Three hospitals sharing one set of doors

Hospitals are usually divided up by specialty or by acuity. Neither of those divisions predicts anything useful about flow. The division that does is much simpler: did the hospital choose when this patient arrives?

Unplanned demand chooses its own time, arrives in clusters, and cannot be told to come back on Thursday. Planned demand arrives exactly when the hospital said, because the hospital said it. Returning demand is the most predictable of all, running on cycles measured in days or months, often for years.

Those three have almost nothing in common operationally, and the consequences of mixing them are severe and entirely predictable. The unplanned stream has variability that has to go somewhere. The planned stream is the only one that can be sent home. So unless something stops it, the variability in the first lands on the second, and it lands as a cancellation.

That is the same mechanism this study has described four times in other currencies. Variability has to be absorbed by spare capacity, by a queue, or by delay. Here the spare capacity is a protected theatre list, the queue is the waiting list, and the delay is the operation that does not happen on the day it was promised.

Twelve entry points, sorted by stream. Two of them are worth pausing on.

Direct admission from a clinician is the entry point most often absent and the most valuable to build. A patient whose diagnosis is already made, sent in by somebody who knows them, does not need to be re-triaged, re-clerked and re-investigated in an emergency department, and routing them through one wastes several hours of everybody's time and a bed space, purely because no other door exists. Building that door is cheap. It also turns the referrer from somebody with no relationship to the hospital into somebody with a phone number, which is the stakeholder argument arriving as a floor plan.

The last row is the one that breaks the tidy three way split. A readmission is a returning patient entering through the unplanned door, and it is recorded as a new emergency attendance rather than as the outcome of a discharge that did not hold. The hospital therefore counts its own failure as fresh demand, which is a measurement problem with real consequences, because the response to rising emergency attendance is to expand the emergency department.

The streams collide at every resource they share, and the list is short enough to state: the front door, imaging, the laboratory, theatres, beds, pharmacy, the lifts, the corridors and the discharge lounge.

Imaging is the clearest and most instructive. An emergency scan cannot wait, so it goes ahead of the outpatient list. The outpatient list runs late. The clinic that was waiting on those results runs late. Patients who took a day off work wait longer, some of them leave, and the follow up appointments generated by that clinic are pushed into next month. One unplanned arrival has propagated into three other streams, none of which will record the reason.

The general form is worth stating because it is not obvious. When a volatile stream and a stable stream share a server, the stable one inherits the volatility of the volatile one. It does not average out. The stable stream simply stops being stable, and everybody downstream of it plans against a schedule that no longer predicts anything.

Which gives the design rule that most of this page is about. Share a resource between streams only when you are willing for the calmest of them to behave like the most volatile.

Put arithmetic on it, because this is a decision that is usually taken on instinct and the instinct is wrong.

Take a theatre suite where unplanned work needs around six sessions in an average week, arriving unpredictably. Ring fence some number of sessions for it. Anything the unplanned stream needs beyond the fence is taken from a planned list that had a named patient on it who has been fasting since midnight.

Read the top row first. With no protection at all, six planned sessions a week are lost, and not one fenced session sits idle, because there is no fence. That configuration has perfect utilisation. It will appear on every report as the most efficient possible use of the theatre suite, and it cancels six operations a week.

Now read down. Fencing at the mean roughly halves nothing and cuts cancellations from six to one. Fencing modestly above the mean takes it to one cancellation every three weeks. The cost of that is two or three sessions a week of protected time that goes unused, which is visible, measurable, attributable and deeply unpopular.

That is the entire trade, and it is the fifth time this study has arrived at the same sentence. The buffer looks like waste, is not waste, and will be removed by somebody reading a utilisation report unless the reason it exists is written down next to it.

There is an apparent contradiction on this site and it is worth resolving directly rather than leaving the reader to notice it.

The sizing page argues that pooling is powerful: a large unit sustains far higher occupancy than a small one at the same service standard, which is why small hospitals are expensive. This page argues that streams should be separated, which means deliberately creating smaller pools and giving up exactly that efficiency.

