The Role of Digital Health Literacy and Socioeconomic Factors in

Colorectal Cancer Screening: Machine Learning Analysis of

HINTS Data

Sujin Kim

, Madhav Dahal

, Avinash Bhakta

and Jihye Bae

Division of Biomedical Informatics, University of Kentucky, 725 Rose Street, Lexington, KY, U.S.A.

Deptment of Electrical and Computer Engineering, University of Kentucky, 512 Administration Drive, Lexington, KY, U.S.A.

Department of Surgery, University of Kentucky, 800 Rose Street, Lexington, KY, U.S.A.

Keywords: Colorectal Cancer Screening, Digital Health Literacy, Socioeconomic Factors, Machine Learning,

Health Information National Trend Surveys (HINTS).

Abstract: While colorectal cancer (CRC) screening rates are on the rise, significant disparities persist, particularly

among underserved populations, highlighting ongoing challenges in achieving equitable access to preventive

care. This study utilizes machine learning models to analyze multi-year data from the Health Information

National Trends Survey (HINTS), identifying critical factors influencing CRC screening adherence across

three distinct time periods (2003–2008, 2011–2013, 2018–2020). Using Random Forest and Logistic

Regression models, interpreted through Shapley Additive exPlanations values, we examine the impact of

sociodemographic characteristics, digital health engagement, and digital literacy on CRC screening behaviors.

Findings reveal that age, prior screening behavior, and digital literacy are key predictors; individuals with

higher digital literacy, for example, exhibited a 22% higher likelihood of adhering to CRC screening

guidelines. Age emerged as a dominant factor, with screening rates peaking at 43% in the 50–64 age group.

These results suggest that interventions targeting digital health literacy and enhancing provider

communication may effectively improve CRC screening rates among underserved populations. This study

underscores the value of data-driven approaches in informing public health strategies to increase CRC

screening adherence and reduce health disparities.

1 INTRODUCTION

Colorectal cancer (CRC) remains one of the most

prevalent cancers globally and is the second leading

cause of cancer-related deaths, even as advancements

in screening and treatment have contributed to

declining incidence and mortality rates (Siegel, 2022;

Bray, 2018). In the United States, CRC

predominantly affects adults aged 65-74, with

screening rates increasing over recent decades due to

the adoption of colonoscopy and non-invasive

methods such as multitarget stool DNA (FIT-DNA)

testing (Keum, 2019). However, a concerning trend is

the rising incidence of early-onset colorectal cancer

(EOCRC) among adults under 50, a rate projected to

https://orcid.org/0000-0002-7878-4322

https://orcid.org/0009-0002-1016-428X

https://orcid.org/0000-0003-2471-3681

https://orcid.org/0000-0002-6609-9782

Corresponding Author: sujinkim@uky.edu

double by 2030 and driven by complex, interacting

risk factors not fully understood (Zhen, 2024; Sun,

2024). Disparities in CRC screening persist,

influenced by sociodemographic factors such as

income, education, and race, as well as health

behaviors. These multifaceted barriers, often studied

in isolation, underscore the need for a more integrated

approach to understanding and addressing CRC

screening uptake.

Digital health interventions, including telehealth,

patient portals, and mobile health (mHealth)

applications, offer promising avenues to address these

complex challenges in CRC screening by making

screening information, reminders, and test results

more accessible to diverse populations (Miller, 2018;

228

Kim, S., Dahal, M., Bhakta, A. and Bae, J.

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data.

DOI: 10.5220/0013310700003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 228-239

ISBN: 978-989-758-731-3; ISSN: 2184-4305

McIntosh, 2024). For instance, studies have shown

that digital health interventions can improve

screening adherence in vulnerable groups compared

to standard care, and patient portal reminders paired

with mailed test kits have increased adherence among

average-risk populations (Miller, 2018; McIntosh,

2024). The Health Information National Trends

Survey (HINTS) provides a valuable, nationally

representative dataset on health behaviors, digital

literacy, and sociodemographic factors that impact

CRC screening. Leveraging machine learning (ML)

such as Random Forest (RF) and Logistic Regression

(LR) to analyze multi-year HINTS data enables the

identification of nuanced relationships among digital

health literacy, socioeconomic variables, and

screening adherence, offering a more holistic, data-

driven approach to improve CRC screening rates and

reduce disparities in an increasingly digital healthcare

environment.

2 BACKGROUNDS

CRC screening uptake is shaped by a wide range of

demographic, psychosocial, and access-related

factors. Insights from the HINTS dataset reveal

crucial predictors that influence CRC screening

decisions, providing a nuanced view of how different

factors affect adherence. For example, Atarere

(2024c) found that smokers are 30% less likely to

adhere to CRC screening protocols than non-smokers,

highlighting the potential of health information

technology interventions to increase participation

among high-risk groups (Atarere, 2024b). Further,

Atarere et al. (2024) reported that patients engaged in

telehealth primary care visits had a 20% higher

likelihood of discussing CRC screening with their

providers, suggesting that Health IT (HIT) tools like

telehealth can significantly improve screening

discussions and adherence in populations typically

resistant to CRC screening (Atarere, 2024a; Atarere,

2024c).

Beyond access to HIT, cultural and social

influences are critical in shaping CRC screening

behaviors. Jun and Oh (2013) found that cancer

fatalism among Asian and Hispanic Americans was

associated with a 15% reduction in CRC screening

likelihood, pointing to cultural perceptions as barriers

to uptake. Similarly, Idowu et al. (2016) observed that

U.S. adults born outside the United States were 18%

less likely to be up-to-date with CRC screening,

underscoring the informational and cultural barriers

faced by immigrant populations. Additionally,

Finney Rutten et al. (2009) found that only 56% of

the general public understood CRC risk and

prevention guidelines accurately, linking low public

knowledge to reduced screening adherence. These

findings highlight a complex interplay of

sociocultural and informational barriers, which

Nawaz et al. (2014) further supported by showing that

CRC screening offered in hospital settings resulted in

a 35% higher uptake rate, suggesting that more

accessible, opportunistic screening efforts could be

effective in improving national screening rates.

