Does Documented Family Support Buffer Readmission? Error-Corrected Moderation Estimates from Clinical Notes

Psychology has long argued that social support works as a buffer: it does not simply make people healthier on average, it weakens the path from a stressor to a bad outcome. In a buffered patient, a shock such as a new cognitive or functional impairment does less damage. Hospitals record this support constantly in free-text notes ("lives with daughter who manages meds," "wife at bedside," "lives alone, no local family"), and clinical language models can now pull such mentions out of notes fairly well. No one has used those extracted mentions to test the buffering claim itself on real clinical outcomes. The obvious attempt, which is to run an extractor over notes and plug its labels into a regression, has a hidden flaw. Extractors fail in systematic ways: they miss negations and credit a relative's situation to the patient. These errors can shrink a real buffering effect toward zero or create a false one.

We propose to measure the buffering effect of documented social support on post-discharge emergency returns in MIMIC-IV, and to make that measurement trustworthy. The plan has three parts. We build a support extractor that separates close, hands-on ties (a co-resident or daily caregiver) from diffuse ties (church, friends), with explicit handling of negation and attribution. We annotate a modest gold sample. Then we use that sample to correct the downstream interaction estimate for extraction error, instead of trusting the extractor's labels. The deliverable is a calibrated, falsifiable estimate of whether, and for which kind of tie and which kind of outcome, documented support weakens the link between discharge-time impairment and return to the hospital.

Core Claim

The claim has two halves. The substantive half predicts that close, functional support moderates the effect of cognitive or functional impairment on 30-day emergency returns, while diffuse support does not. It also predicts that the buffering should be concentrated in low-acuity returns: visits that end with discharge from the emergency department rather than admission. That pattern follows from Grandmothers' role in preventing unnecessary accident and emergency attendances: cohort study, which found that families with low grandmother involvement had an adjusted odds ratio of 8.82 for "no-treatment" pediatric emergency visits. It also follows from Dongdong Jiang et al., who reported that close "bonding" ties protect health-related quality of life while diffuse "bridging" ties sometimes do not. The methodological half is that plugging raw extractor outputs into the moderation model gives a biased interaction estimate, and that a small gold-labeled sample is enough to remove most of that bias.

The model we estimate is

$$\text{logit}\,P(Y=1) = \alpha + \beta S + \gamma B + \delta\, S\cdot B + \theta^\top X,$$

where $Y$ is a 30-day emergency return, $S$ is the stressor (documented delirium, dementia, or functional dependence at discharge), $B$ is the support indicator of a given type, and $X$ is a set of covariates (age, LACE score, comorbidity, insurance, discharge disposition). Buffering means $\delta < 0$. The quantity at risk is $\delta$: because $B$ is inferred from text, errors in $B$ that are correlated with $S$ corrupt $\delta$ directly. Two examples of such correlated errors are sicker patients having longer notes with more family mentions, and cognitively impaired patients having more caregiver-attribution confusions.

Relation to Prior Work

Four lines of work extract social factors from notes, and all of them stop at the label. Annotating Social Determinants of Health Using Active Learning, and Characterizing Determinants Using Neural Event Extraction built SHAC, a corpus of 4,480 annotated social history sections, and an event extractor that includes a "LivingStatus" field. Prompt-based Extraction of Social Determinants of Health Using Few-shot Learning showed that a fine-tuned span extractor (mSpERT) beats one-shot GPT-4 on SHAC. SDOH-NLI recast extraction as natural language inference, meaning the model judges whether a snippet implies a fixed statement such as "the person lives alone." Large Language Models to Identify Social Determinants of Health in Electronic Health Records fine-tuned Flan-T5 to find adverse social support, and found that ICD-10 codes captured only 2.0% of patients whom the notes flagged. These papers supply our extraction methods and training data. None of them asks whether the extracted variable changes how a stressor maps onto an outcome, and none separates caregiving ties from diffuse ones.

One paper constrains the design more than any other: MIMIC-SBDH. Its Bio-ClinicalBERT reached a macro-F1 of 0.96 on "Community-Present," yet failed 44.5% of negation behavioral tests in that category and 70.9% of attribution tests for Education. Attribution failure here means crediting a relative's status to the patient. This is the direct reason we do not trust plug-in labels for a moderation analysis, since an F1 score says nothing about how errors are distributed across stressed and unstressed patients.

