Every winter, a bed manager somewhere walks into the morning huddle with a dashboard that is right. It says the hospital will be over capacity by Thursday afternoon. It said the same thing last Thursday. Last Thursday, the hospital went over capacity anyway.
Predictive analytics for hospital flow has become one of the most marketed uses of data in health care. Admission forecasts, discharge predictions, length-of-stay models, emergency department arrival curves. Many of them are reasonably accurate, because hospital demand is more regular than it feels from the inside. Mondays are busy. Winters are busy. Surgical schedules are known weeks ahead.
The problem is not that hospitals can’t see the crunch coming. It’s that a forecast only has value if someone has the authority and the slack to change what happens next. Most hospitals buy the first without the second.
The math that makes forecasts feel useless
The most important finding in hospital capacity planning is more than 25 years old. In 1999, a BMJ modelling study by Bagust and colleagues simulated emergency admissions to a hypothetical acute hospital in England. It found that risks of having no bed for an emergency patient became discernible once average occupancy exceeded about 85%, and that hospitals running at 90% or more should expect regular bed shortages and periodic crises.
The reason is variability. Admissions arrive in clumps, not evenly. A hospital with spare capacity absorbs the clumps. A hospital running near full has nowhere to put them, so each surge spills into corridors, cancelled operations and ambulances waiting outside.
This has a sharp implication for predictive analytics. A forecast tells you a surge is coming. In a hospital running at 95% occupancy, there is almost nothing left to do with that information. You cannot discharge patients who are not ready. You cannot staff beds you do not have nurses for. The forecast becomes a well-calibrated description of a disaster you have no means to prevent.
Prediction is a lever. It needs something to push against.
What the sepsis model taught us about alerts
There is a second trap, and hospitals have already walked into it with clinical prediction tools.
In 2021, researchers at Michigan Medicine published an external validation of the Epic Sepsis Model in JAMA Internal Medicine. Across 38,455 hospitalisations, the model alerted on about 18% of patients, yet it missed 67% of those who developed sepsis. Its area under the curve was 0.63, a modest level of discrimination for a tool deployed so widely.
Capacity models are not sepsis models, but the lesson carries over. A tool can look good in the vendor’s data and behave very differently in yours. And a tool that fires often without changing outcomes teaches staff to ignore it. A capacity dashboard that shows red every weekday afternoon becomes wallpaper.
Any hospital buying a flow prediction tool should insist on validation on its own historical data before go-live, and should keep measuring after it. Patient mix, discharge practices and referral patterns differ between sites. A model is only as good as its fit to the building it runs in.
Where the value actually sits
If the forecast is rarely the bottleneck, what is? Usually one of three things.
Discharge, not admission. Most flow problems are exit problems. Patients medically ready to leave wait for a pharmacy review, a transport slot, a care home place or a consultant’s round. A prediction that a given patient is likely to be discharged tomorrow is only useful if it triggers the pharmacy, transport and social care work today.
Decision rights. A forecast that goes to an analyst, then a manager, then a meeting, arrives too late. The hospitals that get value from prediction tend to have agreed in advance what happens at each threshold: which elective lists are reviewed, which escalation beds open, who can make the call without asking. The model triggers a playbook, not a discussion.
Slack. This is the uncomfortable one. Every finance director is under pressure to run beds full, because an empty bed looks like waste. But the Bagust finding says that some empty beds are the price of reliability. A predictive tool is most useful in a hospital that has deliberately kept some slack to deploy. That is an expensive message to deliver to a board.
The incentive problem
Here is the trade-off rarely named in vendor pitches. Predictive analytics is sold as a way to run hospitals closer to full, safely. In practice, it is most valuable in hospitals that have accepted they should not run completely full.
A system that buys a forecasting tool to justify cutting beds will get accurate warnings of crises it can no longer avoid. A system that buys one to decide when and where to deploy flexible capacity, staff and discharge support will get something close to what was promised.
The technology is the same. The operating model decides which outcome you get.
A worked example: two hospitals, one forecast
Imagine two acute hospitals of similar size buy the same capacity forecasting tool in the same autumn. Both validate it on their own data. Both find it predicts next-day emergency admissions and discharges within a tolerable margin.
Hospital A runs at an average occupancy in the mid-90s. When the tool flags a surge for Thursday, the site manager sends an alert to ward leaders on Tuesday afternoon. Ward leaders already know they are full. There is no discharge lounge, no escalation ward that can open without agency staff, and the elective surgery list has been booked for weeks. The Thursday meeting discusses the forecast, confirms it, and then does what it always does: cancels some operations on the day and holds patients in the emergency department. The forecast was right. Nothing changed.
