Precision Health and Artificial Intelligence by Arjun Panesar

Author:Arjun Panesar
Language: eng
Format: mobi, epub
ISBN: 9781484291627
Publisher: Apress
Published: 2023-01-27T05:35:40.178000+00:00

Figure 5-1 Relationship between interpretability and explainability

Interpretable, explainable AI facilitates patient-centered care by providing clinicians and patients with the necessary information for shared decision-making. Meaningful conversations about the potential benefits and risks of various courses of action can be supported with data and ensure patients receive treatments best suited to their situation, values, and priorities. One could argue that black-box AI is opposed to the tenets of patient-centered medicine. If clinicians cannot fully comprehend the features and calculations that led to a decision, they can’t explain how specific outcomes or recommendations were derived.

Reliability and Safety

Precision health systems must perform reliably, accurately, and safely. Bodies such as the FDA in the United States and MHRA in the United Kingdom promote safety and effectiveness by requiring providers to demonstrate these features. When applied to precision health technologies, ensuring outputs are technically reliable in that AI systems have learned correlations from previously seen data and are clinically reliable is paramount.

Ethicists must also balance patient empowerment against risks such as inconsistent oversight of at-home tests, the importance of clinically supervised care, false positive and negative results, incorrectly reported data, and unintended consequences. AI systems augment current care rather than replace it and may sometimes require human intervention.

Imagine an AI system trained to triage obese patients into high and low risk for heart attack. If the system understood from the data that researchers used to train the model that South Asian patients were less likely to die of a heart attack, it might recommend that South Asian patients are at low risk. However accurate this correlation from the data, the conclusion is a misinterpretation that research clinically contradicts. Being of South Asian ethnicity, in reality, increases your risk of heart and circulatory disease, type 2 diabetes, and a variety of co-morbidities.1

Self-improving AI presents another challenge in terms of ensuring reliability and safety. Current regulations assume that products should be clinically tested, manufactured, marketed, and used in a consistent, unchanging form. This is challenging with continuous learning AI, which is constantly changing due to the interpretation and analysis of currently available data. Because of the dynamic nature of continuous learning AI, new methods for ensuring the safety and reliability of such systems will be required. Continuously learning AI requires regulation that provides that ostensible improvements the continuous learning system makes to itself do not instead introduce errors into the model that could cause harm. At the same time, regulation must not require near-constant revalidation of the model.

As has been demonstrated with Caucasian-focused medical textbooks used to deliver healthcare education for generations, the evidence base for most genetic tests is limited and needs more data from diverse populations.2 Diversity and cultural differences play a crucial role in how stakeholders interpret safety in precision health. Until data from under-represented communities is collected at a representative scale, everyone should question it. Clinical tests are helpful only when they provide reliable, actionable information that patients and healthcare professionals can use for clinical decision-making, similar to how behavioral change interventions are beneficial only when they engage people in adopting and maintaining lifestyle behaviors.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.