Machine Learning for Pharmacies

The stock problem that eats margin

A chain of eight community pharmacies carries thousands of product lines, a good share of which sell a handful of units a month. Buyers over-order the slow lines to avoid letting a patient down, under-order the seasonal ones because last year's spike was forgotten, and write off expired stock every quarter. Meanwhile branch staff are phoning each other to borrow a pack.

Machine learning for pharmacies is mostly about this. Better forecasts per line per branch, a smarter view of which branch should hold what, and earlier warning about stock heading towards expiry. It is unglamorous and it goes straight to margin.

Where machine learning fits in a pharmacy

• Demand forecasting by line and branch. Accounting for seasonality, local prescribing habits, promotions and the long tail of slow movers.
• Repeat prescription prediction. Estimating when regular patients' next prescriptions will arrive, which smooths dispensing workload and ordering.
• Expiry risk. Flagging stock that will not sell before its expiry date while there is still time to move it to a busier branch.
• Staffing to footfall. Forecasting busy periods for the counter and the dispensary, including the predictable surges around bank holidays and flu season.
• Shortage early warning. Spotting supply problems from wholesaler lead times and failed orders before they become patient-facing.
• Unusual dispensing patterns. Highlighting anomalies in controlled drug records for a pharmacist to review.

Why pharmacy demand is harder than retail demand

Pharmacy forecasting has quirks that ordinary retail tools handle badly. Many lines are intermittent: zero sales for three weeks, then six packs. A standard forecasting method will predict a fraction of a pack every day, which is useless for ordering.

Branded and generic substitution also muddies the data, because a switch in the prescribing system or at the wholesaler moves demand from one line to another overnight. And drug shortages mean past sales sometimes reflect what you could get rather than what patients needed.

1. Group equivalent products so substitutions do not look like demand collapsing
2. Use methods designed for intermittent demand on slow lines
3. Mark shortage periods so the model does not learn from them
4. Forecast repeat prescriptions separately from over-the-counter sales

A worked example on waste

Say the eight-branch chain writes off a few thousand pounds of expired stock a year and ties up far more in slow lines. If a model identifies at-risk stock ninety days before expiry, most of it can be transferred to a branch that will sell it. The value is not the clever forecast; it is the weekly transfer list that a stock controller actually works through.

The arithmetic is simple enough to do before building anything: what did you write off last year, how much of it was predictable, and would anyone act on a list? If the answers are small, unclear and no, save your money.

A forecast that nobody orders from is a report. Build the thing the buyer uses on Monday morning.

Where the line is on clinical use

Everything above is operational. Machine learning that suggests clinical interventions, flags drug interactions for decision-making or triages patients is a different category, closer to regulated medical software. Pharmacy systems already carry clinical checking from approved sources, and we would not build a model to second-guess them.

Patient data still needs care even for operational work. Repeat prescription prediction uses dispensing history, which is health data, so it needs a lawful basis, minimisation and secure handling under UK GDPR. In most cases the model does not need to know who the patient is, only the pattern.

Staffing the counter and the dispensary

Wage costs are the other large line on a pharmacy's accounts, and busy periods are more predictable than they feel on the day. Monday mornings after a bank holiday, the first cold snap of autumn, the week before Christmas when repeat prescriptions arrive early: these patterns repeat every year and are visible in till and dispensing data.

A footfall and dispensing volume forecast by branch and half-day lets an area manager put a locum or a dispensing assistant where the queue will be, rather than where it was last week. It will not be perfect, and an unexpected local surgery closure will throw it. It is still better than a rota copied forward from last month.

When a single branch should not bother

An independent pharmacy with one branch and a good dispenser who knows the regulars probably does not need machine learning. A sensible minimum and maximum per line, reviewed quarterly, will capture most of the benefit. The maths changes with scale.

How we would start

At SpiderHunts we begin with two years of sales, dispensing and ordering data, which usually lives in the pharmacy system and the wholesaler portals. We measure how accurate a simple forecast would already be and where the real errors are. Often the slow lines, not the fast ones, are where the money sits.

Our post on forecasting demand with your own data explains the general approach, and our machine learning service covers building forecasts into the ordering tools buyers already use.

Frequently asked questions