On the outcome side, A Comparative Study of Structured and Narrative EHR Data for 30-Day Readmission Risk Assessment showed that a ClinicalLongformer model over MIMIC-IV discharge summaries (AUROC 0.72) beats structured-data models (AUROC 0.65–0.67). The note text therefore carries outcome signal, but that model is monolithic and cannot say which part of the text matters or whether support interacts with impairment. Dementia-R1 reasons over longitudinal notes using intermediate clinical scores but leaves interpersonal context out entirely.

Two studies motivate the moderation framing and the choice of stressor. Network analysis of cold cognition and depression in middle-aged and elder population: the moderation of grandparenting found weaker coupling between cognition and depression symptoms among grandparents (connectivity 5.41 versus 6.03). Xiyang Zhang et al. found that proximate family ties, but not digital-only ties, protected well-being during an acute shock. Both are survey studies without clinical outcomes, and Yuan et al. caution that the effect may be culture-specific. We therefore treat their findings as hypotheses to test in a US clinical population, not as priors to confirm.

Research Program

The first component is a support schema and gold set. We annotate roughly 2,000 MIMIC-IV discharge summaries at the patient level for five things. The first is close functional support, meaning a named person who lives with the patient or provides daily help. The second is close non-functional ties, such as a spouse or children mentioned without any caregiving role. The third is diffuse ties. The fourth is explicit absence (for example, "lives alone, no family nearby"). The fifth is not documented, which we keep separate from absence because the two are easy to conflate and that conflation is itself a bias source. Each positive label also records whose support it is (the patient's or a relative's) and whether it is current or historical. We select notes by active learning with a surrogate category-presence classifier, following the method of Lybarger et al., so that rare positives such as explicit absence are not starved. We double-annotate 20% of the notes to measure agreement.

The second component is the extractor. We fine-tune a Flan-T5 model in the natural-language-inference format of SDOH-NLI, with new hypotheses matching our schema, starting from SHAC and SDOH-NLI and then adding our MIMIC-IV labels. The architecture is not the contribution. What we add is a behavioral test suite for negation, attribution, and temporality, built in the CheckList style that MIMIC-SBDH used. We report failure rates separately for patients with and without the stressor, because a difference in error rates between those groups is what would bias $\delta$.

The third component is the corrected moderation estimate. We fit the model above on all eligible MIMIC-IV admissions using extracted labels. We then correct $\delta$ with the gold sample: the difference between the estimates computed from gold labels and from extracted labels on the annotated subset serves as a debiasing term. This follows the logic of prediction-powered inference, and it gives valid confidence intervals as long as the gold sample is a random draw from the analysis population. We therefore draw a separate uniform random gold sample of about 800 notes for the correction, apart from the active-learning set. For outcomes, we split emergency returns into low-acuity returns (discharged from the emergency department) and admissions. As a negative control, we use planned readmissions such as scheduled chemotherapy or elective procedures, which support should not buffer.

Baselines

The naive reference point uses structured fields only. We set $B$ from the admission record's marital status and from ICD-10 Z63 and Z60 codes, and fit the same moderation model on the same admissions. Given the 2.0% code capture reported by Chen et al., we expect this baseline to produce a wide, near-null $\delta$. Any substantive text-based finding has to beat it by producing an interval that excludes zero where this baseline's does not.

The strongest extraction competitor is an existing trained extractor used as is, with plug-in labels. Concretely, this is mSpERT trained on SHAC, which reported F1 0.861 in Ramachandran et al., with its LivingStatus outputs mapped onto our schema. We also run fine-tuned Flan-T5 XXL on SDOH-NLI's social-support hypotheses, which Lelkes et al. report at F1 0.8788. Both are retrained on the same source data with the same tuning budget and scored on our held-out gold set. This competitor is strong because it is exactly what a practitioner would reach for. Our extractor beats it only if it reduces the gap in negation and attribution failure rates between stressed and unstressed patients, not merely if it raises F1.

