Patient Phenotyping for Atopic Dermatitis with Transformers and Machine Learning.

Area of Science:

• Dermatology
• Computational Biology
• Clinical Informatics

Background:

• Atopic dermatitis (AD) is a prevalent chronic skin condition affecting millions globally.
• Current clinical trial recruitment for AD is hindered by diagnostic variability and manual patient identification.
• There is a critical need for automated patient phenotyping to streamline cohort recruitment.

Approach:

• Developed a machine learning approach to identify potential AD patients using electronic health records (EHRs).
• Created vectorized patient representations based on diagnostic criteria probabilities or binary values.
• Trained and evaluated supervised machine learning models for AD classification.

Key Points:

• The most effective classifier, XGBoost, achieved a class-balanced accuracy of 0.8036.
• XGBoost demonstrated a precision of 0.8400 and a recall of 0.7500 for AD identification.
• Patient vectors incorporated probabilities or binary indicators for various AD diagnostic criteria.

Conclusions:

• Automated patient phenotyping can accelerate and standardize AD clinical trial recruitment.
• This approach reduces the burden on clinicians in identifying study participants.
• Facilitating recruitment can accelerate the discovery of improved treatments for atopic dermatitis.