As the COVID-19 pandemic began to gain ground globally, Lifemesh commenced an effort to contribute through the dissemination of relevant advanced analytic projects.
We identified three objectives where we can deliver value given our focus on risk stratification and risk mitigation in care plans:
Symptom and Conditional Risk Stratification
The ability to identify the need to be tested based on symptoms is relatively straight forward and requires neither advanced analytics nor machine learning to identify those who should be tested for the coronavirus – or so we thought. In actuality, as we are learning, a large number of people appear asymptomatic or show very mild symptoms, so may, in fact, carry the virus without it ever being detected.
While we continue to work with thought leaders in the healthcare industry to close the gap of data required to proactively identify those who have contracted the illness, the data is not available. Despite this, identifying the risk profile of individuals’ hospitalization, intubation, and potential death is in many ways easier to ascertain via analytic models, particularly as the data continues to evolve. Combining health profiles, comorbidities, initial symptoms, and population impacts can provide a very good understanding of those most at risk, and consequently those in the greatest need of social distancing and other mitigative measures.
The Symptom and Conditional Risk Stratification model created by Lifemesh leverages symptom inputs, health profile data, census, and social determinants of health, in combination with geographic COVID-19 hospitalization and death statistics. The largest challenge to this endeavor, however, has been the timely availability of health-related statistics at the appropriate grain of data that tightly link to the geographic locations (i.e. Counties) and the comorbid conditions of the specific individuals impacted by the virus in terms of hospitalization and or death.
Geographic and Demographic Risk Profile Stratification and Clustering
This project was designed to identify trends in the populations that were most impacted in an attempt to shed light on the cause of spread within different communities. By using coronavirus confirmed cases, hospitalizations, and deaths, and combining them with the complete census data, we were able to identify unique factors and groups of unique factors that positively and negatively correlated to “hot spots” and the spread of COVID-19. We were further able to apply machine learning and clustering models that identified unique geographic clusters of counties that are likely to have a similar likelihood of spread and risk. This model was designed to address the needs of municipalities in assisting in making their decisions related to the risk of re-engagement in day to day activities, business openings, and more.
Unintended Mortality Modeling
This project is currently underway, and the results will be shared shortly. The goal of this model is to identify the potential impact to the public health and morbidity caused by the pandemic, but not directly related to the coronavirus itself. In other words, the impacts of how we have reacted to the virus and the reduced level of care we provide for the ongoing chronic ailments, acute events, and even scheduled preventative interventions. This model focuses on both the physical and mental issues that lead to morbidity by focusing on historical statistics pertaining to the impacts of unemployment on death rates (suicide, drug overdose, other), and by on data that views the morbidity rates caused by less availability to care, which is highly reflective of what we are seeing today.
While the statistics related to the death rates of other factors and classifications of the disease begin to trickle in, we cannot wait until they arrive to bring them into focus. The number of deaths associated with our inattention to these conditions and factors could be quite staggering. In fact, they will likely meet or exceed the COVID-19 deaths if left in the shadows.
Two primary models will be used.
Influences of unemployment on suicide and drug overdoses are based on national and global trends, and adjusted for age groups citing to be provided when published, and identifies social indicators and their impact on suicide and drug overdoses during times of increasing unemployment.
The impacts of inattentive care seeks to identify the increase in death rates by medical condition and uses medical underservice as a proxy to identify the disparities in care as communities cross specific thresholds of service.
Both models are then combined to provide an aggregate projection. It is important to note that these are not intended to be predictions, but rather projections should mitigating actions not be put into place.
As an outgrowth of this project, Lifemesh began building a prototype UI to easily convey the unintended impacts to mortality. While not driven by specific results of our models, we felt value could be provided even in its raw form. So, we have added the ability to perform “what if” scenarios and simulations that calculate increased mortality projections based on the average of 8 quarters of mortality data. This tool will later be updated with the machine learning projections derived from the two models described above, however, it may serve in the short term to provide valuable insight as to potential arising risks to care should we not begin immediate action.
Note: There is an ongoing, and likely escalating, conversation pertaining to COVID-19 attribution to death. Many deaths occurring prior to our knowledge of the coronavirus spread have been reattribute to COVID-19 (this is likely to continue as more deaths are being reported that had been previously unreported). In addition, some deaths have occurred as a result of other factors, but due to COVID-19 positive results also been attributed to COVID-19. Lastly, some percentage of deaths that have occurred due to COVID-19 and comorbid factors may have occurred regardless of the virus. While topics of conversation for another day, it is important to note that the mortality rates for a wide range of diseases will likely be misclassified. As a result, our ongoing mortality statistics may be difficult to sort out for a period of time. As such, it is even more important to recognize that these are overall projections, and their accuracy related to a given classification will be less accurate when compared to the ongoing statistics as they are reported.