COVID-19 Risk Factor Analysis

Summary

Using the CORD-19 dataset, we plan to use NLP/NLU methods to identify risk factors for COVID-19. Time permitting, we will compare risk factors discussed in research with those mentioned in the main-stream media.

We will make 3 passes through the corpus:

1. Train a language model on the CORD-19 research papers.
2. Store the h_t (hidden state/prediction pt of last layer) for each manually defined risk factor (below) like so: {'Diabetes': [.342, ..., -.123], 'Smoking': [0.323, ..., -0.432]}
3. Compare all h_t's in the data with our stored set of h_t's using some similarity metric in order to try to identify new risk factors.

Between steps 2-3, we might do clustering analysis on the stored h_t's to see how well they group. Can also PCA down to 2 dims and visualize.

Manually Identified Risk Factors

CDC Guidance on Risk Factors

• 65 years and older
• Living in a nursing home or long-term care facility
• Chronic lung disease or moderate to severe asthma
• Immunocomprimised:
  • Cancer treatment
  • Smoking
  • Bone marrow transplantation
  • Immune deficiencies
  • Pooly controlled HIV or AIDS
  • Prolonged use of corticosteroids
  • "Other immune weaking medication"
• Severe obesity (BMI >= 40)
• Diabetes
• Chronic kidney disease undergoing dialysis
• Liver disease
• Hemoglobin disorders
  • Sickle cell disease
  • Thalassemia
• Serious heart conditions
  • Heart failure
  • Coronary artery disease
  • Congenital heart disease
  • Cardiomyopathies
  • Pulmonary hypertension

Setting up sqlite

csv to sqlite3