TY  - JOUR
AU  - Singh, Artika 
AU  - Jailia, Manisha 
PY  - 2026
TI  - Pandemic Analysis and Preparedness Using a Three Layer Machine Learning Approach
JF  - Journal of Computer Science
VL  - 22
IS  - 11
DO  - 10.3844/jcssp.2026.3361.3375
UR  - https://thescipub.com/abstract/jcssp.2026.3361.3375
AB  - Outbreak analysis provides research-based information and data to enable scientific decision-making to manage an effective response to a pandemic. Pandemic analysis has its own set of challenges, such as limited data, the need for real time analysis, and the complexity of outbreak patterns. This study proposes a hybrid approach to manage such a spread through a 3-layer machine learning-based outbreak prediction model. The proposed model incorporates pre-processing of data, data standardisation, feature selection, outbreak prediction, and a tool for effective pandemic management. The pre-processing step of the proposed model is to collect the pandemic data, remove irregularities, clean the data, fill in missing values, and format the data in a consistent manner for data analysis. The pre-processing can be accomplished through statistical procedures (e.g., central tendency and statistical dispersion) for data standardisation. A feature selection step, where the most relevant features are tested and selected for outbreak prediction, uses a bioinspired hybrid optimisation algorithm, the Spotted Hyena Optimiser (SHO) and Seagull Optimisation Algorithm (SOA), to minimise the likelihood of overfitting. The final step in the outbreak prediction model incorporates a 3-layer machine learning approach comprised of an O-SVM (Optimised Support Vector Machine) initial layer for boundary detection and classification, a KNN (k-Nearest Neighbors) second layer to perform similarity matching and, lastly, XGBoost as a third layer for prediction of spread severity. The predicted outcome is compared with standard machine learning algorithms, in which the proposed model provides better results in terms of the highest accuracy (0.93), fitness (0.96), and the lowest error (6.9%).