Medical AI Unused in the US, a 'Big Hit' in Africa

Alzheimer's disease begins to accumulate the toxic protein amyloid-beta in the brain long before symptoms appear. This accumulation is typically known to start 15 to even 20 years before a dementia diagnosis. Early treatment to suppress disease progression at this asymptomatic stage can play a decisive role in delaying or preventing the onset of the disease.

The incidence of Alzheimer's disease varies significantly depending on the presence of the APOE E4 gene. Individuals without this gene have a 10-15% chance of developing Alzheimer's by age 85. However, carrying one E4 gene increases the probability to 20-25%, about a 10% increase, and carrying two E4 genes causes the probability to skyrocket to 35-55%.

Traditional statistical methods have limitations in identifying correlations of rare mutations within vast datasets. To overcome this, AI uses a 'window-based agent' approach. This method allows small agents to meticulously search data, increasing the likelihood of discovering rare mutations that might be missed by conventional statistical methods.

Generative AI is already being utilized in various ways in clinical settings. A prime example is 'Ambient AI,' which records a doctor's consultation with a patient to automatically summarize diagnostic content and draft charts. This allows doctors to reduce the time spent on chart documentation and see the next patient more quickly, significantly improving work efficiency.

The US healthcare system uses Medicare's reimbursement schedule as its standard, which requires detailed line items for all medical procedures, including physician interpretation fees and machine usage fees, for insurance claims to be possible. However, there are no separate reimbursement codes established for AI interpretation yet, meaning hospitals have no basis to bill patients for the cost of using AI. This makes the actual adoption of AI medical technology in the field very difficult.

In the United States, the medical association effectively has the authority to determine medical fee schedules. This leads to ongoing concerns about AI potentially replacing doctors and disputes over liability for AI interpretations. Hospitals tend to prefer adopting AI only if it is a profitable and efficient system as a supplementary tool. These conflicting interests and institutional barriers are major factors hindering AI adoption.

Recent survey results show that despite high diagnostic accuracy, only about 30% of radiologists actually use AI tools. The remaining 70% of doctors either believe that direct interpretation is still more accurate or do not feel the need for reimbursement coding because AI adoption does not sufficiently increase diagnostic speed. This low utilization rate indicates that complex factors in actual medical settings are at play, separate from technological completeness.

In the field of rare disease diagnosis, AI has a relative advantage over humans. Vanderbilt University's experiments showed that while human diagnostic accuracy for rare diseases was only about 5%, AI achieved twice that, at 10-13%. This is because AI can remember and not forget vast amounts of medical data, making it advantageous for capturing specific rare cases. However, AI also makes mistakes more than 90% of the time, suggesting that the difficulty of rare disease diagnostic data is very high.

Recently, research has been underway for AI to analyze its own prediction failure rate to improve accuracy. This method involves the AI, when making a prediction, simultaneously calculating and reporting the probability of its own judgment being wrong. Through this, doctors only re-examine cases where the AI expressed uncertainty, thereby improving overall diagnostic accuracy from 90% to 96%.

Contrary to initial concerns about AI adoption, in actual medical settings, AI is found to play a supplementary role in increasing work efficiency rather than replacing doctors. In a situation where the patient population is rapidly increasing due to an aging society and there is a shortage of doctors, AI reduces the burden of treatment and interpretation, allowing doctors to care for more patients. In fact, there has been an increase in demand for radiology professors after AI adoption, confirming that AI is filling healthcare gaps and complementing doctors' expertise.

African healthcare faces an imminent collapse of essential medical services due to severe shortages of medical staff and infrastructure. For example, one psychiatrist may be responsible for as many as 200,000 patients, making proper treatment impossible. Even if specific treatments (like clozapine) exist, there is no system to continuously monitor patients' conditions. Furthermore, national psychiatric hospitals are beyond their capacity, and even basic education on diseases is not properly provided, leading to a severe healthcare gap.

In Nigeria, AI technology successfully overcame the limitations of human resources, achieving significant results in tuberculosis eradication. Using mobile X-ray devices and an AI interpretation system, 20 local personnel, not doctors, screened 94,000 residents over nine months. As a result, 2,414 tuberculosis patients were confirmed through AI interpretation, a remarkable achievement. This case demonstrates that AI can enable large-scale early disease detection even in areas without medical professionals.

In the process of adopting new medical technology, human behavior often poses a greater obstacle than the technology itself. Initially, manual checking by medical staff during examinations often led to diagnostic errors. Later, when an automatic traffic light (visualization) method that intuitively displayed diagnostic results was introduced, diagnostic errors significantly decreased. This suggests that no matter how good the technology, interface design that considers user convenience and behavior patterns is crucial.

In the US healthcare system, even medical equipment approved by the FDA often faces difficulties in adoption due to actual reimbursement issues. For example, for diabetic retinopathy screening equipment, even if the equipment is purchased and tests are performed, the reimbursement is only $44. Such low reimbursement makes it difficult to cover even the equipment operating costs, leading to unprofitability for hospitals. Consequently, even good technology experiences delayed adoption in real-world settings.

In Zambia, where medical personnel are scarce, an experiment successfully demonstrated how AI can dramatically improve the diagnostic capabilities of low-skilled workers. After just 8 hours of training on how to use AI devices, untrained non-professionals showed a diagnostic error of only 0.21 when compared to skilled medical professionals using advanced equipment. This shows that AI can simplify complex diagnostic processes, enabling non-professionals to diagnose diseases with high accuracy, thereby significantly improving access to medical services and reducing maternal mortality.

For AI researchers, errors are a lifelong challenge. Even if AI diagnostic accuracy reaches 99.9%, the remaining 0.1% error always exists. Recently, instead of solving this 0.1% error, 'uncertainty research' is being conducted where AI itself analyzes its prediction failure probability and notifies medical staff of uncertain cases. This is also an attempt to acknowledge AI's limitations and allow medical staff to make final judgments, thereby circumventing liability issues. While AI maximizes the advantage of providing rapid diagnostic results, the societal consensus on who will be responsible in case of an error remains an important task.