Final Project: Predicting COVID-19 Hospitalization
Marvellous Onajobi
2025-12-13
Description
This project examines whether COVID-19 hospitalization can be predicted using demographic and clinical characteristics.
The analysis uses the CDC COVID-19 Case Surveillance Public Use Data obtained from data.cdc.gov.
Due to the large number of observations gathered by the CDC, I filtered the dataset first before extracting it from the website. the filtered dataset is from January 2, 2020 to November 19,2020. The observations total 1 million+ (1048576)
A Logistic Regression model is used because the outcome variable (hospitalization status) is binary.
Objective
To determine whether age group, sex, and presence of underlying medical conditions significantly predict COVID-19 hospitalization.
To interpret model coefficients, odds ratios, model fit, and predictive performance using logistic regression techniques.
Introduction
COVID-19 has resulted in millions of hospitalizations worldwide. Identifying demographic and clinical factors associated with hospitalization risk is essential for improving public health responses and allocating healthcare resources effectively.
Hospitalization status is a binary outcome (Yes/No), making logistic regression an appropriate modeling choice. Logistic regression allows us to estimate the probability of hospitalization and quantify how predictors such as age, sex, and medical conditions influence hospitalization risk.
This project uses publicly available surveillance data from the Centers for Disease Control and Prevention (CDC) to model hospitalization outcomes among confirmed COVID-19 cases.
Fields Used
- age_group -> Categorical
- sex -> Categorical
- medcon_yn -> Categorical
- hosp_yn -> Categorical
library(ggplot2)
library(cowplot)
library(tidyverse)## ── Attaching core tidyverse packages ──────────────────────── tidyverse 2.0.0 ──
## ✔ dplyr 1.1.4 ✔ readr 2.1.5
## ✔ forcats 1.0.1 ✔ stringr 1.5.1
## ✔ lubridate 1.9.4 ✔ tibble 3.3.0
## ✔ purrr 1.1.0 ✔ tidyr 1.3.1
## ── Conflicts ────────────────────────────────────────── tidyverse_conflicts() ──
## ✖ dplyr::filter() masks stats::filter()
## ✖ dplyr::lag() masks stats::lag()
## ✖ lubridate::stamp() masks cowplot::stamp()
## ℹ Use the conflicted package (<http://conflicted.r-lib.org/>) to force all conflicts to become errorslibrary(pROC)## Type 'citation("pROC")' for a citation.
##
## Attaching package: 'pROC'
##
## The following objects are masked from 'package:stats':
##
## cov, smooth, varsetwd("~/Downloads/25_Semesters/Fall/DATA101")
# Load CDC COVID-19 data (example structure)
covid_df <- read_csv("COVID-19_Case_Surveillance.csv")## Rows: 4028318 Columns: 12
## ── Column specification ────────────────────────────────────────────────────────
## Delimiter: ","
## chr (8): current_status, sex, age_group, race_ethnicity_combined, hosp_yn, ...
## date (4): cdc_case_earliest_dt, cdc_report_dt, pos_spec_dt, onset_dt
##
## ℹ Use `spec()` to retrieve the full column specification for this data.
## ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.# Preview dataset
head(covid_df )## # A tibble: 6 × 12
## cdc_case_earliest_dt cdc_report_dt pos_spec_dt onset_dt current_status sex
## <date> <date> <date> <date> <chr> <chr>
## 1 2020-01-02 2020-01-02 2021-01-03 2020-01-02 Laboratory-co… Fema…
## 2 2020-12-20 2020-01-02 2021-01-12 2020-12-20 Laboratory-co… Fema…
## 3 2020-12-21 2020-01-02 2020-12-21 2020-12-21 Laboratory-co… Fema…
## 4 2020-01-02 2020-01-02 2021-01-04 2020-01-02 Laboratory-co… Fema…
## 5 2020-12-24 2020-01-02 2020-12-27 2020-12-24 Laboratory-co… Fema…
## 6 2020-11-02 2020-01-02 2020-01-02 2020-11-02 Laboratory-co… Fema…
## # ℹ 6 more variables: age_group <chr>, race_ethnicity_combined <chr>,
## # hosp_yn <chr>, icu_yn <chr>, death_yn <chr>, medcond_yn <chr>str(covid_df )## spc_tbl_ [4,028,318 × 12] (S3: spec_tbl_df/tbl_df/tbl/data.frame)
## $ cdc_case_earliest_dt : Date[1:4028318], format: "2020-01-02" "2020-12-20" ...
## $ cdc_report_dt : Date[1:4028318], format: "2020-01-02" "2020-01-02" ...
## $ pos_spec_dt : Date[1:4028318], format: "2021-01-03" "2021-01-12" ...
## $ onset_dt : Date[1:4028318], format: "2020-01-02" "2020-12-20" ...
## $ current_status : chr [1:4028318] "Laboratory-confirmed case" "Laboratory-confirmed case" "Laboratory-confirmed case" "Laboratory-confirmed case" ...
