Hospital Demand Forecasting and Seasonal Capacity Planning

Indian hospital demand has strong, largely predictable annual structure. The monsoon and post-monsoon months bring vector-borne illness — dengue and malaria in particular — with associated platelet demand, longer medical ward stays, and pressure on blood bank services. Winter raises respiratory admissions, particularly in northern cities where air quality deteriorates sharply. Summer brings heat-related presentations and gastroenteritis. These patterns repeat annually and are visible in three years of your own admission data.

Outbreak years are the complication. A severe dengue season looks nothing like a mild one, and a forecast built on a three-year average will under-provision in a bad year and over-provision in a good one. The practical response is to forecast the baseline seasonal pattern separately from the outbreak component, and to treat the outbreak component as a scenario rather than a point estimate — a planned response that activates on a trigger, not a number in a plan.

Local public health surveillance is the trigger source. District vector-borne disease reports, municipal case counts, and your own emergency department presentation mix all lead admissions by one to three weeks. A hospital that watches its own ED presentation pattern for the leading edge of a seasonal rise gets a fortnight of warning, which is enough to move rosters and pre-position consumables.

Within the week, hospital demand is highly structured and highly predictable. OPD peaks on Monday and after any closure; elective surgery clusters early in the week so that post-operative days fall on full staffing; discharges concentrate mid-morning and mid-week; emergency presentations rise in the evening and on weekends. These patterns are stable enough that day-of-week factors estimated from a year of data will hold.

Festivals and public holidays are the largest single-day disruptions and the most commonly mishandled, because they move against the Gregorian calendar. Diwali, Eid, regional new year observances, and local festivals each shift elective volume, OPD footfall, and staff availability, and their dates change every year. A forecast that models these as fixed calendar dates will be systematically wrong; they need a holiday calendar table maintained annually, with a flag for the days before and after, which behave differently from the day itself.

The post-holiday rebound deserves separate treatment. Elective demand deferred over a festival period does not disappear; it arrives in the following week, often compressed. Hospitals that staff down for the festival and back up to normal afterwards routinely get caught by the rebound. Modelling the days after a major festival as their own category, rather than as ordinary days, resolves most of it.

For short horizons the simplest methods perform remarkably well. A seasonal naive forecast — next Tuesday will look like the last few Tuesdays, adjusted for trend — is a legitimate baseline and frequently beats more elaborate approaches for a two-week staffing view. Always compute it, because it is the benchmark any sophisticated method must beat to justify its complexity and maintenance cost.

For medium horizons of one to six months, a method that separates trend, seasonality, and calendar effects is usually worth the additional effort. The important property is not sophistication but interpretability: when the forecast is wrong, someone needs to be able to say which component was wrong. A model that produces a number nobody can decompose will be abandoned the first time it misses badly, regardless of its average performance.

For long horizons, forecasting shades into scenario planning and should be labelled as such. Two-year capacity questions depend on catchment growth, competitor openings, empanelment changes, and consultant recruitment — none of which is extrapolable from historical admissions. Present three scenarios with their assumptions stated rather than a single projected line, because the assumptions are what the board should actually be discussing.

A forecast that is never scored will drift into decoration. Record every forecast at the moment it is issued, alongside the horizon and the actual outcome when it arrives, and review accuracy on the same cadence as the forecast itself. The most common failure in hospital forecasting is not inaccuracy but the absence of any record of what was predicted, which makes improvement impossible.

Use error measures that match the decision. Mean absolute percentage error is intuitive and communicates well, but it behaves badly for low-volume series — a ward averaging two admissions a day will produce alarming percentages from trivial absolute errors. For low counts, mean absolute error in units is more honest. Track bias separately from magnitude: a forecast that is consistently ten per cent low is a different and more fixable problem than one that is erratic.

Segment the error. Overall accuracy conceals that the forecast is excellent on weekdays and poor at weekends, or reliable for medicine and unusable for paediatrics. Reporting error by ward, by day type, and by horizon shows where to invest effort, and often reveals that one badly behaved segment is dragging down an otherwise sound forecast.

A forecast becomes operational only when it is converted into the unit the decision-maker works in. Nursing does not roster against predicted admissions; it rosters against required nurses per shift per ward. The conversion runs through explicit assumptions: predicted occupancy, target nurse-to-patient ratio by ward type, expected leave and absence rate, and skill mix requirements. Write those assumptions down, because they are usually where the disagreement actually lies.

The same applies to pharmacy and consumables. A predicted rise in dengue admissions converts to platelet requirement, IV fluid consumption, and specific drug demand through consumption-per-case ratios derived from past seasons. Lead time then determines when the order must be placed. Where the pharmacy module and the admission forecast draw on the same platform — HealUDoc being one such — this conversion can run as a standing calculation rather than a seasonal scramble.

Build in a response ladder rather than a single plan. Level one might be reallocating float staff and drawing on buffer stock; level two, opening overflow beds and activating overtime; level three, deferring elective admissions and invoking mutual aid arrangements. Define the trigger for each level in advance, in terms of observed occupancy or presentation rate, so the escalation decision is made calmly in a planning meeting rather than at eleven at night in a full emergency department.

Every forecast rests on the assumption that the future resembles the past in the relevant respects, and hospitals regularly break that assumption. A new consultant joining or leaving, a competitor opening nearby, a change in scheme empanelment, a new specialty service, or a road closure altering the catchment all invalidate history in ways no model detects. Maintain a change log of these events alongside the forecast so that a miss can be attributed rather than mysterious.

Structural breaks require judgement, not more data. When empanelment for a major scheme changes, the sensible response is to discard the pre-change series for the affected segment and rebuild once a few months of new behaviour exist, using clinician and operational judgement in the interim. Continuing to feed a model data from a regime that no longer exists produces confident and wrong answers.

Finally, be explicit about uncertainty in how the forecast is presented. A range with a stated confidence, or a central case with an upside and downside, invites the right conversation about contingency. A single number invites false precision, and the first time it is wrong, the credibility loss falls on forecasting as a practice rather than on the specific method that failed.