Healthcare analytics: improving patient outcomes with data

Most writing about healthcare analytics is about hospitals: sepsis alerts, readmission models, population health programs run by teams of data scientists. Most healthcare providers are not hospitals. They are physiotherapy practices, dental clinics, GP surgeries, therapy groups and specialist clinics with a practice management system, a payment terminal and a few spreadsheets.

Those providers can improve patient outcomes with data too. The measures are simpler, the data is smaller, and the main risks are misreading small numbers and handling patient information carelessly. This guide covers what to measure, how to calculate it and how to read it honestly.

Start with what "outcome" means for your service

An outcome is a change in the patient's health or function that your care is meant to produce. For a small provider, useful outcome measures usually fall into three groups:

• Clinical or functional change, often measured with a validated questionnaire at the start and end of care. Physiotherapy uses pain and function scores; mental health services use symptom scales. These are called patient-reported outcome measures (PROMs).
• Completion of care, because a treatment plan that is abandoned halfway rarely achieves its outcome.
• Unplanned returns: patients who come back with the same problem soon after discharge, or, for inpatient providers, readmissions.

Alongside those sit access measures, which strongly influence outcomes: how long patients wait, whether they attend, and whether follow-up happens on time.

Choose a small set: two outcome measures, two access measures. More than that and nobody reads them.

Measures and how to calculate them

Share of patients with meaningful improvement

For a PROM with a known minimal clinically important difference (MCID), the change that patients actually notice:

Improvement rate = patients whose score improved by at least the MCID / patients with both a start and end score

Example: 120 patients were discharged last quarter. 80 have both scores. 56 improved by at least the MCID. Improvement rate = 56 / 80 = 70%.

Also report the completion rate of the questionnaire itself (80 of 120, or 67%). If only the patients who did well fill in the final form, the improvement rate is flattered.

Plan-of-care completion

Completion rate = patients who finished the planned course / patients who started one, over the same period.

If 200 patients started a six-session plan and 130 finished, completion is 65%. Break it down by clinician, service and referral source to see where patients drop out.

No-show rate

No-show rate = appointments not attended without notice / appointments booked

400 booked appointments and 36 no-shows gives 9%. Missed appointments delay care and waste capacity. Look at the rate by weekday, time slot, lead time (days between booking and appointment) and whether a reminder was sent. Patterns are often obvious once charted: early Monday slots booked three weeks ahead, for example.

Time to first appointment

Measure the days from referral or first contact to first appointment, and report the median and the 90th percentile, not the average. One patient who waited 90 days distorts an average; the 90th percentile shows how long the slowest tenth waited.

Unplanned return within 30 days

Return rate = patients discharged who came back for the same problem within 30 days / patients discharged

Hospitals track 30-day readmissions the same way: readmissions within 30 days of discharge divided by discharges. For a small practice, "same problem" needs a clear rule written down, or the number will shift with whoever codes the visit.

Reading the numbers without fooling yourself

Small numbers move a lot

A clinician with 10 discharges who has 7 improvements (70%) and one with 12 who has 6 (50%) may not differ at all; with numbers that small, one or two patients swing the rate by 10 points. Before comparing people or sites:

• look at longer periods, a quarter or more
• show the count next to every percentage
• use a run chart (the measure over time, with its median) and act on sustained shifts, not single points

Case mix changes outcomes

A clinician who sees more complex or older patients will have lower improvement rates for reasons that have nothing to do with care quality. Compare like with like: by condition, by starting severity, by age band. If you cannot adjust, do not rank.

Correlation is a question, not an answer

If patients who attend more sessions improve more, that may be because sessions help, or because patients who are improving keep coming. Use correlations to decide what to look at next, not to set policy.

Missing data is a finding

If a third of patients have no end-of-care score, report that as prominently as the outcome itself. Improving data capture is often the first real improvement. Our post on spotting data quality issues early covers common checks.

Handling patient data responsibly

You do not need identifiable health records to do most of this analysis. Work from operational exports with the minimum detail the question needs:

• use a patient ID or pseudonym instead of names, dates of birth and contact details
• leave out diagnoses and clinical notes unless the question needs them; a condition category is usually enough
• keep exports in one place, with access limited to the people who need them, and delete old copies

Follow your own legal and professional obligations for health data, which vary by country.

Where clariBI fits, and where it does not

clariBI is useful for the business and operational side of a clinic, not for clinical records. Be clear about that before you connect anything:

• Do not upload patient health records. clariBI is not built for protected health information, has no HIPAA arrangement and does not sign business associate agreements.
• It has no connection to electronic health record or practice management systems. Use operational exports without clinical details.
• It is not a clinical tool and does not assess care quality on its own; it calculates and charts what is in the data you give it.

Within those limits, it does the counting well. You can upload de-identified booking and discharge exports as CSV or Excel (see uploading files), connect payments from Square, Stripe or PayPal, bookings from Calendly or Cal.com on Starter and up, and marketing from Google Ads, Meta Ads and Google Analytics. Each source gets a dashboard automatically, and you can ask questions such as "What is our no-show rate by weekday and time slot for the last six months?" in plain English. The numbers come from calculations on your data, and any narrated figure the chart cannot back is removed.

The healthcare template category includes operational templates such as patient throughput, no-show rate by provider and clinic, scheduling bottlenecks and billing aging, built for spreadsheet exports. The clinics and wellness page lists what connects and the limits in more detail. To keep a team working toward a target, such as a no-show rate under 8%, set a goal linked to the data so its progress updates from the source.

A 90-day plan for a small practice

Month 1. Pick two outcome measures and two access measures. Write down exactly how each is calculated. Export three to six months of history without names or clinical notes.

Month 2. Build one dashboard with those four measures, each with counts shown. Review it monthly with the clinical lead. Fix the biggest data gap, usually missing end-of-care scores.

Month 3. Choose one change to test, such as reminders two days before early-morning appointments. Track the relevant measure on a run chart for the next three months before deciding whether it worked.

For a structured way to follow patient groups over time, such as everyone who started care in March, see our guide to cohort analysis.

Next step

Export last quarter's appointments with status, date, time and a patient ID only. Upload the file and ask: "What was the no-show rate by weekday and hour?" If one slot stands out, that is your first test.