Emergency department (ED) crowding over the next few hours is mostly decided by people who are already in the building: patients waiting for tests, patients waiting to be seen, and admitted patients "boarding" in ED beds until an inpatient bed opens. Current crowding forecasts mostly ignore this. They treat the hourly patient count, called the census, as an opaque time series and extrapolate it, and they lose accuracy sharply beyond a couple of hours. We propose a "bottom-up" forecast instead. A model estimates, for each patient currently present, how likely they are to still be there in $h$ hours and whether they will leave by admission or by discharge. Summing these per-patient estimates, and adding a model of new arrivals, gives a full probability distribution over future census and boarding counts.
This design changes the privacy question. A model of aggregate counts can be pooled across hospitals because hourly counts are not sensitive. A model of when each patient will leave needs patient-level records, and small community EDs do not have enough of those to train one alone. That is the case where federated learning earns its place. Federated learning trains a shared model by sending it to each hospital and combining only parameter updates, so raw records never leave the site. We therefore propose federated training of the per-patient departure model. Covariate effects are shared across hospitals, and each hospital keeps its own baseline departure-time curve, because boarding delays depend on local bed supply. The claim is testable. It may turn out that bottom-up forecasting adds nothing over shared aggregate counts. In that case federated learning is unnecessary for ED crowding, and we will report that result.
Core Claim
We claim three things. First, at horizons of 2 to 8 hours, a census forecast built from per-patient departure models is more accurate and better calibrated than top-down census models. The extra information comes from patient-level state: acuity, number of pending orders, elapsed time, and whether an admission decision has already been made. Second, the per-patient model benefits from multi-site training, especially at small sites. Third, the right way to share across sites is to pool covariate effects and keep the departure-time baseline local. Full averaging and discarding unusual sites are both worse choices.
Write $S_i(h \mid x_i)$ for the probability that patient $i$, with current features $x_i$, is still in the ED $h$ hours from now. Write $\lambda(s)$ for the arrival rate at future time $s$, and $\bar S(u)$ for the probability that a newly arrived patient is still present $u$ hours after arrival. The expected census is then
$$ \mathbb{E}[C(t+h)] = \sum_{i \in \text{present at } t} S_i(h \mid x_i) + \int_t^{t+h} \lambda(s)\, \bar S(t+h-s)\, ds . $$
The first term depends on patient-level data. The second term needs only aggregate arrival counts. We model departures as independent per patient given the shared context, which makes the first term a sum of Bernoulli variables. The arrival term is a Poisson-type count. Together these give prediction intervals, not just point forecasts. Splitting the same sum by departure type (admission versus discharge) gives a boarding forecast. The first term should dominate at short horizons. The arrival term should dominate at longer ones, which is why the advantage should shrink toward 8 hours and beyond.
Positioning Against Prior Work
Two single-site ED papers motivate the problem. A gradient boosted model using 15-minute census slopes forecasts 2-hour census well, but accuracy drops sharply at 8 hours (Emergency Department Optimization and Load Prediction in Hospitals). That model is the top-down approach we compete against. Interpretable Lasso models with binned vitals predict ED stays over 4 hours and admit-versus-discharge decisions at 10, 60 and 120 minutes after triage (Developing interpretable machine learning models to predict length of stay and disposition decision for adult patients in emergency departments). We reuse their binning to build our per-patient model, but we predict a time-to-departure curve rather than a fixed threshold, and we aggregate the curves into a census forecast. Work on English COVID-19 bed occupancy shows that collapsing length of stay to an average underestimates occupancy, because it erases long delays in multi-stage pathways (Importance of patient bed pathways and length of stay differences in predicting COVID-19 hospital bed occupancy in England). That finding argues for full per-patient distributions, as here.
On the federated side, existing hospital work predicts static, per-patient targets and never forecasts occupancy. One study recruits representative eICU hospitals and drops the rest, reaching centralized-level ICU length-of-stay error (Client Recruitment for Federated Learning in ICU Length of Stay Prediction). We keep every site and personalize the baseline instead. A graph-penalized method personalizes linear length-of-stay models across five hospitals (Comparing Federated Stochastic Gradient Descent and Federated Averaging for Predicting Hospital Length of Stay). We use it as a competing personalization scheme. A third study finds that many local epochs and small batches help FedAvg on eICU (Federated Learning in Multi-Center Critical Care Research: A Systematic Case Study using the eICU Database). We adopt those tuning ranges.
