A Machine Learning Classification Model for Gastrointestinal Health in Cancer Survivors: Roles of Telomere Length and Social Determinants of Health

1. Introduction

Cancer is the second leading cause of death in the United States, after heart disease. Projections for 2024 estimated 2.0 million new cases and 611,720 cancer deaths [1]. Encouragingly, medical advancements have increased survival rates among patients with cancer [2,3]. However, cancer survivors often deal with multiple short- and long-term side effects over the course of their cancer treatments [2,3]. These effects include physical (e.g., pain, neuropathy, functional limitations), gastrointestinal (GI), and mental health (e.g., depression, anxiety) concerns [2-4]. The prevalence and severity of these physical, GI, and mental health concerns can vary widely, impacting survivors’ health-related quality of life (HRQOL), treatment adherence, daily functioning, nutrition, and overall prognosis. Addressing these overall health concerns is essential for enhancing cancer survivors’ well-being in the long term.

Of note, GI symptoms often persist in cancer survivors even after completing treatment. These symptoms include nausea/vomiting, appetite loss, altered bowel movements (e.g., diarrhea or constipation), bloating, indigestion, heartburn, and abdominal pain [5-9]. GI symptoms rank as the most common chronic physical side effects of cancer treatments, after psychological distress and fatigue in cancer survivors with mixed cancer types [10]. In 142 breast cancer survivors, the GI symptom cluster was the second most prevalent after chemotherapy [11]. In 413 colorectal cancer survivors, 81% experienced persistent GI symptoms 8 years post-treatment [8]. In a review of GI toxicity after radiotherapy in rectal cancer survivors, long-term GI toxicity continued for over 3 months and included diarrhea (35%), fecal incontinence (22%), abdominal gas (71%), and abdominal pain (13%) [12].

GI side effects related to cancer treatments are prevalent in older adults cancer patients, impacting physical and social functioning and HRQOL [13]. GI symptoms are significant concerns for older adults cancer patients, with the incidence of overall GI symptoms reported to be as high as 40% in cancer patients on standard-dose chemotherapy and 100% on high-dose chemotherapy [14]. Several factors contribute to the increased prevalence of GI issues in this population. Firstly, the aging process causes clinically significant effects on oropharyngeal motility, upper-esophageal motility, colonic function, and GI immunity [15]. Second, older adults often have comorbidities and long-term exposure to medications, alcohol, and tobacco that may exacerbate GI distress [16]. Further, cancer treatments can induce accelerated aging in individuals with cancer [17]. Mechanisms such as oxidative stress, inflammation, and mitochondrial dysfunction are implicated [17]. This accelerated aging phenomenon can worsen existing GI health conditions. As such, GI health concerns may be associated with the aging process, and cancer survivors can be more vulnerable to these connections.

Recent studies highlight the increased significance of biological over chronological aging in cancer survivors’ physical and psychological well-being [18-20]. Of note, telomere length (TL), which shortens during cell division, is a validated measure of biological aging [21,22]. In individuals of the same chronological age, shorter TL is linked to accelerated biological aging and various health conditions in cancer survivors [20,21]. While the association of TL with survival and mortality is well-studied in cancer survivors [22], its association with HRQOL, including GI health, requires further investigation [18,19]. Social determinants of health (SDOH) significantly impact the physical and mental health of cancer survivors. Factors such as race/ethnicity, socioeconomic status, education, and marital status play crucial roles in the health outcomes of cancer survivors [23,24]. Chronic stress associated with poor SDOH triggers systemic inflammation, exacerbating physical symptoms [25,26]. Moreover, there is a potential link between TL, SDOH, and inflammation [27]. Poor SDOH status was associated with TL shortening due to chronic stress and inflammation in US adults living in the community [27]. Therefore, SDOH and TL may be related to GI health in cancer survivors. Understanding this complex interplay could inform interventions to improve GI health in cancer survivors.

The classification of GI health conditions and identification of contributing factors are crucial steps in choosing and applying personalized interventions for cancer survivors [28,29]. Machine learning (ML) offers substantial advantages in cancer survivorship care, particularly in classification or prediction models [30,31]. Unlike traditional statistical methods, ML can handle small sample sizes and multiple variables with complex relationships by controlling covariates and multicollinearity. It excels at identifying intricate patterns, handling high-dimensional data, and adapting over time [30,31]. This capability is especially beneficial in cancer survivorship research, where the number of survivors for certain types of cancer might be limited, and the relationships among cancer treatments, and health outcomes can be complex [28,29]. While ML has been employed to develop predictive models for cancer diagnosis and survival [32], its application to GI health conditions in cancer survivors remains relatively rare.

