A new Slop University paper matches orthodontic treatment-length estimates against actual removal dates across a 21-clinic chain, finding removal timing tracks the software's first quote almost exactly.
Software-quoted treatment-length estimates predicted 6,842 patients' actual bracket-removal dates far more tightly than case complexity did, and moved with a platform update that changed how the estimate rounds.
The School of Emergent Priorities has spent several years tracing what happens once an institution’s forecast of what will occur quietly becomes the thing it aims for. A new Slop University paper finds the pattern in braces.
Dr Anouk Mensah and Dr Runa Adegoke matched 6,842 patients’ software-quoted treatment-length estimates against the date their brackets actually came off, across twenty-one clinics of one orthodontic chain. Removal dates clustered tightly around the exact month the software had first quoted — almost six times more tightly than case-mix severity alone would predict — regardless of how complicated the software had recorded a case as being. When the chain’s planning platform was updated partway through the study to round its estimates to the nearest three months instead of six, the clustering moved with it, settling onto the finer grid within weeks. A clinician-confirmation step introduced afterwards, requiring staff to sign off on the software’s figure before it reached a patient, left the pattern unchanged.
The estimate was supposed to describe a course of treatment already under way. What it did instead was tell everyone downstream roughly when to stop — and downstream is where the teeth are.
— Dr Anouk Mensah, Postdoctoral Fellow
Dr Runa Adegoke, the paper’s co-author, said the confirmation step’s null result was the one she found hardest to explain away: “we built it specifically to put a clinician between the number and the appointment, and the number won anyway. What that leaves us wanting to know is what a forecast would have to look like for a clinic to actually treat it as one.”
The full paper is available from the University’s research repository under an open licence, doi:10.5555/slop.c6nhcx.