Evaluation of daily eating patterns on overall diet quality using decision tree analyses

Abstract

Background

Preliminary evidence suggests that meal timing is associated with higher quality diets. Less is known about whether types of food consumed during specific eating episodes (i.e., day-level eating patterns) predict diet quality.

Objectives

We investigated the association between day-level eating patterns and diet quality.

Methods

Decision tree models were built using 24-h dietary recall data from the National Health and Nutrition Examination Survey 2015 and 2017 cycles in a cross-sectional study. Sixteen food groups and 12 eating episodes (e.g., breakfast, lunch) were included as input parameters. Diet quality was scored using the Healthy Eating Index-2020 and categorized as higher or lower quality diets based on the median score. Mean decrease in impurity (MDI) ± standard deviation determined the relative contribution that day-level eating patterns had on diet quality; higher values represented greater contributions.

Results

We analyzed 12,597 dietary recalls from 9347 United States adults who were aged 18 y and older with ≥1 complete recall. Meals (breakfast, lunch, dinner) and respective snacking episodes had the greatest variety of dietary groups that contributed to the Healthy Eating Index-2020 score. Any whole-grain intake at breakfast predicted a higher quality diet (MDI = 0.08 ± 0.00), followed by lower solid fat intake (<8.94 g; MDI = 0.07 ± 0.00) and any plant protein intake at dinner (MDI = 0.05 ± 0.00).

Conclusions

Day-level eating patterns were associated with diet quality, emphasizing the relevance of both food type and timing in relation to a high-quality diet. Future interventions should investigate the potential impact of targeting food type and timing to improve diet quality.

Introduction

Sixty percent of United States adults are diagnosed with a chronic condition that can be prevented or managed with dietary modifications (e.g., heart disease, type 2 diabetes) [1]. Expert dietary recommendations, such as the United States Dietary Guidelines for Americans, suggest consuming a high-quality diet as a means to prevent these noncommunicable diseases [1]. High-quality dietary patterns often consist of greater intake of nutrient-dense foods (i.e., fruits, vegetables, whole grains, dairy, protein, oils) and less intake of alcohol, sodium, added sugars, and saturated fat [1]. In the United States, the Dietary Guidelines are updated every 5 y and are used to develop a quantitative scoring metric, known as the Healthy Eating Index (HEI) [2]. The HEI is a quantitative measure that evaluates how well Americans meet these diet recommendations, with scores ranging from 0 to 100. A score of 100 represents that a diet has fully met the Dietary Guidelines and a lower score represents a lower quality diet [3]. A recent study on diet quality in United States adults has revealed that mean overall HEI scores fall between 57 and 61, indicating that an average United States adult diet does not meet current dietary recommendations [4].

Low nutrition literacy has been suggested to be a contributor to lower quality diets [[5], [6], [7], [8], [9], [10]]. Prior studies show an association between having more nutrition knowledge and consuming a higher quality diet [[5], [6], [7], [8], [9], [10]]. Yet, despite the accessibility of publicly available print and electronic resources, ∼80% of United States adults do not recognize MyPlate, the official symbol of the United States Dietary Guidelines [11,12]. Effective communication and its delivery are crucial in emphasizing the importance of consuming high-quality diets, including providing specific strategies on how to achieve these goals [13,14]. When investigating consumer preferences for messaging about legume recommendations—a food group encouraged in the United States Dietary Guidelines—researchers have discovered that statements with a time element and portion size recommendation are strongly preferred by participants [15]. This type of message framing is also applicable when delivering patient-centered behavioral interventions [16]. Behavior change techniques often incorporate established goal-setting principles and can be adapted for use with meal-based dietary interventions, especially those supported by digital technologies [15,17]. A greater understanding of the quantity and types of food consumed throughout the day is crucial for applying techniques that encourage higher quality diets, because it helps in selecting and specifying the target behaviors necessary to improve diet quality in nutrition literacy-tailored interventions [18,19].

One recent narrative review identified 4 studies that investigated associations between government diet recommendations and meal patterns, an approach that provides granular information about meal timing and foods consumed throughout the day [15]. All studies revealed that meal timing and/or frequency were associated with diet quality or food groups [[20], [21], [22], [23]]. Those who engaged in a grazing eating pattern, had irregular eating times, and skipped breakfast were likely to consume a lower quality diet [[20], [21], [22], [23]]. These findings were confirmed by 3 other studies that reported individuals who consumed breakfast or consumed dinner earlier in the day had higher diet quality scores [[24], [25], [26]]. Upon investigating associations between food items and diet quality, 3 studies identified that low-fat dairy, vegetables, fruits, grains, and/or fish consumed during breakfast or other main meals were commonly found in higher quality diets [[23], [24], [25]]. The totality of evidence suggests that both meal timing and foods can predict overall diet quality. There is limited understanding of how the combination of food items, portion size, and specific eating episodes (i.e., day-level eating patterns) is associated with diet quality [[23], [24], [25]]. These types of data can be used to inform dietary interventions that encourage the consumption of health-promoting foods in real time [27,28].

