Clinical prediction from electronic health records (EHRs) is complicated by pervasive missingness and label scarcity, which make performance sensitive to the match between data conditions and pipeline choice. Choosing the best pipeline for a new incomplete dataset still requires costly trial-and-error. We cast this as an algorithm selection problem and address two bottlenecks—instance scarcity and distance quality—that have so far prevented meta-learning from reaching clinical settings. Graph neural networks offer diverse strategies (patient similarity networks, bipartite imputation graphs, attention-driven feature interaction), yet no single architecture dominates across missingness patterns, and selecting the best pipeline for a new dataset remains a trial-and-error approach. Formal algorithm selection could automate this choice but requires many characterized meta-instances—more than clinical settings typically provide. We propose two solutions: (1) constructive instance augmentation, applying controlled quality perturbations (MCAR and MNAR missingness injection, label trimming) to 20 base EHR datasets to expand the meta-knowledge base to 83 characterized meta-instances, each described by a 10-dimensional missingness fingerprint, without additional model training; and (2) dynamic-supervised metric learning, using differential evolution to optimize fingerprint feature weights so that static distances preserve method-performance similarity captured by dynamic fingerprints, which require model sweeps and are unavailable at deployment. Under base-dataset-level leave-one-dataset-out cross-validation over 21 pipelines, the resulting metric-learned kNN recommender attains the highest win rate (20.5%) among non-oracle strategies on the augmented store, selecting the correct pipeline more often than any fixed default. At deployment, the recommender needs only the 10-dimensional static fingerprint with pre-learned weights; no sweep data is required for new datasets. Cross-domain evaluation on 25 external subsets (colorectal cancer, kidney disease, MIMIC-IV) demonstrates framework modularity: when the fingerprint module is adapted (standard meta-features in place of the missingness-specific set), the recommender achieves regret of 0.025 (55% below random selection).
Li et al. (Fri,) studied this question.