Learned imputation finally beats linear baseline on clinical time series

The paper's core contribution is identifying that realistic clinical missingness isn't random but clustered around physiological extremes, and that this pattern itself is learnable rather than something to be handled post-hoc. Most prior work treated missingness as a generic problem; CAIR makes it domain-specific.

This connects directly to the broader pattern visible in recent clinical ML work. The Time-Aware Transformer paper from August tackled deployment constraints by designing around operational realities (lab latency in acute respiratory crises). CAIR follows the same logic: rather than expecting a generic learner to discover why it fails on real data, the researchers baked clinical knowledge into the architecture itself. Both papers reject the assumption that more capacity or better optimization solves domain mismatch. The PerturbRx work on treatment-conditioned transitions also shares this mechanistic framing, where understanding causal structure (treatment effects, missingness patterns) becomes part of the learned model rather than noise to average over.

If CAIR's two-stage approach outperforms simpler baselines on held-out clinical datasets from institutions outside the training set (especially those with different EHR systems and missingness distributions), that confirms the domain knowledge transferred. If performance degrades significantly on out-of-distribution missingness patterns, the curriculum awareness was overfit to the training distribution.