Nutrition messaging should also account for the economic and environmental barriers to consuming a high-quality diet [5]. Recent government initiatives have focused on the advancement of nutrition security in the United States to provide equitable and consistent nutritionally adequate diets and reduce the risk of diet-related diseases [29]. However, dependable access to nutrient-dense foods is closely connected with food security status [30,31]. Current conceptual models propose that the relationship between food security and health is mediated by nutrition security at the individual level [32,33]. The proposed relationship has been confirmed by studies that found that individuals who experience food insecurity have lower diet quality and adverse health outcomes [31,34,35]. There are several ongoing efforts to mitigate food and nutrition insecurity, although many require substantial time and resources from several organized subsystems to impart change [[36], [37], [38]]. There remains a need to develop timely diet interventions that simultaneously address nutrition and food security [39,40]. One immediate approach is to develop strategies that are tailored to the individual’s environment and eating preferences or practices [41,42]. However, it is unclear how diet quality is associated with current day-level eating practices in households with food insecurity.

The primary study objective was to examine if day-level eating patterns predict diet quality using decision tree modeling, a supervised modeling approach that predicts outcomes according to a sequence of decision rules [22]. We hypothesized that breakfast intake would have the biggest contribution to diet quality. We also investigated whether day-level eating patterns differed based on food security status.

Methods

Study population

Data from the NHANES 2015–2016 and 2017–2018 cycles were used to train a decision tree to investigate associations between day-level eating patterns and diet quality. Briefly, NHANES is an ongoing national cross-sectional survey that monitors the health and nutritional status of the United States community-dwelling population every year [43]. A stratified, multistage probability cluster sampling design was used to ensure that the participants in the survey were representative of a noninstitutionalized United States population. Informed consent was obtained for all participants. Study protocols were approved by the National Center for Health Statistics Research Ethics Review Board. Additional details about the NHANES procedures can be found on the Centers for Disease Control and Prevention website [43].

The initial training data set had 19,225 participants (Figure 1). We excluded participants if they were under 18 y (N = 7377), had no recalls (N = 1599), were pregnant or lactating (N = 175), and had implausible daily energy intake, defined as males with <1000 or >4000 kcal and females with <800 or >3500 kcal (N = 727) [44]. The final sample size comprised 9347 participants.

FIGURE 1.

Participant flowchart. Inclusion and exclusion criteria of NHANES 2015 and 2017 cycle participants. The corresponding number of dietary recalls and individual food items for the participants included in the analyses are provided in this figure.

Diet measurements for eating patterns

Diet data from NHANES What We Eat in America were used 1) to determine intake of food groups and dietary elements (i.e., dietary groups) during specific eating episodes and 2) to calculate diet quality. A trained diet interviewer conducted ≥1 24-h dietary recall for each participant [45]. During these recalls, participants reported the frequency and quantity of foods and beverages consumed during the previous 24-h period. Diet information was collected either in-person or through telephone interviews using the USDA Automated Multiple-Pass Method to improve the accuracy of reporting [46]. Before merging the 2 NHANES cycles, the tap water variable from cycle 2015 was renamed to ensure a common variable name with cycle 2017.

Eating episodes and dietary group intake (input; independent variables)

There were 12,597 unique recalls and 192,526 individual food items in the data set (Figure 1). Participants were asked to label when a food or beverage was consumed during the day in either English or Spanish language. Cuban Spanish was used to translate the Spanish eating occasions into English. We collapsed the labeled eating episodes into a 3-meal eating pattern (i.e., breakfast, lunch, dinner) and 3 broad domains: snack, extended meal, and beverage (Supplemental Table 1). We then anchored the 3 domains to the 3-meal pattern, based on when a participant reported that eating episode (e.g., snack reported between breakfast and lunch). A total of 12 eating episodes were included in the analyses: breakfast or brunch, extended breakfast, beverage after breakfast, snack after breakfast, lunch, extended lunch, beverage after lunch, snack after lunch, dinner, extended dinner, beverage after dinner, and snack after dinner. According to NHANES diet recall protocol, participants were given the opportunity to identify each eating episode without being provided with definitions for these categories.

