Cyber Doc: Avoiding Overtreatment

AI Doctors in a Quandary: Can They Deliver on Healthcare’s Promise?

The promise of artificial intelligence in healthcare – AI doctors that diagnose and treat with precision – has generated both excitement and concern. While advanced technology holds the potential to revolutionize medicine, recent studies raise critical questions about whether these “cyber doctors” are living up to the hype and, more importantly, if they are perpetuating or even amplifying existing biases in healthcare.

Imagine a scenario: groundbreaking medical technology exists to treat your specific illness, but your AI-powered doctor, lacking complete access to nuanced patient information, recommends a traditional treatment that proves less effective compared to the innovative methods accessed by patients with different socioeconomic backgrounds. The frustration would be palpable.

Now, consider this: what if the AI system’s recommendation is influenced not by information gaps, but by your gender or income level?

Recent international research suggests that increasingly sophisticated large language models are inadvertently amplifying the “disparities in treatment” issue already present in healthcare.

A study by researchers at the Icahn School of Medicine at Mount Sinai and Mount Sinai Health System, published in a Nature journal, found that patients identified as “high-income” were more likely to receive CT scans and MRIs. Those with lower incomes often received only basic examinations or none at all.

Meanwhile, patients flagged as “homeless,” were often directed towards emergency care, invasive interventions, or mental health assessments. The study evaluated nine natural language models across 1,700,000 consultation outcomes from 1,000 emergency room cases (500 real and 500 synthetic).

Previous research demonstrated that AI could predict aspects like race and gender from X-ray images alone. This raises concerns that AI physicians might be even more adept at “treating patients according to personal details” than their human counterparts.

Researchers suggest that the models themselves drive these “biases”, ultimately contributing to widening health disparities among distinct demographics. On the other hand, some patients potentially pay for unnecessary tests and treatments, not only wasting money but also jeopardizing the patient’s health.

Faced with these troubling findings, experts are increasingly focused on the critical need for the healthcare sector to adapt.

Are Cyber Doctors Learning the Wrong Lessons?

The problem of feeding “dirty” or biased data to large language models is becoming a growing concern for AI companies, and its implications in healthcare could be more severe.

One expert is quoted as saying that AI could, in the future, misdiagnose the common cold as cancer.

A multicenter, randomized clinical study in the US also seems to support that concern: researchers found a significant decrease in diagnostic accuracy of 11.3% when clinicians used AI models with proven systemic biases.

Some observers have jokingly commented: “Smart AI helps with small problems, while bad AI makes big mistakes.”

The Data Dilemma: A Double-Edged Sword

Data quality is, in fact, a crucial factor.

Researchers studying this issue, also pointed out several data-related issues:

* Data sets which don’t truly represent the patient population: For instance, people with low incomes, children, and pregnant women, tend to be underrepresented, making it difficult to conduct drug research.
* Poor data labeling: Biases could be introduced during the data labeling process due to subjective interpretations or inconsistencies in labeling standards.

More significantly, many unintentional biases already exist within the healthcare system.

Researchers have shown that doctors frequently interpret female patients’ pain as “exaggerated or hysterical” while viewing male patients as more stoic.

A study examining 20,000 patient discharge records revealed that women waited an average of 30 minutes longer than men before being seen.. Women also had a 10% lower chance of receiving pain scores in their medical records, and when they did, the pain ratings given to men were significantly higher (on a scale of 1 to 10).

Another British study found that women experiencing heart attacks experienced misdiagnoses at a rate 50% higher than men. Strokes and hypothyroidism are some other medical conditions that women tend to be misdiagnosed with more often.

Furthermore, in terms of over-diagnosis, research from Fudan University in Shanghai indicates that in China, the rate of excessive treatment for female lung cancer patients increased by over double, from 22% between 2011 and 2015 to 50% between 2016 and 2020. Among female lung adenocarcinoma patients, almost 90% experienced an instance of over-diagnosis.

Feeding this kind of information into large language models could have a detrimental effect. But can unbiased data solve this problem? Experts still disapprove this idea.

AI as “Self-Governing”: A Path to Salvation?

“Just teaching the positive things, and not the negative ones, doesn’t necessarily build a person with exceptionally strong morals.” Another expert made this statement when discussing the matter. He believes that finding solutions for bias based on data is not a good strategy. It’s challenging to build completely bias-free databases, and the training of truly unbiased large language models may not succeed even with good data.

