Making the Subjective Objective: Machine Learning and Rhinoplasty

Overview

Facial recognition software can estimate a person's age from a photograph. The authors applied a commercial system to before and after photographs of their rhinoplasty patients to test whether the operation changed how old they appeared to an algorithm.

Sections of note

• Retrospective chart review of all female patients who had open rhinoplasty with the senior author from 2014 through 2018 with postoperative photographs at 12 or more weeks.
• Photographs were analyzed with Microsoft Azure Face API, which crops the face and predicts age through deep neural networks.
• 100 patients met full inclusion criteria.
• Average post-surgical follow-up 29 weeks; median 14 weeks; range 12 to 64 weeks.
• Patient ages ranged from 16 to 72 years; mean 32.75, median 28.00, standard deviation 12.79.
• Preoperatively the algorithm estimated patients 0.03 years older than their actual age; correlation between actual and predicted preoperative age was r 0.91.
• Postoperatively patients were estimated 3.10 years younger than their actual age, P below 0.0001.

What it means for a patient

• An algorithm that estimated age almost exactly before surgery estimated it about three years lower afterward.
• This measures what software infers from a photograph, not how old anyone looks to another person.
• Median follow-up was 14 weeks, when swelling is still resolving, so the postoperative photographs are not of settled results.
• Limits: 100 women, one surgeon, retrospective, and no control group of unoperated faces photographed twice.

Why this paper matters

Perceived age is hard to measure consistently between human raters, and an algorithm gives a repeatable number. The finding suggests rhinoplasty affects perceived age, which is not its stated purpose. Whether the algorithm is responding to the nose or to changes in photography and swelling is not established.

Terms

• Machine learning: Software that derives patterns from data rather than following written rules.
• Convolutional neural network: A machine learning architecture used for analyzing images.
• Face API: A commercial software service that analyzes faces in photographs.
• Correlation coefficient r: A number from -1 to 1 showing how strongly two measures move together.
• Open rhinoplasty: Surgery using an external incision across the columella.