Missing Data Imputation in Healthcare Datasets Using GAIN and KNN

Missing data in health-related datasets poses a significant challenge to data analysis processes and adversely affects the accuracy of decision support systems. In this study, two imputation methods— the statistically grounded K-Nearest Neighbors (KNN) algorithm and a deep learning-based approach known as Generative Adversarial Imputation Nets (GAIN)—are examined and compared. While KNN offers a simple and interpretable non-parametric solution, GAIN employs a complex artificial neural network architecture composed of generator and discriminator networks to estimate missing values. These two methods, differing in structural complexity, were evaluated on 14 publicly available health datasets characterized by small sample sizes and limited dimensionality. The missing rate was set at 10%, and missing values were simulated under the Missing Completely at Random (MCAR) mechanism. Performance was assessed using Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and computational time as evaluation metrics. The results indicate that the KNN method yields more consistent and accurate results on datasets with small sample sizes and low dimensionality. On the other hand, GAIN demonstrated promising potential, particularly in handling larger and more complex datasets.