What this is, and what it is not
An independent portfolio project built entirely from published, aggregate NHS England data. It contains no patient-level or personally identifiable information of any kind, and it is not affiliated with, endorsed by, or produced for NHS England or any charity.
It is a demonstration of analytical method, not a clinical audit or a commissioning tool. A lower recorded diagnosis rate is an indicator to investigate — it can reflect coding and recording practice as much as access to services — and never a judgement on a local service.
The question
Dementia diagnosis rates vary widely across England, but a rate on its own is a poor guide to where effort should go. A small area with a low rate and a large area with a middling rate are very different problems: one is about intensity, the other about scale.
So the project holds two lenses at once — the recorded diagnosis rate, and the estimated gap in absolute people — and then does something less common: it makes the confidence you can place in those numbers a visible part of the product rather than a footnote.
The live Power BI report
The report itself, embedded and interactive — not a screenshot. Five pages: the national picture, local variation, care and equity signals, data confidence, and the metric definitions. Move between them with the tabs along the bottom of the report, and filter by NHS region or Integrated Care Board on the first page.
The Power BI report needs a wider screen. Its canvas is 1280 × 768, so it is not readable inside a phone-width frame.
Prefer something lighter? There is also a self-contained HTML version of the same figures — 44 KB, no external requests, built as a check that the model's numbers reproduce outside Power BI.
How it was built
- Seven published NHS England files, three reporting grains Primary Care Dementia Data for March 2026, including a 224 MB practice-level extract streamed and aggregated so that nothing patient-level is ever retained.
- A reproducible Python pipeline into a star schema Two conformed dimensions — 42 ICBs and 13 reporting months — against two fact tables at 546 and 4,368 rows. Deterministic, and hash-verified against the source files.
- A SQL analysis layer The same logic re-expressed as SQLite views and data-quality checks, so the SQL and the BI model reconcile rather than merely agree.
- A five-page Power BI report A measure-only model with explicit period comparisons, moving from the national headline through local variation and care signals to a deliberate close on data confidence, and finishing with the metric definitions themselves.
- A data-quality and provenance layer Ten rules across uniqueness, completeness, validity, consistency, referential integrity and timeliness, plus a register of every source file with its URL, size and SHA-256 hash.
What the data says
Rate and scale are different problems
The lowest-rate areas are not the largest-gap areas. Somerset has the lowest recorded rate at 56.1%; the largest absolute gap sits in North East and North Cumbria at roughly 13,200 people. Population size drives the gap, so a credible prioritisation needs both lenses.
Care coverage is oddly uniform
Care-plan coverage clusters between about 74% and 78% across most ICBs, while diagnosis rates range from 56% to 76%. That points to the bigger and more uneven opportunity being in getting a diagnosis recorded at all, rather than in what follows it.
One in five records has no usable ethnicity
Nationally 20.1% of the register is coded not defined, not stated or inconclusive — and locally that ranges from 2.5% to 65%. Dropping those records would manufacture a clean representation picture that simply is not true.
So completeness is shown, not hidden
Unknown records are kept and displayed as their own category, and the report states plainly that where completeness is low, no inference about under- or over-representation should be drawn. Completeness is the finding.
Limits worth stating
- The denominator is a modelled prevalence estimate with published confidence intervals. Rankings should be read as tiers, not as a league table.
- Recorded is not delivered. A care plan coded in the last 12 months says an activity was recorded, not that it was good.
- Ethnicity completeness varies enormously, so any local equity conclusion has to be read next to its completeness figure.
- The publication is changing. March 2026 is the final monthly release; NHS England has announced quarterly publication and format changes for 2026/27. That is tracked as an open data-quality risk, which is why the sources are versioned and hashed.
Built with
- Python
- pandas-free stdlib ETL
- SQLite
- SQL views
- Power BI Desktop
- DAX
- TMDL / PBIP
- HTML, CSS, JavaScript
The Power BI project is stored in the PBIP/PBIR format — pages, visuals and the semantic model as plain text files — so changes review and merge like code rather than like a binary.