Unmasking bias in artificial intelligence: a systematic review of bias detection and mitigation strategies in electronic health record-based models
Ref: https://academic.oup.com/jamia/article/31/5/1172/7634193
In this paper, the authors review one of the most determining aspects in the use of AI, Unmasking bias in artificial intelligence, specifically in electronic health record-based models.
Artificial Intelligence (AI) is revolutionizing healthcare by leveraging electronic health records (EHR) to develop predictive models that significantly enhance clinical research and clinical decision-making (CDS). However, the potential for these models to perpetuate or exacerbate healthcare disparities due to inherent biases cannot be ignored.
The review identifies six main types of bias in EHR-based models:
- Algorithmic Bias: Errors introduced by the algorithms themselves.
- Confounding Bias: Bias due to unconsidered variables that influence both the predictor and the outcome.
- Implicit Bias: Bias resulting from the unintentional encoding of social stereotypes.
- Measurement Bias: Inaccuracies in data collection or recording.
- Selection Bias: Bias arising from the non-random selection of samples.
- Temporal Bias: Bias related to the timing of data collection.
The authors conclude that it is crucial to prioritize more in-depth research focused on developing standardized, generalizable, and interpretable methods for detecting, mitigating, and evaluating bias in AI models. This is critical to ensure that these technologies are equitable, thereby minimizing the risk of healthcare disparity due to biases.
