Using a Cloud-Based Machine Learning Classification Tree Analysis to Understand the Demographic Characteristics Associated With COVID-19 Booster Vaccination Among Adults in the United States.

A tree model identified adults age ≤34 years, Johnson & Johnson primary series recipients, people from racial/ethnic minority groups, residents of nonlarge metro areas, and those living in socially vulnerable communities in the South as less likely to be boosted. These findings can guide clinical/public health outreach toward specific subpopulations. Published by Oxford University Press on behalf of Infectious Diseases Society of America 2022.

Coronavirus disease 2019 (COVID-19) booster vaccination increases protection against severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection, including the recently predominant Omicron variant (B.1.1.529), and reduces COVID-19-associated hospitalization and death [1]. During August–November 2021, a series of Emergency Use Authorizations and recommendations, including those for an additional primary dose for immunocompromised persons and a booster dose for persons age ≥18 years, were approved by the Food and Drug Administration [2]. In the United States, as of April 2022, all adults (age ≥18 years) were eligible to receive a booster dose ≥2 months after vaccination with the 1-dose Johnson & Johnson/Janssen (J&J) primary series or ≥5 months after the second dose of the Pfizer-BioNTech or Moderna 2-dose mRNA primary series [2]. Certain populations may have also chosen to receive a second booster dose using an mRNA COVID-19 vaccine ≥4 months after the first booster dose [2].

As of March 2022, ∼47% of persons age ≥18 years who were eligible to receive a booster dose after completing a primary series of COVID-19 vaccine had not yet received a booster [3]. Disparities in COVID-19 vaccine booster uptake have been related to socioeconomic status, insurance status, disability, and social demographic factors, including age, education level, race/ethnicity, and residency in rural or urban areas [4-7]. In the present study, we applied machine learning methods in the form of a classification tree algorithm to identify and describe relationships and interactions of demographic factors associated with the receipt or nonreceipt of a COVID-19 booster vaccine among eligible persons age ≥18 years in the United States.

METHODS

Over 152 million COVID-19 primary vaccine completion records (administered from 12/14/2020 through 09/15/2021) and 81 million first booster dose records (administered through 03/15/2022) reported to the Centers for Disease Control and Prevention (CDC) from 49 states and the District of Columbia (DC) were analyzed using the cloud-based data platform Microsoft Azure DataBricks (Azure Databricks | Microsoft Azure). Texas had data-sharing restrictions on information reported to the CDC; its data were not available for inclusion. Vaccine records from US territories were not included in the present study. Recipients’ primary series and booster dose records were matched. A classification tree model was built to examine factors contributing to receiving a booster dose, with Gini impurity as the classification tree splitting metric [8]. Input variables included primary series vaccine product (Moderna, Pfizer-BioNTech, J&J), age group (18–24, 25–34, 35–44, 45–54, 55–64, ≥65 years), sex (male, female), race/ethnicity (Hispanic/Latino, non-Hispanic Black [Black], non-Hispanic American Indian/Alaska Native [AI/AN], non-Hispanic Asian/other Pacific Islander [Asian/OPI], Non-Hispanic White [White], other/multiracial/unknown [other/unknown]), region (South, Midwest, Mountain, Pacific, Northeast [South Region includes AZ, NM, OK, AR, LA, MS, AL, TN, KY, GA, SC, NC, WV, MD, VA, FL, DE, & DC; Midwest Region includes ND, SD, NE, KS, MN, IA, MO, IL, WI, IN, MI, & OH; Mountain Region includes NV, UT, CO, WY, MT, & ID; Pacific Region includes WA, HI, AK, OR, & CA; Northeast Region includes PA, NY, VT, NH, ME, MA, RI, CT, & NJ]), urbanicity (large central metro, large fringe metro [large fringe metro counties are counties in Metropolitan Statistical Areas of ≥1 million population that do not qualify as large central; for more information regarding urbanicity classification, please see Gaffney et al.] [7], medium metro, small metro, micropolitan, noncore [9]), and CDC/ATSDR Social Vulnerability Index (SVI) of zip code of residence (low, medium, and high). Factors affecting SVI scores include socioeconomic status, household composition, disability, minority status, housing type, and transportation. A lower SVI score means the zip code of residence is less socially vulnerable [10-11]. All the input variables were derived from vaccine records. Gender identity was not available. Descriptive analyses were performed for input variables, and feature importance of input variables and prediction rate of each end node were reported. This study was reviewed by the CDC and conducted in accordance with applicable federal law and CDC policy.