Therefore, by leveraging ML with high precision, we aimed to develop and validate an ML classification model of GI health conditions (better vs. worse), incorporating TL and SDOH indicators as our primary interests, and demographic and clinical characteristics including inflammatory markers as secondary interests. The current study is a pilot to explore and identify the significant features including biological aging markers (i.e., TL in our study), and SDOH indicators to classify GI health conditions in adult cancer survivors, not just limited to those over 65. This approach enhances the performance of ML classification models by increasing sample size and providing a comprehensive understanding of GI health across different age groups.

2. Methods

2.1. Sample and Procedures

In this secondary analysis, we employed data from the National Health and Nutrition Examination Survey (NHANES), conducted by the National Center for Health Statistics (NCHS) under the Centers for Disease Control (CDC) [33]. The NHANES includes cross-sectional, nationally representative health and nutritional data from 21,004 non-institutionalized US civilian aged 2 months and older [33]. We combined NHANES surveys from 1999-2000 and 2001-2002, for which leukocyte DNA samples were obtained to assess TL in participants 20 years of age and above. Of the 10,291 participants 20 or over, 7,827 (76%) had TL data. Among the participants with TL data, 645 with self-reported cancer diagnoses were included in the current study. We also applied sampling weights in the analysis to address oversampling and non-response biases, ensuring the accuracy of estimates reflecting the broader US population [34]. We employed a standard mining approach consisting of four stages: (1) data acquisition, (2) preprocessing (e.g., data cleaning, exploratory data with addressing class imbalances, optimizing dataset classes with feature engineering, and data normalization), (3) model learning with training and testing data, and (4) model evaluation [35] (Figure 1).

2.2. Features

2.2.1. Demographic and clinical data including inflammatory markers.

Information on chronological age (years), sex (male, female), comorbidities including hypertension, diabetes, cardiovascular disease, and history and types of cancer diagnosis were collected from self-reported questionnaires. White blood cell count (WBC) was extracted from complete blood count (CBC) data retrieved from the “Complete Blood Count with 5-part Differential - Whole Blood” category of the 1999-2002 NHANES laboratory data [36]. C-reactive protein (mg/dL) levels were measured by high-sensitivity, latex-enhanced nephelometry by the Immunology Division, Department of Laboratory Medicine, University of Washington Medical Center. As diet is relevant to GI health, we also included diet data [36].

2.2.2. Telomere length (TL) measurement.

The measurement of TL in the NHANES study has been published elsewhere [37-40]. Detailed protocols describing TL measurement for the NHANES study are accessible on the CDC website under the laboratory section [41]. In brief, TL was assessed by isolating purified DNA from whole blood samples using the Puregene (D-50 K) protocol. The TL assay was conducted via polymerase chain reaction. TL was measured relative to standard reference DNA (T/S ratio), with each sample analyzed three times on different days and in duplicate wells, totaling six data points. Potential outliers (<2% of samples) were identified and excluded. The interassay coefficient of variation was 6.5% [27]. The provided values represent the mean (standard deviation) of the T/S ratio. The CDC Institutional Review Board granted human subject approval for TL measurements, which were quality-controlled before being linked with the NHANES database.

2.2.3. Social determinants of health (SDOH).

We included nine selected SDOH variables based on sociodemographic variables and cancer health risk behaviors identified in the CDC Healthy People 2030 SDOH framework [42] and in the National Academy of Medicine [43]. The nine risk factors corresponding to SDOH domains included in our study were: i) Racial/ethnic minorities, ii) Low education achievement (i.e., less than a high school diploma or equivalent), iii) poverty-income ratio (i.e., Family Income/Poverty Threshold; A ratio <1 indicates annual family income is below the poverty level), (iv) Food insecurity as per the Kendall/Cornell scale (i.e., low or very low food security)[46], (v) Current smokers (individuals who have smoked a minimum of 100 cigarettes in their lifetime and who currently smoke either daily or occasionally), (vi) Heavy drinkers (as per the Alcohol Use Disorders Identification Test-Concise [AUDIT-C] screening tool, > 4 for males and > 3 for females indicating a moderate risk of alcohol abuse) [47], (vii) Low physical activity (engaging in less than 10 minutes of moderate or intense activity or strength training in the past month), (viii) lack of a partner (divorced/widowed/single marital status) [45], and (ix) diet quality. In the NHANES 1999-2002 datasets, dietary information was collected using 24-hour recalls [37]. Using this NHANES data, we included major food groups - total energy in kcal, total protein, total carbonate, total saturated fat, total fatty acid, total sodium, and total fruit and vegetables - based on the U.S. Dietary Guidelines for Americans in our study [38]. Then, we further computed overall diet quality using the Healthy Eating Index (HEI) 2015 [37,39]. The overall HEI-2015 score varies from 0, indicating nonadherence, to 100, perfect adherence. There is no established minimal clinically significant difference for the HEI-2015 [37,39].

