Deep Learning-Based Missing Value Imputation for Heart Failure Data from MIMIC-III: A Comparative Study of DAE, SAITS, and MICE+LightGBM

Background: Inadequate data in electronic health records can create problems for clinical decision support systems and predictive modelling tools. ICU patients with heart failure may have incomplete physiological data as a result of irregular patient monitoring, sensor failure, or clinical workflow issues, which highlights the need for strong imputation methods. Objective: The goal of this research is to examine the performance of three different methods of imputing the missing value rows in heart failure patients records (from the MIMIC-III database)the Denoising Autoencoder (DAE), Self-Attention Imputation for Time Series (SAITS) and Multiple Imputation by Chained Equations (MICE) with LightGBM as compared against each other. Methods: We created our own heart failure cohort with 14,090 ICU stays from the MIMIC-III database. After performing a strict 10-step preprocessing pipeline, we utilized Random Forest, correlation analysis, and mutual information, in order to select the final 21 clinically relevant features from the original 81 features. To mimic realistic clinical scenarios, we introduced artificial missing values in three different amounts; 20%, 30%, and 50%. The final three methods used to impute these missing values were DAE (deep learning autoencoder), SAITS (Transformer-based attention mechanism), and MICE+LightGBM (hybrid statistical and machine-learning method). We evaluated the performance of our models using the following three metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Normalised Root Mean Square Error (NRMSE). Results: DAE and SAITS demonstrated superior and comparable performance across all levels of missingness. At 20% missingness, DAE achieved a mean MAE of 0.004967, a mean RMSE of 0.005217 and a global NRMSE of 3.260893; SAITS achieved a mean MAE of 0.005461, a mean RMSE of 0.005797 and a global NRMSE of 3.244695, while MICE+LightGBM showed substantially higher error (mean MAE 0.106452, mean RMSE 0.223782 and global NRMSE 0.222056). At 50% missingness, SAITS exhibited the best performance (mean MAE 0.004538, mean RMSE 0.004954, global NRMSE=2.618904), followed by DAE (mean MAE 0.004992, mean RMSE 0.005307, global NRMSE 3.27805), with MICE+LightGBM showing degraded performance (mean MAE 0.19729, mean RMSE 0.3767273, global NRMSE 0.377731). Feature-level analysis revealed that deep learning methods maintained consistent accuracy across diverse physiological variables, including lactate, pH, bicarbonate, and blood urea nitrogen. Conclusions: Deep learning-based imputation methods (DAE and SAITS) significantly outperform traditional hybrid approaches for missing-value imputation in heart failure ICU data, demonstrating robustness across varying levels of missingness and feature types. These findings support the adoption of DL-based imputation strategies in clinical decision support systems for heart failure management.