Early studies using HINTS 123 data (2003–2008)

identified foundational barriers to CRC screening,

including limited awareness, inadequate knowledge

of guidelines, and socioeconomic and cultural

disparities. For instance, Geiger et al. (2008),

analyzing HINTS 1 data, pinpointed knowledge gaps

and access issues as central barriers. Similarly, Hay et

al. (2006) reported that perceived risk was a decisive

factor in CRC screening uptake, with substantial

demographic differences in risk perception. More

recent HINTS data (HINTS 5, 2018-2020), however,

indicate a shift toward increased digital literacy and

greater use of HIT, both of which are positively

correlated with screening adherence. The growth of

telehealth and online health resources appears to

support CRC screening behaviors, marking a

significant change in how technology influences

preventive health actions.

ML applications in CRC research have

increasingly used electronic health record (EHR) data

to enhance predictive models, especially in EOCRC

studies. For example, studies by Sun et al. (2024) and

Zhen et al. (2024) showed high predictive accuracy

for EOCRC among individuals under the standard

screening age, with area under the curve (AUC)

scores reaching up to 0.888 when using RF models.

These studies illustrate the clinical value of EHR data

in identifying risk factors such as immune and

digestive disorders, allowing for the development of

models that can flag high-risk patients who may

benefit from early diagnostic interventions.

While EHR-based ML models provide valuable

insights into clinical risk factors, they often overlook

behavioral elements influencing CRC screening

adherence, such as health literacy, digital literacy, and

risk perception—areas extensively covered in the

HINTS dataset. Unlike EHR data, which primarily

capture clinical encounters, HINTS data offer a

comprehensive view of health behaviors, allowing for

broader exploration of sociobehavioral factors like

health literacy and digital engagement. Integrating

these broader factors is essential for understanding

screening behaviors, as digital literacy,

socioeconomic status, and perceived cancer risk have

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data

229

become key determinants of screening adherence,

especially with the evolution of HIT over time.

In this study, we examine these broader

behavioral factors by analyzing HINTS data across

three significant temporal milestones—HINTS 123

(2003-2008), HINTS 4 (2011-2013), and HINTS 5

(2018-2020). This approach allows us to track

changes in CRC screening attitudes, HIT

engagement, and health information access over

nearly two decades. By combining HINTS data with

ML, we aim to uncover complex, time-varying

relationships between behavioral factors and

screening outcomes, thereby informing tailored

interventions designed to improve CRC screening

rates across diverse populations.

3 METHODS

3.1 Data Source and Study Variables

The HINTS datasets offer a robust, nationally

representative sample capturing U.S. adults' cancer

knowledge, perceptions, information-seeking

behaviors, and digital health tool adoption over time.

Each HINTS cycle introduces unique variables to

reflect emerging HIT trends, while maintaining core

questions that allow for cross-cycle comparisons on

foundational topics such as cancer knowledge,

perception, and health information access. For this

study, we grouped three survey cycles within each of

three temporal groups, focusing on CRC screening-

related variables while noting differences in HIT

priorities across cycles:

• Group 1: Merged HINTS Cycles 1-3 (2003,

2005, 2008) – This grouping consolidates early

cycles with a focus on CRC-related variables,

including “Awareness of CRC screening,”

“Frequency of CRC screening conversations

with healthcare providers,” and “CRC screening

completion history.” Predating widespread

digital tool adoption, this group establishes a

baseline of public knowledge and screening

behaviors (Ford, 2006; Geiger, 2008; Hay,

2006).

• Group 2: Merged HINTS Cycle 4 (2011, 2012,

2013) – Reflecting the era’s shift toward digital

health adoption, this group includes variables

like “Access to EHRs,” “Use of telehealth for

health information,” and “Frequency of patient

portal usage,” though certain CRC-specific

questions are omitted in favour of HIT-focused

variables.

• Group 3: Merged HINTS Cycle 5 (2018, 2019,

2020) – Emphasizing advanced HIT usage and

digital health literacy, this group includes

variables such as “Confidence in locating reliable

health information online,” “Frequency of

telehealth usage,” and “Patient portal

engagement for preventive care.” Although

CRC-specific questions are less prominent,

consistent HIT variables allow for comparison

across cycles.

This grouping allows for a comprehensive

examination of CRC screening behaviors over time,

highlighting the growing influence of digital tools on

CRC engagement. The primary outcome variable,

Ever_Tested_Colon, is binary, coded as 1 for

individuals who reported undergoing a colorectal

cancer screening test and 0 for those who did not.

Predictor variables include multiple categories

representing demographic characteristics (e.g., age,

gender, race/ethnicity), socioeconomic status (e.g.,

income, education level), health behaviors (e.g.,

smoking status, physical activity), access to

healthcare (e.g., health insurance coverage, regular

healthcare provider), and digital health engagement

(e.g., frequency of internet use for health information,

use of electronic health records, and telehealth usage).

This comprehensive set of predictors allows for a

nuanced analysis of factors influencing CRC

screening uptake, considering both individual health

characteristics and the broader context of healthcare

access and technology use.

3.2 Data Processing and Feature

Engineering

To ensure data consistency across cycles, we

standardized variable names using the HINTS 5

Cycle 4 format and encoded all categorical variables

numerically. Race was expanded into binary

indicators (Hispanic, White, Black, Asian, and

Other), and marital status was binarized, with

“Married” coded as 1 and all others as 0. Missing

values were replaced with -1 to retain cases, with

verification that imputing -1 did not distort

predictions. Group 1’s multiple-response categorical

variables were converted into binary format for

simplicity, and non-informative variables, duplicates,

and weight variables were excluded. Feature selection

focused on variables consistently present across

cycles, including demographics, health information-

seeking behaviors, digital health adoption, and CRC

screening history. Final dataset sizes were 11,710

rows and 1,022 columns for Group 1, 10,534 rows

HEALTHINF 2025 - 18th International Conference on Health Informatics

230

and 475 columns for Group 2, and 12,391 rows and

535 columns for Group 3, where each row indicates

each individual sample, and each column represents

variables extracted from HINTS datasets.