Data sets from the 24-h diet recalls also yielded 1) the name of each food and beverage and 2) the quantity consumed from each USDA Food Pattern Group for each of these items. Sixteen dietary groups were examined in the analyses: whole fruits, total vegetables without legumes, whole grains, refined grains, dairy, fruit juice, seafood, plant protein (i.e., soy, nuts, legumes), meat, cured meat, poultry, eggs, oils, solid fats, added sugar, and alcohol. These 16 dietary groups were selected because of their relationship with chronic disease and for their contribution to the overall HEI score calculation [3]. The seafood and plant protein intake variables for each item were created by summing seafood intake high and low in omega-3 PUFAs and summing soy, nuts, and legumes intake, respectively.

To obtain total dietary group intake at each eating episode, we summed all the food items consumed during the eating episode for each of the 16 dietary groups. Dietary group intake was later dichotomized into “higher/lower” or “yes/no” based on the median intake consumed by the overall sample for each eating episode. If the median intake for a food group at an eating episode was zero, we dichotomized the intake as “yes” or “no” to reflect any or no consumption.

Diet quality (criterion; dependent variable)

The HEI-2020 was used to determine diet quality for each diet recall [3]. In this manuscript, we use “HEI-2020” despite its identical scoring to HEI-2015, to reflect the current Dietary Guidelines [47]. The HEI-2020 score was calculated using 13 dietary components that focus on the adequacy (total and whole fruits, total vegetables, greens and beans, whole grains, dairy, total protein foods, seafood and plant proteins, fatty acids) and moderation (refined grains, sodium, added sugars, % fatty acids of energy). A higher score represents greater adherence to the United States Dietary Guidelines (score range: 0–100). The macro provided by the National Cancer Institute was used to calculate HEI-2020 scores [48]. We used the simple scoring algorithm since we were investigating the association between diet quality with eating episodes and dietary group intake, rather than determining population mean intake or assessing intervention effectiveness [48]. After determining the median HEI-2020 score for the overall study sample, each recall was considered to report either a lower (HEI-2020 score < 50.47) or higher quality diet (HEI-2020 score ≥ 50.47). Diet quality data were dichotomized for 2 purposes: 1) to investigate the daily dietary patterns correlated with scores above the median among United States adults, and 2) to strengthen the clinical applicability of the findings. This approach allowed us to clearly distinguish the key features that contribute to higher quality diets. Notably, the median HEI-2020 score in our study sample was lower than the mean overall HEI scores reported for United States adults [4]. This discrepancy may be attributed to our application of exclusion criteria (e.g., pregnancy or lactation) and differences in data distribution.

Sociodemographic, anthropometric, food security measurements

Sociodemographic data were collected through the NHANES household interview on age, sex, race and Hispanic origin, education, and family poverty-to-income ratio [46]. We preserved the original categories of these variables as defined in the NHANES data sets. The race-ethnicity variable was derived from survey items inquiring about race (“What race do you consider [insert name] to be? Please select one or more.”) and Hispanic origin (“Do you consider [insert name] to be Hispanic, Latino, or of Spanish origin?”). For the race item, the main categories included American Indian or Alaskan Native, Asian, Black or African American, Native Hawaiian or Other Pacific Islander, White, and Other. Height and weight were measured using a stadiometer and digital weight scale, respectively. BMI (kg/m²) was calculated using weight (in kg) divided by height squared (in m).

Household food security status was collected using the 18-item United States Household Food Security Survey Module, assessing food security-related behaviors and experiences [49] Each affirmative response scored 1 point out of 18 questions; based on their score, participants were placed into 4 categories: full (score = 0), marginal (score = 1–2), low (score = 3–7), and very low food security (score = 8–18). If participants were in the full or marginal food security categories, they were considered to be food secure per USDA definitions [49]. If participants had low or very low food security, they were considered to be food insecure. At the time of our analysis, food security data were only available from the NHANES 2015–2016 cycle (N = 4678). Those without food security data were excluded from the exploratory analyses.

