Yesterday I asked Claude to produce a report for a GP surgery.
I wanted it to identify patients with potentially high blood pressure that might require treatment. Important data hidden in plain sight.
Exactly what AIs are supposed to be good at.
Claude understands FHIR. It can read a FHIR bundle, correctly connect resources to each other and make sense of terminologies.
But can it draw relevant conclusions from FHIR data?
I created a new project in Claude and fed it 10 FHIR bundles – one per patient. Each bundle contained Observations, Conditions, Encounters, Medications and more. Full patient histories over many years.
I asked it to produce a HTML page that could be read and understood by doctors and nurses – no hand holding required. I wanted it to document each patient’s blood pressure readings and draw attention to patients that might need treatment or changes to their treatment.
I told it to factor in the patient’s ages, any conditions they had been diagnosed with and any medications they might be on.
Basically, to look at all relevant data provided in the FHIR bundles.
And an abbreviated snapshot.
8 out of 10 patients required some attention.
This number may seem high but I confess to massaging the FHIR Observations a little to generate results that showed borderline or elevated BP readings.
It highlighted in red three patients that needed immediate attention and gave reasons why.
Example for a patient with a BP of 142/92:
BMI 30+ (obese). Stage 2 HTN. Requires immediate pharmacotherapy + aggressive lifestyle intervention.
For a 5 year old patient with a BP of 131/88:
PEDIATRIC ALERT: BP significantly elevated for age. Requires pediatric cardiology referral to rule out secondary causes.
Bear in mind that these were not one-off readings. It had full patient histories to work with.
What did it get wrong?
Initially it identified Stage 1 hypertension as starting at a systolic reading of 120 instead of 130. I only caught this because I’d been reading up on the different stages of hypertension last week. I pointed out the mistake and Claude made corrections.
It also identified Simvastatin as a “BP Medication”. I don’t have the medical knowledge to know if this is correct or not but Googling suggests it’s used to treat high cholesterol and not high blood pressure.
These are the two possible mistakes I identified but there may be more.
This is a perfect illustration of why you need regular clinical involvement and oversight when building tools that merge AI and healthcare data.
Without fully understanding what Claude is telling you, how can you know it’s correct?
The sample patient size was small but picture what this same prompt might produce when looking at selected batches of patient records from a real GP’s surgery.
It demonstrates that given the right prompts and with the right oversight Claude can extract data from FHIR and draw conclusions that have real clinical value.
---