2.3. Outcome

2.3.1. GI health.

We evaluated GI health status (i.e., worse or better status) using the Health Status Questionnaire, which includes the CDC health-related quality of life

(HRQOL)-4. The HRQOL-4 shows high reliability (0.57 - 0.75) in the general population and cancer survivors [48]. For GI health conditions, participants were asked to answer “Did you have a stomach or intestinal illness over the last month?” (No -better GI health versus Yes -worse GI health).

2.4. Data Analyses

2.4.1. Initial data analysis.

For descriptive analyses, categorical variables were presented as counts and percentages, while continuous variables were presented as means and standard deviations. We primarily compared the training and test datasets using two-sample independent t-tests or Chi-square tests. Additionally, we examined the initial associations of input data with GI health (better versus worse) using two-sample independent t-tests or Chi-square tests.

2.4.2. Machine learning model.

Data preprocessing. We only included input variables showing significant associations with GI health in our initial data analyses. We used the permutation feature importance to quantify the impact of each feature (i.e. input variables) on GI health prediction in our ML models [31,49]. Among 645 adult cancer survivors used to build the model, all had TL data and most of the employed features were available with low missing data rates (missing data rates <4%). Despite the rates of missing data in our study being trivial, we applied multivariate imputation with chained equations to address missing data. Prior to model training, continuous features underwent min-max normalization, while categorical features were target-encoded [49] reducing feature dimensions and enhancing the performance and stability of our ML models.

Classificationmodeling. We first created a training dataset by randomly matching 75% of all cancer survivors in the dataset and created a test dataset within the remaining 25% of cancer survivors (Figure 1). Python's train_test_split function was employed to

randomly assign samples to either the training or testing set based on the average prevalences of better GI health, ensuring dataset integrity [30]. The training dataset had 484 cancer survivors, and the test dataset had 161 cancer survivors. We utilized the training dataset for initial modeling and the test dataset to assess the model’s performance in classifying GI health on unseen data. When we considered the number of features in the 645 sample size, we applied 6 supervised ML modeling methods, including logistic regression (LR), supportive vector machine (SVM), decision tree (DT), random forest (RF), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost) [30]. Each method was assessed with a specific learning algorithm to determine its effectiveness in classifying GI health conditions. Logistic regression, a widely used binary classifier, served as the baseline model for comparison. Hyperparameter tuning was conducted through random search with five-fold cross-validation to prevent overfitting [49]. Binary cross-entropy was used as the evaluation criterion. Hyperparameters such as tree complexity, learning rate, and number of trees were adjusted for RF, GBM, and XGBoost. Linear kernel functions were favored over nonlinear ones, like radial basis function kernel, in support vector machine models to avoid overfitting in a small dataset [49]. Feature importance analysis identified and sorted the top influential features for predicting GI health conditions. Finally, the prediction model performance was evaluated for various metrics, including accuracy, precision, recall (sensitivity), specificity, F1 score, and area under the receiver operating characteristic curve (AUC).

We controlled for other input features that showed significant relationships with GI health in our ML models. A significance level of p < 0.05 was applied, and the statistical analyses were carried out using R software, the MLR (Machine Learning in R) R package (version 3.6.3, R Foundation for Statistical Computing, Vienna, Austria), and Python (version 3.10.2, Python Software Foundation, Wilmington, USA).

2.5. Conceptual Framework

The conceptual framework (Figure 2) for this study is based on the original framework, the Integral Conceptual Model of Frailty [50]. Figure 2 displays this framework, which includes various factors (e.g., life course determinants) and diseases (e.g., cancer and cancer treatment) that may impact frailty, which is correlated with accelerated aging. This impact is seen in the model’s sub-dimensions, namely physical, psychological, and social frailty. These three sub-dimensions can be characterized by a decline in various factors. Specifically, a decline in nutrition, mobility, physical activity, and physical function in the GI tract is relevant to GI health. Increases in frailty (specifically physical GI function in our study) ultimately result in adverse events (i.e., GI distress in our study). We used TL as a proxy for accelerated aging, which is impacted by SDOH as well as cancer disease and its treatment. Our review analyzed and synthesized the data by mapping TL (accelerated aging), SDOH (life course determinants), and GI health (adverse events) in adult cancer survivors to elements of the Integral Conceptual Model of Frailty [50]. In our study, we focused on the components represented by the grey boxes in this framework (Figure 2).

