The following is a guest article by Ori Geva, Co-Founder and President, Medial EarlySign.
The driving force of innovation – from vacuum cleaners and washing machines, to telephones, typewriters and computer processors – has been to simultaneously increase productivity and efficiency while saving time. And nowhere is this more true than with the advent of AI and machine learning, in which sophisticated computer systems are able to not only analyze vast amounts of complex data faster and more efficiently than their human counterparts, but can learn and provide insights based on this information.
Therefore, there is somewhat of a dichotomy when it comes to the application of artificial intelligence to healthcare, an industry where speed and efficiency can often be the difference between life and death.
Enabling computers to perform much of the costly and time-consuming work previously done by healthcare professionals is expected to pave the way for improved patient care when it comes to earlier diagnosis, more effective treatment, lowered rates of readmission, and ultimately improving patient outcomes. However, AI health implementations are often very large multi-million dollar projects which may take several years to implement successfully and prove their worth.
The Challenges of Implementation
A typical challenge with AI implementations often stems from over-ambition or a desire to have one solution platform for all needs. These projects tend to have a high bar and lofty goals, requiring large sums of money to execute, and multiple adjustments and reassessments throughout the process. They may lack supporting evidence that they will succeed and are considered, at best, to be highly experimental and costly. The scope of such projects is often so large that it becomes difficult, or even impossible to attain tangible results in a timely and cost-efficient manner.
Other projects may involve healthcare organizations collating data from disparate sources, into vast data warehouses or data lakes with different modalities combining to paint a comprehensive picture of the healthcare landscape. However, such projects require great overhead in terms of time, investment and infrastructure in order to create these repositories, as well as continuous investments to keep them up to date. This strategy is commonly used for population health purposes and has shown considerable potential. However, the cost and time invested can far outweigh the potential benefits gained from such projects.
Another issue with these mega projects is that health systems may become captive clients of large legacy systems that promise to deliver AI solutions and insights without necessarily prioritizing these projects. With economic value still being evaluated, and focus placed on lower hanging fruit and other risk areas, there is often a disconnect and delay in applying clinical AI into actual practice.
Focusing on Manageable Projects
Healthcare providers should focus instead on short-term, condition-specific projects that can deliver concrete results within a short time frame. The truly innovative should begin looking toward adopting AI exploration strategies. The impetus for this change in approach is, of course, the feedback and successes which these projects reveal in real-time. The industry as a whole could benefit from investing more resources into smaller, more focused projects with quicker implementations. The manageable size of these projects will allow us to prove efficacy quickly and make changes accordingly.
Executing a series of smaller-scale projects throughout the healthcare industry is also an effective way to help individuals at risk of serious and high-burden conditions, across diverse populations and with a range of medical concerns and histories. The short feedback loop of each individual project can enable an attention to detail which would not be possible in larger scale projects. Patients would thus benefit from the fast tempo of receiving, processing, and adapting the data.
Areas to Explore
It is imperative for providers to take stock of their resources and understand how to best leverage their existing systems and data. This includes assessing whether there are current screening modalities that can be leveraged to identify patients at risk for a specific disease, or an existing way to predict which patients are likely to experience an adverse event and how that might be prevented. AI becomes practical when seeking to optimize the infrastructure already in place and augment existing strategies, teasing out the absolute most that can be extracted from these resources.
Ensuring a Successful AI Project
The success of an AI project does not depend solely on getting the model right, but on ensuring that it addresses the right challenge and solves a real problem. This requires selecting an issue in need of meaningful intervention, where there is sufficient evidence to suggest that the project will result in improved outcomes. Success will also largely depend on the ability for the model to be integrated into hospital workflow with minimal disruption to the existing environment. This will also be essential to ensuring stakeholder buy-in – from hospital leaders and data science teams, to primary care staff and ultimately the patients themselves. All must be on board and willing to go through the learning curve and be educated.
Above all, we must connect and engage with patients, humanizing the project. In an era of increasing automation and digitization, it is imperative to remember that patients are not machines, hence our focus must remain on the human experience of care. They are complex individuals prone to making decisions which healthcare providers may not anticipate and contradict medical advice, such as a patient with a heart condition returning to their unhealthy habits after experiencing a coronary event. A successful AI project must take all these variables into account and, crucially, engage with patients to ensure that they adhere to advice and treatment therapies. Together, all of these factors are critical to consider for an AI initiative to be successful. While in hospital, patients tend to be compliant. However, once they leave the primary care environment, complications arise and it is vital to take the varying human tendencies into consideration. It is crucial to ensure that the benefits provided by AI can be monetized and translated into real results.
If a project focuses on just connecting pipes and collecting data, it has missed the point. Integrating the model into the existing system and engaging with patients are key to bringing about successful change, and this can only be achieved when initiated on a small scale, with a full understanding of the value chain.
This raises the question of whether AI integration is a vertical or a horizontal effort. And there is no right answer, as it is a learning process that requires experimentation.
However the benefits of pursuing smaller projects are clear, allowing models to be tested and tuned with existing systems to optimize for increased scale in the future. Dividing the challenges into smaller baskets and implementing them on a project-by-project basis can enable us to work smarter, learn faster, and affect real change. It’s not just about technology and who has the greatest algorithm, it’s about finding a more productive solution. Ultimately, it’s about bringing AI down to earth. As the late Herb Kelleher, co-founder of Southwest Airlines said: “think small and act small, and we’ll get bigger.”
About Ori Geva
Ori Geva is Co-Founder and President of Medial EarlySign. The company’s advanced AI-based algorithm platform helps healthcare organizations accurately stratify populations to optimize care for individuals and prevent or delay serious health conditions, by leveraging routine blood test results, and common labs and EHR data. Medial EarlySign creates actionable opportunities for better clinical decision making and early intervention to improve patient outcomes, focus financial resources, and reduce overall costs. Follow Ori on LinkedIn.
In fact, there are much more problems than described in the article. In reality, AI cannot be perfect in a matter that concerns some medical condition. Because if you teach AI thousands of operations, then in 1001 it will be completely different, for example, and only a real professional will be able to react correctly. It seems to me that AI is best used as a reference. Like wikipedia for example.
If you compare it against perfection, you’re right that it’s a problem. However, if you compare it against the alternative (people), then some error like you describe should be ok if it errors less than a human. That’s something we have a hard time doing though since we’re ok with humans not being perfect, but we’re not ok with AI being imperfect.