While examples of ANI exist today, there are larger issues that AI developers and the medical community must face and address before AI becomes mainstream in healthcare.
explainability
Healthcare professionals tend to make decisions using data obtained using technology that they understand or have a basic enough understanding of to trust. This may not be possible with AI. However, millions of learning parameters (weights of connections in the network) determine the output of a deep neural network, making the decision process unintuitive to understand. Even after visualizing the sensitivity of different parts of the network and browsing through thousands of these noisy images, we still don't find any learned rules that are easy to figure out. Inference is not a byproduct of algorithms. Therefore, explainable AI is important to provide enough insight to gain confidence in AI-based algorithms.
augmented intelligence
This is a term often recommended by organizations such as the American Medical Association. It focuses on the complementary role of AI in healthcare and emphasizes that AI is designed to augment, rather than replace, human intelligence. It also talks about the value that AI can provide. This comes from how we can combine the unique capabilities of human experts with the capabilities of AI to provide better care for patients. A similar term related to augmented intelligence is “human-centered AI.” It explores the need to develop AI-based systems that learn from and collaborate with humans in deep and meaningful ways.
Data quality and quantity
AI feeds data. The more quality data you have access to, the better you can perform tasks. Advanced algorithms require annotated data to reliably learn the task for which they are designed. Some medical professionals serve as data annotators, but this is a time-consuming and tedious task. Some medical algorithms can only be improved with large amounts of annotated data. Therefore, the dedicated contributions of data annotators are crucial for the benefits of implementing AI in healthcare settings. Therefore, we can conclude that data annotators are the unsung heroes of the medical AI revolution.27.
privacy issues
Medical AI needs access to medical records, data from health sensors, medical algorithms, apps, and any other source of information that it can learn from. Data may come from a healthcare provider or from an individual. Even if institutions anonymize data, it has been proven that individuals' profiles can often be traced back.
Legal issues and responsibilities
What if a deep learning algorithm misses a diagnosis, the doctor accepts that decision, and the patient suffers the consequences? What if an autonomous surgical robot injures a patient during surgery? There is an ongoing debate about who will be held responsible if moving robots or AI harm patients in the future. The current consensus is that if an expert uses a tool in circumstances outside of its regulatory approval, misuses the tool, or despite serious professional doubts about the validity of the evidence regarding the tool. If you apply the tool, you could be held liable. Use the toolmaker's knowledge to obfuscate negative facts. Otherwise, the onus is on the creators and the companies behind them.
trust
It will take a lot of time to trust self-driving cars and see how they react in familiar situations and whether they would make similar decisions in an emergency. As a result, it will take more time for patients as well as medical professionals to trust AI in medical diagnostics, support medical decision-making, and design new drugs. This must be taken into account when deciding to implement this technology in healthcare settings.
biased AI
One study concluded that commercial companies' facial recognition systems were 11 to 19 percent more accurate for people with lighter skin tones. These produced particularly inaccurate results when identifying women of color. In another example, AI was introduced into the US criminal justice system to predict recidivism. They found that the algorithm predicted that black people were disproportionately likely to commit future crimes, no matter how minor the initial crime. In addition to racial bias, AI algorithms often discriminate against women, minorities, other cultures, and ideologies. For example, Amazon's human resources department was forced to stop using an AI-based machine learning tool that the company had developed to screen the best job candidates. Because a smart algorithm turned out to favor men. Because these algorithms learn from the data they are fed, AI programmers need to understand the problem of algorithmic bias and proactively counteract it by adjusting their algorithms.28.
patient design
When designing algorithms for medical purposes, patients need to be involved at the highest level of decision-making to ensure their needs are met and their concerns and recommendations are incorporated into the technology. An example of its importance is how one startup developed an algorithm that can detect signs of Alzheimer's disease in the phones of patients in Canada. However, patients with a French accent showed different results. Such problems may be avoided by inviting patients early in development.
With active ongoing efforts to solve each of these, it remains an open question whether the algorithms that have become a common part of medical practice can address them all.29.
