Clinical risk models increasingly include a patient's social circumstances, such as food insecurity, unstable housing, or lack of transport. Hospitals rarely ask about these directly, so most models substitute an area-level number: the deprivation index or median income of the census tract where the patient lives. This is the ecological fallacy, assuming a property of a group holds for each member. A deprived person in an affluent tract looks well-off to the model, and a comfortable person in a poor tract looks deprived. The resulting errors show up in the clinic as alerts that fire for the wrong patients and stay silent for the right ones.
Two cheap fixes are available, and neither is enough. Asking every patient a screening questionnaire is accurate but adds friction to visits that are already short on time. Using the tract number alone costs nothing but is wrong in exactly the places that matter. We propose a middle path. A model estimates each patient's individual social needs from their electronic health record (EHR), anchored to the known make-up of their area. It also reports how much it trusts that estimate. When the estimate is uncertain and the answer would change a clinical decision, the system asks the patient one or two screening questions at check-in. The concrete deliverable is a curve: how much of the benefit of universal screening can be recovered at a given fraction of the questioning burden.
Core Claim
Area-level social indices are not uniformly bad proxies. They fail where the area is internally mixed and the patient's record carries little contradicting signal. Both conditions can be estimated before the patient is seen. A model that combines individual EHR features with the area's distribution of social circumstances, rather than only its mean, can therefore flag the minority of patients for whom the neighborhood misdescribes the person. Asking only those patients should recover most of the accuracy and alert-precision gains of universal screening at a small fraction of its questioning cost.
The new element is that the size of the ecological inference error is treated as a per-patient quantity that drives a workflow action, namely whether to ask. Prior work treats it either as a modeling nuisance to be averaged away or as a population-level critique.
Where This Sits
Two lines of work motivate the proposal by assigning area indices directly to patients. Han et al. predict chemotoxicity in colorectal cancer using the Area Deprivation Index as an individual covariate. Hu et al. predict uncontrolled hypertension for 164,041 Boston Medical Center patients. They combine individual answers to the THRIVE social-needs survey with census-tract income and report large racial error asymmetries at the default threshold: a false positive rate of 27.71% for Black patients versus 5.92% for White patients. Neither study asks how much of that error comes from the area proxy disagreeing with the individual. Hu et al.'s dataset matters to us for a second reason: it is the rare setting where the area proxy and the individual answer coexist for the same patients, so the ecological gap can actually be measured.
The empirical case that area numbers can mislead comes from three sources. Sajjad et al. find that more accessible, more advantaged Victorian local government areas had higher diabetes admission rates (IRR 2.32 for type 2 diabetes). This is a supply-and-help-seeking signal that a naive model would read as risk. Lucchini et al. show that neighborhood averages hid individuals who could not afford to self-isolate. These findings constrain the design: area features may encode access rather than need, so they cannot be trusted as need labels.
The method draws on disaggregation work that recovers individual quantities from group summaries. Generalized Linear Models for Aggregated Data fits regression models when only group histograms of the target are known. Its accuracy approaches the fully observed model as the histogram moves from a mean to 5 or 10 bins. CLAM and A-MoGP do the same for causal effects and spatial fields. CLAM openly reports an "identifiability gap": many different local patterns can sum to the same regional total.
We reuse the histogram-constraint idea, and we make that identifiability gap the signal for when to ask rather than a flaw to hide. We also depart from all three in one important way. They assume individual labels are never available, whereas a hospital has labels for a non-random screened subset. That subset sharpens the model. It also introduces selection bias, which the area constraint can partly correct.
The workflow literature constrains what may be built. Wang et al. find that tools requiring separate logins or extra data entry are rejected, and that only 7 of 20 reviewed systems were evaluated in live settings. Fedele et al. find that primary care clinicians want EHR-native automation, not more forms. Keyes et al. describe Stanford's monitoring of 13 deployed models, including subgroup performance tracking and the retirement of four vendor models. That monitoring treats inputs as fixed observed attributes. We add a monitored quantity, cross-level calibration (how well area-anchored estimates match screened answers), which can drift as neighborhoods change.
