From single conventional regression to ensemble modelling: relative importance of the Healthy Eating Index-2015 components in relation to adverse pregnancy outcomes Authors: Julie M Petersen, Lisa M Bodnar, Ashley I Naimi, Sharon I Kirkpatrick Published: 2026 ABSTRACT AbstractThe Healthy Eating Index (HEI) is widely used to assess diet quality, but certain contexts (e.g. pregnancy) may benefit from tailored versions. We evaluated whether the HEI’s current approach of assigning approximately equal weights to all components to compute the total score is appropriate when studying diet quality around conception. Data were from a US prospective cohort of individuals who had not delivered a previous pregnancy past 20 weeks’ gestation (2010–2013, n 7882). Usual dietary intake around conception was estimated from FFQ. Select adverse pregnancy outcomes (gestational diabetes, pre-eclampsia, preterm delivery and small-for-gestational age birth) were abstracted from the medical record. We regressed each outcome on the thirteen HEI-2015 component scores using SuperLearner, an ensemble machine learning method that combines predictions from multiple algorithms and avoids relying on parametric assumptions that characterise standard regression. We assessed the relative importance of each component using two permutation-based metrics: change in negative log likelihood (global influence) and absolute difference in the predicted probabilities (individual-level influence). Six of the thirteen components (Greens and Beans, Saturated Fats, Total Protein Foods, Seafood and Plant Proteins, Fatty Acids and Added Sugars) were important according to at least one metric for at least two of the four outcomes. In contrast, the Refined Grains component was not appreciably important for any outcome. These findings suggest that equal weighting of the HEI components may not be appropriate when evaluating diet quality for studies of pregnancy.Keywords: Diet quality, Healthy Eating Index, Machine learning, Pregnancy, Weighting scheme The Healthy Eating Index (HEI) measures diet quality based on standards of the Dietary Guidelines for Americans (DGA)(1), applicable to all individuals aged ≥ 2 years. The HEI was developed for surveillance into which aspects of diet need improvement and for research on associations between diet quality and health outcomes(2–4). Overall diet quality is represented by the HEI total score, which is the sum of individual component scores – each corresponding to a food group or nutrient from the US Department of Agriculture (USDA) food patterns(5). Slight variation in the composition of the components has occurred across the versions of the HEI. However, consistently most, if not all, of the food groups are equally weighted in the HEI total score(6). Equal weighting is used because it reflects the directive of the dietary guidelines to consider all recommendations as a whole(7). The HEI can be readily applied across a variety of population groups(2). The HEI is commonly used to evaluate associations between diet quality (before and during pregnancy) and pregnancy outcomes(8–10). According to recent research, the average pregnant person in the USA has poor diet quality based on the HEI(2,6,11,12). It has been questioned whether the HEI is truly appropriate for all settings or if certain situations call for more tailored indexes. Some components may be more important than the others for specific conditions or life stages, such as pregnancy(13). Research that suggests differential contributions of each component could inform refinements to the HEI scoring algorithm. Alternative, non-equal weighting schemes have demonstrated improvements to predicting all-cause mortality, CVD and obesity(14,15). However, limited research has considered the individual HEI components in addition to, or in lieu of, total score in relation to pregnancy outcomes(16–23), and even less research has explored the weighting scheme for pregnancy(20,21). In a prior study, we evaluated associations between the HEI-2010 component scores and pregnancy outcomes using generalised linear models(21). The Total Vegetables and Greens and Beans components, but none of the others, were associated with pre-eclampsia, preterm delivery and small-for-gestational age (SGA) birth, and no components were important for gestational diabetes(21). Given how few components were identified, we were curious whether a more flexible modelling approach would yield different results. Ensemble machine learning draws from a variety of algorithms to optimise predictions(24–28) and may overcome some challenges common to dietary data, such as non-linear associations as well as correlations and unspecified interactions among the predictors(29–31). Machine learning has been successfully