Statistical analysis

Statistical analyses were performed using SAS version 9.4 (SAS Institute) and Python 3.6 modules Pandas and Scikit-learn. Independent t-tests and Chi-square analyses were used to determine differences in sociodemographic, anthropometric, and diet data between food secure and insecure groups. Decision tree classifiers are commonly used to predict the category of a particular entry (i.e., qualitative responses). Decision tree model-based classifications were performed to predict which eating episodes and dietary groups were associated with higher quality diets in all participants, as well as those with and without food security. To model the relationship between USDA Food Pattern Groups and eating episodes for HEI-2020 assessment, we incorporated food groups that either contribute to the HEI-2020 score or are associated with health outcomes, such as alcohol. This approach not only trains the model to replicate the HEI-2020 calculations but also enables detailed analysis of interactions among specific dietary components (e.g., eggs, fruit juice) beyond the 13 standard HEI components. Additionally, eating episodes were integrated into the decision tree to add to the model’s ability to examine dietary behaviors predictive of diet quality. The decision tree used binary features for each food category—“0” represented that a person consumed less than the median consumption for that category, whereas a “1” indicated that the individual’s consumption of that category was above the median. The number of nodes (i.e., below or above the median for each dietary group at a given eating episode) and depth of the decision tree (i.e., number of decision splits before a prediction can be made) were tuned to prevent overfitting these models. Each model (overall, food secure, food insecure) was validated using the 10-fold cross-validation to ensure that all data points were tested in an iteration. In this approach, the data are divided into 10 equal groups. The model is trained on data from these 9 groups and tested on the remaining group. The process is repeated 10 times, with each group serving as the test set once to determine the model’s accuracy, true positive rate, and false positive rate. The true positive rate represents the proportion of the positive instances that were accurately detected by the model [True Positive/(True Positive + False Negative)], whereas the false positive rate represents the proportion of negative instances that were incorrectly classified as positive instances by the model [False Positive/(False Positive + True Negative)].

The optimal model was selected if the training and validation performance were ≥70% as consistent with previous nutrition literature [23,50] and/or had higher percentage parameters, and provided the most clinically interpretable results (i.e., shorter depth). Feature importance graphs were generated to show the relative contribution [i.e., mean decrease in impurity (MDI)] that eating episodes and dietary group intake may have on diet quality. Purity is the measure that determines if observations in a region are predominantly from the same class (i.e., HEI-2020 score category); a high impurity (i.e., low MDI value) indicates that observations are from multiple classes. A higher MDI value meant that the dietary group consumed during that eating episode was a stronger predictor of diet quality. The direction of this association was determined based on the HEI-2020 because these dietary groups were accounted for in the scoring metric. Additionally, the dietary groups with the 3 highest MDI values were considered to have higher contributions to diet quality than the other dietary groups. The 3 highest MDI values were selected because prior studies reported that most patients establish ≤3 simultaneous goals to increase one’s likelihood of success [51,52]. Each decision tree model was further tested using data from NHANES 2013–2014, which also contained food security data. We reran the food security decision tree models to assess the robustness of our exploratory results. The MDI values differed between the initial and reanalysis because each analysis used a different random seed, which sets the starting point for the random elements in the decision tree algorithm. Thus, using different seeds will lead to slight variations in the results. Supplemental Table 2 provides the participant characteristics for each NHANES cycle used in our analyses.

Results

Participant characteristics

The overall sample (N = 9347) had a mean age of 49.4 ± 18.4 y and a mean BMI of 29.6 ± 7.1 (Table 1). A quarter of the participants who responded to the food security survey were food insecure (N = 1165 of 4678 participants). The food insecure group had a lower mean age, a higher proportion of Hispanic and non-Hispanic Black participants, lower education level, lower socioeconomic status, and higher BMI compared with the food secure group (P all <0.0001). When investigating differences in dietary intake between these 2 groups, the food insecure group reported lower total energy intake and HEI-2020 scores but with recall 1 only (P all <0.0001) and not recall 2 (Table 1). The food insecure group also consumed fewer servings of many HEI-2020 dietary components during their first recall (P < 0.05), except for added sugar (16.6 ± 13.7 compared with 15.3 ± 13.3 teaspoon equivalents; P = 0.004). There were no statistically significant differences in diet for recall 2 between those with and without food security.

TABLE 1.

