Adaptive MICE Improves Reproducibility, Not Universality, in Pulmonary Function Forecasting
Incomplete longitudinal data remain one of the most persistent problems in predictive medicine, especially when models are built to forecast pulmonary function over time. In this study, investigators evaluated an interpretable initialization rule within a random-forest chained imputation workflow designed for prediction models that rely on repeated pulmonary function test values. The work centered on whether an adaptive choice of imputation parameters could outperform more familiar strategies when both masked-cell accuracy and downstream forecasting performance were tested under controlled conditions.
The dataset was substantial for this kind of modeling exercise: 596 patients contributed 2,112 visits, and a separate controlled cohort of 367 patients supplied 396 eligible prediction windows for evaluation. Patient-disjoint five-fold splits were used, which matters because it limits leakage across training and test partitions and makes the reported performance more credible for real-world deployment. Missingness was introduced under two mechanisms, MCAR and observed-covariate-dependent MAR, at 5%, 10%, and 20%, then each imputer and scaler was fitted only on training rows. In other words, the workflow was built to mirror the constraints of clinical machine learning rather than a tidy laboratory simulation.
The adaptive configuration used fixed variable-level initialization assignments. A fold-wise sensitivity audit then recalculated the rule statistics and cross-validation threshold using training patients only, and all five folds retained the same assignments. The practical consequence is straightforward: the imputed inputs and downstream results did not change across folds, which supports reproducibility and makes the approach easier to audit in future studies or clinical pipelines.
A reproducible rule is useful, but reproducibility alone does not guarantee superiority.
Masked-Cell Accuracy Favored Different Imputers Under Different Conditions
Twelve imputation approaches were compared using direct masked-cell accuracy and downstream predictive utility, with a fixed latest LSTM model handling the forecasting step. Across 360 latest-LSTM runs, the best-performing method for root mean squared error depended on the masking condition rather than on a single universal winner. Mean-initialized MICE ranked first in four conditions, adaptive MICE in one, and LOCF in one.
Those results cut against the assumption that a carefully engineered adaptive rule should dominate across all missingness scenarios. Instead, the performance pattern suggests that the best imputation strategy depends on how values go missing, how much data are absent, and which downstream model consumes the reconstructed inputs. In a pulmonary function forecasting setting, those details are not abstract. Serial lung function values are often used to anticipate decline, monitor disease trajectory, and inform follow-up intensity, so imputing them well matters only if the imputation preserves the information the prediction model actually needs.
The strongest signal in the analysis was not a universal winner but a mismatch between direct imputation quality and end-task performance. Methods can look strong when judged by reconstruction error alone, yet still fail to improve the later prediction model. Conversely, a method that is only modestly accurate at the cell level may still support better forecasting if it preserves the temporal structure the LSTM is learning from. The study design separated those two questions clearly, and that separation exposed the limits of treating imputation accuracy as a proxy for clinical prediction value.
Downstream Forecasting Penalized the Adaptive Rule After Correction
None of the 66 paired comparisons against the variable-specific adaptive configuration remained significant after Holm correction. That result is the central clinical computing message in the paper: the adaptive procedure was interpretable and reproducible, but it was not consistently better than the alternatives once multiplicity was controlled.
For model developers, this is a reminder that elegant logic does not automatically convert into superior prediction. A variable-specific initialization scheme may be attractive because it appears tailored to the data structure, yet the downstream outcome still depends on the interaction between missingness pattern, imputation variance, and the forecasting architecture. In a chained random-forest imputation framework, small shifts in initialization can ripple through subsequent iterations, but those ripples did not produce a durable advantage here. The fact that the fold-wise sensitivity audit preserved the assignments strengthens confidence in the stability of the method, even as the comparative results weaken claims of broad performance superiority.
The study also reinforces a useful separation for clinical informatics teams. Imputation fidelity and predictive utility should not be collapsed into a single metric, especially when the endpoint is not a static label but a longitudinal physiologic measurement. For pulmonary function forecasting, a method can be statistically tidy, reproducible, and interpretable while still failing to improve the actual prediction task across masking conditions.
What the Results Mean for Longitudinal Clinical Modeling
In clinical AI, missing data are often treated as a nuisance layer beneath the “real” modeling problem. This analysis shows why that view is too simple. When pulmonary function trajectories are incomplete, the imputer becomes part of the predictive engine, and its behavior can alter the quality of the final forecast just as much as the choice of recurrent network or training split. The safest conclusion from this work is not that adaptive MICE should be discarded, but that its value must be judged against both reconstruction accuracy and downstream model behavior in the same experiment.
The cohort size, the patient-disjoint evaluation design, and the explicit testing under MCAR and MAR all give the findings practical weight. Even so, the comparison did not yield a universal best method, and the adaptive configuration did not secure an across-the-board advantage after multiple-comparison correction. For groups building pulmonary prediction tools, the next step is to treat imputation as a model component with its own validation burden rather than as a preprocessing footnote.
The unresolved question is not whether imputation matters. It clearly does. The harder issue is which missing-data strategy remains robust when the pattern of missingness changes and the downstream predictor is asked to make a clinically useful forecast from imperfect longitudinal records.