Medical Affairs

The adoption of artificial intelligence (AI) products in the healthcare sector is increasing at an exponential pace over the past few years. Whether it is to assist clinicians in diagnosing ischaemic strokes for early intervention, picking up nasty lung nodules on chest X-Rays, or predicting children’s biological and skeletal maturity by assessing bone age, we can see the increasing number of areas where AI-powered products support the clinicians’ daily work. Clinical AI systems in imaging have also played an important role in combating the COVID-19 pandemic, for example in hospitals to screen for positive cases of infection during the early days in 2020.

In China and everywhere in the world, the number of use cases and products from different vendors has increased tremendously over the last few years, and there may be a consolidation phase as start-ups merge or business owners simply decide against further investment to remain competitive. However, the number of products available in the market is still innumerable, and will continue to grow as innovative clinicians and data scientists work together to provide more precise and targeted diagnosis and treatment to patients.

What is lagging behind is the systematic framework or process in which hospitals assess which AI products, whether clinical or non-clinical, to bring into the organization. The typical pathway seems to be a clinical leader or hospital executive getting connected with a product vendor who demonstrates the AI product with all the bells and whistles, and the awed hospital leadership brings the product into the organization because of the promised potential for increased clinical accuracy, efficiency and productivity, or simply better understanding of the business.

However, it is absolutely critical that hospitals develop a robust evaluation framework to assess AI products that they might wish to adopt, at least certainly for clinical AI products that aid diagnosis or treatment of patients. Most hospitals nowadays have a process for the introduction of new technology or procedures, and similarly clinical AI products should require a parallel evaluation process, albeit with slight differences, such as the need for “explainable AI” – the ability to be able to explain in laymen’s terms what the AI “black box” is doing for you, the patient.

At United Family Healthcare (aka New Frontier Health), a premium healthcare network that operates eleven hospitals in Tier 1 cities in China, some of our hospitals jumped on the AI technology bandwagon early on as well. Instantaneous AI translation is utilized for our medical records, as the regulation in China requires medical records to be documented in Chinese, and a significant proportion of our clinicians are foreign trained and Chinese is not their native language. Robotic process automation (RPA) is also commonly used, such as when vaccination use and product records are required to be entered into the Centre for Disease Control’s database as well as the local hospital’s medical record, which are not directly linked. RPA reduces duplication of work for doctors and nurses. The example above itself saved over 0.4 FTE of these clinicians time in a single hospital.

The above use case examples are relatively lower risk when it comes to quality of care or patient safety, and perhaps it might not be necessary to spend too much time on the initial evaluation (although one might say the accuracy of the medical records following translation is critical!) When AI tools are used in the clinical arena however, such as for clinical decision support in imaging, or for endoscopic procedure guidance, the need to comprehensively evaluate their performance metrics, data security, or to have a mechanism to identify rouge AI results then start to become important. You probably don’t want to find out after implementation of the AI technology that the false positive rate is higher than you would accept, or that for the same cost you could have purchased a superior product!

“It is absolutely critical that hospitals develop a robust evaluation framework to assess AI products that they might wish to adopt, at least certainly for clinical AI products that aid diagnosis or treatment of patients.”

United Family Healthcare has adopted clinical AI products in various imaging modalities, and it has since proven successful in terms of improved increased productivity and reduced misdiagnosis. However, it had followed a similar path to adoption as what is described above. There has been no adverse events as a result of the clinical AI tools, but it is important to develop a mechanism that systematically capture and review concerns that may arise. Opportunities that were forgone by this approach also include the chance to compare similar products that were in the market, and to ensure the vendor provides sufficient information to help us explain the system logic to our consumers.

We have since developed a governance structure and evaluation framework that assesses the clinical AI products in a number of areas, namely regulatory licensing, data integrity and performance metrics, interpretability, data privacy, local IT system and workflow integration, training, facility readiness, financial impact and product competition. The well-embedded incident management system is leveraged to identify, alert and review potential concerns. This is overseen by the Clinical AI Advisory Committee.

New medical devices and technologies undergo a robust process for introduction into a hospital, and we believe a similar approach is required for clinical AI products. Certainly in China, where the introduction of new medical technology or procedures are tightly regulated, it seems that it is only a matter of time before introduction of clinical AI products undergo similar regulatory requirements. Hospitals would do well to establish an evaluation system sooner rather than later, and AI technology companies would be great partners to support the hospitals in this journey.