Streamlining the KOOS Activities of Daily Living Subscale Using Machine Learning

Area of Science:

• Orthopedics
• Sports Medicine
• Biostatistics
• Machine Learning Applications in Healthcare

Background:

• Patient-reported outcome measures (PROMs) like the Knee injury and Osteoarthritis Outcome Score (KOOS) are valuable but burdensome.
• The KOOS, a knee-specific PROM validated for anterior cruciate ligament (ACL) reconstruction, comprises 42 questions across 5 subscales.
• The KOOS Activities of Daily Living (ADL) subscale currently includes 17 questions.

Purpose of the Study:

• To determine if the number of questions in the KOOS ADL subscale can be reduced using machine learning (ML) while preserving data integrity.
• To identify a minimal set of essential questions for accurately predicting KOOS ADL scores post-ACL reconstruction.
• To decrease patient burden and resource requirements associated with PROM administration.

Main Methods:

• A cohort study design was employed using pre- and postoperative KOOS ADL scores from the Surgical Outcome System (SOS) data registry.
• Categorical Boosting (CatBoost) ML models were developed to assess the predictive value of individual KOOS ADL questions.
• A subset of minimal essential questions was identified based on ML model performance.

Main Results:

• The study analyzed data from 2525 patients (aged 16-50) who underwent ACL reconstruction.
• The CatBoost model accurately predicted KOOS ADL scores using only 6 questions (R² = 0.95), comparable to using all 17 questions (R² = 0.99).
• The identified essential questions included descending stairs, ascending stairs, standing, walking on a flat surface, putting on socks/stockings, and getting on/off the toilet.

Conclusions:

• Machine learning algorithms effectively identified a streamlined set of 6 essential KOOS ADL questions (35% of the original 17).
• This reduction in questions maintains high accuracy in predicting functional outcomes after ACL reconstruction.
• Utilizing ML to shorten PROMs can reduce patient burden, potentially increasing compliance and data quality.