3. Results

3.1. Initial Descriptive Analyses

3.1.1. Participant characteristics, clinical data including inflammatory markers, TL, and GI health.

Table 1 describes the participant demographic and clinical characteristics and TL of both the training (N = 484, 75% of the total sample N = 645 adult cancer survivors) and the test (N = 161, 25% of the total sample N = 645 adult cancer survivors) datasets. The mean participant age was 66.3 ± 14.7 years for the training dataset and 65.5 ± 16.2 for the test dataset (p =.102). Approximately half of the participants were women, and skin cancer (approximately half were melanoma) was the most common cancer type in both datasets. WBC, CRP, and TL did not significantly differ between the training and test datasets. 66.7% reported better GI health for the training dataset, while 62.7% reported better GI health in the test dataset (p =.412).

3.1.2. SDOH variables.

Table 2 presents the SDOH variables for both datasets. Most participants were non-Hispanic Whites, married, and had an education level of high school or below. Cancer risk behaviors (such as smoking, heavy alcohol consumption, physical activity, and diet quality score – HEI) were not significantly different between the two datasets. The prevalences of participants’ annual family incomes below the poverty level, and experiencing food insecurity were similar in both datasets (39.7% versus 37.3%, p =0.423; 8.6% versus 8.1%, p =0.879, in the training dataset versus the test dataset, respectively).

3.1.3. Potential risk factors for GI health within the training dataset.

Before developing the ML model for GI health, we examined the associations between potential risk factors, including demographics, clinical data including inflammatory markers, TL, and SDOH, and GI Health in the training dataset. Table 1 and Table 2 summarize the results of these comparisons between the cancer survivors with better and worse GI health. The better GI health group was younger (63.3 years vs. 66.4 years, p = 0.031), contained a smaller number of females (47% vs. 65%, p = 0.013), and had fewer comorbidities (41% vs. 45%, p = 0.043). Mean WBC (k/ul) and CRP (mg/dl) levels were also lower in the better GI health group. The mean TL (kb) was longer in the better GI health group (Table 1). Non-Hispanic Whites were more prevalent in the better GI health group (80.3% vs. 77.3%). Hispanic individuals were more prevalent in the worse GI health group (10.7% vs. 7.5%, p = 0.039) (Table 2). Higher-income levels were associated with better GI health (p = 0.038). In terms of cancer risk behaviors, heavy alcohol users (21.3% vs. 15.2%, p <.001) were more prevalent in the worse GI health group compared to the better GI health group. Cancer survivors with regular physical activity (38.5% vs. 58.3%, p =.035) and better diet quality (7.5% vs. 5.6%, p =.038) were more likely to have better GI health status, compared to those with worse GI health status. The worse GI health group had higher prevalences of income below the poverty level (37.6%) and experiencing food insecurity (8%) than the better GI health group. Marital status, education, and current smoking status were not associated with GI health status.

3.2. Machine Learning Models for GI Health

3.2.1. Performance comparison for classification models.

We present the classification performance of all ML models using the training dataset as a development phase and the test dataset as a validation phase (Table 3). This evaluation was conducted with five-fold cross-validation. Furthermore, Figure 3 illustrates and compares the AUCs for each model in the training and test datasets. In the training dataset, random forest (RF) performed well across multiple metrics with a higher AUC value and F1 score (AUC: 0.9842, F1 score: 0.9489) compared to other models (Table 3). In the test dataset, the gradient boosting machine (GBM) model showed the best performance with the highest accuracy (0.7442), a strong AUC (0.8035), and an F1 score (0.8092), indicating good overall performance in distinguishing between the positive and negative cases. Incorporating TL (a value of feature importance = 0.31 in training dataset) and several SDOH indicators (feature importance values ranges: 0.03 to 0.24) in our ML models proved their relative importance significantly contributing to classifying GI health conditions, achieving good classification performance, and demonstrating potential high predictive accuracy for both the training and test datasets.

