One scholar highlights the problematic ways in which artificial intelligence and health privacy interact.
As artificial intelligence (AI) continues to revolutionize healthcare, the need to protect personal health data becomes even more important. But intervening attempts to protect personal health data can also slow the development of healthcare AI.
Society can establish a new balance between technological progress and data protection, protecting personal health data without stifling innovation. In a recent article, his law professor W. Nicholson Price II explores the complex relationship between enhanced privacy protection and healthcare AI, striking a balance that enables both AI innovation and personal privacy. We offer a unique perspective to achieve.
Privacy issues posed by AI are evolving rapidly. Price discusses new privacy issues created by AI. AI’s ability to find patterns in seemingly unconnected data can lead to unintentional disclosure of information that should not have been made public. Price illustrates the problem through the example of a large company where AI analyzes shopping habits to infer fertility status. This observation is unlikely to be made by a human analyst.
To establish the impact of AI on medical privacy, Price also demonstrates how AI undermines mechanisms such as anonymization used to protect medical data. Anonymization is a commonly used method to protect medical privacy by removing identifiers from personal information, Price notes. The governing law rule on health data privacy, the Health Insurance Portability and Accountability Act (HIPAA), oversees only identifiable health information and ensures a safe place for anonymized information.
By removing the listed identifiers, health data users can avoid HIPAA monitoring. And AI can do just that. Given enough computing power, AI can re-identify anonymous data. You can also make advanced inferences about your personal non-health data.
Price argues that AI has made this anonymization less effective. He notes that researchers are using his AI to re-identify the “majority” of patients using anonymized information. Although AI has the potential to degrade medical privacy, AI will use its ability to process large amounts of data to extract highly specific health information that was too costly to anonymize before AI was introduced. can improve medical privacy by anonymizing Reveal the patient’s identity.
The race between increased and decreased privacy, and the resulting arms race between stronger privacy protections and AI that could defeat those protections, would create a “continuous dysfunction,” This influences Price’s recommendation for a new understanding of medical privacy.
Medical privacy can cause problems for AI, even if AI can cause problems for medical privacy. Price explained how increased medical privacy could slow AI development by making health datasets more costly to use, less accurate, and more difficult to create. increase.
Many modern medical and patient care products, such as telemedicine applications and patient and diagnostic tools, use AI, and the US Food and Drug Administration has approved hundreds of AI-related products over the past few years. While these products are creating new diagnostic, therapeutic, and organizational protocols, Price says they face many challenges, including regulatory burdens of privacy protection, inaccuracies in datasets, and trust issues that undermine data collection efforts. It summarizes the privacy obstacles that hinder more rapid development.
Developing accurate AI requires the use of large amounts of data. However, removing identifiers from the data to make it HIPAA compliant reduces its usefulness in AI development because it makes it more difficult to develop accurate models. When AI training requires long-term health records, privacy regulations can make it difficult to connect data points to form a coherent long-term record. Price concludes that these hurdles have increased his AI development costs and slowed his progress.
Bias in AI training datasets is also a major concern for health AI developers. While some bias is inevitable, Price argues that privacy protections increase data collection costs and reduce the accuracy of datasets, increasing bias. Increased privacy protections can increase the cost of obtaining patient approval and consent. As a result, only a small percentage of hospitals share data with medical AI developers and other medical researchers.
These hospitals tend to be well-resourced hospitals and academic institutions in urban areas rather than rural hospitals or community health centers, Price said. But when a health AI trained only on urban data is applied to rural settings, the AI’s performance may suffer, as is the case with IBM’s Watson for Oncology.
Price demonstrates this deterioration in the example of IBM’s Watson for Oncology, a health AI tool aimed at improving cancer care. Watson for Oncology learned from well-resourced data from Memorial Sloan Kettering Cancer Center in New York. However, IBM chose to stop marketing Watson Health after a data audit found the AI-provided treatment recommendations to be flawed due to its geographically limited training dataset.
Price said of a follow-up study that found that the health AI dataset was “disproportionately trained on the California, Massachusetts, and New York cohorts, with little or no participation from the remaining 47 states.” I am arguing. AIs trained on these datasets reflect flawed data imbalances, and when applied to data in these underrepresented areas, the AI may fail or perform poorly. be the cause.
The data collection process itself can also introduce bias issues, Price elaborates. Different populations have different levels of willingness to allow their data to be used in research. Price argues that the “long history of systemic racism and prejudice that exists within the health care system” undermines trust in it. This systemic distrust makes it even more difficult to collect diverse data. As a result, health-related AI is demographically imbalanced, which can lead to further bias.
The AI’s ability to identify patterns in datasets can also introduce bias. Price illustrates this feature with the example of smartphones with different privacy protections. Smartphones with tighter privacy protections tend to be more expensive, leading to a demographic divide with users of cheaper phones. AI trained solely on data from more expensive smartphones may be less accurate when applied to data collected from users of cheaper smartphones.
Price concludes his analysis by advocating legal changes to the regulatory regime covering health privacy. One of his ways, Price says, is for regulators to revise his HIPAA to mitigate potential detrimental effects on dataset creation, and the biases that can arise from that detrimental effect. I argue that it is. The less inequality HIPAA creates, the less negative impact it will have on the development of medical AI.
Alternatively, Price suggests that by allowing individuals to lower their expectations of health privacy and use personally identifiable health data in research, they increase the amount of information in datasets and have AI learn from those datasets. I suggest that it can be improved. By doing so, people can benefit from better health AI. Balancing health privacy and the need for AI development may be difficult, Price concludes, but the benefits of improving the complex relationship could be significant.
