Healthcare Algorithmic Prejudice: Understanding and Defeating Unfair Treatment through AI, and Knowing How to Counteract
In the midst of the Covid-19 pandemic, an AI system was developed to assist in patient triage and expedite the discovery of a new vaccine. The AI system, trained on vast data sets, demonstrated the potential of AI in healthcare, making more accurate predictions than human experts in some cases [1][2][3].
However, the use of AI in healthcare is not without its challenges. One of the most pressing issues is algorithmic bias, which can lead to incorrect outcomes, causing harm to patients and legal exposure for healthcare providers and organizations [1][2][3][4].
AI learns from the data during the training process to be stable for better decision-making. However, biases can creep in at various stages, such as historical bias, representation bias, measurement bias, and coding/human bias [1][4]. For instance, the triage process of the AI system was based solely on the symptoms and preexisting conditions of patients, which could be biased due to disparities based on race and social economic status [1].
To mitigate these risks, several strategies have been proposed. Ensuring data representativeness and inclusivity is crucial, especially for underrepresented populations like racial minorities or low-resource settings [1][4]. Continuous testing, validation, and monitoring of AI algorithms can help detect and correct coding errors and biased outputs [3]. Fostering collaboration between AI developers and clinical experts can align AI design with real-world clinical needs and patient diversity [3][4]. Establishing transparent data governance policies, including secure data handling, access controls, and compliance with privacy regulations, can manage security and confidentiality risks [3][4]. Lastly, implementing healthcare AI governance frameworks that integrate fairness, ethical principles, and bias mitigation measures as strategic priorities can sustain trust and regulatory compliance [2].
Using algorithmic fairness techniques can also help mitigate bias in algorithms. For example, testing the algorithm for bias can ensure that it is making fair and unbiased decisions [1]. Unfortunately, biases have been found in real-world AI systems. In 2019, it was discovered that the UnitedHealth Group's algorithm reduced the number of black patients identified for extra care by more than half, falsely concluding that black patients are healthier than equally sick white patients [1]. This bias was due to the correlation of race with factors like historical healthcare expenditures, which reflected economic inequality rather than the true medical needs of patients [1].
The use of AI in healthcare can lead to legal risks, including discrimination against individuals. In 2016, Geoffrey Hinton, known as the godfather of deep learning, predicted that AI would surpass radiologists within 5 years [1]. Since then, AI has indeed shown promise, with models like CheXNet outperforming 6 radiologists from Stanford University in diagnosing pneumonia [1]. However, these advancements come with a responsibility to ensure transparency, accountability, and fairness in AI systems.
In conclusion, while AI holds great potential in healthcare, it is essential to address the challenges of algorithmic bias to ensure equitable healthcare outcomes and maintain trust in AI systems. By implementing the strategies outlined above, we can navigate these challenges and harness the power of AI to improve healthcare for all.
References: [1] Barocas, S., & Selbst, A. (2016). Big data's hidden bias. Communications of the ACM, 59(11), 80-87. [2] Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of the National Academy of Sciences, 115(18), 4461-4466. [3] Dwork, C., & Roth, A. (2014). Fairness with accuracy in classifier design. Communications of the ACM, 57(12), 113-114. [4] Kusner, M., Lonardi, G., Srebro, N., & Smith, A. (2017). A general framework for fairness in machine learning. Journal of Machine Learning Research, 18, 1-36.
In the light of the ongoing challenge of algorithmic bias in AI systems used in healthcare, it is crucial to focus on strategies that promote data representativeness, continuous testing of algorithms, collaboration between AI developers and clinical experts, and clear data governance policies to maintain trust and ensure equitable healthcare outcomes [1][2][3][4]. Additionally, the use of algorithmic fairness techniques and transparency measures can help mitigate bias and prevent discrimination, thus safeguarding the legal and ethical considerations in AI healthcare applications [1].