Research Program
The work has three parts.
Measuring the gap. Using screened patients at Boston Medical Center (the THRIVE cohort; access through collaboration), we measure how often tract-level income and deprivation disagree with individual unmet needs. We relate the disagreement rate to within-tract heterogeneity, estimated from census income distributions. We then attribute the hypertension model's subgroup false positives and false negatives to the area features: we re-score the model with each patient's area feature replaced by their own screened answer, holding all else fixed.
An area-anchored estimator of individual need. Let $s_i \in \{0,1\}^8$ be patient $i$'s answers across the eight THRIVE domains. These are observed only for the screened set $S$. Let $x_i$ be EHR features and $a(i)$ the patient's tract. A model $p_\theta(s_i \mid x_i, a(i))$ is trained on the objective
$$\min_\theta \sum_{i \in S} \ell\big(s_i, p_\theta(\cdot \mid x_i, a(i))\big) + \lambda \sum_{a} D\big(h_a, \hat h_a(\theta)\big),$$
where $\ell$ is cross-entropy, $h_a$ is the census histogram of a need-related quantity (for example income bins) for tract $a$, $\hat h_a(\theta)$ is the same histogram implied by the model's predictions over all patients from tract $a$, $D$ is a divergence, and $\lambda$ weights the constraint.
The first term learns from the screened patients. The second term ties predictions for the unscreened majority to what is known about their areas. Following Bhowmik et al., predictions are matched to histogram ranks by sorting, which keeps the constraint cheap. Hospital patients are not a random sample of tract residents (Gap 5 in our file), so $\lambda$ is tuned rather than fixed, and the constraint is checked against held-out screened data. We obtain uncertainty by resampling over the set of individual assignments that fit the area histogram almost equally well. This is where CLAM's identifiability gap reappears, now as a measured quantity.
The asking rule. A screening question is triggered for patient $i$ when
$$\mathbb{E}\big[\,|\Delta d_i|\,\big] > c,$$
where $d_i$ is the downstream clinical decision (for example, whether the hypertension alert fires at the operating threshold), $\mathbb{E}[|\Delta d_i|]$ is the probability that learning $s_i$ would flip it, and $c$ is a cost set by a screening budget. The rule ignores uncertain patients whose decisions would not change. This is the difference between targeting by uncertainty and targeting by consequence.
The output is a single EHR field with an optional one- or two-item prompt at check-in. The design follows the "unremarkable" placement Wang et al. describe, rather than a separate tool.
Baselines
Area-only model. The naive reference point is Hu et al.'s pipeline as published: XGBoost on demographics plus census-tract income, with THRIVE answers used wherever they exist. It reports an AUC of 74.49%, against 73.42% for logistic regression. We refit it on identical splits with the same hyperparameter search budget as our model. For a fair comparison we also apply the per-group thresholds they used to equalize the false negative rate at 25%. Beating it means higher alert precision, and a lower false positive rate for Black patients, at that matched 25% false negative rate. AUC alone does not count.
Universal screening. This is the upper reference point: every patient's individual answers are observed. Every policy is judged by the fraction of the gap between area-only and universal screening that it closes, at each questioning budget.
Plain supervised imputation. This is the strongest competitor. It is the same model family trained on screened patients with $\lambda = 0$, combined with the same asking rule. It isolates the aggregate constraint. We report no Lacuna number for it because none exists for this task.
Uncertainty-ranked asking. This baseline asks the patients with the highest predicted uncertainty, regardless of decision impact. It isolates the decision-flip criterion.
Heuristic area rule. This baseline asks all patients living in high-deprivation tracts, which is what a hospital would likely do without a model.
Random asking. Patients are asked at random, at the same budget as every other policy.
We found no reported result on Lacuna for social-need imputation from area data on any clinical dataset, and no numeric losses for the aggregate-GLM method on DE-SynPUF or the Texas discharge data.
Why not just…?
Why not just screen everyone? Universal screening is our upper bound, not a rival, and it is the most accurate option. The case against it is burden and completeness. Wang et al. and Fedele et al. document clinician resistance to added data entry, and the CURATE.AI pilot (Sumner et al.) saw only 3 of 18 patients generate enough data to reach active recommendations. If our curve shows that recovering most of the benefit requires asking nearly everyone, the proposal reduces to "screen everyone." We would report that outcome as such.
