Giving New Insights:
Machine learning could help identify readmission risk factors that were previously unrecognized or the ones that were once thought to cause an increased risk of readmission as not predictive. One such example was noted at the University of Kansas. When they implemented the model, they found that patients on no medications were actually at higher risk of readmission than those on multiple medications, contrary to their belief. In another study that used ML techniques to predict 90-day readmissions for atrial fibrillation (heart condition involving abnormal heart rhythm), the researchers noted that AI tools led to improved understanding of data and helped identify new features that may contribute to the onset of atrial fibrillation and readmissions. The knowledge obtained from these tools will help identify the new and previously unrecognized patterns and augment physicians’ clinical decision-making.
Improving the Bottom Line:
Many readmissions cost even more than index admission. Reducing readmissions could help hospitals in reducing costs and avoiding penalties and cuts in reimbursement. By taking a personalized approach as guided by these models, resource allocation could be done wisely, potentially reducing waste. One of the predictive models mentioned above on surgical patients reported a potential saving of $20 million annually.
CHALLENGES AND BARRIERS
Despite these success stories and their benefits in handling readmissions, AI-based predictive models face several challenges in their widespread implementation.
Abracadabra:
It sounds like fun! But not in medicine. Many of these models are opaque and cannot explain themselves. One cannot see what is going on behind the scene nor understand how the machine comes up with specific predictions. This lack of transparency and lack of explainability make it hard for physicians to trust these models and their results. While it is impossible to know every step of the algorithm but to trust the model, one has to have at least a basic understanding of how the machine reached a particular result, what factors it took into account and how much weight it is giving to each of these factors.
One size does not fit all:
Another barrier to the broad implementation of these models is the problem with data diversity. Patient populations vary by region and even by the facility, and hospitals treat different patients with distinct approaches and interventions. Therefore, algorithms built on one subset of the patient population may not perform well in another hospital that serves a different patient population, creating what is referred to as “distributional shift” and introducing bias. One way to address the issue is to include a diverse patient population while training the machine. Another approach is to have a risk model that allows the health system to customize the tool by including features unique to its organization; however, this approach requires resources, time, and expertise not available to many hospitals.
CAPTCHA effect:
You might have come across websites where you try to sign up or create a new account, and it asks you to write the numbers and letters from the distorted image or identify pictures with a street sign to make sure that you are not a machine. That is because of the machine’s inability to capture text or image when presented in a different context. The same is true with clinical documentation. The notes written by physicians and other patient care team members, including nurses, pharmacists, physical therapists, occupational therapists, diabetic educators, case managers, speech therapists, and so on, contain a wealth of information about patient’s medical history and course of illness and are rich in vocabulary. AI cannot easily capture these. Hence, the pearls hidden in these notes that could potentially contribute toward patient outcomes cannot be plugged into the algorithm during score calculation, affecting the score’s predictive power and accuracy. Although some AI models can convert this free text, aka unstructured data, to more machine-readable structured data; however, due to the richness in vocabulary in clinicians’ notes and variations in documentation styles, these models are still not flawless.
Validation of Results:
This is another area where the readmission model struggles due to a lack of interoperability. If a patient does not get readmitted to the hospital where (s)he was initially admitted, it does not mean that patient was not readmitted to any other hospital. The model will have no way to figure this out, making the test’s validity a challenge. Improving interoperability is vital for the validation of these results.
Cost:
These models come with a price tag. The cost of collecting the data, preprocessing the data, verifying the results, deployment of these models, modifying the algorithm, and integrating them in a workflow need to be carefully evaluated. Furthermore, every prediction software that uses the magic word AI is not necessarily meant for your organization. Thus, doing a cost-benefit analysis and building and understanding the AI suitability framework is crucial before healthcare organizations commit to such technologies.
Though AI and ML offer several benefits in tackling readmission challenges, the challenges persist. Like many AI applications in healthcare, these models are not yet ready to be put on autopilot. To gain acceptance and make a vast impact, they require addressing the barriers, some of which can be overcome by improving interpretability and transparency to gain physicians’ and patients’ trust, increasing interoperability to validate test results, and including a diverse patient population to reduce bias. All this is not going to be achieved overnight and will require time. However, despite these hurdles and having room for improvement, the use of artificial intelligence to lower readmission rates will likely become the standard in patient care, and these techniques will play an increasingly important role in care management and transition of care.
“We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten.” Bill Gates
References:
