Mitigating Bias in AI in Healthcare

Introduction:

As the technological landscape rapidly evolves, driven by the remarkable progress in artificial intelligence (AI), it creates transformative possibilities across various industries. Among these, the healthcare industry stands at the forefront of attention, with growing interest in AI’s application in patient care. However, the exciting consideration of AI implementation in healthcare comes with its set of ethical considerations and potential pitfalls. This paper aims to delve into the risks associated with implementing AI in healthcare and proposes strategic measures for their effective mitigation. Specifically, when harnessing the potential of AI in care delivery, it is vital to address the challenge of bias amplification in AI diagnosis and treatment recommendations to ensure fairness and equity. The foremost risk to patients, in adopting AI solutions within healthcare, lies in the potential amplification of biases, resulting in unequal treatment and disparities. The opaque nature of AI model training, coupled with the influence of societal dynamics, can lead to inaccurate outcomes for certain groups. Therefore, int the pursuit of improving patient outcomes and equity in healthcare, healthcare leaders must ensure transparency, rigorous evaluation, and an unwavering commitment to ethical considerations, ensuring that the promises of AI are harnessed responsibly and equitably.

Important Definitions:

• Artificial Intelligence (AI): Artificial intelligence (AI) refers to the use of technology to build machines and computers that can mimic cognitive functions associated with human intelligence. These machines can, for example, understand and respond to spoken or written language, analyze data, make recommendations, or perform various tasks that typically require human intelligence. AI encompasses a set of technologies that enable systems to reason, learn, and act, ultimately solving complex problems.
• Generative AI (GAI): type of AI capable of generating new artifacts in response to prompts; GAI is focused on understanding the inherent patterns in data and using that understanding to create new and original content, such as text, images, audio, or other forms of data.
• Large Language Models (LLMs): These models are trained on vast amounts of text, enabling them to imitate understanding, processing, and producing human-like communication.
• AI Hallucination: A situation where an AI model generates outputs that may sound plausible but are not based on accurate or reliable information. This can occur when the model encounters knowledge gaps or lacks sufficient data, leading it to generate fabricated or misleading information.

Bias Risk:

AI systems are inherently socio-technical, meaning they’re deeply influenced by societal dynamics and human behavior. Based on their training set, AI algorithms might perform better for some groups of people than others because they would not be representative of real-world settings. If not carefully designed and equipped with proper controls, these systems can inadvertently amplify, perpetuate, or exacerbate biases, resulting in disparities in patient care and undesirable outcomes. An illustrative example of this bias unfolds in emergency decision-making scenarios, as described in a study conducted by MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT Jameel Clinic. The study reveals that biased AI models can significantly influence human decisions in urgent medical situations. However, the harmful effects of these biases can be mitigated by presenting advice in a descriptive rather than prescriptive manner, as to avoid AI hallucinations. Moreover, language models used for advice are susceptible to bias when fine-tuned with smaller datasets.

The subtle biases inherent in AI have far-reaching implications, particularly in emergency decision-making. Discriminatory AI models can impact decision-making during medical emergencies, emphasizing the need for careful considerations in model training. Label bias, determining the outcome to predict, introduces a layer of complexity, where good predictions may not necessarily lead to good decisions, making it challenging to measure outcomes that truly matter. This complexity is further highlighted in a NIH research article titled “Dissecting racial bias in an algorithm used to manage the health of populations.” The study reveals that racial bias in an algorithm, relying on health costs as a proxy for health needs, reduces the identification of Black patients for extra care by more than half. Less money is spent on Black patients who have the same level of need, and the algorithm thus falsely concludes that Black patients are healthier than they are. Reformulating the algorithm to eliminate the use of costs as a proxy for needs rectifies this racial bias, emphasizing the critical role of algorithmic transparency in healthcare applications of AI.

Conclusion:

As we continue to explore the integration of AI systems in healthcare organizations, it becomes paramount to embed inclusivity at its core. The transformative potential of AI should not inadvertently perpetuate disparities. A healthcare system committed to monitoring and mitigating the risks of bias in AI implementation is essential. By doing so, we not only ensure the responsible use of technology but also contribute to fostering equitable healthcare access for all.

Sources:

· ETHICS AND GOVERNANCE OF ARTIFICAL INTELLIGENCE, World Health Organization, 2021

· Artificial intelligence (AI) vs. machine learning (ML), Google

· Artificial Intelligence Risk Management Framework, NIST, 1/2023

· Subtle biases in AI can influence emergency decisions — MIT, 12/16/22

· https://pubmed.ncbi.nlm.nih.gov/31649194/:

· Wang et al: https://predictive-optimization.cs.princeton.edu