3.3 Predictive Model Development

This study utilized Python (version 3.8) within the

Google Colab environment, leveraging cloud

resources for efficient data processing, model

training, and interpretation. Key libraries included

Pyreadstat for SAS file handling, Pandas and NumPy

for data manipulation, and Scikit-Learn for machine

learning model implementation, pre-processing, and

validation. Model interpretability was achieved using

SHapley Additive exPlanations (SHAP), enabling a

detailed examination of feature contributions to

specific predictions. All data were securely accessed

and processed via Google Drive integration, ensuring

data consistency and reproducibility. Regarding

ethics and data privacy, this study exclusively

analysed publicly available, de-identified data from

the HINTS thereby maintaining strict confidentiality

and compliance with data privacy standards. As no

personally identifiable information (PII) was present,

this secondary data analysis was exempt from

Institutional Review Board (IRB) oversight.

We employed two predictive models, Random

Forest (RF) and Logistic Regression (LR) with

Elastic Net regularization, to predict colorectal cancer

(CRC) screening outcomes (CRC Screened or Not).

Following Min-Max normalization, 15% of the

dataset was allocated for testing. Hyperparameter

optimization was conducted through 5-fold cross-

validation using a grid search approach. For RF, we

searched across key hyperparameters, including the

number of trees (n_estimators: 100–500, step 50),

maximum tree depth (max_depth: 5–20, step 1),

minimum samples required to split a node

(min_samples_split: 2–10, step 1), and minimum

samples per leaf node (min_samples_leaf: 1–5, step

1). For LR, optimization included searching over

regularization strength (C: 0.1–1.0, step 0.1), the

balance between L1 and L2 penalties (l1_ratio: 0.5–

0.9, step 0.1), and iteration limits (max_iter: 100–

1000, step 100). The configurations yielding the

highest average cross-validation accuracy were

applied to the testing set, with performance evaluated

across key metrics: accuracy, precision, recall, F1

score, and area under the curve (AUC) (Table 3).

The optimal hyperparameters for RF varied

across datasets. For HINTS123, the best RF

configuration included 400 trees, a maximum depth

of 12, a minimum of 7 samples per split, and 2

samples per leaf, achieving a cross-validation

accuracy of 98.25%. In HINTS4, the optimal

configuration used 400 trees, a depth of 14, a

minimum of 8 samples per split, and 1 sample per

leaf, achieving a cross-validation accuracy of

81.84%. For HINTS5, the best RF model used 400

trees, a depth of 15, 9 samples per split, and 2 samples

per leaf, with a cross-validation accuracy of 80.61%.

For LR, the optimal configuration across datasets

involved a regularization strength of 0.1, an L1/L2

ratio of 0.5, and 800 iterations, yielding a cross-

validation accuracy of 80.86% in HINTS5. Test set

performance metrics for both models across datasets

were summarized in Table 3, demonstrating the

models' strengths in predicting CRC screening

outcomes under varying data conditions.

3.4 Model Interpretability with SHAP

Analysis

To interpret model predictions and identify top

predictors of CRC screening, we calculated SHAP

values for the optimized RF model. SHAP values

were chosen for their consistency and additive feature

contributions, providing insights into how each

variable influence on the CRC screening prediction.

Focusing on the 15 most impactful variables in each

group, we visualized these features to clarify the key

drivers of CRC screening across different periods and

demographic groups. This interpretability step

enhances our understanding of the predictors behind

screening behaviors and the evolving role of HIT.

This methodology—combining multiple HINTS

cycles, consistent variable standardization, and

advanced modelling with interpretability—offers a

comprehensive, time-sensitive analysis of CRC

screening behaviors. By integrating SHAP analysis,

we achieved transparent, interpretable predictions,

allowing us to pinpoint essential factors influencing

CRC screening adherence across various

demographic and HIT-related contexts.

4 RESULTS

4.1 Patient Characteristics of CRC

Screened or not

Across the three groups, CRC screening uptake

showed a clear upward trend over time. In Group 1

(2003–2008), the screening rates started with 40.12%

tested in 2003, dipping to 33.33% in 2005, and rising

to 37.52% in 2008, resulting in an overall group rate

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data

231

of 36.94% tested. Group 2 (2011–2013) saw a more

substantial increase, with a 50.02% overall screening

rate, increasing steadily from 49.32% in 2011 to

52.28% in 2013. Group 3 (2018–2020) had the

highest screening rates, with a group average of

62.15%. Notably, 60.90% were tested in 2018,

60.02% in 2019, and this rose to 64.49% by 2020.

This progression highlights a positive trend in CRC

screening uptake, suggesting increased awareness

and access to screening over the years.

SES Characteristics Among Three Groups

Table 1 presents a comparison of socioeconomic

(SES) and digital factors among CRC screening-

tested individuals across the three HINTS groups in

abridged format highlighting underserved groups,

while Appendix A provides comprehensive details

on additional variables. The findings from Table 1

reveal significant disparities in colorectal cancer

(CRC) screening rates among underserved

populations, particularly when compared to their

counterparts—groups not explicitly represented in

the table. Individuals aged 18–49 years displayed

consistently lower screening participation, with rates

rising from 5.36% in Group 1 to 10.90% in Group 2

but falling again to 7.56% in Group 3. In contrast,

their counterparts aged 50 and above, particularly the

50–64 age group, dominated screening adherence,

consistent with guideline recommendations targeting

this demographic. This trend highlights a critical age

disparity, emphasizing the need for enhanced

outreach and interventions tailored to younger

populations.

Table 1: Comparison of CRC Screening Tested Amon

Three HINTS group – Underserved.