3.2.2. Feature Importance.

The most significant features of the best-performing models in each dataset (the RF model in the training dataset and the GBM in the test dataset), were separately ranked using the permutation feature importance (Figure 4). Figure 4 illustrates the relative importance of input features included in the ML models. Among the 13 input features excluding marital status, education, and smoking status in each RF and GBM model, the important top features for GI health were similar between the training versus testing datasets. A longer TL was the most influential feature, followed by an annual household income below the poverty level. Physical activity, lower WBC, and food security were among the top-ranked features of importance for better GI health in both datasets. While several other features (e.g., lower WBC levels, no food insecurity, higher income, younger chronological age, non-Hispanice White group) play a role in better GI health, their impact is comparatively modest.

4. Discussion

This study is the first to develop and validate the ML classification models for GI health in adult cancer survivors using supervised ML approaches to account for multiple factors. Although we used cross-sectional data, the ML algorithms used in our study constructed classification models based on demographics and clinical characteristics including inflammatory markers, TL, and SDOH factors for GI health, with good (if 0.5 < AUC < 0.7) to moderate to high (if > 0.7) prediction accuracy [29-31,49]. We also identified the relative importance of features classifying GI health conditions, by demonstrating that TL and some SDOH features (e.g., economic status, lifestyles) significantly influence the outcome classification (Better vs. Worse GI health status). The ML models developed and validated in our study could inform personalized approaches to identify cancer survivors at high risk for long-term GI distress, and thus, provide tailored interventions that address unmet needs triggering GI distress in adult cancer survivors.

Despite various predictive ML models being used in cancer survivors such as cancer diagnosis risk predictions, cancer survival rates, or detection of psychological symptoms [51-54], few studies have applied ML algorithms to classify or predict GI health in cancer survivors. Previous research has identified risk factors for GI distress in cancer survivors [7,8], but these studies did not explore the associations of TL and SDOH with GI distress. Emerging evidence supports the impact of SDOH [55,56] and biological age [56,57] on symptom disparities and HRQOL in cancer survivors. Our study addresses this gap by demonstrating the feasibility of using ML approaches to classify GI health. Specifically, we explore how TL and SDOH factors contribute to GI health in cancer survivors, providing new knowledge in this area. The ML models can handle numerous features effectively, minimizing both Type I and Type II errors in multiple comparisons. This advantage is often not feasible in traditional statistical methods (e.g., regressions and univariate analyses). Furthermore, ML models predict or classify GI health conditions more accurately than traditional statistical methods by leveraging large datasets and complex relationships among multiple input features.

Our findings suggested that not all features contributed equally to classifying GI health conditions. TL was identified as the most influential factor in GI health, independent of chronological age, suggesting a potential role for biological aging in GI conditions. The results of our study identified the positive relationships between better GI health, younger age, and longer TL. Having an income higher than the poverty level and routine physical activity also significantly contributed to better GI health.

Telomeres, protective caps at the ends of chromosomes, play a crucial role in cellular aging [39]. Beyond chronological age, short TL lengths are associated with cellular senescence, where cells lose their ability to divide and function properly. Furthermore, senescent cells release inflammatory molecules, contributing to chronic inflammation associated with GI disorders like inflammatory bowel disease, altered bowel patterns, abdominal pain, indigestion, bloating, nausea, and gastroenteritis [19,58]. Biological aging influences gut health by impairing the integrity of the intestinal barrier, affecting immune cell function, and impacting gut microbial diversity [19,58]. Further, our findings reveal that the biological age might better reflect the functional aging of the GI tract, compared to chronological age [59]. Wang et al. [19] similarly discovered that longer leukocyte TL was associated with better GI function in patients with functional GI disorders. Investigating the mechanisms responsible for the shorter leukocyte TL observed in these settings could provide insights into managing GI health beyond chronological age considerations in cancer survivors.

The poverty-income ratio (PIR), was the second most significant feature of GI health in our study, other SDOH variables—such as lower income levels and racial/ethnic minority groups—were also associated with worse GI health. Previous studies support our findings that socially and economically vulnerable populations are exposed to more chronic stress, which can influence accelerating aging and pro-inflammatory status in the body [60,61]. Furthermore, socially and economically vulnerable populations face challenges in accessing healthcare resources, including community health services, oncology care, and primary care providers (PCPs). Additionally, vulnerable populations are more likely to reside in unsafe environments and neighborhoods, which may contribute to housing and food insecurity [62]. Collectively, all of the aforementioned risk factors can contribute to various forms of GI distress in cancer survivors [61].

