Analytics & BI
Dashboards are the most-built and least-used artefact in health IT. Why most of them fail, and the one question that separates a decision tool from expensive wallpaper.
In one line
Business intelligence turns operational data into something a human can act on. In health it produces more dashboards per decision than any other industry, and the reason is always the same: the dashboard was built to display data, not to answer a question.
The question that decides everything
Before building anything, ask: what decision does this change, and who makes it?
If the answer is vague — "leadership needs visibility", "the board wants a dashboard" — stop. You are about to build wallpaper. Expensive, well-crafted, correctly-coloured wallpaper that somebody opens twice and never again.
A useful analytics artefact has a named person who will do a specific thing differently depending on what it shows. If nobody's behaviour changes based on the number, the number is decoration.
That's the whole test, and it eliminates most dashboard requests before a line of SQL is written.
Why health BI fails specifically
It shows what's easy, not what matters. Bed occupancy is easy — it's a count. Whether the discharge was safe is hard. So dashboards fill with counts, and the counts become the goals.
It measures the numerator. As always in this domain: the patients who attended, the results that were recorded. A "diabetes control" dashboard describing only monitored diabetics is a dashboard about your monitoring, presented as a dashboard about diabetes.
Nobody says what "now" means. A number with no timestamp, no denominator, and no comparison is not information. "412 admissions" — since when? Out of what? Compared to what? Up or down? Three of those cost one line each and most dashboards skip all three.
It's built for the person who asked, not the person who acts. The executive requested it. The ward manager would have to act on it. Nobody asked her what she'd need, so she doesn't use it, and the executive concludes the ward doesn't care about data.
Goodhart's law, which health cannot escape
When a measure becomes a target, it ceases to be a good measure.
This is not a cynical aside; it is the central operational risk of health analytics, and it always works:
- Measure A&E four-hour waits → patients get admitted at 3h55 who didn't need admission.
- Measure documentation compliance → the fields get filled, from memory, at the end of the shift.
- Measure readmission rates → readmissions get recorded as "observation stays".
- Measure surgical mortality → the sickest patients stop being offered surgery.
None of that is fraud. It's people responding rationally to what they're judged on. The lesson for whoever builds the dashboard: you are not neutrally observing the system, you are changing it. Choosing the metric is an intervention, and it deserves the scrutiny you'd give a clinical one.
The partial defence: pair every target with a balancing measure. Four-hour waits alongside inappropriate admissions. Throughput alongside safety. It doesn't eliminate gaming; it makes it visible.
Building it properly
- Model it right. A star schema is what BI tools want; querying the transactional EHR directly is why the dashboard takes 40 seconds and the ward gave up.
- One number, one definition. "Admissions" meaning three different things in three dashboards destroys trust permanently, and trust doesn't come back. A semantic layer or an agreed definitions document is not bureaucracy — it's the product.
- Show the denominator and the date. Always.
- Distribution, not just average. The mean hides exactly the inequity you should be looking for.
- Design for the person who acts. Sit with the ward manager. Watch what she does at 07:30. Build that.
The best analytics work in health frequently isn't a dashboard at all. It's a weekly list of twelve patients who need a phone call — sent to the person who can make the calls.