Participant descriptives for NHANES 2015 and 2017 cycles training data set¹

Characteristic	Overall (N = 9347)	Food security status
		Food secure (N = 3513)	Food insecure (N = 1165)
Age2 (y)	49.4 ± 18.4	49.6 ± 18.5	45.7 ± 17.7
Sex, female	4639 (50)	1743 (50)	595 (51)
Race, ethnicity2
Mexican American	1461 (16)	531 (15)	289 (25)
Other Hispanic	1039 (11)	402 (11)	197 (17)
Non-Hispanic White	3274 (35)	1318 (38)	285 (24)
Non-Hispanic Black	2040 (22)	669 (20)	289 (25)
Other race, including multi-racial	1533 (16)	563 (16)	105 (9)
Education level2
Less than 12th grade	1533 (21)	606 (17)	427 (37)
High school graduate or equivalent	2265 (24)	803 (23)	291 (25)
Some college or associate degree	2890 (31)	1044 (30)	334 (29)
4-y degree or higher	2255 (24)	1059 (30)	110 (9)
Family poverty-to-income ratio2
≤130%	2495 (30)	722 (22)	633 (58)
131%–349%	3389 (40)	1351 (41)	400 (37)
≥350%	2487 (30)	1197 (37)	58 (5)
BMI2 (kg/m2)	29.6 ± 7.1	29.1 ± 6.9	30.5 ± 7.2
Total energy (kcal)
Recall 12	2032.4 ± 709.5	2051.2 ± 697.1	1971.2 ± 701.6
Recall 2	2218.3 ± 793.0	2095.8 ± 719.7	2203.5 ± 785.1
Total Healthy Eating Index-2020 Score
Recall 12	50.4 ± 14.1	51.3 ± 14.1	48.5 ± 13.7
Recall 2	48.4 ± 14.6	48.6 ± 15.3	48.9 ± 16.9
Diet quality category3
Higher quality diet, recall 1	4307 (48)	1723 (51)	477 (43)
Higher quality diet, recall 2	152 (41)	48 (43)	24 (43)
Daily dietary group intake
Recall 1
Whole fruits2 (cup equivalent)	0.7 ± 1.1	0.7 ± 1.1	0.6 ± 1.0
Total vegetables2 (cup equivalent)	1.5 ± 1.2	1.5 ± 1.2	1.3 ± 1.2
Whole grains2 (ounce equivalent)	0.8 ± 1.4	0.9 ± 1.4	0.7 ± 1.3
Refined grains (ounce equivalent)	5.6 ± 3.6	5.6 ± 3.5	5.8 ± 3.8
Dairy (cup equivalent)	1.3 ± 1.2	1.4 ± 1.3	1.3 ± 1.2
Fruit juice (cup equivalent)	0.2 ± 0.6	0.3 ± 0.6	0.3 ± 0.7
Seafood2 (ounce equivalent)	0.7 ± 2.2	0.7 ± 2.2	0.5 ± 2.1
Plant protein2 (ounce equivalent)	1.3 ± 2.3	1.3 ± 2.3	1.2 ± 2.2
Meat (ounce equivalent)	1.6 ± 2.4	1.6 ± 2.5	1.6 ± 2.5
Cured meat (ounce equivalent)	0.9 ± 1.6	0.9 ± 1.6	0.8 ± 1.6
Poultry (ounce equivalent)	1.6 ± 2.7	1.6 ± 2.6	1.7 ± 2.9
Eggs (ounce equivalent)	0.6 ± 1.0	0.6 ± 0.9	0.6 ± 0.9
Oils2 (g)	27.6 ± 19.5	27.4 ± 19.6	24.7 ± 18.2
Solid fats2 (g)	33.1 ± 22.2	33.3 ± 21.9	31.1 ± 21.0
Added sugar2 (tsp3 equivalents)	15.8 ± 13.5	15.3 ± 13.3	16.6 ± 13.7
Alcohol (g)	8.2 ± 22.5	8.6 ± 22.0	7.7 ± 22.7
Recall 2
Whole fruits (cup equivalent)	0.7 ± 1.5	0.6 ± 1.1	1.1 ± 2.2
Total vegetables (cup equivalent)	1.4 ± 1.3	1.3 ± 1.3	1.3 ± 1.3
Whole grains (ounce equivalent)	0.8 ± 1.5	0.9 ± 1.6	1.1 ± 1.9
Refined grains (ounce equivalent)	6.1 ± 4.0	5.9 ± 3.9	6.4 ± 3.7
Dairy (cup equivalent)	1.4 ± 1.4	1.3 ± 1.4	1.5 ± 1.5
Fruit juice (cup equivalent)	0.3 ± 0.7	0.3 ± 0.8	0.3 ± 0.7
Seafood (ounce equivalent)	0.9 ± 2.8	0.8 ± 2.9	0.8 ± 2.2
Plant protein (ounce equivalent)	1.3 ± 2.6	1.4 ± 2.9	1.4 ± 2.5
Meat (ounce equivalent)	1.9 ± 2.7	1.4 ± 2.6	1.9 ± 2.7
Cured meat (ounce equivalent)	1.2 ± 1.9	1.2 ± 1.8	1.0 ± 1.6
Poultry (ounce equivalent)	1.8 ± 3.0	2.1 ± 3.5	1.5 ± 2.7
Eggs (ounce equivalent)	0.8 ± 1.2	0.6 ± 0.9	1.0 ± 1.4
Oils (g)	27.3 ± 20.0	26.7 ± 18.5	25.7 ± 16.4
Solid fats (g)	35.7 ± 24.4	30.8 ± 19.5	37.1 ± 28.5
Added sugar (tsp equivalents)	17.6 ± 15.5	16.3 ± 15.1	16.5 ± 14.9
Alcohol (g)	8.6 ± 26.1	9.6 ± 25.6	6.8 ± 17.7

Abbreviation: tsp, teaspoon.

¹Data presented as mean ± SD or N (% of sample). Percentages may not sum to zero due to rounding.

