AMC Health’s Data Advantage
AMC Health has the distinction of sitting on a trove of hundreds of millions of data points – mostly comprised of near-real-time inputs - over more than 20 years of remote monitoring deployments. This includes daily physiometric data from the home, daily medication use data from smart dispensers, and patient or caregiver reported data from the home on symptoms, behavior, environment, and access to care. To this we have married ADT, claims, lab and pharmacy data, as well as copious notes from the clinician’s daily interactions with their remote charges. Indeed, no entity in the remote monitoring and virtual care space has access to such a wealth of information just begging to be mined, analyzed, and turned into useful knowledge.
One additional advantage unique to AMC is our unrivaled repository of clinical process as it is applied to weaving virtual care into chronic condition care pathways. We would not have been able to distil this knowledge base were it not for our over 14 years of providing direct clinical services to our customers. If you do not fully understand how clinical decisions are made in the context of resource availability, you cannot hope to interpret the machine learning-derived patterns you are finding. Indeed, many companies who have not enjoyed success in real-world applications of machine learning can attribute their failures to the enticement of focusing on the data science itself without deeply learned, practical, clinical considerations. We, on the other hand, live and breathe those workflows every day.
What AMC Health Has Been Doing in Machine Learning
AMC Health has been investing heavily in advanced data science for several years now to make AMC Health’s CareConsole clinical decision support dashboard, and our support services, smarter and more effective. Our unrivaled, mature models include a highly accurate candidate selection model that predicts who is most likely to see benefit from virtual monitoring and care programs, as well as a powerful, machine learning-derived risk score embedded directly within CareConsole. This risk score is predicated on dozens and dozens of variables, and provides clinical users a probability of an adverse outcome – specifically ED, hospitalization, or death – in the next 7 days, sans mitigation. We’ve also honed models to predict those enrollees in virtual, chronic care management programs at risk of voluntary, premature disenrollment, as well as the optimal approaches recruiting and onboarding program candidates. This last, multi-modal model was the first to leverage our enormous repository of unstructured data – in the form of call transcripts and engagers’ notes – for use with large language models combined with traditional, observational predictive tools.
Perhaps most significantly, our effort to distill far more precise alert patterns from the myriad physiometric readings that come into our application every day. This has the potential to not only reduce the overall alert burden impacting program scale, but it is establishing previously undiscovered physiometric patterns that presage risk of patient deterioration that no clinician has ever had access to, precisely because no one has ever had the luxury of this volume of real-time data coming from the home facilitating deep learning to reveal complex patterns of significance. These include inter- and intra-day variability patterns of blood pressures, heart rates and bodyweights that are showing significantly higher correlation to true risk than have legacy alert values. This is truly game-changing as we are doing nothing short of changing how chronic illness is being managed on a daily basis given previously unavailable real time data from where they are living out their chronic illness ‘careers’, namely the home.
This work will soon be folded into the next generation of our proprietary risk score in CareConsole, which will eventually include not just structured, physiometric data, but, thanks to tools like natural language processing and genAI, all of that untapped, unstructured data on which we are sitting, such nurses notes, texts from patients, and transcripts of conversations with patients.
Current Machine Learning Explorations
As this predictive AI work proceeds, AMC is simultaneously developing AI agents that will interact with patients to solicit information required to validate alerts (while simultaneously providing the patient with scripted, self-care information) in preparation for a handoff to a human care team member. When combined – i.e. the above risk prediction tools in concert with the agentic AI tools – remote care managers will enjoy dramatically higher efficiency in tandem with significantly improved clinical outcomes. The patient-facing agents that are being designed for this project represent the fist step toward dynamic, patient-facing AI tools that can supplement the patient education provided by human clinicians, and do so in as human-like a manner as possible. This is not just an efficiency-boosting necessity, but an attempt to address the critical gap in self-care education that human clinicians couldn’t possibly bridge to a satisfactory degree.
Where This is Going
As AMC Health’s data science gets ever more accurate, we will be offering prescriptive analytics that will change established best practices. At that point, we will begin to know much, much earlier not only which patients are heading for trouble, but also which specific virtual interventions (including the most impactful patient education) should be prescribed, and in a manner that is bespoke to each individual patient’s unique morbidity profile. To get there, we intend to incorporate additional data typologies with which we have not yet experimented, including publicly available data (including environmental and epidemiological data), to make our predictive modeling continuously more accurate.