When Patients Hide: Cross-Modal Disclosure Gaps in Long-Horizon EHR Prediction

People with stigmatized conditions such as substance use, self-harm, or depression often tell their doctors less than their bodies, their families, or their lab results reveal. Hospital records inherit that silence. A diagnosis code is missing because the patient denied the problem, not because the problem was absent. Years later, the same patient may stop returning to the health system altogether, and the record then shows "no further events" when the truth is avoidance. Machine learning models trained on these historical records treat both silences as good news. The group with the most at stake is therefore the group whose long-term risk is most likely to be underestimated.

We propose to measure this failure and correct it. The idea is to look for disagreement between the parts of a record the patient controls and the parts they do not. On one side are self-reported history and the diagnoses that follow from it. On the other are objective measurements (toxicology screens, blood alcohol, liver enzymes) and third-party statements in clinical notes ("per family, patient has been drinking daily"; "patient declines to discuss mood"). Where the objective or collateral streams report something the self-report and codes leave out, we call it a disclosure gap. We test whether a model that represents this gap explicitly, and that also predicts whether a patient will quietly disappear from follow-up, gives better-calibrated multi-year risk estimates than current models that fuse several kinds of hospital data.

Core Claim

In retrospective electronic health record (EHR) cohorts, missing documentation of stigmatized conditions is not random, so absence cannot be treated as a negative. Two things follow. First, discordance between patient-controlled and patient-independent modalities carries prognostic signal. That signal is not recoverable from missingness indicators or encounter counts alone, and naive fusion of the same modalities does not capture it. Second, the same patients are more likely to leave the health system without a recorded event. A long-horizon model must therefore predict "silent departure" jointly with the clinical outcome, or it will read avoidance as survival.

The testable claim is specific and falsifiable. Among patients with high disclosure gaps, models that use codes only and standard multimodal EHR models should be miscalibrated: they should under-predict 1-year and longer readmission and mortality. Adding an explicit gap representation together with a joint model of the observation process should reduce that subgroup's calibration error by more than it changes overall discrimination.

Where This Sits

The help-seeking literature supplies the motivation. Up to 95% of distressed people never engage with professional care, and many prefer to manage their symptoms themselves Johnson et al.. Several studies explain why stigma shapes concealment. Affiliate stigma drives families to hide diagnoses, and the drive is strongest among people with more social standing to lose Yi Wang et al.. Mandatory reporting and penalties push people who use drugs away from clinics Jun Tao et al.. The clearest evidence that concealment is selective rather than total comes from search logs: people describe symptoms to search engines that they withhold from clinicians Automatic Identification of Web-Based Risk Markers for Health Events. That work has no verified clinical outcomes and only six months of data. We carry its central observation, that disclosure depends on the channel, into the clinical record itself, where outcomes can be verified.

The conversational line of work treats concealment as something to overcome at the moment of the interview. UPSD⁴ uses normalizing and loading questions to raise diagnostic accuracy on simulated stigmatized users from 25.00% to 34.09% Breaking the Stigma! Unobtrusively Probe Symptoms in Depression Disorder Diagnosis Dialogue. Audiovisual screening shortens assessment time Comparative Efficacy of MultiModal AI Methods in Screening for Major Depressive Disorder. Both are single-session studies, UPSD⁴ runs on synthetic personas, and neither can say what happens to patients over years. We take the opposite stance. We do not try to elicit disclosure. We accept that it has already failed in historical data, and we ask how to model its footprint.

The methods we build on come from multimodal EHR modeling. MEDHMP pretrains a hierarchical model over vitals, codes, medications, and notes, and improves 30-day readmission and 6-month chronic-disease prediction on MIMIC and TriNetX Hierarchical Pretraining on Multimodal Electronic Health Records. UTDE handles irregularly timed vitals and notes with gated interpolation and cross-attention Improving Medical Predictions by Irregular Multimodal Electronic Health Records Modeling. CLIMATv2 forecasts multi-year trajectories with a calibration-aware loss Clinically-Inspired Multi-Agent Transformers for Disease Trajectory Forecasting From Multimodal Data. All three treat missing or irregular data as a byproduct of hospital workflow. CLIMATv2 also assumes standardized research follow-up visits, which never occur in avoidance-driven care. FairReFuse is the closest analogue to our mechanism. It weights modalities per individual to remove gender shortcuts FairReFuse. Its target, however, is demographic fairness on interview corpora, not deliberate withholding in longitudinal records.

Retrospective epidemiology constrains the design. Propensity matching can only balance recorded covariates Long-Term Survival in Men With High Grade Prostate Cancer, and concealment is exactly the kind of confounder that goes unrecorded. The stroke readmission cohort shows that behavioral friction appears in structured data as discharge against medical advice (odds ratio 1.485) and self-pay status Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China. Its 31-day window is also the kind of limited follow-up our design aims to extend.

Research Program

The work has three parts. The first is a disclosure-gap extractor. We will write a clinical annotation guide for three kinds of note evidence: patient denial or refusal, collateral reports that contradict the patient, and explicit hedging ("patient is guarded"). A clinically trained team will label a few thousand note sentences from MIMIC-III and MIMIC-IV admissions involving substance use, self-harm, or mood symptoms. A sentence-level classifier, either a fine-tuned clinical language model or an LLM labeler checked against the gold labels, will then tag all notes.