applied to identify women who are meeting dietary guidelines in pregnancy(32). However, to our knowledge, only one other study has applied machine learning to assess the appropriateness of the HEI’s equal weighting for pregnancy, and that study focused on gestational length only(20). Thus, in the present work, we applied ensemble learning to evaluate the relative importance of the HEI-2015 component scores when studying periconceptional diet quality in relation to four adverse outcomes of pregnancy. Methods Data were obtained from the Nulliparous Pregnancy Outcomes Study: monitoring mothers-to-be (nuMoM2b), a multi-centre prospective cohort study that has been previously described(33). In brief, 10 038 pregnant individuals were recruited in the first trimester (6 0/7 to 13 6/7 weeks’ gestation) from eight US medical centres (2010–2013). The pregnancy needed to be the participant’s first (or no prior pregnancies beyond 20 0/7 weeks’) and an ultrasound-confirmed viable (high likelihood of survival) singleton (pregnant with only one baby). Participants attended three study visits during pregnancy, at which time a wide array of information was collected, including an FFQ, other standardised questionnaires and psychological batteries. Chart abstractions ascertained medical and reproductive history and pregnancy and birth outcomes. During the first trimester, participants completed a modified Block 2005 FFQ in English or Spanish(34,35). Modifications for nuMoM2b included alteration of the time of interest to reflect the 3 months around conception and additional questions to improve fat and carbohydrate estimates. The FFQ assessed usual frequency and portion size of approximately 120 foods and beverages. The food list was developed based on dietary recall data from the National Health and Nutrition Examination Survey 1999–2002(36,37). Visual aids illustrating portion sizes were provided to enhance accuracy. The FFQ has been compared with other self-reported dietary assessments and shown to have acceptable validity for pregnant women (nutrient correlations of 0·4–0·6 compared with food records)(38). NutritionQuest estimated daily intakes of nutrients and food groups based on the USDA Food and Nutrient Database for Dietary Studies (version 1) and Food Patterns Equivalents Database, respectively(39,40). Intakes of food groups and nutrients were used to compute the thirteen HEI-2015 component scores(6). The standards for scoring each component are summarised in online Supplementary Table 1. Nine are adequacy components: Total Fruit, Whole Fruit, Total Vegetables, Greens and Beans, Total Protein Foods, Seafood and Plant Proteins, Whole Grains, Dairy and Fatty Acids; and four are moderation components: Refined Grains, Sodium, Added Sugars and Saturated Fats. Component scoring is density-based to allow for assessment of quality independent of total energy(2,41). For most components, daily intakes of relevant foods are converted to standard units (i.e. cup equivalents, ounce equivalents or grams), which are then divided by daily total energy intake to yield amount equivalents per 1000 kcal. For Fatty Acids, the score is based on the ratio of intakes of PUFA and MUFA to intake of SFA. For Added Sugars and Saturated Fats, scores are based on their proportional contributions to total energy. Each component score has a minimum of 0 and a maximum of 10, unless the set of foods is represented by two components (e.g. Total Fruit and Whole Fruit), in which case each of the two components has a maximum of 5. The standards for the maximum score, except Sodium, correspond to the least restrictive ones of the USDA Healthy US-Style Food Patterns(5,42). Values between the minimum and maximum are proportionately scored. Moderation components are reverse-scored; higher scores indicate greater DGA adherence. Outcomes were abstracted from medical records and included gestational diabetes, pre-eclampsia, preterm delivery and SGA birth. These outcomes were selected because they have been associated with poor periconceptional diet quality(9,10,43). Gestational diabetes mellitus was based on standard clinical glucose tolerance testing (GTT), which typically occurs between 24 and 28 weeks’ gestation, defined as 3-h 100-g GTT with at least two of the following: fasting ≥ 95 mg/dl, 1-h ≥ 180 mg/dl, 2-h ≥ 155 mg/dl, 3-h ≥ 140 mg/dl or 2-h 75-g GTT with one of the following: fasting ≥ 92 mg/dl, 1-h ≥ 180 mg/dl or 2-h ≥ 153 mg/dl or 50-g GTT with a 1-h value ≥ 200 mg/dl, if fasting 3-or 2-h GTT was not conducted(44). Pre-eclampsia was based on the American College of Obstetricians and Gynecologists diagnostic criteria, adapted for nuMoM2b(44,45). The gestational age algorithm was based on local standard clinical practice(33), with preterm delivery