Variable Category

Group

Tested

N (%)

Group

Tested

N (%)

Group

Tested

N (%)

Age 18-49

232

(5.36)

574

(10.90)

582

(7.56)

Gender Female

2735

(63.22)

2969

(56.35)

4327

(56.19)

Educatio

< 12

ears

409

(9.45)

484

(9.18)

519

(6.75)

Income < $20K

661

(15.28)

971

(18.43)

1269

(16.48)

Employ

ment

Unempl

2732

(63.15)

3099

(58.82)

2729

(35.44)

Race

Non-

White

826

(19.07)

2101

(39.88)

3146

(40.88)

Further disparities emerged in education and income

levels. Individuals with less than a high school

education (<12 years) showed persistently lower

screening rates, declining from 9.45% in Group 1 to

6.75% in Group 3. Compared to individuals with at

least a high school diploma or higher, this group

remains significantly underserved, underscoring the

influence of education and health literacy on

screening adherence. Similarly, CRC screening

among individuals earning below $20,000 annually

increased from 15.28% in Group 1 to 18.43% in

Group 2 but slightly dropped to 16.48% in Group 3.

This income disparity reflects barriers such as

affordability and access to preventive healthcare,

which disproportionately affect lower-income

populations.

Racial disparities were also notable, with non-

White populations increasing their representation

from 19.07% in Group 1 to 40.88% in Group 3. While

this improvement is promising, non-White

individuals remain under-screened compared to

White populations, who consistently demonstrate

higher screening rates. Gender and employment

disparities also persisted; although females

consistently represented a higher proportion of those

screened, their participation rates declined from

63.22% in Group 1 to 56.19% in Group 3.

Conversely, males, historically less represented in

screening, may have experienced incremental

improvements over time. Meanwhile, unemployed

individuals saw significant reductions in screening

rates—from 63.15% in Group 1 to 35.44% in Group

3—remaining significantly underserved compared to

their employed counterparts, who likely benefit from

better healthcare access.

In summary, underserved groups—such as

younger individuals aged 18–49, those with lower

educational attainment, lower incomes,

unemployment, and non-White populations—

consistently lag behind their counterparts (older

individuals, those with higher education or income,

the employed, and White populations) in CRC

screening participation. These disparities highlight

the need for targeted public health interventions,

policy reforms, and culturally sensitive outreach

strategies to promote CRC screening among these

vulnerable groups. Addressing these gaps is essential

to reducing health inequities and improving

population-level outcomes.

HEALTHINF 2025 - 18th International Conference on Health Informatics

232

Figure 1: CRC Screening Uptake by Age Groups Among 3

HINTS Groups.

Our further findings on age-specific screening

reflect the impact of recent screening uptake policy

recommendations. Updates to CRC screening

guidelines by the American Cancer Society (ACS)

and the U.S. Preventive Services Task Force

(USPSTF), which now recommend initiating CRC

screening at age 45, may be contributing to a modest

increase in screening rates among individuals aged

45-49, particularly in Groups 2 and 3 (Atarere,

2024b). Figure 1 reveals a gradual increase in testing

rates within the 30-49 age bracket, which could

indicate early adoption of these guideline shifts by

healthcare providers and patients. However,

screening rates for those under 50 remain low

compared to older age groups, highlighting an area

with significant potential for improving adherence.

In the 50-64 age group, screening rates reach their

peak, confirming this as the primary age range for

CRC screening. As shown in Figure 1, Group 2 leads

with a screening rate of 43.88%, followed closely by

Group 1 at 40.92% and Group 3 at 38.06%. This

pattern reflects the age group's alignment with

recommended screening ages and highlights the

effectiveness of CRC screening initiatives targeted

toward this demographic. Recent studies indicate that

high adherence in this group is often driven by regular

healthcare provider recommendations and routine

medical care access (Atarere, 2024b; Wu, 2022).

Among older adults aged 65-74, screening rates

remain substantial, likely due to Medicare coverage,

which facilitates preventive screenings. Figure 1

illustrates that Group 3 has the highest adherence at

31.58%, followed by Group 1 at 28.62% and Group 2

at 25.09%. The sustained adherence in this age group

is likely bolstered by Medicare's support for routine

screenings, as well as a focus on preventive health

among this population (Atarere, 2024a). Group 3’s

elevated rate could reflect improved preventive health

practices among the cohort or more proactive

Medicare utilization.

In the 75+ age group, CRC screening rates drop

markedly, which aligns with current guidelines that

discourage routine screening for older seniors due to

increased procedural risks and lower expected

benefit. In Figure 1, Group 1 has the highest rate at

24.23%, followed by Group 3 at 20.31% and Group 2

at 17.73%. This decline may reflect adherence to

recommendations, though the relatively higher rate in

Group 1 suggests that some seniors continue

screening based on individual decisions or physician

recommendations, despite standard guidelines

(Atarere, 2024c).

Overall, Figure 1, Table 1, and Appendix A

collectively underscore age as a critical factor in CRC

screening uptake, with the highest adherence

observed in the 50-64 and 65-74 age groups. The

variations across groups reveal stronger adherence in

the 50-64 range for Group 2 and the highest rates in

the 65-74 range for Group 3, highlighting the role of

Medicare in supporting screening behaviors. These

findings suggest that CRC screening is highly age-

dependent, following established guidelines, with

Figure 1 clearly depicting age-specific screening

trends. This evidence points to a need for targeted

strategies to improve screening rates among younger

adults, who remain under-screened.

Digital Characteristics Among Three Groups

The findings in Table 2 reveal evolving trends in

digital connectivity, device ownership, and access to

health information resources among CRC-screened

individuals across three HINTS groups. Internet

usage has significantly increased over time, with

Group 3 showing the highest rate at 76.73%,

compared to 69.39% in Group 2 and 54.65% in Group

1, indicating a growing reliance on the internet for

health-related information and resources. Among

connection types, broadband, Wi-Fi, and mobile

internet use have also expanded in recent groups, with

mobile access (“Cell” in Table 2) notably rising from

21.94% in Group 2 to 39.54% in Group 3,

underscoring a shift toward mobile connectivity and

more flexible access to health information.

