Artificial Intelligence (AI) . . . it seems like every IT organization is claiming they use it. Feels like only yesterday when an avalanche of companies were rushing to adopt the descriptive label of dot com or cloud or crypto or artisanal. New day, new name. But are all of the AI claims valid, or are the creators just relabeling traditional Boolean logic systems as a marketing tool?
The definition of Artificial Intelligence (AI) is generally accepted to be “a computer system that is able to perform tasks that ordinarily require human intelligence.” This means contextualizing data (understanding the underlying meaning of that which is being observed, as opposed to just recognizing key words or data elements). It’s about comprehension and assimilation and (machine) learning. Intuitive thought over recognition. Some common examples of AI applications include advanced web search engines (Google), recommendation systems (YouTube, Netflix, Amazon), understanding human speech (Siri, Alexa, ambient transcription), self-driving cars, and competitive gaming (chess, GO).
Within the healthcare sphere, AI has heretofore mostly been applied to revenue cycle activities, cancer research (think IBM Watson) and drug discovery. More recently, however, the technology is increasingly being leveraged for ground-level clinical endeavors like population health management, clinical decision support, genomic therapy and personalized medicine. Below is a quick primer on a few real-world AI healthcare applications currently in use:
Forecasting: Population Health Management
- Risk stratification of patient populations (predicts which patients are at risk of developing a disease or experiencing an adverse event)
- Predicting low-weight, high-risk pregnancies
- Predicting medication adherence (via filled prescriptions)
- Determining the Risk of Hospital-Acquired Infections
- Predicting Clinical Pathway* Efficacy (effectiveness of treatment plan) including the influence of social determinants
*A clinical pathway is a document outlining a standardized, evidence-based multidisciplinary management plan, which identifies the appropriate sequence of clinical interventions, timeframes, milestones and expected outcomes for a homogenous patient group. In other words, a specific treatment game plan.
- Predicting Non-Adherent Patients
- Predicting Errors/Complications (clinical or medication)
- Predicting Sepsis (infections), Atrial Fibrillation, Congestive Heart Failure, NICU need, Transplant Cases
- Determining the Risk of Bed Sores
Real-Time: Clinical and Administrative Support
- Determining Treatment Effectiveness (quality of results)
- Early Disease Detection
- Bed Management
- Automatic removal of PHI (identity anonymization) for analytics sharing
- Robotic Process Automation (RPA) applied to revenue cycle management (coding/posting/denial management), reading images, reducing readmissions, natural language processing (unstructured data)
- Reading x-rays, other images and doppler ultrasound
- Diagnosing skin conditions
Although there are numerous “out-of-the-box” solutions available for the above use cases, each typically requires a degree of customization and tweaking in order to optimize the system to achieve the desired results. This means that an organization must either purchase the solution and dedicate staff to this task, or outsource the entire function to a third party (i.e. a population health management firm).
As previously discussed in other postings, the runway for AI applications contains endless possibilities that will be enabled as the technology evolves, such as autonomous robotic surgery, machine-driven diagnosing and clinical/surgical decision support. But the path to this evolution will no doubt contain multiple friction points, as healthcare providers gradually (and grudgingly) become disintermediated from their traditional roles and responsibilities, and the overall “trust factor” becomes elevated in importance.
Paging Dr. Howard, Dr. Fine, Dr. Howard . . .
Clearly, we can do better with respect to our current healthcare delivery in this country. But are we OK with where all of this is going? I guess time will tell. As with any advancement in technology, AI adoption will be largely contingent upon value perception, utility, trust, user experience and the degree and quality of its impact on society. We’ve learned time and again that new is not always better, but when it is, it’s usually pretty great.