The Dark Matter of Clinical Data: An Identification Limit on Disentangling Physiology from Care Process in electronic health records
Organizations: Laboratory for Computational Physiology, Massachusetts Institute of Technology, Cambridge, MA 02139, USA · Boston University School of Public Health, Boston, MA, USA · Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, USA · Siriraj Informatics and Data Innovation Center, Mahidol University, Bangkok, Thailand ·
Abstract
Representations learned from electronic health record (EHR) data encode patient physiology together with the clinical process that decides what is measured, when, and for whom. We formalize the data-generating process with latent physiology, ordering intent, laboratory values that exist only when measured, a learned representation, and clinical environments, and analyze two strategies for undoing this entanglement. First, no division of the information carried by laboratory "value" and "ordering" channels is identified from a single dataset: when ordering depends on the analyte, its distribution among untested patients is not recoverable, and a sentinel code for the undefined recorded value makes it carry all the entropy of the ordering indicator, fixing partial information decomposition (PID) for every redundancy measure satisfying the Williams-Beer axioms and invariant under injective re-encoding. Second, cross-site invariance of representation-conditional outcomes is a falsifiable implication of biological invariance taken jointly with exclusion and representation-sufficiency conditions, but not an identification strategy: we construct observationally equivalent models that disagree about biological invariance, whether the invariance test fails or passes. For binary physiology this happens exactly when case mix differs, the representation is lossy, and physiology affects the outcome. A worked example yields a 9.6-percentage-point cross-site gap under an unchanged biological mechanism. Tests performed for every patient remove the ordering gate, and measured tests can restore identification: biological invariance is identified when a test exactly summarizes outcome-relevant physiology, or, for binary physiology, when two tests with independent errors are available at sites with different case mix. Cross-site analyses should be reported as tests of a stated joint hypothesis.
Figures & tables
| Assumption | Statement | Role | Checkable from observational data? |
|---|---|---|---|
| A2 Positivity | for all with | Ensures ; restricts completions in Theorem 1(i) | Partly: Theorem 1(ii)–(iii) need only , which is checkable; part (i) uses the per-site condition |
| A3 Domain restriction | undefined on | Forces an encoding choice (Theorem 1) | A matter of definition |
| A4 Biological invariance | for all with | Hypothesis under investigation | No (Theorem 3, Proposition 1) |
| A5 Case-mix heterogeneity | for some | Makes posteriors site-dependent | Implied when varies and the channel is site-invariant |
| A6a Exclusion | Auxiliary condition (Theorem 2) | No | |
| A6b Representation sufficiency | on | Auxiliary condition (Theorem 2); generically fails under A5 and A7 | No |
| Quantity | (“different patients”) | (“different care”) |
|---|---|---|
| Latent state | distinct from | |
| Case mix , | 0.30, 0.70 | 0.34, 0.66 |
| Channel , | 0.9, 0.1 (lossy) | 1, 0 (identity) |
| Mechanism at , | 0.800, 0.800 | 0.676, 0.773 |
| Mechanism at , | 0.200, 0.200 | 0.227, 0.324 |
| Assumption 4 / Assumption 6b | Holds / fails | Fails / holds |
| Analysis | Can establish | Cannot establish | Report | Basis |
|---|---|---|---|---|
| Dividing information between value and ordering in one dataset (including PID) | Nothing about separate physiological and care-process contributions | Any such division: the analyte’s distribution among the untested is not recoverable, and sentinel codes fix PID | How unmeasured values were encoded | Theorem 1, Corollary 1 |
| Cross-site test of , failed | That the conjunction of A4, A6a, and A6b fails | Which assumption failed; that biology differs | Sites, measurements, representation; effect sizes with uncertainty | Theorems 2 and 3 |
| Same test, passed | Corroboration of the conjunction | That A4 holds, even given A6 | That the joint hypothesis, not A4 alone, was tested | Proposition 1 |
| Adding ordering indicators and values | Sharper posteriors if the record pins down relevant physiology | Identification in general | Whether ordering practice differs by site | SI Corollary 2 |
| Sensitivity analysis over case mix and representation loss | Whether A4 is compatible with the observed gap for posited error rates | The error rates, except conditionally on A4 and a parametric latent model | The sensitivity model and the error rate needed to explain the gap | Proposition 2, Section 4.4 |
| Cross-site comparison using measured tests | Whether A4 holds, given an exact summary of outcome-relevant physiology, or two conditionally independent tests and differing case mix | The sufficiency or conditional-independence assumptions themselves | Which tests were used, whether they were routine, and the argument for these assumptions | Propositions 4 and 5, Corollary 3 |
Causal Identifiability
Causal identifiability asks when data and stated assumptions uniquely determine causal structures, effects, or counterfactuals.
See all 142 papers