As physicians, our toolbox for practicing the art and science of medicine hinges on decision-making. Making evidence-based decisions moves society’s health and well-being forward, and thus we are a profession predicated on continuous learning. A top challenge for physicians today is engaging with and applying data, information, and knowledge given the pace at which it grows in the digital age.
Help is on the horizon. The fields of data science, statistics, machine learning (ML), and artificial intelligence (AI) — in their best elaboration — can supply insight engines that bring trusted clinical decision-making support to the bedside of the patient: the right data, at the right place, at the right time, in the hands of the healing profession.
Bayesian Health presents an exciting example of delivering such an integrated bedside tool by applying ML/AI, practiced at a high level with physician participation, to synthesize effective collaboration between medicine and computer science.
The company’s AI-based platform launched commercially this week, revealing its groundbreaking progress in clinical decision support beginning with patient deterioration/sepsis.
Early recognition of sepsis addresses one of the most prevalent outcome, quality, and value metrics in the hospital setting. Led by machine learning expert Suchi Saria and developed at Hopkins Medicine with high quality data linked to real-time EMR, Bayesian Health’s platform emphasizes insight precision and user experience and is tested by top physicians and clinicians. The company’s state of the art machine learning and AI techniques targeting sepsis recognition have delivered exciting preliminary results — reductions in morbidity and mortality, the ultimate outcome.
The Bayesian Health sepsis module is just the initial commercialization expression of the Saria insight engine. Based on the platform’s sensitivity and precision, physician adoption has understandably followed. Moreover, in a key validation achievement, the platform has shown extension to the community hospital setting with comparable results. This validation portends potential to rapidly reach commercial scale — thus the official company launch.
The timing is propitious. Recently, the University of Michigan paused its use of the Epic-embedded sepsis AI model, based on performance results evaluated and published in JAMA Internal Medicine. As STAT News summarized, “A popular algorithm to predict sepsis misses most cases and sends frequent false alarms.” Quality matters, as does recognizing data set shift. Greater transparency in data and methodologies employed in emerging AI applications is a non-negotiable requirement for winning physician confidence and encouraging adoption.
Moreover, greater field testing is a must. Large technology businesses have been experimenting with healthcare market penetration. While AI models built in their labs show promise, implementation has repeatedly disappointed. Limited physician and clinical input is apparent, and trumpeted collaborations with a leading institution and/or healthcare system is often a study “n of 1.” If you know one HC system, you know only one HC system.
By and large, barriers to scale and translation to fragmented settings have been underestimated, and solution design hasn’t reached its expected potential. In addition, solutions which require dedicated on-site resources aren’t easily externally validated, generalizable, or scalable. Literature review hasn’t shown a whole lot of documented innovation implementation, physician engagement, and adoption. Few measure length of stay, morbidity, and mortality reductions in prospective studies. Bayesian Health is working on outcome research that measures just this.
Bayesian Health takes a different approach to clinical AI, one that focuses on quantifying and understanding how physicians use decision support, and they are letting the sun shine on their special sauce. With proof of principle and proof of product, Bayesian’s frequent publication of peer-reviewed results and its transparency to the scientific and medical community address key issues in the space: black box algorithms and unintended consequences.
I’ll be following both the big picture and specific indications as additional use cases emerge and commercial momentum builds. I’m anticipating a sequence of noteworthy developments over the next 12 months as Bayesian Health navigates and orchestrates its product development and expansion choices.
There is clear demand in the community for clinical decision support that can be measured with quality outcomes. Welcome to the toolbox Bayesian Health!
— Doug Given is Board Director at Health2047.
“External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients,” JAMA Internal Medicine. DOI: 10.1001/jamainternmed.2021.2626