A benchmark comparing engineering and statistical federated methods on ED data shows that FedAvg-type training biases coefficients (Federated Learning in Healthcare: A Benchmark Comparison of Engineering and Statistical Approaches for Structured Data Analysis). This constrains our design. If clinicians are to read hazard coefficients, we must also check their stability, not only forecast accuracy. FAIR-EC builds a multinational ED federated network around static admission scores (FAIR-EC: A Global Research Network for Fair, Accountable, Interpretable, and Responsible AI in Emergency Care (Preprint)). It is the natural venue for a real multi-site test, but it does not forecast load.
Federated time-series methods forecast aggregate series (Tackling Data Heterogeneity in Federated Time Series Forecasting, Unlocking Dynamic Inter-Client Spatial Dependencies: A Federated Spatio-Temporal Graph Learning Method for Traffic Flow Forecasting). They represent the top-down alternative, which we argue does not actually need federation in this setting.
Research Program
The per-patient model is a discrete-time competing-risks hazard model. A hazard model predicts, for each hour, the chance that a patient still present leaves during that hour. "Competing risks" means it predicts two such chances, one for leaving by discharge and one for leaving by admission. The features are triage vitals binned as in Song et al., chief-complaint group, counts of pending and completed orders, elapsed time, whether a disposition decision has been made, and ED-wide context such as current census and hour of day.
We will build two versions:
- a binned logistic hazard, which is interpretable;
- a small multilayer network, to measure how much accuracy interpretability costs.
Parameters split into a shared block, the covariate effects, and a site block, the hourly baseline hazards for each departure type. Federated rounds average only the shared block. The site block is fit locally, in the same way that a stratified survival model gives each stratum its own baseline. The arrival model is a Poisson regression with calendar and weather features. It is trained on hourly counts, which hospitals may share directly.
We will run the forecast in two settings. The first is eICU across its 208 hospitals, where the target is ICU occupancy forecast from patients currently admitted. This is real multi-site heterogeneity on public data, though not the ED itself. The second is ED census and boarding on MIMIC-IV-ED, reconstructed from arrival and departure timestamps. MIMIC-IV-ED is single-site, so there we can only study the forecasting question: does bottom-up beat top-down? The federated question needs a true multi-site ED evaluation. We will pursue it through FAIR-EC partner sites, including the Singapore General Hospital and Duke ED data, using the same code. That part depends on the collaboration.
Baselines
The naive reference is a seasonal historical average: the mean census for the same hour and weekday over the previous eight weeks. It is the planning tool many EDs actually use.
The strongest top-down competitor is a reimplementation of the census-slope gradient boosted model from Emergency Department Optimization and Load Prediction in Hospitals. That paper reports 81.52% accuracy and 72.90% of 2-hour forecasts within 4 patients, but on a private trauma-center dataset, so we can only compare against our reimplementation. No reported result was found on Lacuna for census forecasting on MIMIC-IV-ED or for occupancy forecasting on eICU. We give it the same lagged features at every horizon, plus a pooled multi-site variant trained on shared hourly counts. The pooled variant is the strongest privacy-free alternative.
For per-patient length of stay on eICU, the strongest federated method is client-recruited training. Client Recruitment for Federated Learning in ICU Length of Stay Prediction reports an MAE of 2.21 for both Federated-SRC and centralized training, against 2.26 for standard selection. We will rerun it on the same cohort and convert its per-patient predictions into occupancy forecasts. This tests whether discarding sites hurts occupancy forecasts at the sites it drops.
We also include FedAvg and FedProx on the full hazard model, and graph-penalized FedSGD from Comparing Federated Stochastic Gradient Descent and Federated Averaging for Predicting Hospital Length of Stay. The bounds are purely local training and centralized pooling.
