All case studies
  • Artificial intelligence
  • Digital healthcare

Matching bookings and admissions in diagnostic imaging

An agent that matches each booking to its admission, even when the same exam is written differently, and finds who did not show up.

The starting point

In a diagnostic centre every exam is born twice: when it is booked in the schedule and when the patient is admitted at the front desk. It looks like the same data, but often there is no direct link between the two moments.

As a result, answering a simple question, who did not show up?, means comparing lists by hand, line by line. Yet it is information that matters: to contact patients again, to understand how much no-shows weigh and to organise the schedule better.

Why the links get lost

Different wordsThe same procedure is described one way in the schedule and another way at admission.
Different daysThe patient turns up before or after the booked date.
No linkThe admission is recorded without recalling the booking.

Same exam, different words

This is the heart of the problem. The system does not compare words but their meaning: that is why it recognises as equivalent descriptions that a plain text comparison would treat as different.

  • In the scheduleCervical MRI Same procedure At admissionMagnetic resonance of the cervical spine
  • In the scheduleChest X-ray Same procedure At admissionTwo-view radiograph of the chest
  • In the scheduleFull abdominal ultrasound Same procedure At admissionUpper and lower abdomen sonography

Illustrative examples.

How it reasons

  1. 1CollectBookings and admissions for the period to analyse.
  2. 2CompareFor each possible pair: same patient, distance between dates, same day or not.
  3. 3Read the proceduresTurns the descriptions into semantic embeddings and measures how similar they are.
  4. 4DecideA model estimates the match probability and picks the most likely admission.

Two models, two jobs

The first is a language model (Sentence Transformers): it turns each description into a vector of numbers, an embedding, that represents its meaning. Two descriptions of the same exam produce nearby vectors, even with no words in common; closeness is measured with cosine similarity.

The second is a classification model that combines all the clues (patient, dates, similarity of procedures) and returns a probability. It is trained on already verified matches, so it learns how that facility works, and it can be retrained as more matches are verified.

What you get

  • For each booking, the most likely admission with its match probability: uncertain cases stand out and can be checked.
  • The list of bookings without a match, meaning patients who did not show up.
  • No more comparing schedule and admission by hand.

What the result looks like

Booking Admission Probability Outcome
12/03 · Cervical MRI 12/03 · Magnetic resonance of the cervical spine 0,97 Matched
12/03 · Chest X-ray 14/03 · Two-view radiograph of the chest 0,93 Matched
13/03 · Full abdominal ultrasound 13/03 · Upper and lower abdomen sonography 0,88 Matched
13/03 · Brain CT No admission – No-show

Example with made-up data.

More case studies

All case studies

We are always curious to hear how your needs are changing

Half an hour on Teams: you tell us where you are heading, we tell you frankly whether and how we can be of use.

Book a meeting