²P < 0.05 between food secure and insecure groups. Descriptive differences between food security status were analyzed using independent t-tests or Chi-square analyses.

³Higher quality diet was defined as a Healthy Eating Index-2020 score ≥50.47.

Performance of decision tree model to predict diet quality

The different parameters of the decision tree model were compared in the main analysis, progressing from a depth of 5 up to a depth of 30 in 5-unit increments. A depth of 10 was determined to be optimal because the training, validation, and testing parameters were >70% and had a shorter depth (Table 2). Although depths of 5 and 15 also met these criteria, the training performance at a depth of 5 was substantially higher than at a depth of 15 where the model began to overfit (80% compared with 71%). The optimal parameters for the decision tree at a depth of 10 included a training percentage of 80%, a validation percentage of 72%, and a testing percentage of 71%—values that were consistent with other nutrition-related studies [23,50]. The true positive rate was 75% for the training set and 67% for the validation set, whereas the false positive rate was 15% for the training set and 23% for the validation set.

TABLE 2.

Model parameters for decision tree to determine optimal models

Depth	Train (%)	Validate (%)	Test (%)
Overall analysis (N = 9347)
5	71.3	71.0	70.1
10	80.0	71.8	70.8
15	90.2	70.2	70.8
20	97.1	68.3	69.6
25	99.2	68.4	67.8
30	99.8	68.5	68.3
Food secure analysis (N = 3513)
5	71.1	68.2	70.2
10	83.2	70.5	69.2
15	95.0	67.7	68.3
20	99.7	66.8	68
25	99.9	66.8	67.1
30	100.0	67.3	67.1
Food insecure analysis (N = 1165)
5	74.2	68.6	71.9
10	88.6	65.5	71.4
15	97.5	67.3	66.7
20	99.5	65.9	66.3
25	100.0	68.2	65.5
30	100.0	66.4	66.2

Decision tree models were implemented using Python 3.6 with Pandas and Scikit-learn modules.

A depth of 10 was also selected for decision trees for those with and without food security because these models had parameters that met the >70% threshold and/or had a higher training percentage than other decision trees (Table 2). Therefore, the decision tree for the food secure group had a training percentage of 83%, validation of 71%, and testing of 69%. The true positive rate for the food secure decision tree was 79% for the training set and 65% for the validation set; the false positive rate was 13% for the training set and 27% for the validation set. The parameters for the selected food insecure model included a training percentage of 89%, validation of 66%, and testing of 71%. The food insecurity model’s true positive rate was 81% for the training set and 53% for the validation set, whereas the false positive rate was 7% for the training set and 25% for the validation set.

Daily eating pattern predictions on diet quality

At least 69% of the examined dietary groups (range: 11–16 dietary groups) had an MDI value >0 during the following eating episodes for the overall sample: breakfast, lunch, dinner, and respective snacking episodes (i.e., after breakfast, lunch, and dinner), indicating a wide variety of foods contributing to diet quality (Table 3). The same pattern emerged for the food secure group, with more than half of the dietary groups predicting overall diet quality during the 3 meals and respective snacking episodes (range: 9–16 dietary groups). However, the food insecure decision tree revealed that only breakfast, lunch, dinner, and snacking after dinner had at least half of the dietary groups with an MDI value >0 (range: 10–16 dietary groups).

TABLE 3.

Number of dietary groups that predict diet quality for each eating episode

Eating episodes	Total number of dietary groups with MDI >0
	Overall	Food secure	Food insecure
Breakfast, brunch	14	12	12
Extended breakfast	0	2	1
Beverage after breakfast	5	2	2
Snack after breakfast	12	9	4
Lunch	16	16	15
Extended lunch	0	2	1
Beverage after lunch	3	3	2
Snack after lunch	13	12	5
Dinner	16	16	16
Extended dinner	1	2	5
Beverage after dinner	4	4	1
Snack after dinner	12	12	10

Abbreviation: MDI, mean decrease in impurity.