defined as a live birth or stillbirth before 37 0/7 weeks’ gestation. SGA birth was defined as birth weight < 10th percentile based on fetal weight standards for gestational age(46,47). Statistical analyses For each pregnancy outcome, we computed the median and interquartile range (IQR) of the HEI total score and each component score among participants with the outcome and those without the outcome. When comparing these groups, we applied an effect size of 0·5 (half of the standard deviation of the sample) and therefore considered ≥ 5·5, ≥ 0·6 and ≥ 1·3-point differences in the medians of the HEI total score and components with maximum scores of 5 and 10, respectively, to be meaningful differences between those with v. without a given outcome(2). To evaluate whether the individual HEI-2015 component scores were differentially related to the pregnancy outcomes, we applied SuperLearner (or stacking). We regressed each pregnancy outcome on the thirteen HEI-2015 component scores (R sl3 package), resulting in four ensemble learners. Each SuperLearner was built using a prespecified library of candidate learners, including: (1) random forests (ranger) with 250, 500 and 1000 trees, minimum of 2, 3 and 4 predictors selected at random for each split, and minimum terminal node sizes of 25, 50 and 100; (2) extreme gradient boosting (xgboost) with maximum tree depth of 2, 4, 6 or 8 and shrinkage parameter of 0·3; (3) elastic-net regularised generalised linear models (glmnet) with mixing parameter α = 0·0 (ridge penalty) and 0·25, 0·50, 0·75 or 1·0 (Lasso penalty); (4) multivariate adaptive regression splines (earth) with backward pruning and three knots; (5) generalised linear regression and (6) simple mean. Using a library of diverse candidate algorithms enables SuperLearner to exploit the strengths of each to create a cross-validated ensemble of learners. For instance, tree-based learning algorithms (e.g. random forests and gradient boosting) and multivariate adaptive regression splines are non-parametric, meaning that they do not rely on strict assumptions regarding the distribution of the data and thus are more flexible than conventional approaches like generalised linear regression. Tree-based learners can more easily account for complex interactions and non-linearities without a priori specification. Elastic-net regularisation shrinks less important model coefficients towards zero to optimise prediction accuracy while avoiding overfitting and is better suited to dealing with multicollinear design spaces(48). Lastly, generalised linear models and simple mean are beneficial in that they are simple and not prone to overfitting. Since scaling is recommended for some of the candidate learners, we standardised the HEI-2015 component scores by subtracting the mean and dividing by the standard deviation. Each ensemble learner was fit using 10-fold cross-validation. Our primary goal was to use the SuperLearner to assess the relative importance of each HEI component. As diagnostic measures, we also computed accuracy statistics, including the receiver operating curve and AUC, accuracy (number of correct predictions/total number of predictions), precision (i.e. positive predictive value; number of correctly classified condition positive/total number of predictions classified as condition positive), recall (i.e. sensitivity; number of correctly classified condition positive/number truly condition positive), negative predictive value (number of correctly classified condition negative/total number of predictions classified as condition negative) and specificity (number of correctly classified condition negative/number truly condition negative). To evaluate the model’s classification performance for imbalanced datasets (not a 50/50 split among those with v. without the outcome), we also computed the area under the precision-recall curve (AUC-PR), balanced accuracy (recall + specificity/2), Matthews correlation coefficient (MCC) and the F1-score. With respect to interpretation, an AUC or AUC-PR value of 0·7–0·8 is considered acceptable, 0·8–0·9 is excellent and ≥ 0·9 is outstanding; an MCC of 0·3–0·5 is considered moderate and ≥ 0·5 is considered strong; and an F1-score of ≥ 0·7 is considered good, whereas a lower F1-score suggests the model is failing with respect to precision and/or recall. For the statistics that required a cut-off to classify an observation as condition positive based on their predicted probability, their predicted probability would need to have a value greater than or equal to the overall prevalence of that outcome in the original dataset. In addition, we reported the weights of each candidate algorithm in each SuperLearner model. To evaluate the relative contribution of each HEI-2025 component, we estimated two