Forecasting Hospital Deaths and ICU Load from the Patients Already Admitted

During a respiratory surge, most of the intensive care unit (ICU) demand and in-hospital deaths of the next ten days will come from people who are already in a hospital bed today. Standard epidemic forecasts ignore that information. They treat hospitalizations and deaths as aggregate time series, or as compartments with population-average transition rates, and they extrapolate counts forward. The intended application is a single hospital system rather than a national hub. The proposal is to forecast these outcomes from the bottom up instead. Each current inpatient, and each new admission, gets a calibrated individual risk of ICU transfer and of death, computed from routine admission labs and vital signs. The hospital-level forecast is the probabilistic sum of those individual risks, plus a smaller term for patients not yet admitted.

The case for this rests on a known failure. Operational forecasts break down exactly when they matter. In the US COVID-19 Forecast Hub, 95% prediction intervals covered the truth less than half the time during the growth phases of major waves (Challenges of COVID-19 Case Forecasting in the US, 2020-2021). Four-week mortality forecasts carried about 76% more error than one-week forecasts (Evaluation of individual and ensemble probabilistic forecasts of COVID-19 mortality in the United States). One unmodeled driver is the severity mix of patients being admitted, which drifts within and across waves as variants, ages and vaccination status shift. That drift is visible in admission bloodwork days before it shows up in ICU or death counts.

Core Claim

At horizons of 1 to 14 days, a hospital system's ICU census and deaths are better forecast by summing calibrated individual risks over the current inpatient population than by any model of the aggregate count series. This holds most strongly during wave growth, when the severity mix moves fastest. Two ingredients carry the gain.

The first is that the full distribution of individual risk is used, not a summary. The second is that the individual risk model is re-calibrated online despite the fact that its labels, the outcomes, arrive with a delay of days to weeks.

The three directions play unequal roles. Epidemic forecasting supplies the target, the evaluation protocol and the scoring rules. Automated prognostic stratification supplies the mechanism: a per-patient risk model. Molecular risk stratification acts as a cautionary methodological lens. That literature routinely compresses biomarkers into an additive weighted score and then splits patients at the cohort median. Examples are the four-gene lung adenocarcinoma model (Predicting the survival of patients with lung adenocarcinoma using a four‑gene prognosis risk model) and the ferroptosis signature (Establishment of a ferroptosis-related gene signature for prognosis in lung adenocarcinoma patients). We test directly whether that habit destroys the population-level signal. We also test whether routine lab analytes can stand in for costlier molecular measurement, which is the bet made by OMICmAge quantifies biological age by integrating multi-omics with electronic medical records. Its EMRAge component is a Cox mortality model trained on 19 routine labs.

Relation to Prior Work

Two lines of work already put stratification inside epidemic models, and both motivate this proposal.

The first is Population stratification enables modeling effects of reopening policies on mortality and hospitalization rates. It splits a compartmental model into low- and high-contact groups by employment sector, and it reports a large drop in hospitalization error relative to unstratified models in Harris County: mean squared error of 8.04 versus 20.98 to 649.63. That supports the idea that heterogeneity matters. However, its strata are binary, static and non-clinical.

The second is Interpretable Sequence Learning for Covid-19 Forecasting. It learns time-varying transition rates into hospital, ICU and ventilator compartments, driven by mobility and policy covariates. Its rates remain population averages, with no patient-level information.

We differ from both in where the stratification comes from: measured patient physiology rather than demographics or policy. We also differ in what is stratified: the people actually in the hospital, not the general population.

Three works supply the individual-risk methods and constrain how we build them.

Characterization of non-monotonic relationships between tumor mutational burden and clinical outcomes shows that a log-linear Cox model and a median split can miss inverted-U hazard shapes that a small neural survival network captures. In melanoma, stratification reached p = 4.9e-5 with the network versus p = 3.7e-2 with the median split. Admission labs such as white cell counts and sodium plausibly behave this way, so we compare additive and flexible risk models.