RESULTS

As shown in Figure 1, the classification tree model had a depth of 5 branches, with 17 end nodes and 32 nodes in total. Detailed information (eg, sample sizes, prediction rates, etc.) about the 17 end nodes is presented in Supplementary Table 1. The model generated a feature importance score for each input variable; a higher score meant that the specific feature had a larger effect on the model that was being used to predict the outcome variable [12]. In sum, age group had the highest feature importance score (0.739), followed by region (0.168), primary series vaccine product (0.071), race/ethnicity (0.010), SVI ranking (0.009), urbanicity (0.004), and sex (0.000). Overall, the model correctly predicted the booster status of 61.5% of individuals. In general, adults aged ≤34 years, J&J primary series recipients, persons belonging to racial/ethnic minority groups, residents of nonlarge metro areas, and those living in socially vulnerable areas were less likely to be boosted.

Figure 1.

Classification tree diagram depicting demographic characteristics associated with COVID-19 booster vaccination among adults completing the primary series before September 15, 2021, by social demographic factors, March 15, 2022, United States. Detailed information (eg, sample sizes, prediction rates, etc.) about the 17 end nodes is listed in Supplementary Table 1. Abbreviations: AI/AN, American Indian/Alaska Native; COVID-19, coronavirus disease 2019; J&J, Johnson & Johnson/Janssen; OPI, other Pacific Islander; SVI, Social Vulnerability Index.

The first partition or split in the classification tree was between adults age ≤54 and ≥55 years. Then, the model split South apart from all other regions (Midwest, Mountain, Northeast, and Pacific), and different branches were developed for residents of the South and non-South regions. Among persons aged 35–54 years in non-South regions, those who received a Pfizer or Moderna primary vaccine series and were non-Hispanic White were more likely to be boosted. Among persons aged ≥55 years in non-South regions, those who received a primary series of Moderna or Pfizer vaccines were more likely to be boosted. Among Southerners age 35–54 years, those who resided in low-SVI areas (ie, less socially vulnerable) and large fringe metro areas were more likely to be boosted. Among Southerners age ≥55 years, those who received a Moderna or Pfizer primary vaccine series were more likely to be boosted. Among Southerners age ≥65 years who received a J&J primary vaccine, those who resided in large fringe metro areas were more likely to be boosted.

Table 1 presents results from descriptive analyses of COVID-19 vaccine booster dose status by social demographic factors. Lower booster coverage was observed among J&J primary series recipients, younger age groups (eg, 18–34 years), residents of areas that are more socially vulnerable, people from racial and ethnic minority groups, and residents of the South.

Table 1.

COVID-19 Vaccine Booster Dose Status for Adults Completing the Primary Series Before September 15, 2021, by Social Demographic Factors, March 15, 2022, United States

Variable	Booster Dose Status				Total, No.
	Not Boosted, No.	%	Boosted, No.	%
Total	71 056 071	46.71	81 060 169	53.29	152 116 240
Primary series completion dose vaccine product
Pfizer-BioNTech	37 312 813	46.88	42 272 769	53.12	79 585 582
Moderna	26 062 915	43.48	33 886 214	56.52	59 949 129
Johnson & Johnson	7 680 343	61.04	4 901 186	38.96	12 581 529
Age group
18–24 y	8 953 407	64.28	4 975 323	35.72	13 928 730
25–34 y	13 177 374	60.43	8 628 063	39.57	21 805 437
35–44 y	12 453 119	53.82	10 685 458	46.18	23 138 577
45–54 y	11 763 997	48.61	12 438 296	51.39	24 202 293
55–64 y	11 470 661	40.69	16 717 198	59.31	28 187 859
≥65 y	13 237 513	32.40	27 615 831	67.60	40 853 344
Sex
Male	34 863 037	48.91	36 418 962	51.09	71 281 999
Female	36 193 034	44.77	44 641 207	55.23	80 834 241
Urbanicity
Large fringe metro	22 259 023	46.06	26 065 686	53.94	48 324 709
Large central metro	18 549 122	45.38	22 324 584	54.62	40 873 706
Medium metro	15 313 549	47.85	16 688 091	52.15	32 001 640
Small metro	6 127 185	47.80	6 691 771	52.20	12 818 956
Micropolitan	5 413 237	49.11	5 608 530	50.89	11 021 767
Noncore	3 393 955	47.97	3 681 507	52.03	7 075 462
Social Vulnerability Index
High	22 625 719	50.05	22 583 831	49.95	45 209 550
Medium	28 684 159	46.96	32 396 900	53.04	61 081 059
Low	19 746 193	43.09	26 079 438	56.91	45 825 631
Race/ethnicity
Hispanic	10 212 284	58.70	7 186 269	41.30	17 398 553
Non-Hispanic Black	5 880 660	52.48	5 325 513	47.52	11 206 173
Non-Hispanic American Indian/Alaska Native	571 057	57.20	427 329	42.80	998 386
Non-Hispanic Asian/OPI	3 214 046	38.12	5 218 318	61.88	8 432 364
Non-Hispanic White	30 949 706	42.83	41 305 029	57.17	72 254 735
Other/Unknown	20 228 318	48.36	21 597 711	51.64	41 826 029
Region
South	26 809 572	53.52	23 284 388	46.48	50 093 960
Midwest	13 871 561	41.70	19 396 829	58.30	33 268 390
Mountain	3 481 795	45.87	4 108 301	54.13	7 590 096
Pacific	12 059 872	41.07	17 307 608	58.93	29 367 480
Northeast	14 833 271	46.65	16 963 043	53.35	31 796 314

