Related Experiment Video
Updated: Apr 15, 2026
12:18
A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment
Published on: January 11, 2020
8.2K
Enhancing Model Generalizability in Medical Artificial Intelligence: Systematic Comparison of Categorical Encoding
Chien-Wei Chuang1,2, Chung-Kuan Wu3,4,5,6, Chao-Hsin Wu1,2
1Graduate Institute of Business Administration, Fu Jen Catholic University, 510 Zhongzheng Rd, Xinzhuang District, New Taipei City, 242062, Taiwan, 886 2905-3895.
JMIR Medical Informatics
|April 13, 2026
Summary
Preprocessing choices significantly impact machine learning model performance in clinical research. Combining one-hot encoding with Random Over Sampling Example (ROSE) improved model accuracy and stability across different datasets.
Area of Science:
- Clinical research
- Machine learning
- Data preprocessing
Background:
- Clinical datasets often have missing values, complex categorical variables, and imbalanced classes.
- These data issues complicate machine learning model development and interpretation.
Purpose of the Study:
- Introduce a structured framework for preprocessing medical tabular data.
- Assess how preprocessing choices influence predictive model stability and portability.
Main Methods:
- Developed a modular workflow including imputation, categorical encoding (one-hot, frequency, target), and resampling (SMOTE, ROSE).
- Evaluated 6 classification algorithms (LGR, DT, Random Forest, XGBoost, CatBoost, LightGBM).
- Assessed cross-dataset portability using end-stage renal disease (ESRD) and Behavioral Risk Factor Surveillance System (BRFSS) datasets.
Main Results:
- One-hot encoding with ROSE consistently improved Area Under the ROC Curve (AUC) to 0.940 and accuracy to 0.932.
- ROSE enhanced sensitivity without significantly altering data structure.
- Performance trends were reproducible across datasets, indicating model portability.
Conclusions:
- Data preprocessing is crucial for machine learning model outcomes in clinical settings.
- The proposed framework provides a reproducible tool for health care prediction data preparation.
- This work can inform standardization efforts in clinical machine learning preprocessing.