Predicting the efficacy of immune checkpoint inhibitors monotherapy in advanced non-small cell lung cancer: a machine learning method based on multidimensional data found that cheap inflammatory blood ratios (neutrophil-to-lymphocyte and similar) carry much of the prognostic signal. Its clinical-only model reached an AUC of 0.89. This supports restricting inputs to labs that are drawn routinely.

Revisiting Clinical Outcome Prediction for MIMIC-IV warns that high discrimination can coexist with poor detection of rare events: an AUROC of 86.86% alongside a precision-recall AUC of 17.24%. We therefore score our individual models on calibration, not only discrimination, because a sum of miscalibrated probabilities is a biased forecast.

Research Program

The census-sum forecast. Each patient $i$ in the hospital at forecast time $t$ receives, from a survival model on routine labs, vitals, age and comorbidity codes, a probability $p_i(h)$ of dying (or entering ICU) within the next $h$ days. The forecast for deaths over $(t, t+h]$ has two parts:

$$D_{t,h} = \sum_{i \in \text{census}(t)} B_i\big(p_i(h)\big) + \sum_{s=t+1}^{t+h} A_s \, \bar{p}_s(t+h-s).$$

Here $B_i$ is a Bernoulli draw for patient $i$. $A_s$ is the forecast number of admissions on day $s$, taken from any admissions forecaster. $\bar p_s$ is the mean risk of future admissions, extrapolated from the recent risk distribution of admissions. Monte Carlo draws of both terms give quantiles directly. The first term needs no epidemic model at all, and it should dominate at short horizons. The second term is where uncertainty about epidemic dynamics enters, and it grows with $h$.

Calibration under delayed labels. The individual model is trained on earlier data, and its calibration will drift when a new variant arrives. The fix is to refit a recalibration layer each day, consisting of an intercept and a slope on the log-hazard. It is fit only on patients whose outcome window has closed, or treated as censored. Because those labels are 1 to 3 weeks old, we compare this delayed recalibration against two alternatives. One recalibrates using early proxies such as respiratory support escalation within 48 hours. The other does no recalibration. The research question is how much of the drift the faster proxies recover.

Score form. The same pipeline is run with four risk models: an additive linear Cox score, the same score dichotomized at the median (the molecular-signature convention), gradient-boosted survival trees, and a small neural survival network. A fifth variant uses age and sex only. This separates two questions: how much comes from the physiology, and how much from the shape of the risk model.

Data. Individual models are developed on MIMIC-IV, using its ED and ICU encounters. MIMIC-IV shifts dates per patient, so it cannot be aligned to calendar epidemic waves. For that reason, the forecasting evaluation needs a health system's EHR with true admission dates spanning several respiratory waves (COVID-19 2020–2022, influenza seasons). Securing such a partner extract is the main practical dependency of the project.

Before that extract is available, we run a semi-synthetic test on MIMIC-IV alone. We resample real patient trajectories into synthetic admission waves whose severity mix drifts on a known schedule. This keeps real lab–outcome relationships while letting us control the drift, so it can falsify the mechanism cheaply. Aggregate admission and death counts for the partner system's state can be paired with the US COVID-19 Forecast Hub dataset to put the gains in operational context.

Baselines

Naive flat baseline. This mirrors the COVIDhub-baseline: the last observed weekly value projected forward, with quantiles from historical changes. It is the right floor because this baseline beat 10 of 28 Hub mortality models (Evaluation of individual and ensemble probabilistic forecasts of COVID-19 mortality in the United States) and was a strong competitor during inflection points (Challenges of COVID-19 Case Forecasting in the US, 2020-2021). All results are reported as relative weighted interval score (WIS), the Hub's quantile-based probabilistic error, against this floor. For the strongest existing operational reference, the COVIDhub-ensemble reached a relative WIS of 0.61 for state and national mortality. That number is not directly comparable to a single hospital system. It indicates the magnitude of gain an aggregate ensemble achieves over naive. No reported result was found on Lacuna for hospital-system-level ICU census forecasting.

