{"id":50,"date":"2021-07-13T19:01:30","date_gmt":"2021-07-13T19:01:30","guid":{"rendered":"https:\/\/baliga.systemsbiology.net\/see-interns\/hs2021\/?page_id=50"},"modified":"2021-08-21T08:53:20","modified_gmt":"2021-08-21T08:53:20","slug":"kalea-and-yannell-temp","status":"publish","type":"page","link":"https:\/\/baliga.systemsbiology.net\/see-interns\/hs2021\/summer-projects\/kalea-and-yannell-temp\/","title":{"rendered":"Multi Comorbidity Respiratory Health SARS-CoV-2 Index"},"content":{"rendered":"\n

Abstract <\/span><\/strong><\/p>\n\n\n\n

Intern partners, Kalea Velasco-Cosare<\/span><\/a> and Yannell Howard<\/span><\/a>, along with mentors Ryan Roper <\/span><\/a>and Jennifer Hadlock<\/a><\/span>, in the Hadlock Lab<\/span><\/a> worked on creating a Multi Comorbidity Respiratory Health SARS-CoV-2 Index <\/strong>over this past summer!<\/p>\n\n\n\n

\"\"
The pyramid above breaks down the research process that we took to develop our project showcased throughout this page.<\/span><\/figcaption><\/figure><\/div>\n\n\n\n

Introduction<\/strong><\/p>\n\n\n\n

For the first three weeks, we spent much of our time brainstorming a project both of us would like to pursue that would also be an important question to research for the larger community and further serve an interest for the Hadlock Lab at ISB. This was the most intense part of our internship as we built a solid foundation for our project while being exposed to and learning new terms and concepts!  We eventually decided on creating an index that assesses a person\u2019s risk for a severe COVID-19 outcome resulting in death, if they were to get a positive test result for COVID-19 and had certain underlying comorbidities.  <\/p>\n\n\n\n

At first, we looked at analyzing specific diseases such as sickle cell anemia and behcet\u2019s disease and creating a clinical concept around each disease such as a diagnosis flowchart. However, after reading the research paper Characteristics and Factors Associated with COVID-19 Infection, Hospitalization, and Mortality Across Race and Ethnicity – PubMed (nih.gov)<\/a> <\/span>of which both our mentors helped create, we were fascinated by the term Charlson Comorbidity Index (CCI). The Charlson Comorbidity Index is an index used to assess a person’s 10 year survival rate if they have certain comorbidities. The original version was published in 1987 by Dr.Mary Charlson and colleagues and is still used today. Inspired by the CCI, we decided to pursue something similar, however with adjustments to address the current times we live in including adding social demographics, patient information, and new morbid conditions such as SARS-CoV-2 to the variables of the index. <\/p>\n\n\n\n

Implications<\/strong><\/p>\n\n\n\n

Since it\u2019s hard to tell whether or not a variable is influential to a certain outcome, our project aimed to gather more information on influential COVID-19 clinical variables while providing a way to measure those influential variables simultaneously. The goal of our project is to provide patients and general everyday people a tool to assess their risk of dying from a severe COVID-19 outcome if they get COVID-19 and have underlying comorbidities at the same time. This will allow people to quantitatively gauge their risk further allowing them to analyze appropriate preventative measures they should take. <\/p>\n\n\n\n

Methods <\/strong><\/p>\n\n\n\n

Once the groundwork was decided, we had to find a dataset that we would be able to use and run through our model, while also deciding what type of model to use. We ended up using a Logistic Regression Model. <\/p>\n\n\n\n

<\/p>\n\n\n\n

\n\n\n\n

A logistic regression model can be used to estimate the probability (a value between 0 and 1) of an outcome or event (such as patient death) based on one or more independent variables. Each model input has a coefficient that indicates the weight or significance of the variable in predicting the outcome and it can be either negative or positive.<\/strong><\/p>Ryan Roper, MS. Data Scientist and our Mentor<\/span><\/cite><\/blockquote><\/figure>\n\n\n\n

\n\n\n\n

<\/p>\n\n\n\n

The Hadlock Lab works with Electronic Health Records (EHR) which has information on over 20,000,000 patients that require training to be able to work with. Due to time restraints, we instead used ISB\u2019s INCOV dataset which has information on 212 patients that tested positive for COVID-19. <\/p>\n\n\n\n

\"\"
This timeline shows the recruitment process for the INCOV dataset that we used. Once a patient in the dataset was diagnosed with a positive test result for COVID-19, they had their blood drawn three times. Blood Draw Times: T1 (after initial clinical diagnosis), T2 (few days after T1), and T3 (several weeks after T1).<\/span>
<\/figcaption><\/figure><\/div>\n\n\n\n

