Does the Neighborhood Food Environment Change the ICU Obesity Paradox?

In intensive care, heavier patients have fewer dangerous blood-sugar crashes and slightly better survival. Clinicians call this the "obesity paradox," and it is usually read as a story about individual physiology: extra body mass acts as a metabolic buffer. Public-health research tells a different story about where obesity comes from. In that literature, obesity is largely produced by the local environment: how close supermarkets are, how dense fast-food outlets are, how educated and how poor a neighborhood is. These two stories have never been tested against each other. This proposal asks a narrow question. Among ICU patients with the same body mass index (BMI, weight relative to height), does the protective association between obesity and fewer hypoglycemic events depend on the food environment of the neighborhood the patient came from? To answer it, we would build a privacy-preserving, neighborhood-level food-environment extension to MIMIC-IV, the large public critical-care database from Beth Israel Deaconess Medical Center in Boston, and run a pre-registered effect-modification analysis on glucose outcomes.

This matters for two reasons. If obesity that arises in a food desert is metabolically different from obesity that arises elsewhere, then the BMI coefficient in ICU glucose models averages over two populations, and a patient group defined by environment, which is already disadvantaged, may be miscalibrated. If the food environment adds nothing once individual chronic markers are known, that is also useful: it would show that the metabolic legacy of the environment is already written into the blood tests, and that geographic linkage is unnecessary for this purpose.

Core Claim

We hypothesize that tract-level food access modifies the association between BMI and ICU hypoglycemia and glucose variability. A census tract is a Census neighborhood unit of a few thousand residents. The specific prediction is that the protective odds ratio for obesity is weaker for patients from low-access, high-deprivation tracts. We also expect that this modification survives adjustment for county-level social determinants, insurance type, diabetes status, and admission HbA1c (a lab measure of average blood sugar over recent months). The new element is not "add social determinants to an ICU model." It is a targeted metabolic outcome, a neighborhood resolution that actually varies within a single hospital's catchment, and an interaction test rather than an overall gain in predictive accuracy.

Where This Sits

Three studies define the gap, and each plays a different role.

The first constrains the design. Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit built MIMIC-IV-SDOH by joining MIMIC-IV stays to county-level indices: County Health Rankings, the Social Vulnerability Index, and the AHRQ social determinants database. It found no AUROC or AUPRC gain for in-hospital mortality or readmission, although it did find better calibration and lower false-positive rates for diabetic and Black diabetic patients. We take this negative result seriously. Generic mortality is the wrong target, and county is the wrong resolution. Most BIDMC patients come from a few Massachusetts counties, so a county feature barely varies across the cohort. The authors themselves note that county aggregation cannot capture local food environments. Their calibration finding in diabetic subgroups is the motivating hint that metabolic patients are where community signal shows up.

The second supplies the clinical method. The obesity paradox and hypoglycemia in critically ill patients analyzed MIMIC-III, eICU, HiRID and AUMCdb using logistic regression over sliding 6-hour windows. It adjusted for insulin, nutrition, dextrose, steroids, SOFA (a standard organ-failure score), lactate and septic shock, and reported hypoglycemia odds ratios of 0.72 for overweight and 0.65 for obese patients, with weaker protection in diabetics. We reuse that windowed design almost exactly. The departure is that the study treats BMI as a purely individual trait with no upstream context.

The third supplies the community-side hypothesis and the choice of exposures. Community-level determinants of obesity linked an Ohio health system's EHR to ZIP-level business counts and reported obesity odds ratios of 0.58 for grocery density and 0.80 per 10% more college graduates. Its authors note that ZIP codes hide food-desert pockets. Large-scale diet tracking data reveal disparate associations between food environment and diet and its companion How Food Environment Impacts Dietary Consumption and Body Weight used genetic matching on the USDA Food Access Research Atlas. They found that grocery access shifts diet far more in Black- and Hispanic-majority areas than in White-majority ones. That heterogeneity is why we pre-specify race-stratified interaction estimates. Neither of these studies touches acute care.

Research Program

The first component is the data. MIMIC-IV publishes no addresses, so tract linkage cannot be done by outside researchers. It has to be done by the data custodians, as the county linkage in MIMIC-IV-SDOH must have been. We would propose that the custodians geocode each stay to a census tract inside their secure environment and attach a small set of USDA Food Access Atlas and Census variables: a low-income low-access flag, distance-to-supermarket band, and tract poverty and education. They would then release only a coarsened composite: a food-environment tertile plus the binary food-desert flag. Before release, they would suppress any cell defined by tertile, race, age band and sex that holds fewer than k patients. If release is refused, the identical analysis code would run inside the enclave and only model coefficients would leave. Neither path changes the science.

The second component is the outcomes. Following Plecko et al., the primary outcome is any glucose below 70 mg/dL in the next 6-hour window, with below 40 mg/dL as a severity check. The secondary outcome is the within-stay coefficient of variation of glucose. What is the ideal glucose range for a patient with sepsis in the ICU? reports a U-shaped relationship between mean glucose and mortality, with the lowest risk at 160–190 mg/dL, so we add time outside that band as a tertiary outcome in sepsis patients.

