Although machine learning already is widely used in the healthcare system, researchers in a new report are calling for vigilance in its use by health insurers.
Machine learning is used by health insurers to predict everything from early disease onset and medication noncompliance to the likelihood of future hospitalizations. However, there is growing concern the algorithms could perpetuate biases.
The researchers from the study published in Health Affairs in February evaluated three types of machine learning used by payers and identified key biases within each.
Among the study's findings:
- For rare diseases or diseases that predominantly affect minority groups, such as sickle cell anemia, predictive modeling is less common. Therefore, patients with these diseases might not have specific targeting interventions.
- To predict hospitalizations, machine learning models use data on current hospitalizations. However, given socioeconomic disparities of patients who use hospitals, this data could be biased and could in turn entrench these disparities into its algorithm.
- Because of differences in prescriptions for medication along racial lines, models for medication adherence can be influenced by these biases instead of clinical evidence.
Read the full study here.