A Custom Machine Learning Classifier for Cosmetic Data Mapping

The client is DermCare, the aesthetics industry’s leading managed e-commerce and patient-engagement platform, serving dermatologists, plastic surgeons, and medical spas across the United States. Much of DermCare’s value to those practices rests on analytics — the ability to show a practice what it sells, to whom, and how its business is truly performing.

That promise, however, depends entirely on the quality of the underlying data. To turn raw practice records into meaningful insight, DermCare needed the charge data flowing in from hundreds of practices to be clean, consistent, and comparable. SCIMUS was engaged to solve the problem standing in the way of that goal — and to build, end to end, the machine learning system that would make reliable analytics possible at scale..

1. Every practice records its charges in its own words. In the PMS and EMR systems aesthetics practices rely on, a treatment is entered as free text — and free text is endlessly inconsistent. A single procedure might appear as “Botox,” “BOTOX Cosmetic,” “botox 20u,” or a dozen other variations, each typed by a different person on a different day. Multiply that across hundreds of practices and thousands of procedures, and the same handful of treatments fragments into a sprawl of near-duplicate descriptions.
2. For analytics, this was fatal. Data that cannot be grouped cannot be measured. Before any meaningful insight could be drawn, every one of those free-text descriptions had to be mapped to a standardized category so that like could be compared with like — and that mapping was being done by hand.
3. Key challenges included:
4. 1. Unbounded inconsistency. Free-text entry meant the same treatment was described in countless different ways, with no shared vocabulary across — or even within — practices.
5. 2. Duplication that distorted the numbers. Near-identical descriptions splintered a single procedure into many, making accurate aggregation impossible and any analysis built on the data unreliable.
6. 3. A heavy manual onboarding cost. Bringing a single new practice online required between twenty and sixty hours of painstaking work to map its descriptions to standard categories by hand.
7. 4. A burden that never ended. The problem did not stop at onboarding. Practices continually added new descriptions to their systems, each of which had to be mapped in turn — making this an open-ended, perpetual support task rather than a one-time effort.

SCIMUS built the Cosmetic Data Mapper (CDM) end to end — a custom machine learning system that classifies free-text charge descriptions into standardized categories automatically, in real time, and without human review.

The foundation of any classification model is the data it learns from, and here there was no existing dataset to draw on. So the team built one. SCIMUS assembled a training corpus of more than two million charge-description variations and hand-labeled each to a standardized category taxonomy developed specifically for the aesthetics domain. This labeled corpus — the product of deliberate, large-scale data engineering — became the ground truth from which the model would learn to recognize the countless ways a single treatment might be written.

What we delivered:

• Developed an ML-based classification model capable of categorizing free-text treatment descriptions into standardized categories.
• Trained the model to recognize semantic patterns, enabling accurate classification of unfamiliar descriptions beyond exact string matching.
• Implemented confidence scoring for each classification to identify ambiguous cases and improve data reliability.
• Automated data standardization at scale, transforming inconsistent free-text data into structured, consistent, and analyzable information.

A well-designed — meaning cost-efficient and result-oriented — team was assigned to the project.