The third component is the model. It has the form

$$\text{logit}\,P(H_{w}=1) = \beta_{B} B + \beta_{F} F + \beta_{BF}\,(B \times F) + \gamma^\top X_{w},$$

where $H_w$ indicates hypoglycemia in window $w$, $B$ is the BMI category, $F$ is the food-environment tertile, $X_w$ holds the Plecko covariates plus county SDOH, insurance, diabetes and HbA1c, and standard errors are clustered by patient. The claim lives entirely in $\beta_{BF}$.

A separate measurement issue must be handled. Racial differences in laboratory testing shows that testing intensity in the MIMIC-IV emergency department differs by race even after exact matching on complaint and triage score. Glucose measurement frequency could therefore differ by neighborhood and manufacture an apparent effect. We will restrict the analysis to ICU stays where glucose is measured routinely, model the number of measurements per window as its own outcome, and inverse-weight by the probability of being measured.

Baselines

The naive reference point is the Plecko et al. specification refit on MIMIC-IV, with BMI category and clinical covariates but no community information. The file reports its odds ratios on the four pooled databases: OR 0.72 for overweight and 0.65 for obese. There is no reported result found on Lacuna for this model's discrimination on MIMIC-IV hypoglycemia. We will report our replication's odds ratio and AUROC first, so every later comparison has a concrete anchor.

The strongest existing competitor is the MIMIC-IV-SDOH approach of Yang et al.. It adds county-level indices to clinical features in logistic regression and XGBoost and evaluates discrimination, expected calibration error and false-positive-rate parity. Its exact aggregate AUROC values are not given in the file. It is the strongest competitor because it is the only published linkage of community data to MIMIC-IV, and its finding of no improvement is the default expectation. We will rerun it on our identical cohort, splits and outcome, with the same hyperparameter search budget as our tract-level model. Beating it means two things. On the effect-modification question, the BMI-by-food-environment interaction must be significant and must keep its direction when the county indices are in the model. On prediction, calibration error must improve within the food-desert and diabetic strata by more than the bootstrap interval of the county model. We do not require an AUROC gain, and we will say so in the pre-registration.

The key ablation replaces tract food access with the individual-level proxies MIMIC-IV already holds: insurance type, language, race, and admission HbA1c and albumin. If these absorb the interaction, the geographic linkage has no added value.

As a sequence-model reference, we will also train a Triplet-tokenized transformer pretrained with value forecasting via MEDS-torch. That pipeline reports a hospital-mortality AUC of 88.2 ± 0.1 on MIMIC-IV, but has no reported result for hypoglycemia. We will compare it with and without the food-environment features, to check whether a high-capacity model already recovers the effect from the clinical record.

Why not just…?

Why not just use individual BMI, diabetes and HbA1c? This is the most serious threat. HbA1c is a direct biological trace of months of diet, so it may already carry everything the neighborhood would add. Plecko et al. show diabetes status already modifies the paradox. The ablation above is designed to settle the question directly. If $\beta_{BF}$ disappears once HbA1c and albumin are included, we will report that the environment's metabolic legacy is readable from blood tests and that linkage is unnecessary for ICU glucose risk. HbA1c is missing for many stays, and missingness may itself be socially patterned, so we will also report the subset with HbA1c measured.

Why not just use the existing county-level MIMIC-IV-SDOH features? They are already available and validated against All of Us. However, within a Boston-area catchment, county takes very few values, so it cannot express the within-county contrasts the food-environment literature identifies. Roth et al. note that even ZIP codes mask food-desert pockets. Our head-to-head comparison on identical data will show directly whether the finer resolution matters.

Why not just extract social context from clinical notes? Notes are already in MIMIC-IV and need no new linkage. But MIMIC-SBDH found that Bio-ClinicalBERT fails 44.5% of negation tests in its Community category and 70.9% of attribution tests for Education. Notes also rarely mention grocery access at all. We will include note-derived community and economics flags as an additional comparison arm, but we expect low coverage for food access specifically.

Proposed Experiments

The primary experiment estimates $\beta_{BF}$ on all adult MIMIC-IV ICU stays that have recorded BMI and routine glucose monitoring, using the window design above. We will pre-register the direction (weaker obesity protection in the lowest-access tertile) and a minimum effect of interest: a 0.10 difference in the obesity odds ratio between the extreme tertiles. We will report results stratified by race, following the heterogeneity in Althoff et al., and by diabetes status.

Three checks decide whether the effect is credible. First, a negative-control outcome: the food environment should not modify the BMI association with a non-metabolic event such as unplanned extubation. If it does, the signal is generic deprivation, not diet. Second, the measurement-intensity analysis: if neighborhood predicts how often glucose is measured, the main estimate is reported only under inverse weighting. Third, the proxy ablation: the interaction must survive adjustment for HbA1c and insurance. The predictive arm compares calibration error and false-positive-rate parity across the naive, county, proxy, note-derived and tract models, using the same five-fold patient-level splits.

There is no second MIMIC-like database with neighborhood geography in the file. External replication would therefore come from All of Us, whose geographic resolution we would first need to confirm, and we state this as a limitation. Any of three outcomes is informative and publishable: a robust interaction, an interaction fully absorbed by HbA1c, or a null under adequate power.