Why not just train a classifier on the patients who were already screened? This is the real threat, and the plain-supervised baseline tests it directly. Screened patients at a safety-net hospital are unlikely to be a random sample. If they are close to random, the area constraint should add little and we expect to lose this comparison. Our claim is narrower: the constraint helps on the unscreened population, and most of all for patients in tracts poorly represented among the screened. We test this with held-out evaluations in which screening is made deliberately non-random. If supervised imputation matches the constrained model even there, the disaggregation component is dropped, and the contribution becomes the decision-targeted asking rule alone.
Why not just recalibrate thresholds per group, as Hu et al. did? Group-specific thresholds equalize error rates across race. They cannot move an individual misdescribed by their tract, because the error lies inside the group. The attribution analysis in the first part of the program tests this. We measure how much of the residual false positive rate at equalized false negative rate disappears when area features are replaced by individual answers. If very little disappears, the ecological gap is not a meaningful driver of alert errors, and the motivation fails.
Why not just extract social history from notes with a language model? Fine-tuned extraction works well on the SDOH-NLI benchmark: Lelkes et al. report an F1 of 0.8788 with Flan-T5 XXL, versus 0.67 zero-shot. But it can only find what clinicians have written down. We include note-extracted features as inputs where notes exist, which makes extraction complementary to our approach. A sensitivity analysis checks whether asking remains valuable once notes are used.
Proposed Experiments and Decisive Checks
Main experiment. On the Boston Medical Center EHR and THRIVE cohort, split by time to mimic deployment, we compute the burden–benefit curve for every asking policy at budgets from 5% to 100% of visits. We report three things. The first is individual need-estimation quality: AUROC and calibration per domain, broken down by tract heterogeneity. The second is downstream hypertension alert precision and subgroup false positive rate at a 25% false negative rate. The third is the number of questions asked per visit.
Method sanity checks. We test the constrained estimator on DE-SynPUF and the Texas discharge data, where individual targets can be hidden and recovered, using the setups of Bhowmik et al.
Silent-mode prospective test. Following the Stanford monitoring approach, the estimator runs in the background on new visits. We log the cross-level calibration gap monthly, and drift beyond a band triggers review. A short vignette study with primary care clinicians then checks whether the flagged prompts are judged acceptable.
Decisive outcomes. Four results would settle the question. The idea works if, at a 20–30% budget, decision-targeted asking closes most of the area-only-to-universal gap in alert precision, and clearly beats both the high-deprivation rule and uncertainty-ranked asking. The disaggregation ingredient earns its place only if the constrained model beats plain supervised imputation under non-random screening. The motivation fails if replacing area features with individual answers barely changes subgroup false positive rates. The cost-benefit claim fails if no budget below roughly half of visits recovers most of universal screening's benefit, since that pattern would mean area indices carry too little individual information to be worth targeting around.
References
- AI-driven chemotoxicity prediction in colorectal cancer: impact of race, SDOH, and biological aging
- Accounting for racial bias and social determinants of health in a model of hypertension control
- Association between area-level socioeconomic status, accessibility and diabetes-related hospitalisations: a cross-sectional analysis of data from Western Victoria, Australia
- Socioeconomic disparities in mobility behavior during the COVID-19 pandemic in developing countries
- Generalized Linear Models for Aggregated Data
- Aggregated Multi-output Gaussian Processes with Knowledge Transfer Across Domains
- Human-centered design and evaluation of AI-empowered clinical decision support systems: a systematic review
- Development of a Clinical Decision Support Tool to Implement Asthma Management Guidelines in Pediatric Primary Care: Qualitative Study
- Monitoring Deployed AI Systems in Health Care.
- Preliminary insights into artificial intelligence guided dosing in hypertension and diabetes: challenges and lessons learnt in a pilot feasibility study
- SDOH-NLI: a Dataset for Inferring Social Determinants of Health from Clinical Notes