Using mixture cure models to address algorithmic bias in diagnostic timing: autism as a test case.

Area of Science:

• Health Informatics
• Biostatistics
• Clinical Prediction Modeling

Background:

• Algorithmic bias in clinical prediction models can arise from variations in diagnostic timing.
• Traditional models struggle with accurate predictions when outcomes are not uniformly observed.

Purpose of the Study:

• To evaluate mixture cure models for predicting diagnoses and addressing algorithmic bias.
• To compare mixture cure models against traditional time-to-event and classification models.

Main Methods:

• A simulation study and analysis of North Carolina Medicaid data (children born 2014-2023).
• Evaluation of traditional models versus mixture cure models under varied diagnostic timing and censoring scenarios.

Main Results:

• Mixture cure models provided unbiased estimates, unlike traditional models which showed increased bias with wider diagnosis timing differences.
• Real-world data revealed racial/ethnic disparities in autism diagnosis rates; mixture cure models adjusted for these, improving fairness.
• Mixture cure models demonstrated robust performance across varying censoring times.

Conclusions:

• Mixture cure models effectively mitigate algorithmic bias in predictive modeling, particularly for conditions like autism.
• This methodology enhances prediction accuracy and fairness, especially when outcomes are not universally observed.
• The approach aligns clinical needs for early detection of pediatric developmental conditions.