The second part is the gap itself. For each admission we compute, for each condition, the difference between two signals. One is the objective or collateral signal: positive toxicology, blood alcohol, collateral mentions. The other is the self-report and coded signal: history-of-present-illness statements and ICD codes. The model receives these differences as explicit inputs, not only the raw modalities, so the key ingredient can be switched off in an ablation.

The third part is a joint long-horizon model. We attach two prediction heads to a MEDHMP-style admission encoder. One predicts the clinical outcome (readmission or death at 1 year and beyond). The other predicts silent departure, meaning no further contact with the system. Where linked death records exist, silent departure is separable from death. A simple competing-risks formulation lets the model stop counting unobserved patients as event-free.

Baselines

The naive reference point is a gradient-boosted model on structured codes, demographics, and insurance status, with no notes and no missingness handling. This is roughly the variable set of the stroke readmission study Wen et al.. Every other result is measured against it.

The strongest simple competitor is the same model plus missingness indicators, encounter counts, discharge-against-advice flags, and a flag for inter-visit gaps. Any structured-data team would try this first, and it may capture much of the avoidance signal cheaply. It gets the same features as our model except the note-derived gap terms, and the same hyperparameter search budget.

The strongest existing method is MEDHMP fine-tuned on our long-horizon targets, using the released pretraining recipe on the same MIMIC splits. MEDHMP reports a 7% AUROC and 29% AUPR relative gain for 30-day readmission over 18 baselines MEDHMP. I found no absolute AUROC for MEDHMP on Lacuna, and no reported result on Lacuna for any model on 1-year-plus readmission in stigmatized subgroups. We will therefore reproduce MEDHMP ourselves before comparing. UTDE Zhang et al. is a second fused baseline for the within-admission part. It reports relative F1 gains of 4.3% for fused 48-hour mortality, but no absolute figure appears in the file.

Two ablations isolate the proposal's ingredients. The first gives the full model the same notes and labs but removes the explicit gap inputs, which tests whether the discordance representation matters or whether fusion alone suffices. The second removes the silent-departure head and uses standard censoring.

Beating a baseline means two things. First, a lower expected calibration error, with confidence intervals from 1,000 bootstrap resamples, in the top gap-score quintile at 1 year and at the longest available horizon. Second, AUROC no worse than the baseline overall. We state in advance that a discrimination gain without a subgroup calibration gain does not confirm the claim.

Why not just…?

Why not just add missingness indicators and encounter counts to a structured model? This is the most serious threat. If avoidance leaves a strong trace in visit patterns, this cheap model could match ours. The difference lies in the ambiguity of absence. An uncoded condition in a patient with a negative toxicology screen and no collateral report means something different from an uncoded condition alongside a positive screen and "per wife, drinks heavily." Only the notes and labs can tell those apart. The first ablation settles the question directly. If the indicator model matches the full model in the high-gap quintile, we will report that concealment is adequately captured by structured proxies and scale back the claim.

Why not just fine-tune MEDHMP, which already fuses notes and codes? MEDHMP might learn the discordance implicitly through its contrastive alignment across modalities. Two things make this uncertain. Its text encoder is kept frozen during pretraining for memory reasons MEDHMP, which limits how well it can represent hedged denial phrasing. It was also evaluated only up to 6 months. We will probe MEDHMP's admission embeddings with a linear classifier for our gold gap labels. High probe accuracy combined with equal subgroup calibration would mean the explicit gap adds nothing, and we would report that.

Why not just apply inverse-probability-of-censoring weighting for attrition? Weighting fixes informative censoring only if the censoring depends on recorded covariates. Concealment is the textbook unrecorded confounder described in the propensity-matching literature Tewari et al.. We will still run weighting with the structured covariates as a baseline for the silent-departure component. If it matches the joint head once the gap features are added to its weighting model, the simpler estimator wins and we will adopt it.

Why not just use better intake interviews such as UPSD⁴ going forward? That approach is worth pursuing, but it answers a different question. It improves disclosure in future encounters. The historical records that current risk models are trained on cannot be re-interviewed. Its evidence is also synthetic and limited to a single session Breaking the Stigma!.

Proposed Experiments

The primary cohort is adults in MIMIC-IV with any admission involving substance use, self-harm, or mood symptoms, identified by codes, toxicology orders, or note mentions. It is deliberately defined broadly so that it does not require the very codes that concealment suppresses. Outcomes are readmission within 1 year and across the multi-year span of repeat MIMIC-IV admissions, plus mortality where linked. MIMIC-III serves as a temporal replication. For external validity and longer follow-up we will seek the multi-center network TriNetX used by MEDHMP. There we test whether the silent-departure head generalizes across health organizations, where moving between systems is the dominant form of attrition.

The decisive checks come in three steps. The first is validity. Clinician-annotated gap labels must reach acceptable inter-rater agreement, and the extractor must reproduce them on held-out notes; otherwise the rest of the study does not proceed. The second is the existence of the failure. In the top gap quintile, the code-only and MEDHMP baselines must show systematic under-prediction on calibration plots, and patients who look event-free in-system must show excess mortality in linked records. If neither appears, the premise is wrong. The third is the fix. The full model must reduce subgroup calibration error beyond both the indicator baseline and the fused model without gap inputs.

As a secondary check tied to the stroke readmission findings, we will test whether the gap score mediates part of the known association between discharge against medical advice and readmission. We will also report calibration across sex and race within the high-gap group, following the concern raised by FairReFuse that fusion can amplify demographic shortcuts.