Aggregate time-series models. These are ARIMA and the parallel ARIMA–LSTM hybrid of Prediction of COVID-19 Data Using an ARIMA-LSTM Hybrid Forecast Model. On China incidence data that hybrid reported an MSE of 4049.9 against 4681.4 for ARIMA. Both models are fit to the same hospital-system series, with the same training windows and the same hyperparameter search budget as our model's tunable parts.

Ratio-convolution estimator. This is the strongest competitor on this specific question. Deaths are forecast as a fitted delay distribution convolved with admissions, multiplied by a rolling in-hospital fatality ratio. It already absorbs severity drift, but with a lag. It receives the same admissions forecasts our second term uses. That way, any gain we show comes from the census and risk information, not from better admissions forecasting.

Ensemble. An equally weighted quantile median of all the baselines, following the Hub's practice.

Ablations. Four ablations isolate the key ingredient: risk from age and sex only, the median-dichotomized score, mean risk instead of the full distribution, and no recalibration.

What counts as beating the baselines. Lower WIS than the ratio-convolution estimator and the ensemble at 7- and 14-day horizons, with a paired bootstrap over forecast dates that excludes zero. Equally important is 95% interval coverage within 10 points of nominal during growth weeks, defined in advance as weeks where admissions rose more than 20% week over week.

Why not just…?

Why not just use the rolling fatality ratio times admissions? This is the real threat. If severity mix drifts slowly relative to the outcome delay, the ratio estimator will match us. The advantage we claim exists only during fast drift, because the ratio is computed from patients whose outcomes are already observed. Those patients were admitted one to three weeks earlier, whereas admission labs are available on day one. The semi-synthetic experiment varies the drift rate directly and maps where the crossover happens. If the census-sum model does not beat the ratio estimator in the real growth weeks, the claim is wrong and we will say so.

Why not just add the age mix of admissions as a covariate? Age is the cheapest severity proxy, and the age-and-sex-only ablation is exactly this test. In hospitalized respiratory patients, a large share of severity variation plausibly lies in physiology rather than age; inflammatory ratios carry prognostic signal in other settings (Predicting the efficacy of immune checkpoint inhibitors monotherapy in advanced non-small cell lung cancer: a machine learning method based on multidimensional data). If the age-only variant captures most of the gain, the right recommendation is the simpler model.

Why not just summarize risk as the fraction of high-risk admissions, as molecular signatures do? Median splitting discards information about the tails. The tails drive deaths and ICU load, and the split thresholds shift between cohorts. Non-monotonic marker effects can also be hidden by additive scores (Characterization of non-monotonic relationships between tumor mutational burden and clinical outcomes). The dichotomized and mean-risk ablations measure this loss directly.

Why not just submit to an ensemble and let averaging fix overconfidence? Ensembles improved coverage at the state level but still had less than 50% coverage among many models during surge growth. Averaging models that all lack the same information does not supply that information. Our model is designed as an ensemble member, so we report whether adding it to the baseline ensemble improves WIS.

Proposed Experiments

The first decisive check is semi-synthetic. On MIMIC-IV resampled waves with a known severity drift schedule, the census-sum forecast should beat the ratio-convolution estimator by a margin that grows with drift speed and shrinks with horizon. If there is no advantage under fast drift, the project stops there.

The second check is the real-calendar evaluation on the partner system's EHR. We use rolling-origin forecasts made daily, for horizons of 1, 7 and 14 days, for ICU census, ICU admissions and in-hospital deaths. Scoring uses WIS, 50% and 95% interval coverage, and mean absolute error. Results are reported separately for growth, peak and decline weeks.

The third check covers calibration. Individual-model calibration slope and intercept are tracked by wave, comparing delayed-label recalibration, early-proxy recalibration and none, and these individual-level errors are linked to forecast bias.

The fourth check covers score form. The linear additive, median-split, tree-based and neural risk models are compared both on per-patient C-index and calibration and on downstream hospital-level WIS. The aim is to show whether the conventions of molecular signature building help or hurt when individual scores are aggregated into a population forecast.

A positive result would be lower WIS than both the ratio estimator and the ensemble at 7 days during growth weeks, together with restored interval coverage. That result would establish physiology-based census aggregation as a missing input to operational hospital forecasting.