statistical metrics referred to as ‘variable importance’. The first variable importance metric is the change in the negative log likelihood (CNLL), a global measure of the influence of each HEI component on the overall SuperLearner model fit when a given component score was randomly permutated (keeping all other component scores fixed). A relatively larger CNLL indicates greater ‘importance’ of that component in terms of model fit. A value just above or below 0 indicates the component is unimportant. The second variable importance metric assessed the relative contribution of each component to the predicted probability of the outcome at the individual level. Specifically, it reflected the mean absolute difference in the predicted probabilities (ADPP) of the outcome based on the SuperLearner model before v. after randomly permuting the scores for a given component, while keeping all other component scores fixed. This metric is comparable to the absolute mean of Shapley values for each component(49,50). With respect to interpretation, it should be noted that ‘important’ does not imply ‘healthy’ or ‘unhealthy’, as we were not evaluating the directionality of associations between each HEI component and each outcome, but rather their relative contribution at the global and individual level. We did not conduct formal sample size calculations as this study was a secondary analysis of data from a large observational cohort. For reproducibility, we set the seed for the main analysis. In a post hoc analysis to evaluate the robustness of our findings to the original seed selection, we re-analysed the variable importance results for each outcome under twenty different seeds and summarised their distributions across the iterations. Results We excluded participants without FFQ data (n 1784), whose pregnancies ended before 20 0/7 weeks’ gestation due to miscarriage or termination, or who were otherwise missing information to define outcomes of interest (n 372). The final analytic sample included 7882 pregnancies. Online Supplementary Figure 1 describes the participant flow, and online Supplementary Table 2 describes the distribution of characteristics among participants within the analytic sample v. those excluded. Within the analytic sample, the majority were 25–34 years old, identified as non-Hispanic White, had a college education, had a BMI of 18·5–24·9 kg/m2, had private insurance coverage and planned the pregnancy. Excluded participants tended to be younger, less likely to be married, identify as a racial/ethnic minority, have lower educational attainment and less likely to have private insurance; there were not major differences based on BMI or pre-existing health conditions. Gestational diabetes occurred in 5 % of pregnancies, pre-eclampsia in 9 %, preterm delivery in 8 % and SGA birth in 12 %. The median HEI-2015 total score was 65·4 (IQR 56·7–73·3). The HEI-2015 total and component scores tended to be lower among participants with an adverse outcome compared with participants without them. However, the only meaningful differences based on an effect size of ≥ 0·5 occurred for Greens and Beans among participants with pre-eclampsia, preterm delivery or SGA birth compared with those without these adverse outcomes (Table 1). Table 1.Distribution (median and interquartile range) of the Healthy Eating Index-2015 component scores by pregnancy outcome (Nulliparous Pregnancy Outcome Study: monitoring mothers-to-be, n 7882) Table 1 long description. Gestational diabetesPre-eclampsiaPreterm deliverySGA birthPresent (n 394)Absent (n 7488)Present (n 681)Absent (n 7201)Present (n 633)Absent (n 7249)Present (n 915)Absent (n 6967)MedianIQRMedianIQRMedianIQRMedianIQRMedianIQRMedianIQRMedianIQRMedianIQRTotal score64·456·2–72·565·456·7–73·362·554·1–71·265·657·0–73·463·354·4–71·865·556·9–73·463·354·4–72·265·657·1–73·4Adequacy components Total fruit3·92·5–5·04·22·5–5·03·92·4–5·04·32·5–5·03·92·4–5·04·32·5–5·04·12·3–5·04·32·5–5·0 Whole fruit4·92·7–5·05·02·8–5·04·62·5–5·05·02·8–5·04·92·7–5·05·02·8–5·04·72·5–5·05·02·8–5·0 Total vegetables4·02·8–5·04·02·8–5·03·72·6–5·04·12·9–5·03·72·6–5·04·12·9–5·03·72·6–5·04·12·9–5·0 Greens and beans3·92·2–5·03·92·2–5·03·31·8–5·04·02·2–5·03·41·7–5·04·02·2–5·03·41·8–5·04·02·2–5·0 Whole grains2·91·5–4·93·11·5–5·22·81·2–4·73·11·5–5·22·91·3–4·93·11·5–5·22·71·2–5·13·21·6–5·2 Dairy products6·74·8–8·86·64·9–8·96·74·8–8·66·64·9–8·96·64·8–8·86·64·9–8·96·34·5–8·86·75·0–8·9 Total protein foods5·04·1–5·04·83·8–5·04·83·8–5·04·83·8–5·04·93·9–5·04·83·8–5·04·73·8–5·04·83·8–5·0 Seafood and plant proteins4·32·6–5·04·62·8–5·04·22·3–5·04·72·8–5·04·32·5–5· 04·72·8–5·04·52·5–5·04·72·8–5·0 Fatty acids5·23·5–7·55·23·4–7·55·03·3–7·05·23·5–7·55·13·3–7·45·23·5–7·55·13·7–7·45·23·4–7·5Moderation components Refined