The consumption of whole grains—regardless of serving size—at breakfast had the highest MDI value compared with other eating episodes in the main analysis (MDI = 0.08 ± 0.00; Figure 2). Any intake of whole grain at breakfast (i.e., >0 ounce equivalent) predicted a higher quality diet. Solid fat and plant protein intake at dinner had the next 2 highest MDI values (MDI = 0.07 ± 0.00 and 0.05 ± 0.00, respectively). Specifically, low solid fat intake (i.e., <8.94 g; ∼1.3 tablespoons of butter) and any plant protein intake later in the day predicted higher quality diets. Both the initial and reanalysis decision trees revealed differences in dietary group intake predictions by food security status. In the initial analysis, refined grain intake (>0 ounce equivalent) at breakfast had the highest MDI value in the food secure group, followed by plant protein and higher solid fat intake at dinner (MDI >0.05). The reanalysis showed a similar pattern, with refined grain intake at breakfast and higher solid fat intake at dinner (MDI = 0.09 ± 0.00 and 0.07 ± 0.00, respectively). Unlike the initial analysis, higher vegetable intake at breakfast had the same MDI value as plant protein at dinner (MDI = 0.06 ± 0.00 for both). Figure 3 presents the results from the reanalysis. Initial analyses in the food insecure group found that lower refined grain intake at breakfast and lower solid fat intake at dinner predicted higher diet quality, followed by higher plant protein intake at lunch (MDI >0.04). In the reanalysis, lower solid fat intake at dinner continued to predict higher quality diets (MDI = 0.14 ± 0.01; Figure 3). The reanalysis also identified lower solid fat intake at lunch and higher whole-grain intake at breakfast as predictors of higher diet quality (MDI = 0.05 ± 0.00 for both).

FIGURE 2.

The relative contribution of day-level eating patterns (dietary group intake during eating episodes) on diet quality in the main analysis. The contribution of each examined dietary group during meals (blue colored bars) and snacks (red colored bars) on diet quality are presented using MDI values and SDs. Extended meals and beverage-only eating episodes are not shown as these eating episodes did not have ≥50% of the examined dietary groups with MDI values >0. Hyphens represent either extremely low or null MDI values. Decision tree models were implemented using Python 3.6 with Pandas and Scikit-learn modules. MDI, mean decrease in impurity.

FIGURE 3.

The relative contribution of day-level eating patterns (dietary group intake during eating episodes) on diet quality for those with and without food security. The contribution of each examined dietary group during meals and snacks on diet quality by food security status are presented using MDI values and SDs. The top bar graph presents results from the food secure group, whereas the bottom bar graph presents results from the food insecure group. Extended meals and beverage-only eating episodes are not shown as these eating episodes did not have ≥50% of the examined dietary groups with MDI values >0. Decision tree models were implemented using Python 3.6 with Pandas and Scikit-learn modules. MDI, mean decrease in impurity.

Discussion

The study investigated how eating episodes and types of foods eaten throughout the day predict diet quality. In the main analysis, the 3-meal pattern and snacking episodes contained a variety of foods that contributed to diet quality. The following eating pattern emerged as the greatest predictor of diet quality: any whole grains at breakfast with lower solid fat and any plant protein intake at dinner. Exploratory analyses revealed minor differences in day-level eating predictors of diet quality between those with and without food security.

Our hypothesis that breakfast would have the largest contribution to diet quality was partially supported by the finding that breakfast included a wide variety of foods contributing to diet quality and that whole-grain intake at breakfast was the greatest predictor of diet quality in the main analysis. The finding of whole-grain intake at breakfast was consistent with earlier nationally representative studies [25,53,54]. The timing of whole-grain intake might be driven, in part, by nutrition literacy and individual food choices [[55], [56], [57], [58]]. Consumers were often able to distinguish whole-grain breads and cereals and commonly consume these items earlier in the day [[55], [56], [57], [58]]. These results also aligned with previous reports that skipping breakfast contributed to a lower quality diet; by skipping breakfast, fewer servings of whole grains were consumed by participants [24,59,60]. Whole-grain foods accounted for ∼16% of total daily grain intake among United States adults despite the MyPlate recommendation, “Make half your grains whole grains” [61]. Although whole-grain intake has increased since 2003, fewer than 50% of households purchased whole-grain products, such as breads and cereals [55,62]. The mean whole-grain intake in this study (0.8-ounce equivalents/d) was equivalent to <1 slice of bread—an amount much less than the maximum score in the HEI-2020 Whole Grains scoring standard (i.e., 1.5-ounce equivalents per 1000 kcal of intake). Our study and others suggest that whole-grain intake—especially at breakfast—should continue to be encouraged during clinician-patient discussions [55,63].

The main analysis also revealed that dinner had the greatest variety of foods that predicted diet quality, including 2 major contributors: lower solid fat (i.e., sources of saturated, trans fatty acids) and higher plant protein intake (i.e., soy, nuts, legumes) [64]. Many Americans reported dinner to be their largest meal of the day, a meal characteristic that has been associated with greater diet variety [[65], [66], [67]]. These results agreed with evidence that types of dietary fat and grains contributed to diet quality [68]. The point distribution of the HEI scoring metric among its diet components may be partly responsible for these results [47]. Solid fat intake was included in previous HEI iterations under the “Empty Calories” component with added sugar and alcohol; however, the component was removed because of the addition of an “Added Sugars” component in subsequent HEI scoring metrics [3]. Although directly removed, solid fat was included in the HEI-2020 scoring metric as saturated fat in both the Fatty Acids and Saturated Fats components [47]. As a result, solid fat intake was represented twice in the overall HEI-2020 score, with each component having a maximum score of 10 to align with the Dietary Guidelines for Americans. A similar effect may have occurred with plant protein intake. The HEI-2020 scoring metric included legumes under 4 diet components, each with a maximum score of 5: total vegetables, greens and beans, total protein foods, and seafood/plant proteins [47]. This study only created a single plant protein group as an input for the decision tree models, which may explain why this group became a major predictor of diet quality.