It is argued that the fact that AI often has difficulty meeting people’s ethical requirements is primarily caused by the differences between large language models and human goals.

Medical practitioners, for instance, may strike a balance between patient care and illness management, while AI could disregard a patient’s suffering in its quest to “cure the patient.”

Human alignment aims to provide large language models with better ethical guidance and embed human values.

Methods that are frequently utilized include eliminating data throughout the training phase, adding instruction adjustments to help large language models understand what people say, and employing a reward function, in which human doctors can score the data and create a “reward model”, using reinforcement learning to iterate this methodology and lead the model to give responses that match human values. Retrieval-Augmented Generation (RAG) and Reinforcement Learning with Human Feedback (RLHF) are all tools used for human alignment.

In some ways, this approach also gives large language models an AI supervisor, continually controlling their actions and words.

However, this method is often thought to be a short-term fix. According to some researchers, human alignment might increase the administrative risks connected to artificial intelligence, among other things.

One researcher has cited that alignment’s expenditures and unavoidable losses would be hefty financial burdens on corporations. Before the team disbanded after a year in operation, OpenAI created a “super alignment” team that was scheduled to handle the alignment problem by 2027. The project was anticipated to utilize 20% of computing power, according to the project’s lead, OpenAI’s former chief scientist Ilya Sutskever.

In the research by the Icahn School of Medicine, the researchers modified the model, but “bias” persisted. The challenges and complexity of healthcare biases and over-diagnosis are hard to imagine. Though new technologies, like such AIs, can solve some of these problems, they are not solutions to the fundamental concerns.

The reality is that generative AI is essentially a probabilistic model, and it is difficult to prevent damage from low-probability events. This is a significant problem in the healthcare industry, where the margin for error is almost nonexistent.

Healthcare’s Evolution: The Ultimate Solution

Over-diagnosis and bias in healthcare can, objectively, be seen as related to the advancements in medical science.

According to the director of the Beijing Union Rockey Function Medical Center, there is a grey area between “precise medical care and over-diagnosis.”

To shift this grey area toward precision medicine, a crucial strategy is to define the borders more clearly using an abundance of data.

For example, considering the prediction of diseases, such as Alzheimer’s, some illnesses would affect almost everyone if given a long enough time horizon, but there is no purpose in treating a patient with the illness since they have passed away from other conditions before symptoms arise.

In order to resolve this problem, it is imperative to understand, in the opinion of experts, the two thresholds of harm and death. Intervening is undoubtedly necessary if the prediction indicates that a person may sustain damage from dementia before passing away. Otherwise, it is useless if damage occurs after death.

Finding the “threshold” or the border requires ongoing medical progress as well as sufficient data support and medical professionals. He stated this process is dynamic; physicians must continuously improve themselves and accumulate sufficient evidence to support their clinical practice.

In reality, the quantity of data required to completely analyze a person’s health may be staggering. For example, the functional medicine that this expert introduced to China in 2007 calls for more than 200 indicators to be tested to map out a person’s health.

During a patient’s treatment at the hospital, the balance is struck between medical expenses, the risk of examinations, and the need for tests. Doctors sometimes struggle to find the underlying cause and address the issue at its root due to the excessive division of specialties.

The functional medicines pay attention to the entire medical fields in order to support the advancement in medicine, and AI can also play a significant part in them. But to make this happen, a shift in perspective is needed for more patients and healthcare professionals, from a disease-centered approach to one that focuses on the overall patient.

Research has shown that breaking the information monopoly will help to reduce over-diagnosis. Patients seeing different doctors in different medical institutions to cross-validate their diagnoses and treatment plans can also deter doctors from over-treating, and the reduction may even approach 40%!

It is anticipated that over-diagnosis would continue to be reduced as new technologies like wearable devices and AI enable people to have a higher understanding of their own health.

It’s still important for the general public to remember that the human body has a remarkable capacity for self-healing, and that many physiological changes do not constitute a disease and do not require over-concern or treatment. For instance, the current anxiety surrounding lung nodules, thyroid nodules, sinus arrhythmias, and cervical erosion.

In this sense, new issues will develop around interpreting and managing an ever-growing amount of physiological changes and early-stage diseases as humans gain a better understanding of their health.