Borrowing Risk Coefficients from Similar Hospitals: Case-Mix-Weighted Regression for Transporting Clinical Models

Clinical risk models usually lose accuracy when moved to a new hospital. Retraining at the new site is rarely an option, because one study found that a hospital needs about 2,400 of its own patients before a locally trained model catches up with an imported one (Generalizability of predictive models for intensive care unit patients). We propose a different route. Before any outcomes are observed, describe the new hospital by the unlabeled distribution of its patients. Then give it regression coefficients borrowed mostly from training hospitals whose patients look similar. The method comes from geographically weighted regression, a technique that fits a separate weighted regression at each location so that effects can change across space. We replace geographic distance with distance between hospitals' patient populations. Each risk factor also gets its own similarity radius, so the data can show which effects are truly universal and which are hospital-specific.

This targets a known dead end. Shifting the intercept or rescaling predicted probabilities at a new site fixes overall calibration, but it cannot fix ranking. Johnson et al. showed that recalibration corrects the standardized mortality ratio but leaves the area under the ROC curve unchanged. Ranking within a hospital can only improve if the slopes change, meaning how much each risk factor counts. Our method changes slopes at the new site using no outcome labels. Because the mapping from case mix to coefficients is fixed in advance, the "lock the model, then validate externally" discipline still holds.

Core Claim

Some risk factors have effects that vary from hospital to hospital in a way that patient case mix can predict. When that holds, a weighted logistic regression over the space of hospital case mixes will improve discrimination at held-out hospitals over a single global model. The gains should be largest at hospitals whose case mix lies far from the pooled average. The method also supplies a pre-deployment diagnostic: a target hospital's case-mix distance from the training hospitals should predict how much the global model will degrade there, before any labels arrive.

The method works as follows. Each hospital $h$ is summarized by a vector $z_h$ of standardized, label-free statistics of its first-24-hour covariates, such as age, admission-diagnosis mix, laboratory medians, and missingness rates. For a target hospital $t$ and covariate $k$, training patient $i$ from hospital $h(i)$ receives the weight

$$w_{ik}(t) = \exp\!\left(-\frac{\lVert z_t - z_{h(i)} \rVert^2}{2 b_k^2}\right),$$

where $b_k$ is a covariate-specific bandwidth, the radius over which hospitals are treated as similar for that effect. The coefficients $\beta_k(t)$ are fitted by backfitting a weighted logistic likelihood, one covariate at a time, as in multiscale geographically weighted regression (the variant that gives each covariate its own bandwidth). A very large $b_k$ means the effect is effectively global. A small $b_k$ means it should be borrowed only from close neighbours in case-mix space. Both the bandwidths and the map from $z_t$ to coefficients are fixed before external testing.

Relation to Prior Work

Two papers supply the method and show its limits. Bayesian Cluster Geographically Weighted Regression fits Bayesian local regressions across 526 Queensland areas and clusters the resulting coefficients. It found, for example, that the protective effect of preschool attendance fades in remote regions. It stays in-sample, uses Gaussian outcomes only, and names logistic extensions as future work. We take its central idea, that effects form regimes not tied to geographic contiguity, and make it predictive and binary.

Varying-coefficient models for geospatial transfer learning places Gaussian process priors on coefficient surfaces over physical coordinates and treats transfer as multi-task learning. It is the closest existing work. It differs from ours in three ways: it works on earthquakes and rents, it relies on smooth stationary kernels over geography, and it never runs a locked external validation. The file's own framing of GWR states its failure mode plainly: kernel weights collapse to zero when the target lies outside the training region. Placing hospitals in case-mix space removes that failure, because every new hospital has a case-mix vector even when it is geographically remote.

Four papers motivate the problem. Johnson et al. report held-out hospital AUROCs from 0.688 to 0.933 under a single global model and treat hospitals as categorical IDs. All models are local documents the Epic sepsis model falling to AUROC 0.60–0.64 and argues that each site should validate and fine-tune locally. That advice is impractical for small hospitals; our method lets them borrow strength from similar hospitals instead. Point-to-point transfers (sepsis transferability; OHDSI hemorrhagic transformation) show large drops but cannot relate them to any measure of distance between sites. The OHDSI German database's collapse to AUC 0.50 is a warning we take seriously. When a feature is structurally missing at the target, no coefficient borrowing can rescue it. Our missingness components in $z_t$ should flag such sites as distant rather than hide the problem.

Cross-validation for geospatial data and Model Evaluation for Geospatial Problems constrain how we evaluate. They show that removing nearby training data distorts error estimates through covariate shift, and they propose importance-weighted buffering as a correction. We borrow that buffering idea to stress-test transport at hospitals that are genuinely distant in case-mix space.

Research Program

We will first build hospital case-mix embeddings from eICU-CRD. The baseline embedding is a vector of summary statistics. A second variant represents each hospital's whole covariate distribution in a learned feature space, to test whether richer descriptions help. If the release exposes hospital region and size descriptors, we will build a parallel geography-and-descriptor embedding so that "where the hospital is" can be separated from "who its patients are."

