The following provides an overview of the purpose and solution components of the Lifemesh Analysis Foundation.
Proactive Health Risk Stratification
The stratification of risk is a critical component to improved health. Risk stratification serves as a predictive process related to a specified event. It is often used to define broader policies, focus areas for research, clinical study cohorts, and other “macro” level initiatives, however Lifemesh extends these broader concepts to also include the active (proactive) intervention at the individual level.
Predictive modeling and risk stratification lie at core of our solutions. While only a component of the overall foundation, they warrant special mention.
To accomplish any predictive model, and thus the ability to stratify risk in any manner, one must first analyze existing data for a broad population and identify correlations, patterns, and health trajectories. This is done by analyzing a wide range of data inputs ranging from health history, social determinants of health, demographic information, and health interactions and patient protocols. It is through this aggregation of population data that one can derive the appropriate segmentation, and ultimate risk scoring needed to implement changes.
To be clear, analysis is not performed for the sole purpose of providing insights into the past, but rather to address the end goal of proactive intercedence in individual health journeys to optimize health and to avoid critical events. In other words, we must look at the long view for forming and refining predictive models but focus on the immediate to implement change.
Risk stratification must remain fluid. Specifically, risk stratification must incorporate a “learning” approach to self-refinement by providing the flexibility to introduce new changes to lifestyle, care, therapeutics, etc. (protocols). As the primary goal is to intervene at key times in one’s health, the tracking and success scoring of that intervention must also be tracked to validate success, and to ultimately identify changes in overall risk.
Augmented Care Effectiveness TM
The Augmented Care Effectiveness (ACE) platform represents the predictive analytics and machine learning (ML) components of the Lifemesh Analytics Foundation. ACE is comprised of two unique elements:
Our predictive model starts with population analysis that identifies possible segmentations that maximizes informative correlations from demographic and medical variables chosen for analysis. This allows our predictive models to capture nuances of a specific segment to more accurately project how risk levels for a particular event progresses over time.
In addition to risk stratification during our predictive modeling process, time to event analysis allows us to quantify the urgency of an intervention. Time-to-event analysis, also referred to as survival analysis, strives to predict the time horizon for a given event to occur. While often focused on mortality, our model identifies time to event analysis for a wide range of critical events, over varying observation windows. Ultimately, this approach quantifies risk in a way that allows stratification that leads to the identification of high-risk individuals in need of urgent attention, or at-risk individuals requiring ongoing observation.
To facilitate the goal of actionable information, Lifemesh has created a rapid experimentation solution known as the Healthcare Experimental Learning Platform (HELP). HELP allows for the dynamic creation of cohorts that can be broken into control and study groups for the delivery of new and existing protocols to care. HELP then tracks the ongoing success of a change in care and automatically defines the success of a new protocol with statistic accuracy. Through HELP rapid experiments can be run in parallel with small clusters of a given cohort to demonstrate promising (or not) interventions to care. While not intended to serve as a clinical study, it does provide a rapid quick view into the feasibility of a given change in care, and the cost/benefit likely derived from such change.
While HELP can be quite granular and clinical in its approach (i.e. therapies for a targeted condition, or the impact of changes in diet plans or sleep on targeted vitals), it can also be focused on the impact of non-critical protocols such as the introduction of dog walking to reduce the number of falls in an aging in place community. In fact, it can even be used to validate the effectiveness of larger initiatives such as introduction of Remote Monitoring and the subsequent reduction in admissions, readmissions, or other critical events.
HELP ultimately becomes the “proactive” component of the broader Proactive Health Risk Stratification approach described above. By combining the predictive risk models with a tool to enable proactive care, and to dynamically track the effectiveness of that change in care, Lifemesh can deliver a complete analytics solution to optimize health and wellness.
While a majority of the Lifemesh analytics focus has been on predicting critical events related to specific chronic diseases (Diabetes, CHF, COPD, Hypertension, Cancer, and HIV), we recently completed a project of the COVID-19 pandemic. We leveraged our risk models in unique ways to identify risk of infection, healthcare needs escalation, and mortality based on a combination of COVID-19 testing, infection, and mortality data in combination with census data, and population health and mortality data provided by the CDC. We further modeled the data to identify geographic and demographic profile stratification clusters to define propensities of infection and spread to specific communities across the country (think community cohorts). Lastly, we developed two models that identified the unintended consequences of COVID-19 in terms of impacts to the overall healthcare system while the country was hyper-focused on the COVID-19 cases at the unfortunate expense of the massive ongoing healthcare needs. The first model is based on historic responses to unemployment and crises in terms of suicide and drug overdose events. And the second focused on the likely implications to the lack of attention to care for those not affected by the coronavirus itself (missed screenings, reduction in ER visits for heart attacks and stroke, missed cancer treatments, etc.). As described above, as much of the data capture is only now taking place on these non-coronavirus missed care opportunities, we leveraged a sophisticated algorithm to define the degradation of care and impact of this lack of focus based on Health Resources & Services Administration data comparing impacts to mortality measure due to a sliding scale of medically underserved populations.
While COVID-19 is not the primary focus of our research, this exercise allowed us to extend our models to the global population, and to provide capabilities that can be reused to rapidly model for other epidemic/pandemic issues such as the flu, a second wave of the coronavirus, or other communicable diseases.
As stated earlier, however, the primary goal of Lifemesh is to make the data actionable, and to then further track the success of these actions and develop broader strategies of care interventions at the individual level. So, while the above COVID-19 project has been quite valuable and enlightening, the end goal must be on turning the information into action.
The Importance of Data in Predictions
Data sparsity is the enemy of all predictive models. Few datasets have all the information one might want to develop highly accurate models, however, the amount of data available from disparate sources is quite immense. It is the consolidation of data that can provide the greatest value, when data can be viewed with a similar level of granularity (i.e. individual, city, county, state, etc.). Frequently elements of data may still not be available to ascertain a specific goal. In this case the use of a proxy for that data may often be used to represent the information in question. Used alone, or in combination with other data, a proxy can provide strong indicators that closely resemble the missing data element.
Use of a combinatory proxy may in fact provide greater value than having had the data itself. The proxy may provide a much larger set of data to work with and may also provide a less targeted variable to be used in the analysis that might otherwise skew results too dramatically based on this single factor. This is of course true with actual data elements as well and is in fact the purpose to the development of member cohorts.