Machine-Learning Model Identifies Patients With Alpha-1 Antitrypsin Deficiency Using Claims Records

Area of Science:

• Medical Informatics
• Rare Disease Diagnosis
• Machine Learning Applications

Background:

• Identifying patients with rare diseases, such as alpha-1 antitrypsin deficiency (AATD), presents significant diagnostic challenges.
• Electronic medical records often contain insufficient confirmed cases for training diagnostic models.
• Large-scale, real-world databases offer a viable alternative for developing predictive models.

Purpose of the Study:

• To apply a machine-learning model to a large US claims database to identify undiagnosed symptomatic patients with AATD.
• To validate the model's performance using expert clinical review.
• To gain insights into the diagnostic journey of AATD patients.

Main Methods:

• A machine-learning model was trained on deidentified data from a US claims database (April 2016–January 2023).
• The model identified high-probability candidates for AATD, with 80 records independently validated by clinical experts.
• Model optimization was performed based on expert feedback, and informative features were identified.

Main Results:

• Clinical experts validated that a high percentage of identified candidates (81% and 78%) should be tested for AATD.
• The optimized model identified symptomatic patients with probable AATD.
• Specific claims data features and unique medical event cadences distinguished AATD patients from those with COPD or chronic liver disease.

Conclusions:

• Machine learning models trained on large claims databases can accurately identify symptomatic patients with AATD.
• This approach enhances the diagnosis of rare diseases.
• The study provides valuable insights into the diagnostic pathways for AATD.