Device ownership also demonstrates upward

trends, particularly in smartphone ownership, which

reached 71.5% in Group 3, and tablet ownership,

rising from 9.47% in Group 2 to 53.6% in Group 3.

This increased adoption of mobile and digital devices

likely supports easier access to health resources,

potentially influencing CRC screening behaviors.

Access to electronic health information similarly

improved across groups, with 65.03% of CRC-tested

individuals in Group 3 accessing digital health

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data

233

resources, up from just 15% in Group 1, reflecting an

increase in both digital engagement and the

availability of electronic health information systems

in recent years. Social media engagement saw a

dramatic rise as well, from 13.44% in Group 2 to

57.17% in Group 3, suggesting an emerging role of

social media in health information dissemination and

community support for CRC-screened individuals.

Table 2: Digital Factors Comparison Across Three HINTS

Groups - CRC Screening Tested.

Variable Category

Group

Tested

4326

(100%)

Group

Tested

5269

(100%)

Group

Tested

7701

(100%)

Use Internet Yes

2364

(54.65)

3656

(69.39)

5909

(76.73)

Internet

Type

Dial-Up n/s

272

(5.56)

166

(2.16)

Cell n/s

1156

(21.94)

3045

(39.54)

Broadband n/s

2551

(48.42)

1915

(24.87)

Wi-Fi n/s

1968

(37.35)

4221

(57.41)

Where Use

Internet

Home

1630

(37.68)

n/s

3323

(43.15)

Work

(1.8)

n/s

1535

(19.93)

Public

Place

n/s n/s

(0.43)

Mobile n/s n/s

3286

(42.67)

Electronic

Health Info

Accessed

649

(15)

2860

(54.28)

5008

(65.03)

Health

Device

Owner

Tablet n/s

499

(9.47)

4128

(53.6)

Smartphone n/s

633

(12.01)

5506

(71.5)

Social

Media

Visited n/s

708

(13.44)

4403

(57.17)

Electronic

Medical

Record

Maintained

1225

(28.32)

4684

(88.9)

6543

(84.96)

Provider

Access

Online

Offered n/s n/s

3374

(43.81)

Health

Tracking

Devices

Wearable n/s n/s 1222

(15.87)

Healthcare provider-driven digital health

initiatives also saw substantial growth. By Group 3,

84.96% of respondents reported that their provider

maintained EHRs, and 43.81% had online access to

their health records, indicating expanding digital

infrastructure that may be contributing to preventive

health engagement, including CRC screening.

Additionally, wearable health-tracking devices were

more commonly used in Group 3, with 15.87%

reporting usage, which suggests an increasing

emphasis on self-monitoring and preventive actions

that could positively impact CRC screening

adherence.

However, an important limitation is the presence

of several "n/s" (not surveyed) entries, indicating that

specific questions were omitted in certain HINTS

cycles. This inconsistency limits our ability to

conduct longitudinal comparisons for some digital

factors, such as types of internet connections, usage

locations, and wearable health-tracking device

adoption. While trends are apparent, these gaps

suggest caution when interpreting results as fully

representative across all cycles. This limitation

emphasizes the need for more consistent data

collection in future cycles to enable comprehensive

trend analysis over time.

Figure 2a: SHAP graph of Group 1 (HINTS123).

4.2 Feature Selection – Critical Factors

Influencing CRC Screening Uptake

Overall Features Among Three Groups

The SHAP analysis across Groups 1, 2, and 3

highlights the most influential factors impacting CRC

screening adherence over time. In Group 1 (2003–

2008), recent fecal occult blood test (FOBT) behavior

(FOB1Yr) emerged as the most impactful feature,

with individuals who had previously undergone

FOBT screening being significantly more likely to

adhere to CRC screening recommendations. This

aligns with the importance of reinforcing preventive

behaviors through past engagement. Related features,

such as stool blood test timing

(BR88WhenStoolBlood, BR89WhyStoolBlood), and

endoscopic procedure timing (EndoYrs), also showed

substantial contributions, emphasizing the role of

HEALTHINF 2025 - 18th International Conference on Health Informatics

234

recent screening experiences and provider

recommendations in fostering adherence during this

earlier period. Demographic factors like age

(AgeGrpB) and socioeconomic elements, while

included, played a less dominant role. Figure 2a

shows 15 important features from Group 1.

Extended

Features for further analysis are included in

Appendix B.

In Group 2 (2011–2013), age (AgeGrpB) became

the most significant predictor, reflecting the growing

adherence to age-specific screening guidelines. The

importance of healthcare provider discussions about

CRC screening (DrTalkColCaTest) was also

prominent, highlighting the critical role of provider-

patient communication in increasing screening

uptake. Preventive health behaviors, such as

mammogram participation (WhenMammogram) and

PSA test history (EverHadPSATest), emerged as

secondary influencers, suggesting that individuals

engaged in other preventive health measures were

more likely to comply with CRC screening. Notable

contextual features included household composition

(ChildrenInHH) and routine healthcare checkups

(MostRecentCheckup), which indicated that family

settings and regular medical care contributed to

adherence during this period. Figure 2b shows 15

important features from Group 2.

Figure 2b: SHAP graph of Group 2 (HINTS4).

In Group 3 (2018–2020), the relative importance of

digital health literacy and access to healthcare

resources became increasingly evident. Age

(AgeGrpB) remained the most significant predictor,

but health insurance coverage through Medicare

(HealthIns_Medicare) and preventive health

behaviors, such as mammograms

(WhenMammogram), gained prominence. Features

related to chronic health conditions

(MedConditions_HighBP) and consistent provider

relationships (RegularProvider) also demonstrated

meaningful contributions. Moreover, digital

engagement variables like internet use through

mobile devices (WhereUseInternet_MobileDevice)

and electronic medical record (EMR) maintenance by

providers (ProviderMaintainEMR) highlighted the

increasing role of digital tools in influencing

screening behaviors. These shifts reflect the growing

integration of digital health technologies and access

disparities into screening decision-making. Figure 2c

shows 15 important features from Group 3.