## $ sex : chr [1:4028318] "Female" "Female" "Female" "Female" ...
## $ age_group : chr [1:4028318] "10 - 19 Years" "60 - 69 Years" "60 - 69 Years" "60 - 69 Years" ...
## $ race_ethnicity_combined: chr [1:4028318] "White, Non-Hispanic" "White, Non-Hispanic" "Hispanic/Latino" "Multiple/Other, Non-Hispanic" ...
## $ hosp_yn : chr [1:4028318] "No" "No" "No" "No" ...
## $ icu_yn : chr [1:4028318] "Missing" "Missing" "Missing" "Missing" ...
## $ death_yn : chr [1:4028318] "No" "Missing" "Missing" "No" ...
## $ medcond_yn : chr [1:4028318] "No" "Missing" "Missing" "No" ...
## - attr(*, "spec")=
## .. cols(
## .. cdc_case_earliest_dt = col_date(format = ""),
## .. cdc_report_dt = col_date(format = ""),
## .. pos_spec_dt = col_date(format = ""),
## .. onset_dt = col_date(format = ""),
## .. current_status = col_character(),
## .. sex = col_character(),
## .. age_group = col_character(),
## .. race_ethnicity_combined = col_character(),
## .. hosp_yn = col_character(),
## .. icu_yn = col_character(),
## .. death_yn = col_character(),
## .. medcond_yn = col_character()
## .. )
## - attr(*, "problems")=<externalptr>df_clean <- covid_df |>
select(age_group, sex, medcond_yn, hosp_yn) |>
mutate(across(
c(age_group, sex, medcond_yn, hosp_yn),
~ na_if(na_if(as.character(.x), "Missing"), "Unknown")
)) |>
drop_na() |>
mutate(across(
c(age_group, sex, medcond_yn, hosp_yn),
as.factor
))
str(df_clean)## tibble [1,030,327 × 4] (S3: tbl_df/tbl/data.frame)
## $ age_group : Factor w/ 9 levels "0 - 9 Years",..: 2 7 6 8 5 5 6 4 7 8 ...
## $ sex : Factor w/ 3 levels "Female","Male",..: 1 1 2 1 1 1 1 2 2 2 ...
## $ medcond_yn: Factor w/ 2 levels "No","Yes": 1 1 2 2 1 2 2 1 1 2 ...
## $ hosp_yn : Factor w/ 2 levels "No","Yes": 1 1 1 1 1 1 1 1 1 2 ...Levels of each factor
levels(df_clean$age_group)## [1] "0 - 9 Years" "10 - 19 Years" "20 - 29 Years" "30 - 39 Years"
## [5] "40 - 49 Years" "50 - 59 Years" "60 - 69 Years" "70 - 79 Years"
## [9] "80+ Years"levels(df_clean$sex)## [1] "Female" "Male" "Other"levels(df_clean$medcond_yn)## [1] "No" "Yes"levels(df_clean$hosp_yn)## [1] "No" "Yes"xtabs(~ hosp_yn + sex, data = df_clean)## sex
## hosp_yn Female Male Other
## No 484909 428068 27
## Yes 56469 60854 0xtabs(~ hosp_yn + age_group, data = df_clean)## age_group
## hosp_yn 0 - 9 Years 10 - 19 Years 20 - 29 Years 30 - 39 Years 40 - 49 Years
## No 38674 113012 178239 150048 145306
## Yes 881 1448 4893 7939 11818
## age_group
## hosp_yn 50 - 59 Years 60 - 69 Years 70 - 79 Years 80+ Years
## No 133534 89962 41953 22276
## Yes 19881 25202 23307 21954xtabs(~ hosp_yn + medcond_yn, data = df_clean)## medcond_yn
## hosp_yn No Yes
## No 548280 364724
## Yes 19566 97757logistic <- glm(
hosp_yn ~ age_group + sex + medcond_yn,data = df_clean, family = "binomial")
summary(logistic)##
## Call:
## glm(formula = hosp_yn ~ age_group + sex + medcond_yn, family = "binomial",
## data = df_clean)
##
## Coefficients:
## Estimate Std. Error z value Pr(>|z|)
## (Intercept) -4.234062 0.034629 -122.271 <2e-16 ***
## age_group10 - 19 Years -0.672428 0.043334 -15.517 <2e-16 ***
## age_group20 - 29 Years -0.004936 0.037263 -0.132 0.895
## age_group30 - 39 Years 0.527492 0.036257 14.549 <2e-16 ***
## age_group40 - 49 Years 0.816849 0.035747 22.851 <2e-16 ***
## age_group50 - 59 Years 1.279404 0.035335 36.208 <2e-16 ***
## age_group60 - 69 Years 1.803887 0.035292 51.113 <2e-16 ***
## age_group70 - 79 Years 2.409035 0.035571 67.724 <2e-16 ***
## age_group80+ Years 2.939019 0.035959 81.732 <2e-16 ***
## sexMale 0.282340 0.006867 41.116 <2e-16 ***
## sexOther -7.406809 21.666266 -0.342 0.732
## medcond_ynYes 1.302220 0.008752 148.794 <2e-16 ***
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
##
## (Dispersion parameter for binomial family taken to be 1)
##
## Null deviance: 730564 on 1030326 degrees of freedom
## Residual deviance: 574727 on 1030315 degrees of freedom
## AIC: 574751
##
## Number of Fisher Scoring iterations: 9Fitting a logistic regression model to predict the probability of COVID-19 hospitalization using age group, sex, and presence of a medical condition as predictors.
