JMIR Publications

ABSTRACT

Background:

Feature normalization is a critical but often underreported component of clinical machine learning pipelines. In diabetes prediction, preprocessing choices may substantially affect model performance and reproducibility.

Objective:

This study aimed to compare normalization functions across model families for binary diabetes classification using public Hugging Face datasets and to identify model-dependent preprocessing effects.

Methods:

We conducted a benchmarking study on two public datasets (GB2024/diabetes and khoaguin/pima-indians-diabetes-database-partitions). Six normalization strategies were evaluated: none, standard, min-max, robust, quantile-normal, and power (Yeo-Johnson). Classifiers included Logistic Regression, SVC (RBF), KNN, and Random Forest. We used a stratified train/test split and stratified 5-fold cross-validation on the training set. Primary metrics were macro-F1 and AUC, with accuracy reported as a secondary metric.

Results:

Normalization effects were clearly model- and dataset-dependent. On GB2024/diabetes, tree-based performance was comparatively stable across normalization choices, while distance- and margin-based models were more sensitive to scaling. On the Pima partitions dataset, min-max scaling improved performance for several models. The best-performing configuration in the benchmark achieved test macro-F1 close to 0.90 on the larger cohort-level dataset-model setting, while no single normalization method was universally optimal across all model families.

Conclusions:

Normalization should be treated as an explicit model-selection factor rather than a fixed preprocessing default in clinical AI workflows. Reporting standards for diabetes ML studies should include transparent normalization rationale and comparative preprocessing evaluation.