The key ablations target the proposal's two ingredients:
- Patient-level information. The bottom-up model is restricted to features derivable from aggregate counts, such as elapsed time and census, with no acuity or orders.
- Site-specific baseline. The full model is compared with full averaging of all parameters.
Every method gets the same splits, the same hyperparameter search budget (equal trial counts), and the same number of communication rounds. Local epochs and batch sizes come from the ranges in the eICU case study. Beating the strongest baseline means two things. Census MAE must be lower at 4 and 8 hours, with confidence intervals from a block bootstrap over days excluding zero. Intervals must also be better calibrated, without a loss at 1 to 2 hours.
Why not just…?
Why not just share hourly census counts and train one pooled time-series model? This is the most serious threat. Counts carry little privacy risk, so if this alternative works the federated machinery is unnecessary. It does not know who is in the building, however. Twenty patients at 2 a.m. who are mostly admitted boarders behave very differently from twenty who are mostly low-acuity patients awaiting discharge. The patient-level ablation settles the question. If the aggregate-only bottom-up model matches the full model, the threat wins and we will say so.
Why not just train each hospital's model locally? Large sites may need nothing more. The eICU case study (Federated Learning in Multi-Center Critical Care Research: A Systematic Case Study using the eICU Database) found that small hospitals cannot train stable models alone. We will report error as a function of site size, and we expect federated gains to be concentrated in the smallest quartile of sites.
Why not just recruit representative sites, or average everything with FedAvg? Recruitment drops unusual hospitals, and those are often the community EDs that FAIR-EC argues must be included. Full averaging forces one baseline departure curve onto hospitals whose boarding delays differ for structural reasons. The comparison that settles this is per-site occupancy error at the dropped or atypical sites, with the shared-effects/local-baseline split set against both recruitment and full averaging.
Proposed Experiments
Experiment 1: forecasting value of patient-level state. On MIMIC-IV-ED we forecast census and boarding counts at 1, 2, 4 and 8 hours. We report MAE, the fraction of forecasts within 4 patients (one nurse at 1:4 staffing), and 80% interval coverage. This tests the first claim with no federation involved.
Experiment 2: federated occupancy across many sites. On eICU we split hospitals by time into train and test periods and forecast occupancy at 4, 8 and 24 hours for all sites. We report per-site error, with emphasis on small and atypical hospitals.
Experiment 3: coefficient stability. For the interpretable model, we compare federated coefficient estimates with the centralized fit, following the bias analysis in Federated Learning in Healthcare: A Benchmark Comparison of Engineering and Statistical Approaches for Structured Data Analysis. This checks whether clinicians could trust the effects they read.
Experiment 4: real multi-site ED validation. Through FAIR-EC sites, we test whether the site-specific baseline transfers to EDs with different triage scales.
Decisive outcomes. The proposal succeeds if three things hold together: bottom-up beats pooled top-down at 4 to 8 hours, the full model beats its aggregate-only ablation, and federated training beats local training at small sites. If the first two fail, the honest conclusion is that ED crowding forecasting needs shared counts, not federated learning.
References
- Emergency Department Optimization and Load Prediction in Hospitals
- Developing interpretable machine learning models to predict length of stay and disposition decision for adult patients in emergency departments
- Importance of patient bed pathways and length of stay differences in predicting COVID-19 hospital bed occupancy in England
- Client Recruitment for Federated Learning in ICU Length of Stay Prediction
- Comparing Federated Stochastic Gradient Descent and Federated Averaging for Predicting Hospital Length of Stay
- Federated Learning in Multi-Center Critical Care Research: A Systematic Case Study using the eICU Database
- Federated Learning in Healthcare: A Benchmark Comparison of Engineering and Statistical Approaches for Structured Data Analysis
- FAIR-EC: A Global Research Network for Fair, Accountable, Interpretable, and Responsible AI in Emergency Care (Preprint)
- Tackling Data Heterogeneity in Federated Time Series Forecasting
- Unlocking Dynamic Inter-Client Spatial Dependencies: A Federated Spatio-temporal Graph Learning Method for Traffic Flow Forecasting