New · sees around the corner

Know where the score is heading

NEDOCS tells you where your ED is right now. The 2-hour forecast projects where it's going — a point estimate with an honest uncertainty band, the band it's likely to land in, and the probability you escalate before the next handoff.

+2h
Default horizon
80%
Prediction interval
0
Black boxes
2-hour forecast · projected NEDOCS Overcrowded
128
from 96 now  ·  rising +16/hr
80% range 108–148  ·  method: model
62% chance of crossing into the Overcrowded band within 2 hours.
  • ↗ Rising trajectory (+16/hr over recent scores).
  • 🛏 ED at 134% of bed capacity.
  • 🌧 Active severe-weather alert: Winter Storm Warning.
  • 🚗 Major nearby incident: Accident on I-95 — possible trauma/EMS inflow.
  • 🫁 Regional respiratory activity: High.

A forecast you can actually trust on a busy floor

Most "predictions" hand you a single number with no error bars and no reasoning. This one shows its work, quantifies its own uncertainty, and refuses to overpromise.

Interval, not just a point

Honest uncertainty band

Every forecast carries an 80% range. A tight band means commit; a wide one means watch. You always know how confident the model is.

Decision-ready

Probability of escalation

The number charge nurses act on: the modeled chance of crossing into a higher NEDOCS band before the next 2-hour handoff window.

Explainable

The drivers are right there

Trajectory, occupancy, boarding, time of day, weather, respiratory pressure — the forecast lists what's pushing it, in plain language.

Self-checking

Must beat the naive baseline

The machine-learning model is backtested every night. It only goes live if it beats a simple "score persists" baseline — otherwise we serve the transparent trend.

Always-on fallback

Works from day one

A least-squares trend baseline runs immediately on your recent scores, so you get a forecast before there's enough history to train a model.

Context-aware

Looks outside the building

Live weather, air quality, nearby traffic incidents, and regional respiratory surveillance nudge the forecast — because surges rarely start inside the ED.

Two layers, always honest

The forecaster never bets everything on one model. A transparent baseline is always running; a learned model only takes over when it has earned the right.

1

Trend baseline

A least-squares slope over your recent scores, projected forward with a band derived from how noisy that trajectory has been. Transparent, instant, and the floor every fancier model must clear.

2

Gradient-boosted model

Trained on lagged scores, velocity and acceleration, operational inputs, occupancy, time-of-day and day-of-week, plus natural-event features — to predict the score one horizon ahead. Backtested on a held-out tail of your real history.

3

Environmental nudge

Live weather, AQI, nearby traffic incidents, and respiratory pressure apply a small, bounded, fully-explained adjustment on top — never a hidden coefficient, always a named driver.

Built on openly available data

Surges are driven by what's happening in your community. We pull from public, free feeds keyed to your hospital's location — and tell you exactly what we used.

Free · no key

NWS weather & alerts

National Weather Service api.weather.gov — current conditions and active severe-weather alerts. Heat, hard cold, and storms all move ED volume.

Free key · optional

AirNow air quality

EPA/AirNow AQI by lat/long. Poor air quality drives respiratory and cardiac presentations.

Free key · optional

TomTom traffic incidents

Crashes, road closures and jams within a radius of your ED (global coverage). A serious wreck nearby can mean trauma and EMS inflow within the hour.

Regional feed

Respiratory surveillance

A 0–3 activity level designed for CDC NSSP / RESP-NET regional ED-visit signals — respiratory pressure leads local surges by days.

Computed locally

Natural events

Lunar phase, day length, and solar position — pure math from date and location, no API call.

Control feature

Yes, we tested the full-moon myth

Every ED has a full-moon legend. The peer-reviewed evidence is essentially null — so we include lunar phase deliberately, as a control. If it earns no weight in the model, that's a credibility win for everything else. Meanwhile day length and the overnight window — the genuinely useful "natural" signals — stay in the mix. We let the backtest decide, not folklore.

🌕

Included, measured, and honestly reported.

Administrators stay in control

The whole pipeline is visible and tunable from Admin → Forecasting: a backtest scoreboard (model MAE vs. baseline, sample sizes, last-trained), a live per-hospital forecast and external-context table, and one-click retraining and data refresh.

  • Toggle forecasting, external data, and natural-event features
  • Drop in the NWS contact string and AirNow key without a redeploy
  • See whether the ML model is active or the trend baseline is serving
  • Retrain on demand and watch the improvement-over-baseline number

Model status

Active

Beat baseline by 21%

Backtest MAE

11.4

vs baseline 14.5

Training samples

3,420

684 held out

Horizon

120 min

Configurable

Common questions

Is this a clinical prediction?

No. It's an operational projection of the NEDOCS score from your own trajectory and context. It's a planning aid for charge nurses, bed management, and leadership — always paired with local policy and clinical judgment, never a substitute for it.

What if there isn't enough history yet?

The transparent trend baseline runs from your very first recent scores. The machine-learning model only activates once there's enough labeled history and it has beaten the naive persistence baseline on a backtest — until then you still get a useful forecast, just from the baseline.

Do I have to send any data to third parties?

No patient data ever leaves your facility. The external feeds are inbound — we fetch public weather/AQI for your hospital's location. Natural-event features are computed locally with no network call at all.

Why include lunar phase if it doesn't work?

Precisely because it's the famous myth. Including it as a control lets the backtest show, with your own data, whether it carries any weight. If it doesn't, that's evidence the rest of the model isn't chasing folklore.

How accurate is it?

Accuracy is reported honestly on the admin Forecasting page as mean absolute error on a held-out tail of your real history, alongside the naive baseline it must beat. The model is retrained on a regular cadence so that scoreboard stays current.

See the next two hours, today

The forecast lights up as soon as your ED is scoring regularly — and gets sharper as your history grows.

Request access Try the dashboard