Artificial intelligence use is growing rapidly, according to every metric you can track. Recently, StudyFinds chronicled how prospective patients are feeling about this new modality, particularly as it affects reaching a medical diagnosis. It’s clear people still yearn for the time when the patient-doctor relationship was one of medicine’s most powerful tools for successful healthcare. So, where does this technology stand in the field of medicine right now?
AI is still in its infancy, but already some algorithms are performing on par with, and sometimes better than, human-developed, standard-of-care protocols for disease management. AI has matched human performance in the efficiency and accuracy of skin cancer classification and in medical imaging diagnostic procedures. A weakness of these findings, however, is that most of these models have been tested retrospectively and lack randomized controlled medical studies. The same is problematic with the hundreds of AI-enabled medical devices that have been approved by regulatory agencies, although lacking randomized controlled studies.
An example is the Epic Sepsis Model, a proprietary prediction model for sepsis — an infection throughout the entire body, detectable in blood tests. The model is used in hundreds of U.S. hospitals, yet it has not been adequately tested. It showed a poor ability to identify septic patients.
A literature review on AI use in radiology (X-ray, imaging) demonstrated that few studies have addressed radiologists’ and doctors’ attitudes toward adopting AI innovations. These are just some of the issues of uncertainty regarding AI systems use in clinical practice.
In the interest of a clearer understanding of AI in medicine, an extensive review analyzed what randomized studies were available for AI algorithms that had been used in clinical practice prior to 2023. The analysis examined the potential of AI to improve care, incorporate patient behavior and symptoms in management, and clinical decision-making efficiency. It also identified areas that require more research. The multi-center collaborative review was published in The Lancet Digital Health in May of 2024.
The review included 86 randomized studies that predominantly focused on gastroenterology (43%), followed by 11 on radiology (13%), five on surgery (6%), and five on cardiology (6%). Of the 86 studies analyzed, only 18 randomized studies assessed the effect of AI on quality-of-care management in clinical practice.
There were some favorable findings in disease management. Implementing AI systems:
- Optimized insulin dosage, improving the average time spent within target ranges for blood glucose.
- Improved the average time patients spend within target ranges for blood pressure.
- Reduced prostate tumor volume in radiation therapy.
- Predicted the risk for diabetic retinopathy, a condition primarily found in diabetics, which can lead to blindness. It caused patients to follow their physicians’ instructions more closely to limit disease progression.
- Reduced postoperative pain scores.
- Provided cancer mortality predictions that increased disease discussions between oncologists and patients.
- Facilitated the identification of patients with atrial fibrillation who were at high risk for stroke.
Randomized AI studies primarily investigated the fields of gastroenterology, radiology, surgery, and cardiology. There are relatively few studies on primary care, the foundation of all specialties. The lack of information in this field points to what may be the greatest area of need for future research.
Most studies were conducted in individual countries, primarily the United States, followed by China. According to the review authors, these data emphasize the need for more international collaboration and multicenter trials. It would contribute to ensuring the generalizability of AI systems across different populations and healthcare systems.
What do these findings tell us? AI can enhance healthcare, but there’s a very long way to go.