By Amer Saati, MD, MS, Physician Informaticist, Northwell Health
One of the new trends in health IT is the use of Artificial Intelligence (AI) in Healthcare settings. It has been a black box to many healthcare professionals and front end caregivers as they try to conceptualize it at point of care. Many hospitals, ambulatory clinic and post-acute settings are using decision support systems in some shape or form though some of these techniques don’t really fit in Artificial intelligence definition or methodologies and many people define AI in completely different ways.
- Identify potential of leveraging AI in population health management
Let’s walk through a population health management typical workflow. Population health is the philosophy of having a holistic view about individuals in certain groups that support the transformation from fee for service model to a value-based. It includes interacting with patients during hospitalization or ambulatory visits in addition to monitoring targeted patients at any point pre-admit or post-discharge. It starts with identifying emerging or high risk patients whether they are grouped by specific clinical conditions, co-morbidities or driven by predictive risk models. Then the next step is to outreach targeted patients either actively during their hospital encounters or proactively before events occur with an aim to do proper care coordination to keep patients healthy. The ultimate goal is to reduce unnecessary hospitalizations, improve readmission rates, shortening length of stay and assign a care navigator who will follow up with the patients to do better care coordination to achieve these goals. Additionally part of the workflow is to identify gaps in care, social determinants of health, link patients to the right services whether internally if we have the right type of service/ resources or externally to community based organizations (CBO).
“In a nutshell, the aim is to connect all the dots about our patients across the continuum by making data actionable.”
- Do we need to use AI in population health management?
The question is not whether we need to use AI or not but is really to understand the problem we are trying to solve then evaluate if AI is the right solution. It is like any other technology we try to implement in clinical settings. Clinical workflow comes first and AI needs to support that model otherwise it will always be considered as an extra effort for our clinicians and would face a resistance otherwise. Here are common opportunities to think about in population health management:
Patient Identification, Risk Stratification: We have used many direct approaches and predictive analysis to identify high risk patients mostly from EMR data that includes a combination of length of stay, chronic conditions, previous ED visits, etc. In recent years, many AI vendors started to leverage more advance technologies that consume external data from community to fill social determinants of health, other public data that feeds the machine to give a comprehensive view about targeted patients. The new vision is to uncover emerging risk patients in addition to focusing on high risk patients.
Identifying care plans and Improving interventions: This is an interesting area that requires some automation if prior knowledge exists about the targeted patients and ability to suggest certain interventions by using machine learning approach that gives probability of successful outcome by performing interventions or enhancing efficiencies by doing tasks in a specific order that proved positive outcome previously.
Care Coordination, Patient Engagement: Shifting from silos to a holistic view about the patient. Leveraging wearable devices to better understand physiological changes while the patient is at home and leverage that data to predict high risk events and bring that data to action.
Identifying gaps in care, social determinants of health: There is a huge need to fill that gap. Some data are embedded within progress notes, discharge summaries or other nursing notes that would add a great value to fill that gap. AI has the potential to assist in searching internal notes and look for unstructured data leveraging Natural Language Processing (NLP) or also by linking to community related data that gives more insight on surrounding circumstances then use that to identify certain trends about individuals and communities.
Internal/ external referrals with ability to close the loop: How do we match the patient with the right service? How do we pick from a list of choices based on quality of service, response rate, and financial agreements. How do we close the loop once we refer patients out of network? Many of these questions could be answered by simple algorithms but also in some area by applying AI or a combination of AI and other products.
Patient Engagement, personalized data: This area overlaps with population health as we think out of the box and try to leverage data driven by patients directly including texting platforms, mobile devices or accessibility of data through social media. There is a huge debate about using that information to impact treatment plan (personalized approach).
Operational and technical considerations: Since we are still in an exploratory phase, it is recommended to use an agile implementation methodology and continues evaluation/ validation cycles of machine learning algorithms throughout development cycle and post go live. Also to make AI solution scalable, there is a need for a good representation of data even if that means more investment to make data available, data interoperability and data warehousing.
Most of the time in order to get buy-in from executive leaderships on using AI there is a need to show a proof of concept and potential return of investment especially that many AI solutions are not financially feasible and some organizations have that on their wish list but still have a skepticism on how to mobilize funding on big scale projects.
The last recommendation is to keep it simple. Whether to partner with vendors, optimize EMR AI tools or build internal teams, always think about a feasible solution that fits smoothly in existing workflows and start addressing problems one at a time to make sure AI is the right solution for problems we are trying to solve.