Analisis Perbedaan Prediksi Penyakit Liver Berdasarkan Gender Menggunakan Machine Learning

Abstrak

The evaluation of machine learning (ML) models in clinical prediction typically relies on aggregate metrics, which can obscure performance disparities across subgroups. This study investigates gender-based disparities in liver disease prediction using the Indian Liver Patient Dataset. Three algorithms—Logistic Regression, Random Forest, and XGBoost—are evaluated under three experimental settings: (1) a global model with gender as a feature, (2) a global model without gender, and (3) gender-specific models trained separately for each subgroup. The results show that the best overall model is XGBoost without the gender feature, achieving an F1-score of 0.8402 and an AUC of 0.7782. However, all configurations consistently yield higher recall for male patients than for female patients, indicating a systematic bias in disease detection. Removing the gender feature does not eliminate this disparity, suggesting that bias is embedded within the distribution of clinical features. A subgroup-specific modeling approach using Random Forest substantially reduces this disparity, achieving the lowest female false negative rate (FNR) of 0.1111 and the smallest recall gap (+0.0649). These findings highlight that subgroup-based evaluation is essential for validating AI models in healthcare and demonstrate that population-aware modeling strategies can improve clinical fairness without significantly compromising overall performance.