Comparing Missing Data Imputation Methods for Patient-Reported Outcomes in Esophageal Cancer Research

This study evaluates and compares various statistical and machine learning imputation methods for handling missing data in esophageal cancer patient-reported outcomes to provide evidence-based recommendations for improving research validity.

The Gist

The "Missing Puzzle Piece" Problem: A Simple Guide to the Research

Imagine you are trying to put together a massive, 1,000-piece jigsaw puzzle of a beautiful landscape. But when you open the box, you realize some pieces are missing. Some are gone from the factory, some were lost under the couch, and some were never even in the box to begin with.

You can still try to guess what the picture looks like, but your final image might look blurry, weird, or just plain wrong.

This is exactly what happens in medical research.

When doctors study how cancer patients are feeling (their "Quality of Life"), they use surveys. But patients are human—sometimes they are too tired to finish a survey, sometimes they skip sensitive questions (like questions about their sex life), or sometimes the clinic is just too busy. These "missing pieces" of data can lead to wrong conclusions about how well a treatment is actually working.

The Goal of the Study

A team of researchers wanted to find the best "Master Artist" to step in and paint those missing puzzle pieces so the final picture looks as real as possible. They tested seven different "Master Artists"—some were old-school mathematicians (Traditional Methods) and some were high-tech robots (Machine Learning).

The "Artists" (The Methods)

To make sense of the different methods, let’s imagine them as different types of restorers:

1. MICE (The Experienced Detective): This artist looks at every other piece in the puzzle to make an educated guess. "If the sky is blue here and blue there, this missing piece is probably blue too."
2. KNN (The Neighbor Watch): This artist looks for the most similar completed puzzles and copies what they did. "This patient is a lot like Patient B, so I’ll assume they feel like Patient B."
3. SoftImpute (The Minimalist): This artist looks for the simplest, smoothest patterns to fill the gaps quickly.
4. Deep Learning/Autoencoders (The Sci-Fi Robots): These are super-complex AI systems. They try to learn the "soul" of the data to recreate it. Some are very smart, but some are like robots that try too hard and end up painting something that looks nothing like the original picture.

What They Found (The Results)

After putting these artists to the test, here is the "report card":

• The Winner: MICE (The Detective). Even though it was the slowest worker (it took a long time to think!), it was the most accurate. It didn't just fill the holes; it made sure the colors and shapes matched the rest of the picture perfectly. It was the best at predicting how a patient would actually be classified clinically.
• The Speedster: SoftImpute. If you had a mountain of data and needed it done now, this was your best bet. It wasn't quite as perfect as the Detective, but it was incredibly fast and "good enough" for most tasks.
• The "Try-Hard" Fail: The Specialized Deep Learning Model. This was like a high-tech robot that got "overconfident." It tried so hard to find patterns in individual patients that it actually started inventing fake information. It created "glitches" in the data that didn't exist in real life.

Why Does This Matter?

If we use a bad "artist" to fill in missing data, we might tell a doctor, "This treatment is working great!" when, in reality, the patients are actually struggling.

By identifying that MICE is the most reliable "Detective" for this specific type of cancer research, the scientists are giving doctors a better toolkit. This ensures that when we look at the "big picture" of cancer care, we are seeing the truth, not just a blurry guess.

---

The Bottom Line: When data goes missing in cancer research, don't just guess. Use a method that respects the complexity of human life. For now, the "Old-School Detective" (MICE) is still the gold standard.

Technical Summary

Technical Summary: Comparing Missing Data Imputation Methods for Patient-Reported Outcomes in Esophageal Cancer Research

1. Problem Statement

In oncology research, particularly within real-world clinical registries, Patient-Reported Outcomes (PROs)—such as quality-of-life surveys—frequently suffer from missing data. This missingness can stem from technical errors (Missing Completely At Random - MCAR) or patient incapacity/reluctance (Missing Not At Random - MNAR), leading to reduced statistical power, systematic bias, and compromised clinical validity.