grains8·76·9–10·08·76·8–10·08·66·7–10·08·76·8–10·08·56·5–10·08·86·8–10·08·76·7–10·08·76·8–10·0 Na4·12·4–6·14·32·5–6·24·42·5–6·54·32·5–6·14·32·4–6·24·32·5–6·14·62·7–6·44·32·4–6·1 Added sugars8·26·1–9·68·26·1–9·67·95·8–9·58·26·2–9·67·85·6–9·48·36·2–9·67·85·6–9·48·36·2–9·6 Saturated fats6·04·1–7·86·34·3–8·25·83·9–7·96·34·3–8·36·23·9–8·36·34·3–8·26·24·1–8·36·34·3–8·2SGA, small-for-gestational age. Distribution (median and interquartile range) of the Healthy Eating Index-2015 component scores by pregnancy outcome (Nulliparous Pregnancy Outcome Study: monitoring mothers-to-be, n 7882) Table 1 long description. SGA, small-for-gestational age. The full set of accuracy performance metrics is provided in online Supplemental Table 3. Here, we focus on the metrics for imbalanced data. The AUC-PR curve values for the SuperLearner models were considered excellent for gestational diabetes and pre-eclampsia (≥ 0·82) and acceptable for preterm delivery and SGA birth (≥ 0·73). In contrast, the MCC values were moderate for preterm delivery and SGA birth (≥ 0·36) and poor for gestational diabetes and pre-eclampsia (≤ 0·26). All models had good recall (≥ 0·75) but poor precision (≤ 0·31), as reflected in their low F1-scores (≤ 0·47). The weight of each candidate algorithm in the SuperLearner varied based on the outcome (online Supplementary Table 4). Nearly all the weight (91 %) was given to elastic-net regression for gestational diabetes. The majority of the weight was given to generalised linear regression and simple mean for preterm delivery (82 %), pre-eclampsia (66 %) and SGA birth (51 %). The algorithms with the next highest weights were random forests for preterm delivery (18 %) and extreme gradient boosting for pre-eclampsia (26 %) and SGA birth (17 %). With respect to the variable importance results from the SuperLearner models (Figures 1–4, estimates listed in online Supplementary Table 5), six of the thirteen components were important according to at least one importance metric for at least two of the four outcomes. Specifically, the Greens and Beans and Saturated Fats components were important according to both CNLL and ADPP for pre-eclampsia, preterm delivery and SGA birth. The Total Protein Foods component was important for gestational diabetes and pre-eclampsia according to both metrics and for SGA birth according to the CNLL. According to both metrics, the Seafood and Plant Proteins component was important for gestational diabetes and pre-eclampsia, and the Fatty Acids component was important for pre-eclampsia and SGA birth. The Added Sugars component was considered important for preterm delivery according to both metrics and for pre-eclampsia according to the ADPP. Figure 1 long description. Variable importance of the thirteen component scores of the Healthy Eating Index-2015 for gestational diabetes outcomes. The components are listed in order according to importance based on the change in the negative log likelihood. The right axis and dots provide the values for the change in the negative log likelihood. The left axis and grey bars provide values for the absolute difference in the predicted probabilities (ADPP). A value (change in negative log likelihood or ADPP) for a given component that is relatively higher than those for the other components indicates greater importance of that component according to that metric. The actual variable importance values are provided in online Supplementary Table 5. Figure 4 long description. Variable importance of the thirteen component scores of the Healthy Eating Index-2015 for small-for-gestational age birth outcomes. The components are listed in order according to importance based on the change in the negative log likelihood. The right axis and dots provide the values for the change in the negative log likelihood. The left axis and grey bars provide values for the absolute difference in the predicted probabilities (ADPP). A value (change in negative log likelihood or ADPP) for a given component that is relatively higher than those for the other components indicates greater importance of that component according to that metric. The actual variable importance values are provided in online Supplementary Table 5. Other HEI components were important for only one outcome. Specifically, for preterm delivery, the Whole Fruit component was important according to both metrics, and the Sodium component was important according to the ADPP. For SGA birth, the Total Vegetables and Dairy components were important according to both metrics and the Whole Grains component was important according to the ADPP. Some components were more moderately important. For instance, for preterm delivery, the Total Vegetables, Fatty Acids and Total Fruit components were moderately appropriate according to both metrics and the