Figure 2c: SHAP graph of Group 3 (HINTS5).

The SHAP analysis underscores the evolving

predictors of CRC screening adherence, with a shift

from prior screening behaviors and demographic

factors in earlier groups (Group 1) to greater

emphasis on healthcare provider interactions,

preventive health engagement, and digital health

access in later groups (Groups 2 and 3). This

evolution highlights the importance of adapting

public health strategies to leverage digital tools and

target demographic disparities while reinforcing the

role of consistent healthcare provider engagement in

promoting screening adherence.

Digital Features Among Three Groups

In order to focus on critical digital factors, Figure 3

provides a comparative overview of key digital

engagement trends across three HINTS groups,

highlighting significant shifts in digital health

utilization over time. Our findings include:

• Mobile Device and Internet Usage: "

MobileDevice" and "Internet_Cell" consistently

rank high in importance across all groups,

reflecting a growing reliance on mobile and

cellular internet access for health information

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data

235

and engagement across Group 1 (HINTS123),

Group 2 (HINTS4), and Group 3 (HINTS5).

• Social Media Engagement: "SocMed_Visited"

shows persistent significance across all groups,

underscoring social media’s role as a major

platform for health-related digital interactions,

particularly for sharing and seeking health

information.

• Device Ownership: Variables like

"HaveDevice_SmartPh" and

"Tablet_AchieveGoal" score higher in Group 3

(HINTS5) and Group 1 (HINTS123), suggesting

a substantial increase in smartphone and tablet use

for health purposes, which supports broader

digital access and convenience in recent years.

• Provider Digital Interactions:

"ProviderMaintainEMR" and

"HCPEncourageOnlineRec" reach their highest

scores in Group 3 (HINTS5), indicating a

strengthened focus on provider-supported digital

health tools, particularly electronic medical

records (EMR), demonstrating deeper integration

of digital interactions within healthcare over time.

• Electronic Health Information Access:

"Electronic_HealthInfo" and

"Electronic_TalkDoctor" have significantly

higher scores in recent groups, particularly in

Group 3 (HINTS5) and Group 2 (HINTS4),

reflecting an upward trend in patients’ access to

electronic health information and digital

communication with healthcare providers.

Figure 3: Comparison of Critical Digital Factors Across

three HINTS groups.

Overall, Figure 3 illustrates a progressive shift

toward digital health resources across the HINTS

groups, marked by increasing mobile internet usage,

social media engagement, provider-supported digital

tools, and enhanced electronic health information

access. This trend underscores the expanding role of

digital tools in facilitating health engagement and

access over time, and subsequent analyses will focus

on these key digital factors to evaluate their impact on

health behaviors and outcomes.

4.3 Comparison of the Machine

Learning Models’ Performance for

CRC Screening Uptake Prediction

This analysis evaluates the performance of LR and

RF models across three HINTS datasets—

HINTS123, HINTS4, and HINTS5—by examining

metrics such as accuracy, precision, recall, F1 score,

and AUC. Each dataset represents unique

characteristics that challenge the models differently,

providing insight into the models' suitability for

various data complexities.

Table 3: Comparison of LR and RF Models on HINTS

Datasets.

Group Model Pr Re F1 AUC

Group 1

HINTS123

LR 99.55 99.1 99.32 99.41

RF 97.55 95.93 96.74 97.23

Group 2

HINTS4

LR 81.2 83.14 82.16 81.84

RF 80.26 85.41 82.75 82.08

Group 3

HINTS5

LR 86.45 82.27 84.31 80.54

RF 83.75 90.92 87.18 80.95

In Group 1, using the HINTS123 dataset, LR

performed exceptionally well across all metrics,

outstripping RF. LR achieved an accuracy of 99.49, a

precision of 99.55, and a recall of 99.1, indicating that

it could classify instances with remarkable accuracy

and minimal error. Its F1 score of 99.32 and AUC of

99.41 further underscore its capability in

distinguishing between classes effectively. In

contrast, while RF also showed strong performance,

its lower recall (95.93) and AUC (97.23) metrics

indicate that it was slightly less effective than LR in

managing this dataset’s characteristics.

In Group 2, with the HINTS4 dataset, RF had a

slight advantage over LR, particularly in metrics such

as recall and AUC. While both models performed

comparably in accuracy—RF at 82.1 and LR at

81.85—RF excelled in recall, achieving 85.41

compared to LR’s 83.14. This advantage in recall

suggests that RF was more sensitive in identifying

positive instances within this dataset, a key benefit for

applications prioritizing true positive detection. RF’s

F1 score (82.75) and AUC (82.08) also outpaced

HEALTHINF 2025 - 18th International Conference on Health Informatics

236

Logistic Regression, highlighting its suitability for

more complex data structures like those found in

HINTS4.

In Group 3, which used the HINTS5 dataset, RF

continued to outperform LR, particularly in recall and

F1 score, demonstrating its strength in identifying

positive cases in this dataset. RF achieved an

accuracy of 83.28 and a recall of 90.92, compared to

Logistic Regression’s accuracy of 80.96 and recall of

82.27. Additionally, the F1 score for RF was

significantly higher at 87.18 compared to Logistic

Regression’s 84.31, indicating that RF maintained a

better balance between precision and recall. This

improved performance highlights RF’s ability to

capture nuanced, non-linear patterns in the HINTS5

dataset, making it a more suitable model for datasets

with complex relationships.

In summary, LR showed exceptional performance

in Group 1, making it ideal for datasets with

straightforward, linear relationships like HINTS123.

