Predicting the covid test result

Abstract

• The goal is to use classification models to predict when covid test results (positive/negative).
• Key questions to be answered
  1. Can we reliably predict covid test results based on the symptom?
  2. If the prediction is possible, what is the contribution/importance of each feature?
• An interactive dashboard/app was built. click here

Data

• Dataset from this published article and corresponding GitHub.
• 1 target: corona_results (positive, negative, other)
• 8 features: age, sex, cough, shortness_of_breath, fever, sore_throat, Headache, contact_with_confirmed
• About 3 million rows: each row is one test (all 3 million rows are used in the data explanatory but only 10,000 rows are used for training the classification models for the sake of computation time)

Tool/Packages used

• Data exploration: pandas, numpy, matplotlib
• Classification: scikit-learn, XGBoost
• Application: Streamlit, Heruko

Modeling

• Different models are tested: KNN, LogisticRgression, NaiveBayes (Bernoulli), DecisionTree, RandomForest, Ensemble, and Boost are tested.
• Models are tuned by GridSearchCV.
• Class imbalance techniques (over/under sampling, class weight adjustment, and threshold tuning) are used to tune the models.
• Feature importance was examined in the LogisticRgression, Forest, and XGBoost.
• Final model: The simple LogisticRgression with a threshold of 0.1 generates balanced results
  • P-threshold: 0.1
  • Accuracy: 0.924
  • Precision: 0.58
  • Recall: 0.64
  • F1: 0.61
• Modeling conclusions
  • Predicting covid test results is possible
  • Recall is rather limited (~ 0.6), likely due to the asymptomatic cases

App

• The model has been incorporated into an interactive app by using Streamlit and deployed on Heroku.
• Click here to check it out

Acknowledgement

• I sincerely appreciate great suggestions from the instructors (Kim and Joan), TAs (Tawney and Ethan), and the group team (Matthew, Leslie, Jason)!