Intro article draft

Summary of Survey Results

Top performing model of family XGBoost (trained on MIMC IV data). Trained using h2o.automl() ** 5-fold cross-validation on training data (Metrics computed for combined holdout predictions): AUC = 0.790; precision = 0.193; recall = 0.415. Dashboard created by passing 1,000 synthetic samples (generated using Synthpop() package).

Top performing model of family XGBoost (trained on MIMC IV data). Trained using h2o.automl() ** 5-fold cross-validation on training data (Metrics computed for combined holdout predictions): AUC = 0.790; precision = 0.193; recall = 0.415. Dashboard created by passing 1,000 synthetic samples (generated using Synthpop() package).

Prototype model derived from synthetic (Analytic dataset --> synthpop() in R) data predicting Pressure Injuries from candidate features.

Implementing H2O automl binomial classifier with balanced classes (for low proportion of PrIs). Using restricted feature set (listed before super learner) of values captured in first 48 hours of final visit. Summary of performance on 5-fold cross-validation on training data (Metrics computed for combined holdout predictions). Also produces synthetic cohort (copy) of MIMIC 4 viable (formatted) cases for use in prototype explainer dashboards.

Ensemble model (h2o automl) predicting HAPrI -> explainer created on synthetic data

Ensemble ML generated from `viable` cases with an explainer model using synthetic cases.

Simple logistic regression model (AUC: 0.71)

Compare the target parameter estimated from MIMIC IV to the parameter estimated from synthetic data.

Interactive and comprehensive analysis of the xgboost model trained on the synthetic dataset derived from the MIMIC dataset. The explainer includes various visualizations and diagnostic tools that help users understand and interpret the model's behavior, performance, and predictions.

Convert the MIMIC IV sample to synthetic and run prediction models with explainers.

Interactive and comprehensive analysis of the random forest model trained on the synthetic dataset derived from the MIMIC dataset. The explainer includes various visualizations and diagnostic tools that help users understand and interpret the model's behavior, performance, and predictions.

Using open-source modeltime.gluonts R package (which connects to the Python GluonTS DeepAREstimator()) Using 'Our World in Data" data and y variable is "smoothed new cases/million" mydata <- read.csv(url("https://covid.ourworldindata.org/data/owid-covid-data.csv"), na.strings = c("", "NA"))

Using open-source modeltime.gluonts R package (which connects to the Python GluonTS DeepAREstimator()) Using 'Our World in Data" data and y variable is "smoothed new cases/million" mydata <- read.csv(url("https://covid.ourworldindata.org/data/owid-covid-data.csv"), na.strings = c("", "NA"))

COVID parametric modeling. Example used: Utah (Some parameters are tuned to Utah so figures won't look nice & double-logistic model won't work without modifying inputs for a particular state)

For CG and team

Courtesy of MIT Delphi: https://www.covidanalytics.io/projections

National Immunization Survey–Teen (NIS-Teen), United States, 2017

Layers: Vaccination rates, insurance rates, annual income, test positivity

JHU data

Prototype risk score ~ BRFSS health scores, COVID cases per 100000, % change

Still artifact of testing, but interesting nonetheless.

Interactive map of Sweden (Level 1)