Table 3.

Recommended Best Practices for Machine Learning to Understand Cognitive Phenotypes in People with HIV

Well-curated high-quality data

Cohort size should take into account study goals and quality, dimensionality, and completeness of dataset (ie, no “one size fits all”)

Larger studies can be complemented by more focused studies with smaller cohorts

Make decisions on best methods to handle missing data

Choose models and metrics that take into account the balance of classes

Internal/external validation to demonstrate model stability and reproducibility

Representative training sets from diverse cohorts (including people with common comorbidities) to increase generalizability

Accurately report and account for biases and confounders

Measures that can be used across international settings