Hospital B runs a little lower on average, having decided to keep a small buffer. It has agreed, in advance, a set of triggers. When the tool forecasts occupancy above a threshold two days out, three things happen automatically. Pharmacy prioritises discharge medications for patients predicted to go home. Two elective lists that were provisionally booked with lower-urgency cases are swapped to day cases. The escalation ward is pre-staffed from a bank rota. On Thursday, the surge arrives. It is uncomfortable, but it is absorbed.
The difference between the two is not the model. It is what the organisation built around it: a buffer, pre-agreed actions and people authorised to take them. Hospital A bought a forecast. Hospital B bought a forecast and changed its operating model.
The strongest counterargument
The obvious objection is financial, and it is serious.
Empty beds cost money. Staff, heating, cleaning and overheads do not disappear when a bed is unoccupied. In systems paid by activity, an empty bed is lost revenue; in systems funded by budgets, it can look like inefficiency to an auditor or a minister. Waiting lists for planned surgery are long in many countries. Holding beds in reserve for emergencies that may not come means fewer planned operations for patients who are already waiting.
Seen this way, predictive analytics is precisely the tool that lets a hospital run closer to full without crisis, and asking for slack defeats the purpose.
The argument partly holds. Better forecasting can reduce the buffer a hospital needs, because it replaces blanket caution with targeted readiness. A hospital that knows Mondays and January are heavy can plan for them and use capacity more fully on quieter days. That is a real gain.
But it cannot abolish the buffer. The Bagust modelling showed that the problem is variability, not just averages, and forecasts reduce uncertainty without eliminating it. There will always be days that exceed the prediction. A hospital running with essentially no slack converts every forecast error into a crisis. The honest version of the pitch is: forecasting lets you run somewhat fuller, safely. Not completely full.
Second-order effects
Forecasts can shift risk to other parts of the system. A hospital that improves its own flow by discharging faster may push pressure onto community services, care homes and families. If those services are also full, patients may come back. Readmission rates are a useful check on whether a flow programme is solving a problem or moving it.
The dashboard becomes the job. In some hospitals, flow management turns into monitoring: more screens, more huddles, more status reports. Staff time spent discussing the forecast is time not spent unblocking discharges. Tools should reduce meetings, not multiply them.
Clinicians may game predictions. If a predicted discharge date becomes a target that managers chase, clinicians may feel pressure to discharge patients to meet it, or may set conservative dates to avoid being chased. Either way, the prediction starts to shape the outcome it was meant to forecast.
Data quality becomes a clinical issue. Forecasting depends on accurate, timely data about admissions, bed status and discharges. In many hospitals, bed status is updated late or inconsistently. A tool built on stale data will be confidently wrong.
The question vendors rarely get asked
There is one more test worth applying to any capacity tool: does it make the hospital’s own knowledge easier to act on, or does it replace that knowledge with a number nobody can interrogate? Experienced bed managers and charge nurses already carry a rough forecast in their heads. They know which consultant discharges late on Fridays and which ward always has a patient waiting for a care home place. A good tool should surface those patterns in data and let staff argue with them. A poor one produces a single red number that staff are expected to trust.
The difference matters for adoption. People act on forecasts they understand and help refine. They ignore forecasts that feel imposed, especially when the forecast is wrong on a day that matters. Explainability here is not an academic nicety. It is the difference between a tool staff use at 7 a.m. and one they close.
What this means for decision-makers
For hospital boards and executives, the key decision is not which vendor to choose. It is what occupancy level the organisation is willing to run at, and whether it will protect a buffer when finances are tight. Without that decision, a forecasting tool is an expensive way to watch the same crisis.
For operational leaders, it means building a trigger playbook before go-live: specific thresholds, specific actions and named people who can act without escalation. Test it on historical surges to see whether it would have helped.
For clinicians, it means engaging with discharge prediction as a planning aid rather than a target. The most valuable use is often identifying the small tasks, a scan report, a medication review, a transport booking, that can be done a day early.
For ministries and funders, it means recognising that funding models that penalise spare capacity undermine the value of any flow technology they encourage hospitals to buy.
What to watch next
Watch for evaluations that measure outcomes, not accuracy: fewer cancelled operations, shorter emergency department waits, fewer patients boarded in corridors. Accuracy metrics are the vendor’s concern. Outcome metrics are the hospital’s. The tools that prove themselves on the second will be worth scaling. The ones that can only show the first are dashboards, not interventions.
The takeaway
Before buying the next capacity model, a hospital leadership team should ask a blunt question: if this tool told us with certainty that we would be over capacity on Thursday, what would we do differently on Tuesday? If the answer is “nothing we are not already doing,” the problem isn’t predictive. It’s structural.
Good forecasting will help hospitals. But only the ones that have built the discharge processes, decision rights and spare capacity to act on it.
If your hospital uses capacity forecasting, has it changed a decision, or only described the outcome? We’d like to hear from you.