Both are correct, and the rule that reconciles them is precise. Pool demand that has the same variability. Separate demand that does not.

Combining two streams with similar arrival patterns genuinely averages them out, and the larger pool is better for everybody in it. Combining a volatile stream with a stable one does not average anything: it transfers the volatility onto the stable stream, which had none, and the stable stream loses far more than the volatile one gains. The pooling benefit is real and it is smaller than the contamination cost.

So the answer is not one pool and it is not twelve. It is a pool per variability class. Pool all the unplanned work together, because it shares a pattern and benefits enormously from size. Pool the planned work together and protect it. And give the returning stream its own space entirely, because it is the most predictable demand in the building and putting it anywhere near an emergency department wastes that predictability completely.

Exits get a fraction of the attention entrances do, and they determine more. Seven of them, and three deserve comment.

Discharge to a relative is the commonest significant exit and it is a clinical handover to somebody with no training, usually delivered in a corridor on the day, while a bed is being waited for. The hospital records a completed episode. The family records the beginning of a fortnight they were not prepared for.

Leaving without being seen is close to a pure measure of whether the front of the hospital works, it is trivially countable, and it is almost never treated as a clinical figure. The people who do it walk out carrying an unassessed problem and cost the hospital nothing, which is exactly why it is not managed.

And the sixth exists because of this market specifically. Where care is paid for out of pocket, a patient who is recovering slowly can watch the bill grow, and some of them will leave before they should. It is recorded as self discharge against advice, which reads as a patient choice and is frequently a price. The countermeasure is not a policy statement. It is cross referencing that figure against length of stay and bill size and looking honestly at what comes back. Detaining a patient over an unpaid bill is ruled out in writing by this group, and that rule means nothing at all unless somebody is watching the number that would reveal the softer version of the same thing.

The neat division into three streams is useful and it conceals something, which is that the streams feed each other.

A discharge creates a follow up, so the planned exit generates returning demand. A poor discharge creates a readmission, so the planned exit generates unplanned demand. A cancelled follow up creates a deterioration, which arrives months later through the emergency door. The quality of the exits determines the size of the entries, on a lag long enough that nobody connects the two.

This has a consequence that runs against ordinary intuition. The highest leverage work on emergency demand is not in the emergency department. It is in discharge quality, follow up reliability and the returning stream, all of which are the calmest and least urgent parts of the hospital and therefore the first places anybody takes staff from when the emergency department is under pressure.

Which is the classic failure loop and it is worth naming as one. Pressure at the front door leads to resources moving from the back. Resources moving from the back degrades discharge and follow up. Degraded discharge and follow up increase demand at the front door, some months later, by which time nobody remembers the decision that caused it.

All of this is operating policy except the part that is not, and the part that is not is the part that matters most, because it closes at massing.

Separate streams need separate routes. That means a distinct ambulance approach, a public entrance that is not the emergency entrance, a maternity route that does not pass through a general emergency department, a day unit reachable without walking through the acute hospital, and a mortuary route a family can be walked along that does not cross the route a surgical patient takes to theatre.

Every one of those is frontage, circulation and position. Every one of them is cheap on a drawing and unavailable afterwards, because where clean and dirty, public and patient, or emergency and elective share a corridor, they will share it for sixty years. Policy can schedule a shared corridor. It cannot unshare it.

So the three stream analysis is not primarily an operations exercise. It is an input to the massing gate, and it is one of the few inputs at that gate that can be worked out properly before anybody knows anything else about the hospital, because it depends on the shape of demand rather than on the size of it.

Three, and the first is the reason good flow designs get dismantled in year four rather than being rejected in year one.

Every fence produces idle time that is measured precisely, appears monthly and can be attributed to a named person. The cancellations it prevents never happen, so they appear nowhere. Given long enough, and a cost pressure, the fence comes down, cancellations resume, and nobody connects the two events because eighteen months separate them. The only defence is to publish the cancellation rate beside the utilisation rate, permanently, so that the pair is read together and a theatre suite running at a hundred per cent while cancelling six lists a week is recognisable as the failure it is.