AI Bias
Artificial intelligence (AI) bias occurs when an AI system outcome reflects systematic errors related to either the source data or the analysis and understanding of that data. A common concern in health care applications of AI relates to patient subgroup (e.g., gender, age, ethnicity, economic circumstances, geography) outputs that reinforce past inequities and can compound health process or outcome disparities.
Content Bias
Systematic error or deviation from the true outcome that originates from a model's data.
Bias may occur when insufficient (inadequate sample size), selective or otherwise unbalanced training datasets underrepresent specific patient populations or determinants (e.g., missing social determinants of health data, non-randomly distributed missing data, etc.).
Training and reinforcement data may contain surrogate markers for true health risks, conditions or outcomes (e.g., laboratory test results used to infer presence of a health condition). Connections between surrogate and true targets can change over time, location, setting or level of care in ways that bias AI model outcomes.
Analytic Bias (Computational or Machine Bias)
Systematic error or deviation from the true outcome that originates from a model's assumptions or design
Algorithmic elements of AI models may distort or over-simplify complex clinical phenomena.
Interpretive Bias (Cognitive Bias)
Systematic error or deviation from a true outcome that originates with inaccurate judgements or assessments embedded in a model's data.
Health records may contain faulty or predudiced interpretations of health data (e.g., problem list or medical history entries) that are prioritized by an AI model in ways that perpetuate misinterpretation and error.
Amplification Bias
When AI systems are configured for continuing improvement by re-learning from real-time datasets, biases may be amplified if the AI model is recursively contributing to the data used for ongoing calibration. Amplification of bias contributes to AI model degradation (senescence) over time.
Presentation Bias
The presentation of AI model outcomes to clinical decision-makers may occur in ways that bias decisions or miss opportunities for decisions.
Performance Bias
Over-reliance on statistical measures of AI model performance may not match the clinical utility of model outcomes.
AI bias can be mitigated with well-selected training datasets, AI feedback about impactful data needing continuing quality improvement (e.g., assessing and recording patient frailty to improve disposition predictions), regular re-training, model updates, surveillance of subgroup outcomes and collaboration with human oversignt in continuing AI model improvement.