Total Protein Foods and Seafood and Plant Proteins components were moderately important according to only one metric. The Refined Grains component was the only one that was not appreciably important for any outcome. In our robustness analyses, we observed some variation in the rank ordering of the components based on variable importance value when we varied the seed, notably for the CNLL (refer to online Supplementary Figures 2–9). Generally, our conclusions about which variables are relatively more important did not change from the main analysis. However, these additional post hoc analyses suggested that the variable importance findings should not be interpreted as strict rankings, but instead general patterns, of relative importance of the components. Also, it seemed clearer that the Total Protein Foods component, followed by the Seafood and Plant Proteins component, was identified as the most important ones for gestational diabetes; the Saturated Fats component was identified as the most important one for pre-eclampsia; the Added Sugars and Greens and Beans components were identified as the most important ones for preterm delivery; and the Total Dairy, Total Vegetables, Saturated Fats, Whole Grains, Fatty Acids and Greens and Beans components were identified as the most important ones for SGA birth. Discussion The HEI-2015 components representing diets rich in dark green vegetables and protein foods (including seafood and plant-based sources), high in healthy fats and low in saturated fats and low in added sugars were considered of greater relevance than components representing other aspects of diet quality for at least half of the pregnancy outcomes under study, often based on both variable importance metrics. Other components were considered important to a lesser degree or for only one outcome. In contrast, the Refined Grains component was not identified as important for any outcome. There is theoretical support that certain foods and nutrients may be particularly beneficial for pregnancy(9,10,51). Many vegetables, plant-based oils and nuts and seeds contain antioxidants that are necessary for pregnancy(52,53); certain fats may enhance their bioavailability(54,55). Diets before and during pregnancy with higher amounts of vegetables, fruits, whole grains, nuts, legumes, fish and vegetable oils have been associated with lower risks for gestational diabetes and hypertensive disorders of pregnancy (including pre-eclampsia)(9,56). Our analysis pointed to the importance of protein foods, including seafood and plant-based proteins, for gestational diabetes and diets rich in green leafy vegetables, beans and peas and low in saturated fats relative to healthier fats for pre-eclampsia. However, we did not observe associations with fruit or whole grains. Further, diets during pregnancy characterised by high intakes of vegetables, fruits, whole grains, low-fat dairy products and lean protein foods have been associated with lower risks for preterm birth and SGA birth, whereas diets characterised by high intakes of refined grains, processed meat and foods high in saturated fat or sugar were associated with higher risk of preterm birth(10,57). Similarly, we found that diets lower in added sugars and saturated fats and higher in green leafy vegetables, beans, peas and whole fruit were associated with preterm birth, and saturated fats, dairy products and vegetable intake were among the most important components for SGA birth. In our prior analysis, using generalised linear regression(21), we found that the Total Vegetables and Greens and Beans components, but none of the others, were associated with pre-eclampsia, preterm delivery and SGA birth, and no components were associated with gestational diabetes(21). In the current study, we again identified the Greens and Beans component as important for those three pregnancy outcomes, and the Total Vegetables component was important for SGA birth. However, several additional components were also identified as important. Differences between the prior study and our current work are most likely explained by the conventional v. SuperLearner modelling approaches. Correlations between dietary covariates could lead to generalised linear model convergence issues, which we encountered when modelling gestational diabetes in our prior, but not the current, study. Nearly all the contribution to the gestational diabetes SuperLearner was provided by elastic-net regression, which is better at handling co-linear variables, such as Total Protein Foods and Seafood and Plant Based Proteins. Further, even though the majority of the SuperLearner weight went to conventional models for pre-eclampsia, preterm delivery and SGA birth, the contribution of the more flexible machine learning algorithms