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

This study demonstrates that deep learning-based imputation methods, specifically Denoising Autoencoders (DAE) and Self-Attention Imputation for Time Series (SAITS), significantly outperform the traditional MICE+LightGBM approach in handling missing data for heart failure mortality risk prediction using MIMIC-III EHR data.

The Gist

Imagine you are trying to bake a perfect cake (predicting if a heart failure patient will survive), but your recipe card is missing several key ingredients because the measuring cups broke or the notes got wet. In the real world, this is what happens in hospitals: patient data often has "missing pieces" because monitors glitch, sensors disconnect, or nurses forget to log a number.

This paper is like a cooking competition where three different chefs try to guess the missing ingredients so the cake can still be baked successfully.

The Setting: The Hospital Kitchen

The researchers looked at a massive digital cookbook called MIMIC-III, which contains records of over 14,000 heart failure patients in the ICU. They picked 19 important "ingredients" (like blood pressure, heart rate, and oxygen levels) that matter most for survival.

To test the chefs, they deliberately "broke" the recipe cards by hiding 20%, 30%, or even 50% of the numbers. Then, they asked three different methods to fill in the blanks:

1. The Traditional Chef (MICE+LightGBM): This is like a smart, experienced baker who uses old-school logic and statistics. They look at the ingredients they do have and say, "Well, usually when sugar is missing, flour is high, so I'll guess the sugar based on that." It's a reliable, step-by-step approach.
2. The Pattern Detective (DAE): This is a deep learning method. Imagine a detective who has seen thousands of cakes. Even if a piece of the recipe is torn, they can look at the whole picture and "clean up" the noise to reconstruct what the missing part should look like based on the overall pattern.
3. The Time-Traveling Chef (SAITS): This is another deep learning method, but it's special because it understands time. It knows that a patient's heart rate at 2:00 PM is related to what it was at 1:00 PM and what it will be at 3:00 PM. It pays attention to the sequence of events, like a conductor listening to a symphony rather than just individual notes.

The Competition: Who Guessed Best?

The researchers tested these chefs by seeing how close their guesses were to the actual numbers they had hidden. They used a scoring system where lower scores meant better guesses (less error).

• When only a few numbers were missing (20%): Both the Pattern Detective (DAE) and the Time-Traveling Chef (SAITS) did an amazing job. They were slightly better than the Traditional Chef. The Traditional Chef made more mistakes, like guessing the wrong amount of salt.
• When half the recipe was missing (50%): This is where the real magic happened. The Traditional Chef got confused and made big errors. However, the Time-Traveling Chef (SAITS) actually became the best guesser, followed closely by the Pattern Detective (DAE). Even when the data was messy and half-gone, the AI methods could "hallucinate" the missing pieces with surprising accuracy because they understood the complex relationships between the numbers.

The Takeaway

The paper concludes that AI-based methods (Deep Learning) are the new gold standard for fixing broken medical records.

Think of it this way: If you have a torn map, a traditional map-reader might just guess the missing road is a straight line. But an AI with a "super-vision" can look at the surrounding terrain, the river, and the mountains to perfectly redraw the missing road, even if half the map is gone.

Why does this matter?
For doctors using Clinical Decision Support Systems (computer tools that help them make life-or-death decisions), having a complete, accurate picture of the patient is vital. If the computer guesses the missing numbers poorly, the doctor might make the wrong call. This study proves that using these advanced AI "guessers" makes the computer's advice much more reliable, potentially saving lives by ensuring no critical data point is left out.

Technical Summary

1. Problem Statement

Electronic Health Records (EHR) are foundational for Clinical Decision Support Systems (CDSS), particularly in Intensive Care Units (ICUs) managing heart failure patients. However, EHR data is frequently compromised by missing values caused by monitoring device errors, transmission failures, or inconsistent documentation. This data incompleteness poses a significant challenge to the accuracy of mortality risk prediction models. The core problem addressed by this study is the lack of robust, comparative methodologies for imputing missing time-series data in critical care settings, specifically determining whether modern deep learning approaches outperform traditional statistical methods in this high-stakes domain.

2. Methodology

The study utilized a rigorous experimental design based on data from the MIMIC-III database:

• Dataset: The analysis focused on 14,090 ICU admissions for patients diagnosed with heart failure.
• Feature Selection: To ensure clinical relevance and reduce dimensionality, the authors selected 19 clinical features. This selection process employed a hybrid approach combining:
  • Random Forest analysis (for feature importance).
  • Correlation analysis.
  • Mutual Information.
• Experimental Design (Missingness Simulation): To objectively evaluate performance, the researchers artificially introduced missing values into the complete dataset at three distinct rates: 20%, 30%, and 50%.
• Imputation Models Compared: Three distinct methodologies were evaluated:
  1. Denoising Autoencoder (DAE): A deep learning architecture designed to learn robust representations by reconstructing input data corrupted by noise.
  2. Self-Attention Imputation for Time Series (SAITS): A deep learning model specifically leveraging self-attention mechanisms to capture temporal dependencies in time-series data.
  3. MICE + LightGBM: A traditional statistical approach using Multiple Imputation by Chained Equations (MICE) combined with the LightGBM gradient boosting algorithm for the imputation process.
• Evaluation Metrics: Performance was quantified using three standard error metrics:
  • Mean Absolute Error (MAE)
  • Root Mean Square Error (RMSE)
  • Normalized Root Mean Square Error (NRMSE)

3. Key Contributions

• Comparative Framework: The paper provides a direct, controlled comparison between state-of-the-art deep learning imputation techniques (DAE, SAITS) and a robust traditional machine learning pipeline (MICE+LightGBM) specifically within the context of heart failure ICU data.
• Scalability Analysis: By testing across varying degrees of missingness (up to 50%), the study demonstrates how different algorithms degrade or maintain performance under severe data scarcity.
• Clinical Validation: The study bridges the gap between theoretical imputation accuracy and clinical applicability, focusing on features directly relevant to heart failure mortality risk.

4. Key Results

The experimental results demonstrated a clear superiority of deep learning-based approaches over traditional methods:

• Overall Performance: Both DAE and SAITS consistently outperformed the MICE+LightGBM model across all missingness percentages (20%, 30%, and 50%).
• Low Missingness (20%):
  • DAE achieved the lowest error rates with a Mean MAE of 0.004967, RMSE of 0.005217, and NRMSE of 3.260893.
  • SAITS followed closely with a Mean MAE of 0.005461 and RMSE of 0.005797.
  • MICE+LightGBM resulted in significantly higher error counts compared to the deep learning models.
• High Missingness (50%):
  • As the data became sparser, the performance gap widened. SAITS demonstrated the best overall performance, followed by DAE.
  • The MICE+LightGBM method showed a marked decrease in performance, indicating its limitations in handling high levels of missing data in complex time-series clinical records.
• Consistency: The deep learning models maintained consistent accuracy across different clinical variables, whereas traditional methods struggled with the complexity of the data distribution.

5. Significance and Conclusion

The study concludes that deep learning-based imputation methodologies (specifically DAE and SAITS) significantly outperform traditional statistical methods for handling missing data in ICU heart failure datasets.

• Implication for CDSS: The findings strongly support the integration of deep learning imputation modules into Clinical Decision Support Systems. By providing more accurate reconstructed data, these systems can generate more reliable mortality risk predictions, ultimately aiding clinicians in making better-informed treatment decisions for heart failure patients.
• Future Direction: The success of SAITS and DAE suggests that leveraging temporal dependencies and non-linear feature interactions via deep learning is critical for modern EHR data processing, moving beyond the assumptions of linearity often found in traditional imputation techniques.