To decide what specific clinical variables and outcomes we wanted to look at, we had to narrow down what variables we included from the dataset. This meant looking for variables we thought would have a direct or indirect impact on a person\u2019s COVID-19 outcome while also considering previous research on known influential clinical variables. We limited it down to 44 clinical variables from 196, however not all the variables had enough patient information or were recorded in a way that was unable to be used in a logistic regression model. In the end, we assessed 21 clinical variables and 4 outcomes including previously known influential COVID-19 clinical variables such as age, sex, ethnicity, and certain underlying comorbidities, and other clinical variables that affect body systems impacted by COVID-19 such as your respiratory, immune, and cardiovascular system. Unlike the scientific method, we didn\u2019t start our project with a clear outcome we wanted to look at, so we ran four different logistic regression models to see what model generated more accurate results. We decided to look at death, days in hospital, mechanical ventilation, and hypoxia resolved as the outcomes for our models. <\/p>\n\n\n\n

\"\"
This chart shows all the clinical variables and outcomes (highlighted in aqua) we used in our model. The field label shows what information was collected and what each clinical variable represents. We used mainly patient information and medical conditions for our variables and used observations and outcomes in the INCOV dataset as outcomes for our logistic regression model. <\/span>
<\/figcaption><\/figure><\/div>\n\n\n\n
\"\"<\/figure>
\n

Shown to the right is a loopy model. Loopy is a program that is used to model the impact of variables on a certain outcome. The minus signs show variables that appeared to be less influential according to our model in comparison to the other variables. While the positive sign shows variables that appear to be more influential according to our model.<\/p>\n<\/div><\/div>\n\n\n\n

Results <\/strong><\/p>\n\n\n\n

\"\"<\/figure>
\n

This shows the total error results for each of our four models. The total error is the collected score of the prediction error for all the variables in our model and provides a good ballpark figure regarding the relative accuracy of our models, or how well the model predicts the outcomes for this set of patients. Ideally, you would like your total error to be as close to zero as possible. <\/p>\n<\/div><\/div>\n\n\n\n

Since our death model turned out to be the most relatively accurate model in comparison to the others, we based our index off of that. For clinical variables to be deemed influential for an outcome, they have to have a positive coefficient model above 0.5 and below 2. To further explain our results: <\/p>\n\n\n\n

<\/p>\n\n\n\n

\n\n\n\n

A negative coefficient indicates an inverse relationship between the input variable and the model outcome, while a positive coefficient indicates a direct or positive relationship. <\/strong><\/p>

The magnitude of the coefficient generally indicates the strength of the association. However, the coefficients must be estimated using a dataset (known as a “training” dataset) in which the input variable values and the outcomes are known. The reliability of the estimated coefficients depends on the size of the training dataset. It is possible that a large negative or positive coefficient might not accurately reflect the actual strength of the association.<\/strong><\/p>Ryan Roper, MS. Data Scientist and our Mentor<\/span><\/cite><\/blockquote><\/figure>\n\n\n\n

\n\n\n\n

<\/p>\n\n\n\n

\"\"
Here is the results for our death model. Highlighted in light blue are variables that turned out to be influential for our model and highlighted in dark blue are variables that can be looked at as \u201coutliers\u201d for this model. This can be caused from the small size of our patient cohort, or from the small size of data available for that variable.<\/span><\/figcaption><\/figure><\/div>\n\n\n\n

Analysis<\/strong><\/p>\n\n\n\n

Out of the four models, 12 clinical variables turned out to be influential. <\/p>\n\n\n\n

\"\"<\/figure>
\n

The first chart represents the clinical variables we expected to be important, the second chart are the actual clinical variables that appeared to be important from our expected results (first chart), and the last chart are all the clinical variables that appeared to be important in one model or more. Highlighted in azure are clinical variables that are known to affect your COVID-19 outcome from past research and highlighted in aqua are clinical variables we hypothesized could be influential due to their effects on the body and its systems. <\/p>\n<\/div><\/div>\n\n\n\n

\"\"
This box chart shows all clinical variables that appeared to be influential in at least one or more models. Each clinical variable is categorized into how many models it appeared to be an important variable in (i.e. sex appeared to be an influential variable in three out of our four models). <\/span><\/figcaption><\/figure><\/div>\n\n\n\n

Interesting results: <\/strong><\/p>\n\n\n\n