In contrast, RF proved advantageous in Groups 2 and

3, excelling in datasets with greater complexity, as

seen in HINTS4 and HINTS5. These findings suggest

that LR is highly effective for simpler datasets, while

RF is better suited for complex datasets requiring

high sensitivity and the ability to handle intricate,

non-linear patterns. This comparison serves as a

guide for model selection based on dataset

characteristics and the importance of specific metrics

such as recall or precision in future applications.

5 DISCUSSIONS

This study explored the impact of digital health

literacy and socioeconomic factors on CRC screening

behaviors across different time periods using data

from the HINTS datasets. By applying machine

learning models, we identified critical determinants

influencing CRC screening adherence and uncovered

variations in the influence of socioeconomic and

digital health literacy factors over time.

Our analysis confirms several important trends in

CRC screening behavior, as observed in prior studies,

while also highlighting new insights specific to digital

health engagement. Age, socioeconomic stability,

and digital literacy emerged as consistent predictors

of CRC screening uptake across the HINTS cycles.

This indicates that while digital health interventions,

such as patient portals and telehealth, have gained

prominence, traditional demographic factors continue

to play a substantial role in CRC screening adherence.

The finding aligns with studies like Atarere et al.

(2024a, 2024b, 2024c), which emphasize that while

digital health tools may enhance adherence,

addressing fundamental socioeconomic and

demographic barriers is essential for achieving equity

in CRC screening rates.

The SHAP analysis demonstrated that prior CRC

screening behaviors, age, and patient-provider

interactions were among the strongest predictors of

screening adherence, especially in the earlier HINTS

groups. This reinforces the idea that positive prior

experiences with CRC screening and effective

communication with healthcare providers are crucial

in fostering long-term adherence to screening

recommendations (Wu, et al.2023). Specifically, our

results underscore the role of healthcare providers in

reinforcing CRC screening messages, particularly for

at-risk populations who may be less engaged with

digital health tools.

Chronic conditions, including hypertension and

diabetes, also play a role in shaping CRC screening

behaviors across the HINTS groups. These conditions,

often requiring routine medical attention, may increase

patients’ engagement with healthcare providers,

creating additional opportunities for providers to

recommend CRC screening as part of comprehensive

preventive care. The consistent significance of

variables such as "MedConditions_HighBP" and

"MedConditions_Diabetes" across multiple groups

indicates that individuals managing chronic illnesses

may be more attuned to preventive health measures.

Additionally, health coverage factors, particularly

Medicare (represented by "HCCoverage_Medicare"),

are associated with higher screening rates among those

with chronic conditions, likely reflecting the expanded

access to preventive services Medicare offers to older

adults with chronic health needs.

Collectively, these insights suggest that

psychological readiness, chronic condition

management, and continuous healthcare engagement

are influential in CRC screening decisions.

Addressing psychological barriers, reinforcing

supportive patient-provider communication, and

leveraging routine chronic care visits for screening

recommendations could help enhance screening

adherence, especially among high-risk or less-

engaged populations.

The influence of digital health literacy on CRC

screening behaviors appeared most prominent in the

later HINTS cycles, particularly HINTS 5 (2018-

2020). This suggests that as digital health

technologies become more integrated into routine

healthcare, the ability to navigate these tools may

increasingly shape preventive health behaviors. For

instance, confidence in locating reliable health

information online and regular telehealth usage were

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data

237

associated with higher screening adherence,

indicating that digital health literacy is becoming an

important factor in promoting preventive behaviors

like CRC screening. These findings suggest a need for

targeted interventions to enhance digital literacy

among populations with historically low screening

rates, such as rural communities and lower-income

groups.

Our comparison of LR and RF models across

different HINTS datasets highlighted that model

performance varies with data complexity. LR

outperformed RF on the HINTS123 dataset (2003-

2008), likely due to the dataset's simpler, more linear

structure, while RF excelled in HINTS4 and HINTS5

datasets, which introduced more complex, non-linear

relationships as HIT variables became more

prevalent. This performance variation indicates that

non-linear models like RF may be more effective for

analyzing recent datasets where digital health literacy

factors play a larger role. Future studies aiming to

incorporate digital health engagement factors should

consider leveraging non-linear models to capture the

nuanced behaviors associated with these variables.

This study has several limitations. First,

HINTS data are based on self-reported responses,

which may introduce reporting bias. Additionally, the

focus on the U.S. population limits the

generalizability of our findings to other regions where

digital health adoption and socioeconomic structures

differ significantly. The machine learning models,

while effective in identifying predictive factors, may

not fully capture the dynamic and complex

interactions that influence health behaviors over time.

Further studies could benefit from incorporating

longitudinal data or using more advanced modelling

techniques, such as neural networks, to capture

temporal patterns in digital health engagement.

Future research should explore integrating EHRs

with survey data to enhance the predictive accuracy

of CRC screening models. Additionally, examining

the role of social determinants of health, such as

social support and community engagement, could

provide a more holistic view of factors influencing

CRC screening adherence.

6 CONCLUSIONS

CRC screening rates in the United States have shown

improvements over time, yet significant disparities

persist, particularly among underserved populations

such as younger individuals, non-White racial groups,

and those with lower socioeconomic status or limited

digital health literacy. Leveraging machine learning

models and SHAP analysis on HINTS data across

three temporal groups revealed evolving predictors of

CRC screening adherence. Key findings emphasize

the persistent influence of traditional

sociodemographic factors like age, income, and

education, alongside the growing importance of

digital health literacy and access to health

technologies. Targeted interventions focusing on

enhancing digital health engagement, improving

access to preventive care, and addressing

socioeconomic barriers are critical to bridging these

disparities. These findings highlight the importance

of integrating digital tools with equitable public

health strategies to improve CRC screening uptake

and reduce health inequities across diverse

populations.