Abbreviations: COVID-19, coronavirus disease 2019; OPI, other Pacific Islander.

DISCUSSION

This study used 233 million COVID-19 vaccination records to construct a classification tree model that assessed demographic characteristics associated with receipt or nonreceipt of COVID-19 booster vaccination among US adult populations. The classification tree model provides a framework to consider the impact of each input variable on vaccination outcomes within specific subpopulations; it would be prohibitively time-consuming to investigate outcomes at this granularity using other analytical approaches.

Age group was the most important characteristic, with a feature importance score of 0.739, and persons age 18–34 years in all regions were less likely to have received a booster vaccination. Previous studies have identified attitudes and beliefs corresponding to low intent to receive primary series vaccination and low primary series coverage among young adults age 18–39 years [13].

The South had lower booster coverage than the other 4 regions and was split by the model from all other regions to form its own branches. SVI and urbanicity were important predictors of booster status in the South. Southerners residing in less socially vulnerable areas or large fringe metro areas were more likely to have received a booster dose. Residents within these areas report higher household income, which has been linked with higher COVID-19 vaccine uptake [9, 14]. In addition, marginalized populations within rural or socially vulnerable areas may have limited transportation options, less paid time off, and reduced ability to access vaccination providers [15-16]. Our finding that SVI is an important predictor of booster dose status among Southerners age 35–54 years is consistent with the observation of greater income-associated health disparities in the South than in other regions [17]. Among non-Southerners, age, primary vaccine type, race/ethnicity, and urbanicity determined the outcome. For persons age 35–54 years who received a primary series of Moderna or Pfizer, the tree model identified non-Hispanic White persons as more likely to be boosted; however, this pattern of race and ethnicity was not found among persons in other age groups or in residents of the South.

Regardless of age or region, recipients of a J&J primary series were less likely to have received a booster dose. Given lower vaccine effectiveness of a J&J primary series compared with an mRNA vaccine primary series, this population would particularly benefit from the increased effectiveness conferred by a booster dose [18]. More information is needed to understand factors contributing to low booster uptake among J&J recipients. Some J&J recipients may have chosen the 1-dose primary series because they were less likely to complete a 2-dose mRNA vaccination series, whether due to vaccination-related anxiety (eg, needle aversion), to concerns about mRNA vaccines due to health conditions or personal beliefs, or to barriers to accessing health care or vaccine providers (eg, transportation, limited time off, reduced availability of specific vaccines in certain geographic areas) [19-21].

These findings are subject to at least 3 limitations. First, Texas data were not included in this analysis, and given Texas’ large population size, lack of data from Texas could have impacted these findings. Second, the booster status of a small portion of individuals may have been misclassified if the booster dose record was not able to be linked to the primary series completion record, such as if vaccinations were received in different jurisdictions. Third, the current tree model yields a 61.5% prediction rate, which may limit the application of these findings. A single classification tree model is often reported to have relatively low prediction accuracy; we found during the process of model selection that replacing a single tree with a random forest of trees or growing the tree model to a depth of >5 branches could improve prediction rates but would dramatically reduce interpretability [12].

The classification tree diagram is a novel approach to analyzing public health vaccination data. One advantage of the classification tree approach is its use of a splitting metric to identify partitions in input variable responses, which describes variability across a population in a way that is easy to understand. By structuring certain demographic characteristics into paths, the classification tree was able to describe the relationships (or lack thereof) between the many input variables used in the model. The paths described possible intersections between demographic characteristics that may have contributed to low access and acceptance of vaccinations and identified specific subpopulations that would be likely to have a higher burden of health disparities. Despite the challenge of seeking to increase the prediction rate, the paths in the tree diagram can inform clinical and public health interventions and outreach toward specific subpopulations. The use of the classification tree model to identify subpopulations that would be less likely to receive a booster vaccine can inform public health efforts and other strategies on a broader scale, such as efforts that involve other vaccinations. The model presented here indicates that low booster vaccination coverage was seen among young adults, J&J primary series recipients, people from racial and ethnic minority groups, residents of nonlarge metro areas, and those living in socially vulnerable communities in the South.

Click here for additional data file.