Abstract
Biases within probabilistic electronic phenotyping algorithms are largely unexplored. In this work, we characterize differences in sub-group performance of phenotyping algorithms for Alzheimer’s Disease and Related Dementias (ADRD) in older adults. We created an experimental framework to characterize the performance of probabilistic phenotyping algorithms under different racial distributions allowing us to identify which algorithms may have differential performance, by how much, and under what conditions. We demonstrate that some algorithms have performance variations anywhere from 3 to 30% for different populations, even when not using race as an input variable. We show that while performance differences in subgroups are not present for all phenotypes, they do affect some phenotypes and groups more disproportionately than others.
Introduction
In modern medicine, phenotyping algorithms are essential for scanning Electronic Health Records (EHRs) to identify cohorts of patients with specific conditions. Traditionally, these were rigid, rule-based systems. Recently, the field has shifted toward probabilistic phenotyping—using machine learning (ML) to handle data sparsity and complexity. However, these ML models often act as "black boxes" that may inherit or amplify biases present in the underlying clinical data.
This study addresses the critical need to understand how probabilistic models perform across different racial subgroups, specifically for high-stakes conditions like Alzheimer’s Disease and Related Dementias (ADRD). Even "race-blind" models—those that do not use race as an input feature—can harbor significant disparities, potentially exacerbating existing health inequities if an algorithm's accuracy varies significantly between groups.
Materials and Methods
The researchers utilized the APHRODITE framework, which employs weak supervision to build phenotype models without requiring extensive manual labeling. The study analyzed clinical data from the Stanford Medicine Research data Repository (STARR), covering over 3.1 million patients from 2008 to 2018, standardized via the OMOP Common Data Model (CDM).
Four phenotypes were targeted: Dementia, Mild Cognitive Impairment (MCI), Alzheimer’s disease, and Parkinson’s disease. The experimental framework evaluated three algorithms—LASSO, Random Forests, and Support Vector Machines (SVM)—under various data compositions:
- Full Model: All races merged.
- Balanced Model: Equal distribution of patients from each race.
- Single-race only: Training on one race and testing on others.
- Leave-one-out: Systematically removing one race from the training set.
To quantify bias, the authors focused on variation, defined as the absolute difference in performance when a model encounters a subgroup it was not primarily trained on.
Results
The study found that general accuracy for "full" models typically ranged between 70% and 90%. However, significant performance gaps emerged depending on the phenotype and training configuration.
Phenotype-Specific Volatility
Bias was not a universal constant across diseases. Models for Dementia and Alzheimer’s were relatively stable, showing less than 8% variance across racial groups. In contrast, models for Parkinson’s disease and MCI were highly sensitive, with performance drops reaching up to 30% in certain scenarios. This variance effectively reduces a sophisticated ML model to near-random performance for specific subgroups.
Data Dominance and Representation
The "Data Dominance" effect was prominent: removing the majority group (White patients) from training data caused the most significant performance crashes, indicating that models often overfit to majority clinical patterns. Furthermore, groups with extremely low representation (e.g., Native Americans at <0.25%) resulted in models with poor generalizability, particularly in sensitivity (recall). This suggests that clinical features like comorbidities or visit frequency act as proxies for race, even when race is excluded from the model.
Discussion and Conclusion
The findings establish a clear need for robust evaluation frameworks to detect subgroup differences. The underlying patient populations for algorithms showing high variance often have significant differences in model features compared to more stable phenotypes.
While performance differences are not universal, their disproportionate impact on certain groups necessitates careful, ongoing monitoring. Future work should address limitations such as the single-site nature of the study and the noise in self-reported race data. Potential solutions like federated learning or ensemble models may help mitigate these disparities without degrading overall performance. In conclusion, this framework successfully identifies systematic differences in phenotyping performance, highlighting the need for developers to actively measure and mitigate subgroup performance gaps to ensure equitable healthcare outcomes.
Summary updated