ACKNOWLEDGEMENTS

Funding: This publication was supported by the

University of Kentucky College of Medicine

Artificial Intelligence in Medicine Research Alliance

and the Center for Clinical and Translational Science,

which is funded by the National Center for Research

Resources and the National Center for Advancing

Translational Sciences, National Institutes of Health,

through Grant UL1TR001998. The content is solely

the responsibility of the authors and does not

necessarily represent the official views of the NIH.

REFERENCES

Atarere, J., Chido-Amajuoyi, O., Mensah, B., Onyeaka, H.,

Adewunmi, C., Umoren, M., & Kanth, P. (2024a).

Primary care telehealth visits and its association with

provider discussion on colorectal cancer screening in

the United States. Telemedicine and e-Health, 30(5),

1325-1329.

Atarere, J., Haas, C., Akhiwu, T., Delungahawatta, T.,

Pokharel, A., Adewunmi, C., & Barrow, J. (2024b).

Prevalence and predictors of colorectal cancer

screening in the United States: Evidence from the

HINTS database 2018 to 2020. Cancer Causes &

Control, 35(2), 335-345.

Atarere, J., Haas, C., Onyeaka, H., Adewunmi, C.,

Delungahawatta, T., Orhurhu, V., & Barrow, J.

(2024c). The role of health information technology on

colorectal cancer screening participation among

smokers in the United States. Telemedicine and e-

Health, 30(2), 448-456.

Bray, F., Ferlay, J., Soerjomataram, I., Siegel, R. L., Torre,

L. A., & Jemal, A. (2018). Global cancer statistics

2018: GLOBOCAN estimates of incidence and

mortality worldwide for 36 cancers in 185 countries.

HEALTHINF 2025 - 18th International Conference on Health Informatics

238

CA: A Cancer Journal for Clinicians, 68(6), 394–424.

De La Garza, Á. G., Blanco, C., Olfson, M., & Wall, M. M.

(2021). Identification of suicide attempt risk factors in

a national US survey using machine learning. JAMA

Psychiatry, 78(4), 398-406.

Finney Rutten, L. J., Hesse, B. W., Moser, R. P., McCaul,

K., & Rothman, A. J. (2009). Public understanding of

cancer prevention, detection, and survival/cure:

Comparison with state-of-science evidence for colon,

skin, and lung cancer. Journal of Cancer Education,

24(1), 40-48.

Ford, J. S., Coups, E. J., & Hay, J. L. (2006). Knowledge of

colon cancer screening in a national probability sample

in the United States. Journal of Health Communication,

11(Suppl 1), 19-35.

Geiger, T. M., Miedema, B. W., Geana, M. V., Thaler, K.,

Rangnekar, N. J., & Cameron, G. T. (2008). Improving

rates for screening colonoscopy: Analysis of the health

information national trends survey (HINTS I) data.

Surgical Endoscopy, 22(2), 527-533.

Hay, J., Coups, E., & Ford, J. (2006). Predictors of

perceived risk for colon cancer in a national probability

sample in the United States. Journal of Health

Communication, 11(Suppl 1), 71-92.

Idowu, K. A., Adenuga, B., Otubu, O., Narasimhan, K.,

Kamara, F., Hunter-Richardson, F., Larbi, D., Sherif, Z.

A., & Laiyemo, A. O. (2016). Place of birth, cancer

beliefs, and being current with colon cancer screening

among US adults. Annals of Gastroenterology, 29(3),

336-340.

Jun, J., & Oh, K. (2013). Asian and Hispanic Americans’

cancer fatalism and colon cancer screening. American

Journal of Health Behavior, 37(2), 145-154.

Keum, N., & Giovannucci, E. (2019). Global burden of

colorectal cancer: emerging trends, risk factors and

prevention strategies. Nature reviews Gastroenterology

& Hepatology, 16(12), 713-732.

McIntosh, J. G., Jenkins, M., Wood, A., Chondros, P.,

Campbell, T., Wenkart, E., ... & Emery, J. D. (2024).

Increasing bowel cancer screening using SMS in

general practice: the SMARTscreen cluster randomised

trial. British Journal of General Practice, 74(741),

e275-e282.

Miller Jr, D. P., Denizard-Thompson, N., Weaver, K. E.,

Case, L. D., Troyer, J. L., Spangler, J. G., ... & Pignone,

M. P. (2018). Effect of a digital health intervention on

receipt of colorectal cancer screening in vulnerable

patients: a randomized controlled trial. Annals of

Internal Medicine, 168(8), 550-557.

Mirzaei, A., Aslani, P., & Schneider, C. R. (2022).

Healthcare data integration using machine learning: A

case study evaluation with health information-seeking

behavior databases. Research in Social and

Administrative Pharmacy, 18(12), 4144-4149.

Nawaz, H., Via, C., Shahrokni, A., Ramdass, P., Raoof, A.,

Sunkara, S., & Petraro, P. (2014). Can the inpatient

hospital setting be a golden opportunity to improve

colon cancer screening rates in the United States?

Health Promotion Practice, 15(4), 506-511.

Siegel, R. L., Miller, K. D., Fuchs, H. E., & Jemal, A.

(2022). Cancer statistics, 2022. CA: A Cancer Journal

for Clinicians, 72(1), 7–33.

Sun, C., Mobley, E., Quillen, M., Parker, M., Daly, M.,

Wang, R., & Xu, J. (2024). Predicting early-onset

colorectal cancer in individuals below screening age

using machine learning and real-world data. medRxiv,

2024-07.

Wu, S., Zhang, X., Chen, P., Lai, H., Wu, Y., Shia, B. C.,

& Qin, L. (2022). Identifying the predictors of patient-

centered communication by machine learning methods.

Processes, 10(12), 2484.

Zhen, J., Li, J., Liao, F., Zhang, J., Liu, C., Xie, H., & Dong,

W. (2024). Development and validation of machine

learning models for young-onset colorectal cancer risk

stratification. NPJ Precision Oncology, 8(1), 239.

The Role of Digital Health Literacy and Socioeconomic Factors in Colorectal Cancer Screening: Machine Learning Analysis of HINTS Data

239