Adaptive Selection of MICE Algorithm Parameters: A Case Study on Pulmonary Function Value Prediction Models

Technical Summary: Adaptive Selection of MICE Algorithm Parameters for Pulmonary Function Prediction

Problem Statement
Artificial intelligence-based medical prediction models often suffer from reduced reliability and predictive performance due to incomplete datasets containing missing values. In the context of Chronic Obstructive Pulmonary Disease (COPD), pulmonary function test (PFT) data presents specific challenges: measurements are taken at irregular intervals, variables exhibit non-linear relationships over time, and missing values arise from selective testing or discontinued follow-ups. While Multiple Imputation by Chained Equations (MICE) is a standard imputation method in medical research, its default parameters (initialization strategy and iteration count) are often applied uniformly. This study addresses the lack of a tailored approach for setting MICE hyperparameters based on the specific statistical characteristics of pulmonary function indicators (FEV1, FVC, FEF25–75%, and PEF).

Methodology
The study utilized a retrospective longitudinal dataset of 598 subjects collected over two years from Chungnam National University Hospital. The research followed a three-stage process:

1. Statistical Characterization of Variables: The authors analyzed the statistical properties of key PFT indicators to inform imputation strategies:

   • FEV1 and FVC: Exhibit a strong linear relationship and low variability.
   • FEF25–75%: Characterized by high inter-individual variability and large residual variance.
   • PEF: Displays periodic, time-series variability dependent on time of day and patient effort.
   • Multivariate Structure: The indicators share a covariance structure, justifying multivariate imputation over univariate approaches.

2. Adaptive MICE Parameter Selection: Based on the statistical properties, an adaptive selection method was proposed for MICE hyperparameters:

   • Initialization Strategy: Mean-based initialization was applied to low-variability, normally distributed variables (FEV1, FVC), while median-based initialization was applied to high-variability or skewed variables (FEF25–75%, PEF) to reduce outlier sensitivity.
   • Iteration Count: The maximum iteration count was adjusted based on variable-specific variability. A baseline of 14 iterations was set for FEV1, FVC, and PEF, while FEF25–75% was assigned 20 iterations to account for its higher coefficient of variation and slower convergence.

3. Experimental Validation:

   • Imputation Comparison: The proposed adaptive MICE was compared against Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) imputation.
   • Missing Rates: Artificial missing rates of 5%, 10%, and 20% were introduced to test robustness.
   • Prediction Model: An LSTM-based regression model was used to predict future pulmonary function values (T+1) based on the two preceding time points (T-1, T).
   • Evaluation: Performance was measured using MAPE, RMSE, $R^2$, and MAE via 5-fold cross-validation.

Key Contributions

• Statistical Justification for Parameter Tuning: The paper provides theoretical grounds for initializing MICE and setting iteration counts based on the specific variance and distribution characteristics of pulmonary function indicators, rather than using uniform defaults.
• Multivariate Interpretation: It treats PFT data as multivariate time-series with explicit correlation structures (e.g., the linear dependence between FEV1 and FVC) rather than independent univariate series, validating the use of MICE over univariate interpolation.
• Adaptive Configuration: The study proposes a specific configuration (Mean/14 for FEV1/FVC; Median/20 for FEF25–75%; Median/14 for PEF) that balances imputation accuracy with computational efficiency.

Results

• Robustness to Missingness: The adaptive MICE strategy maintained stable predictive performance as missing rates increased from 5% to 20%. At a 20% missing rate, the adaptive MICE approach achieved a Mean Absolute Percentage Error (MAPE) of 5.831%, demonstrating that multivariate correlation-based imputation can provide valid input information even in high-missingness environments.
• Comparison with PCHIP: MICE and PCHIP yielded similar performance levels across all missing rates. However, MICE's ability to reflect structural relationships among variables offered a distinct advantage in preserving multivariate dependency structures compared to PCHIP's focus on time-series continuity.
• Parameter Sensitivity: The optimal combination of initialization method and iteration count varied slightly depending on the missing rate (e.g., the "adaptive" method performed best at 5% and 10%, while "mean-only" performed slightly better at 20% in specific configurations), suggesting that no single fixed setting is universally superior, but adaptive strategies remain robust.

Significance and Claims
The authors claim that their study demonstrates the feasibility of using an adaptive MICE parameter selection strategy to mitigate missing data problems in pulmonary function time-series analysis. By reflecting the specific statistical characteristics of indicators like FEF25–75% (high variability) and the linear correlation between FEV1 and FVC, the proposed method provides reliable predictive information without degrading model performance as data completeness decreases. The study concludes that MICE, when properly tuned, is a competitive and appropriate imputation strategy for long-term follow-up pulmonary function data, capable of utilizing structural relationships between variables to support clinical prediction models.

Limitations
The authors acknowledge that the study is based on a retrospective dataset from a single institution, which may limit generalizability. Missing values were artificially introduced at fixed rates, which may not fully replicate real-world "missing-not-at-random" mechanisms. Furthermore, the results are specific to the LSTM architecture used and were not externally validated on independent multi-institutional datasets.