- Log odds of hospitalization = −4.2341 + age group effects + 0.2823(male) + 1.3022(medical condition = yes)
The reference group are female patients aged 0-9 years with no medical condition.
Intercept
Intercept (−4.23): It is the log(odds) of hospitalization for females aged 0–9 years without a medical condition.
On the probability scale, that equals p = 1/(1 + e^4.23) = 0.014
This corresponds to about a 1.4% probability of hospitalization for this baseline group.
Age Group Effects
- All age group coefficients are interpreted relative to children aged 0–9 years, handling sex and medical condition as a constant.
Hospitalization risk is shown to increase with age. Compared to children:
Ages 10–19 for every positive increase in this ag range their chances of hospitalization decreases by 67%.
Ages 20–29 do not differ significantly.
Ages 30–39 and older show for every positive increase in this age range their chances of hospitalization increases by 53%
Ages 40–49 and older show for every positive increase in this age range their chances of hospitalization increases by 81%
Ages 50–59 and older show for every positive increase in this age rage their chances of hospitalization increases by 127%
Ages 60–69 and older show for every positive increase in this age range their chances of hospitalization increases by 180%
Ages 70–79 and older show for every positive increase in this age range their chances of hospitalization increases by 240%
Patients aged 80+ years show for every positive increase in this age range their chances of hospitalization increases by 294%.
This confirms that age is a strong risk factor for hospitalization.
Sex Effect
The coefficient for males is 0.2823, representing the change in log odds when moving from female to male, holding age and medical condition constant.
The chance of men begin hospitalization for every women increases by 28%
This effect is statistically significant with a p values less than 0.005.
Medical Condition Effect
- The coefficient for having a medical condition is 1.3022.
The chance of those with previous medical condition begin hospitalization for every person without increases by 130%
This is the strongest predictor in the model and is highly significant with a p values less than 0.005.
Model Fit and Deviance
Null deviance: 730,564
Residual deviance: 574,727
The reduction in deviance indicates that age group, sex, and medical condition together explain a substantial portion of the variability in hospitalization.
exp(coef(logistic))## (Intercept) age_group10 - 19 Years age_group20 - 29 Years
## 1.449340e-02 5.104676e-01 9.950763e-01
## age_group30 - 39 Years age_group40 - 49 Years age_group50 - 59 Years
## 1.694677e+00 2.263356e+00 3.594496e+00
## age_group60 - 69 Years age_group70 - 79 Years age_group80+ Years
## 6.073205e+00 1.112323e+01 1.889730e+01
## sexMale sexOther medcond_ynYes
## 1.326229e+00 6.071048e-04 3.677450e+00Interpretation
Age Group Effects
In comparison to children aged 0–9 years:
Ages 10–19 have a Odd Ratio of about 0.51 which indicates they have about half the odds of hospitalization.
Ages 20–29 have the same odds as children with an Odd Ratio of about 1.
Ages 30–39 with an Odd Ratio of about 1.69 have a 1.7 times higher odds.
Ages 40–49 with an Odd Ratio of about 2.26 have 2.3 times higher odds.
Ages 50–59 with an Odd Ratio of about 3.59 have 3.6 times higher odds .
Ages 60–69 with an Odd Ratio of about 6.07 have about 6.1 times higher odds.