The decision trees for food secure and insecure groups revealed grain intake to be one of the greatest predictors of diet quality, which provided additional support that the type of grain contributed to diet quality [68]. Differences in day-level eating patterns were observed between food secure and food insecure groups, agreeing with previous findings that food environments can influence diet quality and meal patterns [69,70]. The food insecure group had few eating episodes that contained a variety of dietary groups that contributed to overall diet quality. This finding may be attributed to those with food insecurity having limited access to nutrient-dense foods, snacking more often, and/or consuming higher kcal snack foods than those who are food secure [71,72]. Although results were mostly consistent between the initial and reanalysis for the food secure group, higher solid fat intake at dinner was the only consistent predictor for the food insecure group between the analyses. This result may be due to the smaller sample size used to train the food insecurity decision tree. Collectively, these preliminary findings suggest that tailored nutrition messaging that accounts for food security status should consider eating schedules and foods consumed during these episodes. Additional studies should investigate whether these day-level eating pattern differences persist among those with and without food security using other nationally representative samples.

Strengths of this study include the use of decision tree modeling with aggregated NHANES data sets. Decision trees in nutrition research are recommended as a data-driven approach because of their transparency and ability to provide high accuracy when predicting diet quality [23]. Data from NHANES provide a unique opportunity to develop these types of decision tree models because the multistage, probability sampling design is used to select adults who are representative of the noninstitutionalized United States adults. These data sets also provide granular diet information for a large participant sample to train, validate, and test the models. Despite the cross-sectional nature of NHANES, the risk of reverse causation is minimal because intakes of these dietary groups are collected throughout the day and most are used to generate HEI-2020 scores. The 24-h dietary recalls used in this study can introduce measurement bias because of the reliance on self-reported diet data, which lack objective confirmation. Furthermore, conclusions from the exploratory food security analyses are limited by the small participant numbers used in the analyses; only the 2013 and 2015 NHANES cycles had available food security data at the time of analysis, not the 2017 NHANES cycle. Thus, unaccounted temporal trends between the cycles may have influenced the observed differences in day-level eating patterns between the main analysis with the food secure/insecure decision trees. Lastly, diet quality was only assessed by the HEI scoring metric. A "one-size-fits-all" approach to diet quality runs counter to increasing adoption of all types of healthy diet patterns, especially as other diet quality scores also reduce risk of all-cause mortality and noncommunicable disease [73,74]. Further investigation of associations between day-level eating patterns and other diet quality definitions is warranted to communicate effectively.

Effective health messages often require specific strategies on how to achieve behavior goals, such as providing a time element during goal setting [13,14]. This study revealed that any grain intake at breakfast, in addition to low solid fat and any plant protein intake at dinner, is a major predictor of higher quality diets. The study adds to the current literature investigating the impact of specific dietary groups on overall diet quality, while also providing fundamental knowledge into how meal timing and portion size influence dietary intake. The totality of evidence can be used to inform the development of specific recommendations to improve diet quality during patient-clinician consultations, as well as public health messaging. The study also found preliminary evidence that the impact of meal timing may differ by food security status, lending support to the idea that these public health messages may need to be tailored to the food environment.

Author contributions

The authors’ responsibilities were as follows – AWL, JM, FK, SS: conceptualized the project; AWL, CC, HK, SS: contributed to data management and conducted formal analysis; AWL, CC: wrote the first draft of the manuscript; and all authors: reviewed, commented, and approved subsequent drafts of the manuscript.

Funding

AWL is partially supported by grant NIDDK R01DK129843 although the source was not involved with the publication. FK was supported by a grant from the National Institutes of Health/National Cancer Institute during the manuscript preparation (T32CA193193; PI: B. Spring).

Data availability

Data described in the manuscript, code book, and analytic code will be made available upon request pending application and approval.

Declaration of Generative AI and AI-assisted technologies in the revision process

The initial submission was written without assistance from generative AI technology. For the resubmission, the authors used ChatGPT 4o to review the grammar of the revised manuscript. After using this tool, the authors conducted a final review and edited the content. The authors take full responsibility for the content of the publication.

Conflict of interest

The authors report no conflicts of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: