A Comparative Study of Feature Selection in Machine Learning for Survey-Based Lung Cancer Classification Utilizing SHAP

Lung cancer is a fatal malignancy requiring early detection. While machine learning aids in diagnosis using survey data, existing studies frequently suffer from extreme class imbalance and methodological flaws, specifically data leakage during feature selection, leading to optimistically biased performance. This study aims to determine the most robust feature selection and classification methodology for lung cancer prediction while strictly eliminating information leakage and addressing class imbalance. Duplicate entries were removed from a 309-record survey dataset, resulting in 276 unique patients (86.23% positive, 13.77% negative). An unbiased evaluation framework was established using Nested Stratified 10-Fold Cross-Validation. Within this architecture, nine classifiers were optimized across five dynamic feature selection techniques (ANOVA + Chi-Square, Forward Selection, Backward Elimination, Mutual Information Gain, and Recursive Feature Elimination) strictly within the inner training loops. The K-Nearest Neighbors algorithm paired with Recursive Feature Elimination (KNN-RFE) achieved the highest overall predictive accuracy (91.28% ± 4.91%) and F1-score (95.02% ± 2.81%). However, Logistic Regression optimized with Backward Elimination (LR-BE) demonstrated superior stability in the presence of skewed data, achieving the highest balanced accuracy (89.95% ± 8.42%) and precision (98.25% ± 2.85%). Additionally, SHAP (SHapley Additive exPlanations) analysis was integrated, transparently identifying 'Swallowing Difficulty' and 'Coughing' as the most critical predictive biomarkers. In practice, KNN-RFE is recommended for maximizing overall diagnostic discrimination, whereas LR-BE is strictly preferred for handling severe class imbalances to minimize false positives. The rigorous nested validation ensures these metrics reflect realistic clinical generalization rather than algorithmic overfitting. Furthermore, the SHAP integration provides necessary transparency, translating complex predictive pathways into interpretable, evidence-based screening insights for healthcare professionals.

Authors

Institutions

Publication Details

Journal
Indonesian Journal of Electronics Electromedical Engineering and Medical Informatics
Published
2026-11-19
DOI
https://doi.org/10.35882/ijeeemi.v8i4.367
Primary Topic
Artificial Intelligence in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Comparative Study of Feature Selection in Machine Learning for Survey-Based Lung Cancer Classification Utilizing SHAP

Rafie Rafie, Annisa Rizqiana, Irwan Budiman, Nanda Putri Rahmawati
Indonesian Journal of Electronics Electromedical Engineering and Medical Informatics
Artificial Intelligence in Healthcare
article

A Comparative Study of Feature Selection in Machine Learning for Survey-Based Lung Cancer Classification Utilizing SHAP

Rafie Rafie, Annisa Rizqiana, Irwan Budiman, Nanda Putri Rahmawati
article en

Abstract

Lung cancer is a fatal malignancy requiring early detection. While machine learning aids in diagnosis using survey data, existing studies frequently suffer from extreme class imbalance and methodological flaws, specifically data leakage during feature selection, leading to optimistically biased performance. This study aims to determine the most robust feature selection and classification methodology for lung cancer prediction while strictly eliminating information leakage and addressing class imbalance. Duplicate entries were removed from a 309-record survey dataset, resulting in 276 unique patients (86.23% positive, 13.77% negative). An unbiased evaluation framework was established using Nested Stratified 10-Fold Cross-Validation. Within this architecture, nine classifiers were optimized across five dynamic feature selection techniques (ANOVA + Chi-Square, Forward Selection, Backward Elimination, Mutual Information Gain, and Recursive Feature Elimination) strictly within the inner training loops. The K-Nearest Neighbors algorithm paired with Recursive Feature Elimination (KNN-RFE) achieved the highest overall predictive accuracy (91.28% ± 4.91%) and F1-score (95.02% ± 2.81%). However, Logistic Regression optimized with Backward Elimination (LR-BE) demonstrated superior stability in the presence of skewed data, achieving the highest balanced accuracy (89.95% ± 8.42%) and precision (98.25% ± 2.85%). Additionally, SHAP (SHapley Additive exPlanations) analysis was integrated, transparently identifying 'Swallowing Difficulty' and 'Coughing' as the most critical predictive biomarkers. In practice, KNN-RFE is recommended for maximizing overall diagnostic discrimination, whereas LR-BE is strictly preferred for handling severe class imbalances to minimize false positives. The rigorous nested validation ensures these metrics reflect realistic clinical generalization rather than algorithmic overfitting. Furthermore, the SHAP integration provides necessary transparency, translating complex predictive pathways into interpretable, evidence-based screening insights for healthcare professionals.

Indonesian Journal of Electronics Electromedical Engineering and Medical InformaticsVol. 8(4)
Lambung Mangkurat University (ID)
Artificial Intelligence in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon