Exploring Hyper-Parameters and Feature Selection for Predicting Non-Communicable Chronic Disease Using Stacking Classifier

Pooja Yadav

doi:10.1109/ACCESS.2023.3299332

2023-01-01

Pooja Yadav

S. C. Sharma *

Rajesh Mahadeva *

Shashikant P. Patole *

DOI：10.1109/ACCESS.2023.3299332

原文链接

原文求助

摘要

En 中文

Non-communicable disease, especially chronic disease, is the most common factor of complication of deteriorating physical health and the state of one's mind. It is also a prominent cause of illness and mortality around the world. Primarily chronic disease is preventable at a particular stage though its occurrence is critical. To make clinical decisions, these illness prediction models were created to assist clinicians and patients. A chronic disease prediction model takes into account many risk variables to determine an individual's illness risk. Machine learning approaches have made it possible to predict chronic disease early by collecting Electronic Health Record (EHR) data. This paper focuses on the diabetes dataset extracted from Kaggle and two unseen real datasets. In this paper, we have implemented Synthetic Minority Over-Sampling Technique (SMOTE) algorithm to balance the dataset. We have also explored Boruta as the feature selection method. To tune hyper-parameters of different algorithms, we have proposed an improved technique by combining the Grid Search method with the Grey Wolf Optimization algorithm. The Grid Search method requires extensive searching, while the Grey Wolf Optimization algorithm is easily linked, rapid to seek, and extremely exact. Nine conventional classification techniques have been evaluated in this paper. This research concentrates on the Stacking Classifier to assess the performance of the prediction model that produces the best results. The Proposed Model gave the highest F1-Score 98.84% on PIMA dataset, 98% after validation on the Synthetic dataset, 97.3% on ADRC dataset, 96.20% on FHD dataset. To the best of our knowledge, no previous work has focused on such a sort of technique and these two datasets. The outcomes of the comparison experiment on the PIMA dataset reveals that the proposed strategy performs better. This study also provides the interpretation of the proposed model. It conducts an ethical assessment of what explainability means for the use of Machine Learning models in clinical practice.

Keyword:

Chronic disease

feature selection

hyperparameter tuning

machine learning

non-communicable diseases

prediction model

stacking classifier

interpretability