# The medication history is the weakest link

> Pre-operative medication histories are usually assembled from patient recall and a GP summary that may be months old. The drugs patients are least able to name — anticoagulants and SGLT2 inhibitors described as 'a blood thinner' or 'something for sugar' — are precisely the ones whose management must be planned days ahead, which is why they cause cancellations.

*Source: https://getpathways.ai/insights/the-medication-history-problem*

"A blood pressure tablet and something for sugar."

That sentence, or a version of it, is how a large share of pre-operative
medication histories are taken. It is not a failure of care and it is not the
patient's fault. It is the predictable output of asking someone to recall, from
memory, over the phone, a list of items that look alike, are named
unmemorably, and were prescribed by someone else.

The problem is what that sentence conceals. Behind "something for sugar" may
sit an SGLT2 inhibitor. Behind an unmentioned item may sit an apixaban. Both
require a management plan made days before surgery. Neither will get one.

## The drugs most likely to be missed are the ones that matter

There is an unhappy correlation here. The medications whose perioperative
management is most time-critical are also among the least likely to be
accurately reported.

[SGLT2 inhibitors](/glossary/sglt2-inhibitor-perioperative) — empagliflozin,
dapagliflozin, ertugliflozin — carry a risk of euglycaemic diabetic
ketoacidosis, a presentation in which ketoacidosis develops with a normal or
near-normal blood glucose. The diagnostic signal clinicians are trained to look
for is absent. Withholding has to begin days ahead. And patients almost never
name the drug, because to them it is one of the diabetes tablets.

[Anticoagulants](/glossary/perioperative-anticoagulation) are the same shape.
The withholding interval depends on the specific agent, the bleeding risk of the
procedure, and renal function — none of which can be worked out for a drug
nobody knows the patient is taking. "A blood thinner" is not enough to plan
from, and a patient who has been on the same tablet for three years may not
think to mention it at all.

## Why the referral letter does not close the gap

The obvious answer is to take the medication list from the record rather than
the patient. In practice the record has its own failure mode: it is a snapshot,
and it ages.

A GP summary written eight months ago predates any prescribing change since. A
referral letter reflects the referrer's view at the time of referral, which may
itself have been copied from something older. Neither is wrong, exactly. Both
are stale in a way that is invisible — a list with no gaps looks complete
whether or not it is current.

So the two available sources fail differently and, importantly, they fail
independently. Patient recall misses drugs the record has. The record misses
drugs the patient has started since. Reconciling them requires actually
comparing them, which is work, and which is why it is usually skipped in favour
of trusting whichever one is to hand.

## Verification, not collection

The useful reframing is that this is not a collection problem. Adding another
place to write the medication list does not help if every source is
reconstructed from the same unreliable memory.

It is a verification problem. What changes the reliability of the list is
having the patient photograph the boxes.

This sounds mundane and is the substantive shift. A photograph of a medication
box is primary evidence: the drug name and strength are printed on it. It does
not depend on the patient knowing what the drug is or why they take it — only
on their being able to find the box, which almost everyone can. It works for
handwritten pharmacy labels and for the patient who brings six boxes and
describes none of them.

Three properties follow from working off images rather than recall:

- **Extraction can be checked against its source.** Every value is verified
  against the image before it enters the record, rather than transcribed and
  trusted.
- **Uncertainty becomes visible.** A value read at 0.97 confidence and one read
  at 0.62 are not presented identically. What was hard to read is flagged for
  human review rather than smoothed over.
- **Conflicts surface rather than resolving silently.** A drug in the photos and
  absent from the referral history is a finding, not a discrepancy to be
  quietly reconciled in favour of whichever source was consulted last.

That last one is the case worth dwelling on. The staged example in the product
shows exactly it: an apixaban extracted from a box photo, checked against a GP
summary eight months old, absent from it, and flagged. Under recall-based
history-taking that patient has no anticoagulant. Under photograph-based
verification they have one, and the discrepancy is the thing the clinician is
shown.

## What this is worth

Reliable medication data is not an end in itself. It is the input on which the
time-critical decisions depend — and those decisions have a deadline that
arrives well before the day of surgery.

An SGLT2 inhibitor identified nineteen days out is a withholding plan. The same
drug identified at 07:00 on the morning of surgery is a
[cancellation](/glossary/day-of-surgery-cancellation), because no decision
available that morning can undo the last dose.

The difference between those two outcomes is not clinical judgement. It is
whether anyone knew.