Ages 70–79 with an Odd Ratio of about 11.12 have 11 times higher odds.
Ages 80+ with an Odd Ratio of about 1.9 have nearly 19 times higher odds of hospitalization.
Sex Effects
- Men have about 1.33 times higher odds of hospitalization compared to women.
Medical Condition Effect
- Patients with a pre-existing medical condition have about 3.7 times higher odds of hospitalization compared to those without a medical condition with a Odd Ratio of 3.68.
r_square <- 1 - (logistic$deviance / logistic$null.deviance)
r_square## [1] 0.2133098Interpretation
- This indicates that the model explains about 21% of the variation in hospitalization risk.
1 - pchisq(
logistic$null.deviance - logistic$deviance,
df = length(logistic$coefficients) - 1
)## [1] 0- Age, sex, and medical conditions significantly increase hospitalization risk.
predicted_data <- data.frame(
probability_hosp = logistic$fitted.values,
hosp_yn = df_clean$hosp_yn
)
head(predicted_data)## probability_hosp hosp_yn
## 1 0.007344077 No
## 2 0.080900435 No
## 3 0.202603940 No
## 4 0.372196166 No
## 5 0.031761822 No
## 6 0.107648058 No# Convert outcome to numeric
df_clean$hosp_num <- ifelse(df_clean$hosp_yn == "Yes", 1, 0)
predicted_class <- ifelse(logistic$fitted.values > 0.5, 1, 0)
confusion <- table(
Predicted = predicted_class,
Actual = df_clean$hosp_num
)
confusion## Actual
## Predicted 0 1
## 0 894200 97058
## 1 18804 20265894,200 people were actually healthy, and the model correctly predicted them as healthy. This is a True Negative.
97,058 people were actually hospitalized, but the model mistakenly predicted them as healthy. This is a False Negative.
18,804 people were actually healthy, but the model mistakenly predicted them as hospitalized. This is a False Positive.
20,265 people were actually hospitalized, and the model correctly predicted them as hospitalized. This is a True Positive.
TN <- confusion[1,1]
FP <- confusion[2,1]
FN <- confusion[1,2]
TP <- confusion[2,2]
accuracy <- (TP + TN) / (TP + TN + FP + FN)
sensitivity <- TP / (TP + FN)
specificity <- TN / (TN + FP)
precision <- TP / (TP + FP) # positive predictive value
f1_score <- 2 * (precision * sensitivity) / (precision + sensitivity)
# Print results
cat("Accuracy: ", round(accuracy, 4), "\n")## Accuracy: 0.8875cat("Sensitivity: ", round(sensitivity, 4), "\n")## Sensitivity: 0.1727cat("Specificity: ", round(specificity, 4), "\n")## Specificity: 0.9794cat("Precision: ", round(precision, 4), "\n")## Precision: 0.5187cat("F1 Score: ", round(f1_score, 4), "\n")## F1 Score: 0.2592The model has high accuracy and is very good at identifying healthy individuals (high specificity of 97.9%).
However, it struggles to detect hospitalized patients (low sensitivity of 17.3%).
Overall, the model reliably rules out healthy patients, but additional features or threshold adjustments would be needed to improve its ability to catch hospitalizations.
ROC and AUC
roc_obj <- roc(
response = df_clean$hosp_yn,
predictor = logistic$fitted.values,
levels = c("No", "Yes")
)## Setting direction: controls < casesauc(roc_obj)## Area under the curve: 0.8254plot.roc(roc_obj, print.auc = TRUE)
The AUC = 0.825 means your model is good at distinguishing between healthy and unhealthy patients.
On the plot, the curve is above the diagonal “random guess” line, which shows the model is better than chance.
In plain words: if you randomly pick one unhealthy and one healthy patient, the model has about an 82.5% chance of ranking the unhealthy one higher (more likely to have disease).
Summary
This project modeled COVID-19 hospitalization using demographic and clinical characteristics from CDC surveillance data. Age, sex, and presence of pre-existing medical conditions were all significant predictors of hospitalization. Hospitalization risk increases with age, is higher for males, and is substantially higher for patients with underlying medical conditions. The logistic regression model explained about (21%) of the variation in hospitalization risk. Model performance shows high specificity (97.9%) but low sensitivity (17.3%), meaning it accurately identifies healthy individuals but struggles to detect hospitalized patients. ROC analysis yielded an AUC of 0.825, indicating the model is good at distinguishing between healthy and hospitalized patients, performing much better than random chance.
Overall, the model highlights key risk factors and can guide public health strategies, although improvements are needed to better detect hospitalizations.