Small wins vs. big losses: AI in healthcare

All journeys begin with a single step. The journey to value-based care is no different. One foot in front of the other. The steady accumulation of those steps and one finds oneself a thousand miles from the starting point. The same will be the story of artificial intelligence in healthcare.

For AI in healthcare, there is more focus on the high-profile failures than the small successes — the incremental steps. Yet those small wins offer a vibrant story of transformation, re-invention, and improved patient experience.

It is these small wins, in concert with each other, that will alter the trajectory of healthcare in the U.S. and beyond. Let’s take a look at the range of some of these wins and consider the collective implications if adopted broadly across a major U.S. system.

Computer-assisted image/waveform diagnosis
AI has made some of its earliest strides in healthcare in automated assessment and diagnosis from medical images, such as MRI, CT imaging scans and pathology images, and waveform “images,” such as EKGs.

Advancements in imaging technology and the increasing prevalence of chronic diseases and cancers have led to petabytes of image data from increasing use of diagnostic imaging for early detection.  The exploding generation of medical images has demanded more automated image analysis. Deep learning image analysis algorithms have demonstrated remarkable accuracy and precision surpassing the best radiologists in detecting breast cancer from digital mammograms.

Equally impressive results have been achieved in predicting the risk of cardiovascular events and associated liver disease.  Leading institutions like Intermountain Healthcare have incorporated AI-driven automated image analytics engines as part of its enterprise imaging analytics infrastructure to assess patient disease risk and optimize care for a variety of conditions.

A few different technology vendors are in advanced stages of obtaining FDA approvals for regular clinical use of automated diagnosis tools.

Optimization of hospital operations
Hospital Emergency Departments are overwhelmed — regularly facing overcapacity in inpatient wards. The implications of care in this environment have financial implications. For example, in pay-for-performance arrangements, throughput measures, such as door-to-doc wait time, percent “left without being seen” (LWBS), disposition to inpatient admit time; and quality and outcome-related measures, such as length of stay and 30-day readmissions impact the bottom line.

In particular, LWBS has an immediate consequence of revenue loss. AI and machine learning solutions are examining clinical and administrative data to predict variations in demand enabling operations teams to respond accordingly.  Physicians, triage nurses and other ED personnel are alerted in real-time to ED backlogs and prompted to open additional rooms, adjust priorities, accelerate tasks and make other adjustments to reduce the backlog.  Hospitals are reporting double-digit percentage improvements in key ED throughput metrics through the use of predictive and prescriptive analytics.

Optimization of clinical pathways
As bundled payments make a comeback alongside other value-based alternative payment arrangements, there is a renewed focus on eliminating unnecessary, expensive care.

Healthcare organizations are challenged with managing and affecting optimal care at the right cost as more care is delivered outside hospital walls.  AI is starting to become an integral tool for health systems to understand how to optimally deliver care within the full context of individual patients’ situations.

AI solutions are analyzing data across the continuum to identify justified and unwarranted variation and drive toward standardized clinical pathways within the local population and ecosystem context. Payers are uniquely positioned to engage multiple providers participating in extended episodes of care, such as bundled episodes extending into significant post-acute episodes or even more extended palliative care periods.

Patient risk management
With the rise of “at-risk” financial arrangements at the heart of the shift to value-based care, AI solutions are increasingly employed to continuously and proactively assess future risk trajectories of individuals and populations.  Risk vectors include longer-term trajectories, such as disease progression, utilization and cost risks, and more tactical risk vectors, such as adverse patient events like falls within the hospital, readmissions after heart failure, and cardiac surgery episodes, infection, and surgical complications.

From a population health perspective, it is well recognized that today’s high-cost, high utilizers may not be tomorrow’s high-risk patients. It is also well recognized that there are a multitude of clinical- and non-clinical drivers of risk. AI solutions are providing unprecedented capabilities to examine a virtually unlimited set of variables around patients’ health, lifestyle, and socio-economic situations to predict, with high accuracy and precision, the likelihood of chronic disease progression and associated complications.

AI solutions are also providing transparency into the drivers of risk to inform personalized- and patient-level regimens.  Such predictive systems are now inserted into point- of-care decision-support systems and population surveillance systems.

Clinical Documentation Improvement (CDI), Automated coding
Natural Language Processing is a critical branch of AI and one that has improved immensely over the past couple of years spurred by the vast corpus provided to Alexa, Siri and their myriad friends.

While we are in the early stages of adoption, we are already seeing the use of speech recognition to automate clinical documentation, and reduce the cost and latency of medical transcription. Further, NLP will create a massive corpus for additional research around incorrect and missing diagnoses, ultimately prompting physicians and coders with recommendations to improve documentation and code/charge capture.  Improved clinical documentation and coding have direct consequences in performance on quality measures and billed services.  Since risk-adjusted payments are based on coded attributes, this is already having a direct impact on revenue/reimbursements.

The key here is that the list of use cases for which AI is actually making progress is far longer than the time-worn readmissions or sepsis examples. The technologies that we collectively call AI are having an impact today — and will compound themselves over time, dramatically scaling the price/performance characteristics of healthcare. Their impact far outweighs a few public failures.

Furthermore, this is the future of healthcare — better care and lower costs.


This article was written by Jonathan Muise from MedCity News and was legally licensed through the NewsCred publisher network. Please direct all licensing questions to