Existing literature shows a problematic reliance on single imputation methods or simple complete-case analyses, which fail to account for the complex, non-linear, and ordinal nature of PRO data (e.g., Likert scales with ceiling/floor effects). There is a lack of systematic comparison between traditional statistical methods and modern machine learning (ML) architectures specifically tailored for these high-dimensional, sparse, and ordinal clinical datasets.

2. Methodology

The researchers evaluated seven imputation methods across diverse paradigms using data from the Esophageal and Gastric Data- and Bio-Bank (EGDB Bank). The study focused on 44 variables from the FACT-E (Functional Assessment of Cancer Therapy–Esophageal) instrument at the baseline time point.

Imputation Methods Evaluated:

• Traditional/Baseline: Multiple Imputation by Chained Equations (MICE) using LightGBM; K-Nearest Neighbors (KNN).
• Matrix Factorization: SoftImpute (iterative soft-thresholded SVD) and Bayesian Principal Component Analysis (BPCA).
• Deep Learning (Generative/Reconstructive): Variational Autoencoder (VAE), Denoising Autoencoder (DAE), and a specialized Deep Autoencoder (incorporating patient-specific embeddings and temporal pattern modeling).

Comprehensive Evaluation Framework:
To ensure a holistic assessment, the authors used a multi-dimensional validation strategy:

• Direct Dataset Assessment: Execution time, distribution preservation (Kolmogorov-Smirnov tests), and correlation maintenance.
• Sparse Cross-Validation: Artificial masking of data to calculate Mean Absolute Error (MAE) and Root Mean Square Error (RMSE).
• Clinical Classification Performance: Converting imputed continuous values back to discrete ordinal categories to measure Accuracy, AUC, Sensitivity, Specificity, PPV, and NPV.
• Stability & Bias Testing: Bootstrap resampling (1,000 iterations) and complete-case simulations to assess robustness and systematic bias.

3. Key Contributions

• First Comprehensive Benchmark: The study provides the first large-scale comparison of statistical, matrix factorization, and deep learning paradigms specifically for PRO data in a real-world cancer registry.
• Multi-Metric Validation: Unlike studies that focus solely on error metrics (MAE/RMSE), this work includes clinical utility (classification performance) and structural integrity (correlation and distribution preservation).
• Practical Decision Framework: The paper moves beyond "black-box" testing to provide evidence-based recommendations based on specific research priorities (e.g., accuracy vs. speed).

4. Results

• Superiority of MICE: MICE emerged as the best overall performer. It achieved the highest classification accuracy (0.512), the best multiclass AUC (0.635), and superior distribution preservation. It was highly effective at capturing the non-linear relationships inherent in the FACT-E domains.
• Efficiency vs. Accuracy Trade-offs:
  • SoftImpute was the most computationally efficient (0.04s) and offered a strong balance of accuracy and speed, making it suitable for large-scale routine use.
  • MICE was the slowest (453.46s) due to its iterative ensemble nature.
• Deep Learning Performance: Results were mixed. While DAE and BPCA showed moderate competence, the specialized Deep Autoencoder (Da Xu et al.) performed poorly, exhibiting severe systematic bias, poor distribution preservation, and "catastrophic" correlation distortions. This was attributed to overfitting on patient-specific embeddings and the inappropriate application of temporal loss to cross-sectional data.
• Correlation Robustness: While most methods preserved correlations well, MICE showed sensitivity to extreme sparsity, whereas VAE and KNN remained more robust in maintaining the underlying data structure under high missingness.

5. Significance and Recommendations

The study highlights that while machine learning is promising, it is not a "silver bullet" for PRO data; complex architectures can actually introduce more bias if not carefully designed for the specific data constraints (sparsity and ordinality).

The authors provide a tiered recommendation for clinical researchers:

1. For High-Stakes Accuracy: Use MICE (prioritizing symptom severity categorization).
2. For Large-Scale/Routine Clinical Use: Use SoftImpute (balancing efficiency and accuracy).
3. For Uncertainty/Correlation Focus: Use Bayesian PCA.

Ultimately, the research underscores the necessity of considering MNAR (Missing Not At Random) patterns in clinical practice and encourages researchers to perform sensitivity analyses when using imputed data to ensure the validity of oncology research findings.