Next we will extend multiscale backfitting to the logistic likelihood with ridge penalties inside each local fit. Ridge is the simplest remedy for the local collinearity problem the GWR literature flags. Bandwidths will be chosen by leave-one-hospital-out log loss computed on training hospitals only.

The third component is the transport diagnostic. We will regress per-hospital AUROC loss of the global model on the target's minimum and average case-mix distance to training hospitals, and test whether the relationship holds prospectively on new targets.

The last component is interpretation. The fitted bandwidths $b_k$ show which physiological effects are global and which depend on the hospital. For example, an effect might be local because of differing triage thresholds, which is one of the shift mechanisms the file describes.

Baselines

The naive reference is a single pooled L2-regularized logistic regression on the same 82-style feature set, locked and applied unchanged to each held-out hospital. This replicates Johnson et al., who report aggregate AUROC 0.854 and per-hospital AUROC 0.688–0.933 with SMR 0.560–1.43. We will rerun it on our exact splits rather than rely on those figures.

The second baseline is the same pooled model with intercept recalibration using the target's own labels. It is given an advantage our method lacks, namely access to target labels. Our method must match its SMR and beat its within-hospital AUROC.

The strongest flexible competitor is gradient boosting. We will use CatBoost as in E-CatBoost, with $z_h$ appended to every patient's features so that trees can learn case-mix interactions themselves. E-CatBoost reports AUROC between 0.86 and 0.92 across disease groups, but not under leave-one-hospital-out. Our method has no per-hospital comparison for this baseline in the file, so we will run both. All models receive identical folds, identical feature sets, and the same hyperparameter search budget, with 50 configurations each selected by inner leave-one-hospital-out.

The key ablation is the identical weighted model with case-mix coordinates randomly permuted across hospitals. A second ablation forces a single shared bandwidth, which isolates the value of covariate-specific scales. Beating the baselines means a higher median per-hospital AUROC and a better worst-decile AUROC than both the pooled model and boosting, with paired bootstrap intervals over hospitals excluding zero. Calibration slope must stay no worse than that of the pooled model.

Why not just…?

Why not just recalibrate the global model at each site? Johnson et al. show directly that recalibration repairs the SMR but leaves AUROC unchanged, because a monotone transformation cannot reorder patients. It is the right fix for calibration and the wrong fix for discrimination. Our second baseline measures this split explicitly.

Why not a hierarchical model with random slopes predicted by hospital-level covariates? This is a real threat. It is essentially the linear, parametric version of our idea. For a new hospital, plain random slopes default to the population mean, but slopes regressed on $z_h$ do not. We will fit it as an additional baseline. If it matches the kernel method, the honest conclusion is that a linear meta-regression on case mix is enough, and the contribution reduces to the case-mix-distance diagnostic and the per-covariate scale analysis.

Why not let gradient boosting absorb hospital descriptors as features? Boosting can in principle learn any interaction between case mix and patient features. With only 46 to 208 hospitals, though, the number of distinct case-mix values it sees is tiny, and trees extrapolate poorly beyond them. The boosting-plus-$z_h$ baseline settles this. If it wins on far-from-centroid hospitals, we will abandon the kernel model in favour of boosting and keep only the diagnostic.

Why not reweight training patients to match the target's covariate distribution? Density-ratio reweighting, as used in importance-weighted buffered cross-validation, corrects shift in $P(X)$ but keeps one set of slopes. It helps only if the outcome-given-covariates relationship is stable across hospitals. We include a reweighted pooled model. If it equals our method, the observed heterogeneity was covariate shift rather than varying effects.

Proposed Experiments

The primary experiment is leave-one-hospital-out over eICU-CRD, using the 46-hospital cohort definition of Johnson et al. and the broader 208-hospital set. We report per-hospital AUROC, calibration intercept and slope, and SMR. Results are stratified by quintile of target case-mix distance from the training hospitals.

The external test locks everything learned on eICU and applies it to MIMIC-III. Its case-mix vector places it somewhere in eICU's case-mix space without using any labels. Yin et al. report an external MIMIC-III AUC of 0.741 for logistic regression and 0.855 for an attention LSTM trained on eICU. Their feature sets differ from ours, so we treat these as context and rerun our own pooled logistic baseline on the same features.

A stress test applies importance-weighted buffering. When testing a target, we exclude its nearest case-mix neighbours from training, which mimics a truly novel hospital.

Three checks decide the question. First, the AUROC gain over the pooled model must be positive and must grow with case-mix distance; if it is flat, the varying-slope story fails. Second, the permuted-coordinate ablation must erase the gain. Third, an intercept-only version of our model must recover none of the AUROC gain, which confirms that improvement comes from slopes, as the recalibration result predicts. As a secondary outcome, the transport diagnostic should predict held-out per-hospital AUROC loss with a meaningful rank correlation across hospitals. That would give small hospitals a warning before deployment, independent of whether the coefficient borrowing itself wins.