The strongest predictive competitor is ClinicalLongformer on full discharge summaries, which reported AUROC 0.72, F1 0.68, and Brier score 0.2138 for 30-day readmission in MIMIC-IV in the comparative study. We rerun it on our cohort and outcome definitions. We do not claim to beat it on AUROC. The claim is that a model built from interpretable stressor and support variables loses little discrimination; we would accept an AUROC within about 0.03 while yielding an estimable, testable interaction.

The ablation that isolates our key ingredient compares, on identical extracted labels, the plug-in $\delta$ with the gold-corrected $\delta$. The comparison is anchored by a semi-synthetic experiment in which we plant a known $\delta$ into simulated outcomes built on real extracted and gold labels.

Why not just…?

Why not just use structured marital status and social-problem codes? Code capture of 2.0% against 93.8% from notes, reported by Chen et al., makes Z-codes nearly useless. Marital status is better populated, but it cannot tell a caregiving spouse from an estranged one, and it cannot mark a widowed patient who lives with a daughter. The structured-field baseline tests this directly. If it recovers the same $\delta$ as the corrected text estimate, the text pipeline is unnecessary for this question, and we would report that.

Why not just prompt GPT-4 or a similar model zero-shot and regress on its labels? One-shot GPT-4 reached F1 0.625–0.652 on SHAC against 0.861 for a fine-tuned model (Ramachandran et al.). Its main failure was linking attributes to the right person, which is precisely the attribution error that corrupts $\delta$. Privacy rules also block sending MIMIC text to commercial APIs, so we would test an open-weights instruction model under the same correction. Because the gold-sample correction is agnostic to which extractor produced the labels, a weaker extractor mainly widens the confidence interval. This alternative is therefore a real, cheap option, and we measure how much interval width it costs.

Why not just interpret the ClinicalLongformer readmission model, for example with attributions on support phrases? Attribution scores on a monolithic model show that a phrase moves the prediction. They do not show that the phrase changes the effect of impairment, which is the buffering claim, and they carry no error-corrected uncertainty. We also fit a ClinicalLongformer that receives the stressor indicator as an explicit input and probe whether its predicted risk gap between impaired and unimpaired patients shrinks when support phrases are present. If that probe matches the corrected $\delta$, it is a legitimate cheaper route and we would say so.

Why not just skip the correction, since a good extractor's errors are small? This is the most serious threat, and the semi-synthetic experiment settles it. If plug-in and corrected estimates agree within their intervals across planted effect sizes, and the stratified failure rates show no gap between stressed and unstressed patients, then the correction is dead weight. We would then report the simpler pipeline as sufficient.

Proposed Experiments

The first experiment checks the extractor. We report per-category F1 on the gold set against mSpERT and the SDOH-NLI-tuned Flan-T5. More importantly, we report stratified negation and attribution failure rates, measured against the 44.5% and 70.9% figures from MIMIC-SBDH. The decisive quantity is the difference in error rates between stressed and unstressed patients.

The second experiment is the semi-synthetic check. We plant values of $\delta$ from zero to strongly negative into simulated outcomes. For each, we measure the bias and confidence-interval coverage of the structured-field, plug-in, and corrected estimators. The correction earns its place only if plug-in coverage falls clearly below the nominal 95% while the corrected estimator's coverage stays near it.

The third experiment is the substantive test on real outcomes. The buffering hypothesis is supported if the corrected $\delta$ for close functional support is negative, with an interval excluding zero, for low-acuity returns. Three further results strengthen it: a smaller or null $\delta$ for admissions, a null $\delta$ for diffuse ties, and a null $\delta$ on the planned-readmission negative control. A negative $\delta$ on the negative control would indicate that documentation bias or confounding by frailty, rather than buffering, drives the result. We also report a sensitivity analysis that treats "not documented" as absence, as presence, and as its own category.

The fourth experiment checks transfer. We apply the extractor to MIMIC-III notes, since SHAC and MIMIC-SBDH were both drawn from MIMIC-III, and report how much the stratified error rates shift. Chen et al.'s finding that performance degrades from oncology to the ICU warns that the correction must be re-estimated for each site. Because MIMIC is a single-center dataset, any positive finding should be read as a demonstration of the method, not as a generalizable estimate.