms to explore alternative predictive approaches that may further improve decision-making accuracy.

Moreover, the use of hyperparameter optimization techniques, such as Randomized Search with cross-validation

of 5-fold, should be applied systematically to enhance model performance and ensure generalizability across

different disaster scenarios.

Ethical Approval

This study did not involve human participants or animal subjects. Therefore, ethical approval was not required

for the conduct of this study.

Conflict of Interest

The authors declare that there are no conflicts of interest regarding the publication of this research. No financial,

personal, or institutional relationships influenced the outcomes or interpretations presented in this study.

Data Availability

The data supporting the findings of this study are available to the community in the NCR, Philippines by upon

request. No proprietary or confidential data were used in this research.

ACKNOWLEDGMENT

The researchers express their sincere gratitude to La Consolacion University Philippines and the Graduate School

faculty for their guidance and support throughout this study. Appreciation is also extended to colleagues and

peers who provided valuable insights, and to our families for their encouragement during the completion of this

research.

REFERENCES

1. Rodríguez-Espíndola, O., Ahmadi, H., Gastélum-Chavira, D., Ahumada-Valenzuela, O., Chowdhury, S.,

Dey, P. K., & Albores,

2. P. (2023). Humanitarian logistics optimization models: An investigation of decision-maker involvement

and directions to promote implementation. Socio-Economic Planning Sciences, 89, 101669.

3. Bastani, P., Sadeghkhani, O., Bikine, P., Mehralian, G., Samadbeik, M., & Ravangard, R. (2023).

Medication supply chain resilience during disasters: exploration of causes, strategies, and

consequences applying Strauss and Corbin's approach to the grounded theory. Journal of Pharmaceutical

Policy and Practice, 16(1), 99.

4. Bag, S., Pretorius, J. H. C., Gupta, S., & Dwivedi, Y. K. (2021). Role of institutional pressures and

resources in the adoption of big data analytics powered artificial intelligence, sustainable manufacturing

practices and circular economy capabilities. Technological Forecasting and Social Change, 163, 120420.

5. Ma, K., Yan, H., Ye, Y., Zhou, D., & Ma, D. (2022). Critical Decision‐Making Issues in Disaster Relief

www.ijltemas.in

Page 176

INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,

MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)

ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025

Supply Management: A Review. Computational intelligence and neuroscience, 2022(1), 1105839.

6. Al Alawi, S. J. S., Jamil, J. M., & Shaharanee, I. N. M. (2022). Predicting student performance using

data mining approach: A case study in Oman. Mathematical Statistician and Engineering Applications,

71(4), 1389-1398.

7. Ma, Z., Zhang, J., & Gao, S. (2023). Research on emergency material demand based on urgency and

satisfaction under public health emergencies. PLoS One, 18(3), e0282796.

8. Florek, P., & Zagdański, A. (2023). Benchmarking state-of-the-art gradient boosting algorithms for

classification. arXiv preprint arXiv:2305.17094.

9. Ayodele, A. (2023). A comparative study of ensemble learning techniques for imbalanced classification

problems. World Journal of Advanced Research and Reviews, 19(1), 1633-1643.

10. El Guabassi, I., Bousalem, Z., Marah, R., & Qazdar, A. (2021). Comparative analysis of supervised

machine learning algorithms to build a predictive model for evaluating students’ performance.

11. Li, J. (2024). Area under the ROC Curve has the most consistent evaluation for binary classification.

PloS one, 19(12), e0316019.

12. Owusu-Adjei, M., Ben Hayfron-Acquah, J., Frimpong, T., & Abdul- Salaam, G. (2023). Imbalanced

class distribution and performance evaluation metrics: A systematic review of prediction accuracy for

determining model performance in healthcare systems. PLOS Digital Health, 2(11), e0000290

13. A Ilemobayo, J., Durodola, O., Alade, O., J Awotunde, O., T Olanrewaju, A., Falana, O., Ogungbire,

A., Osinuga, A., Ogunbiyi, D., Ifeanyi, A., E Odezuligbo, I., & E Edu, O. (2024). Hyperparameter tuning

in machine learning: A comprehensive review. Journal of Engineering Research and Reports, 26(6), 388-

395.

14. Ali, Y. A., Awwad, E. M., Al-Razgan, M., & Maarouf, A. (2023). Hyperparameter search for machine

learning algorithms for optimizing the computational complexity. Processes, 11(2), 349

www.ijltemas.in

Page 177