Cancer risk behaviors, including lifestyle choices, smoking, and alcohol consumption, have well-documented associations with physical and psychological symptoms and HRQOL in cancer survivors [24,63,64]. However, limited research has explored the specific relationships between these risk behaviors and GI health in cancer survivorship. Our findings reveal that cancer risk behaviors play a significant role in GI health conditions. Although previous research has primarily focused on other aspects of survivorship such as HRQOL and psychological symptoms [24,63,64], our study highlights the need to consider GI-specific factors. The identification of risk behaviors associated with GI health provides actionable insights for survivor care. Of note, food security was a more significant feature of GI health compared to self-reported diet quality as measured by HEI. This discrepancy in feature importance for GI health could be due to several reasons. First, self-reported diet quality may not fully capture nutrient intake or align with actual dietary behaviors [65]. Some individuals may report high diet quality despite lacking essential nutrients [65]. Second, food security is prevalent among cancer survivors in the U.S. (from 4% to 83.6%) and directly influences nutrient intake, beyond broader social determinants such as poverty, and health literacy [66,67]. Furthermore, food insecurity induces stress, which can exacerbate the risk of GI diagnoses including GI cancers and GI disorders by promoting impaired gut mobility, immune responses, and barrier function [68,69]. Access to diverse healthy foods ensures essential nutrients and greater microbiome diversity, which are vital for overall well-being including GI health [68,69].

Clinical Implications. ML plays a crucial role in classifying or predicting GI health, particularly for socially vulnerable cancer survivors. ML models analyze data from cancer survivors to pinpoint those at greater risk of GI distress. Once identified, targeted interventions can address their unmet needs, whether through pharmacological or non-pharmacological approaches. Integrating ML algorithms into platforms like mobile apps or websites (such as MyChart) is a practical approach. Furthermore, users can access personalized insights about their GI health, receive recommendations, and make informed decisions based on ML-driven risk classifications. ML models can help to further tailor interventions for high-risk groups by considering their specific social needs and vulnerabilities. For example, routine assessment of accelerated aging in cancer survivors could be essential for overall well-being and GI health. Addressing smoking cessation, promoting healthy lifestyles (healthy diet and physical activity), and minimizing alcohol consumption directly could also impact GI health and serve as an anti-aging strategy. Lastly, routine screening for socio-economic needs may contribute to optimal GI health in cancer survivors. For example, oncologists or PCPs can refer to nutritional education or food assistance programs. Increasing multidisciplinary collaboration with social workers, nutritionists, and community resources is warranted not only for overall HRQOL but also for GI health.

Strengths and Limitations. The strengths of our study lie in the inclusion of a variety of input data, specifically inflammatory markers, TL, and SDOH features. Additionally, our focus on GI health—an unexplored area in cancer survivorship—along with the application of ML models, contributes to the development of powerful classification models for GI health that consider both biological and social mechanisms. The findings of our study also reflect the importance of biological age in GI health conditions, applicable to all adult cancer survivors, not just older adults. Furthermore, our ML model was validated using an independent test dataset. Our study has several limitations. First, NHANES is a cross-sectional survey, which may limit the predictability of our ML model. To enhance predictability, longitudinal studies with predictors and GI health conditions measured at different time points are needed. Second, the usefulness of inflammatory markers (WBC and CRP) for classifying GI health remains unclear in our study. One possible reason is that mean WBC and CRP levels fell within the clinically normal range in our samples. Third, findings regarding prediction performance should be interpreted with caution due to the overall small sample size of the test dataset. Further studies with larger sample sizes are warranted to prevent model overfitting. Fourth, using a single question to ask about GI health may have limitations in fully capturing the GI health conditions. Furthermore, the roles of SDOH on the relationships between TL and GI healthy is unknown. Lastly, cancer-related clinical characteristics, such as cancer stages, years since diagnosis, and types of treatments, were not available in our samples, although they are potential covariates for our ML models related to GI health.

5. Conclusions

Using an ML approach to develop and validate GI health classification (better vs. worse) models inclusive of TL and SDOH is feasible among cancer survivors. Overall, the RF- and GBM-generated models showed the best accuracy for GI health classification. This finding suggests the potential of ML to further develop a longitudinal prediction model for GI health. TL and poverty status were the most significant features used to classify GI health, and could be implemented to prevent and manage GI conditions in cancer survivors. We suggest including biological markers and SDOH in ML GI health models to optimize classification accuracy. Future longitudinal studies and external clinical validations are warranted to confirm our findings and improve model predictability.