Racial and Ethnic Disparities in Predictive Accuracy of Machine Learning Algorithms Developed Using a National...

Area of Science:

• Orthopedic Surgery
• Data Science in Healthcare
• Health Equity

Background:

• Machine learning (ML) algorithms have shown promise in predicting outcomes for total joint arthroplasty (TJA).
• However, their performance in racial and ethnic minority populations remains understudied.
• This research addresses the gap by evaluating ML predictive capabilities in diverse TJA patient groups.

Purpose of the Study:

• To assess the performance of ML algorithms in predicting 30-day complications following TJA.
• To specifically evaluate these algorithms within racial and ethnic minority patient cohorts.
• To identify potential disparities in ML model accuracy across different demographic groups.

Main Methods:

• A retrospective cohort study analyzed 267,194 patients undergoing primary TJA from 2013-2020.
• Two ML algorithms, histogram-based gradient boosting (HGB) and random forest (RF), were developed and tested.
• Model performance was evaluated across racial and ethnic subgroups using discrimination, calibration, and accuracy metrics.

Main Results:

• ML models achieved excellent performance (AUC > 0.8) in the non-Hispanic White population.
• Discrimination decreased significantly in White Hispanic, Black, Black Hispanic, and Asian non-Hispanic cohorts (AUC 0.75-0.79).
• The poorest performance was observed in the American-Indian cohort (AUC 0.67-0.68), despite good calibration across most minority groups.

Conclusions:

• ML algorithms exhibit inferior predictive ability for TJA complications in racial and ethnic minorities when trained on current healthcare data.
• Underrepresentation of minority groups in large datasets likely contributes to reduced model accuracy.
• Equity-conscious development and larger, more diverse datasets are crucial for accurate and fair ML predictions in TJA.