may have enabled us to identify the additional components as important. The only other study that we are aware of that has applied supervised learning to interrogate the relative contributions of the HEI-2015 components for pregnancy evaluated the associations with length of gestation(20). The investigators used weighted quantile sum regression – a supervised statistical mixtures approach that is frequently used in environmental epidemiology and can be useful with highly correlated predictors(58). Both theirs and the current study supported the importance of Added Sugars and Greens and Beans in relation to gestational length. However, their study also suggested Total Protein Foods, Seafood and Plant Proteins and Dairy as among the most important(20), whereas our findings suggested these were more modestly important, and that Sodium, Saturated Fats and Whole Fruit may be more important. These discrepancies could be explained by differences in the designs and approaches. Their sample size was substantially smaller (n 421) than ours. Their dietary data approximated usual intake in the first trimester, whereas ours reflected the periconceptional period. Further, they modelled gestational length as continuous, whereas we dichotomised as preterm; their choice was partly driven by the rarity of preterm delivery (< 5 %) in their sample. Lastly, their study controlled for potential confounders, whereas we did not (beyond the other component scores). Strengths of our study include the sample size, rigorous outcome classification and a diverse set of machine learning algorithms. While the HEI-2020 is the latest version, there were no changes to the components or standards from 2015; the index was renamed to clarify that it aligns with the 2020–2025 DGA(1,59). Our study also has limitations. The variable importance metrics were computed using scales where it is challenging to pinpoint an absolute cut-off to define whether a given covariate is important (or not). While we looked at variability due to seed selection, we could not estimate traditional 95 % CI for variable importance. The FFQ data are impacted by measurement errors(60); bias attributed to systematic under- or over-reporting is thought to be lessened by the density- and ratio-based HEI scoring. The distributions of some baseline characteristics differed among the nuMoM2b participants included in our analysis compared with those who were excluded; it is unknown whether their dietary quality differed. These results are limited to this sample and should be replicated to confirm they are generalisable to other populations. Lastly, our associational findings do not indicate direction and should not be interpreted as causal, as we did not adjust for potential confounders. Conclusions Our research adds to growing evidence that the current weighting scheme of the HEI may not be appropriate when studying pregnancy. To date, it is unclear exactly how much weight certain components should receive relative to the others in the total score computation when safeguarding pregnancy health is the focus. The next step is to evaluate the HEI components in relation to pregnancy outcomes with a causal framework. If it becomes clear that a more limited set of dietary components are the primary nutritional contributors to pregnancy outcomes, a change to the operationalisation of dietary indexes (including the HEI), and more tailored dietary guidance, for pregnancy could follow. Supporting information Petersen et al. supplementary material Petersen et al. supplementary material Acknowledgements The authors thank Sara Parisi for her contribution to data management, data cleaning and derived variable creation. The following institutions and researchers compose the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-be (nuMoM2b) Network: Case Western Reserve University/Ohio State University – Brian M. Mercer, MD, Jay Iams, MD, Wendy Dalton, RN, Cheryl Latimer, RN, LuAnn Polito, RN, JD; Columbia University/Christiana Care – Matthew K. Hoffman, MD, MPH, Ronald Wapner, MD, Karin Fuchs, MD, Caroline Torres, MD, Stephanie Lynch, RN, BSN, CCRC, Ameneh Onativia, MD, Michelle DiVito, MSN, CCRC; Indiana University – David M. Haas, MD, MS, Tatiana Foroud, PhD, Emily Perkins, BS, MA, CCRP, Shannon Barnes, RN, MSN, Alicia Winters, BS, Catherine L. McCormick, RN; University of Pittsburgh – Hyagriv N. Simhan, MD, MSCR, Steve N. Caritis, MD, Melissa Bickus, RN, BS, Paul D. Speer, MD, Stephen P. Emery, MD, Ashi R. Daftary, MD; Northwestern University – William A. Grobman, MD, MBA, Alan M. Peaceman, MD, Peggy Campbell, RN, BSN, CCRC, Jessica S. Shepard, MPH, Crystal N. Williams, BA; University of California at Irvine – Deborah A. Wing, MD, Pathik D. Wadhwa, MD, PhD, Michael P. Nageotte, MD, Pamela J. Rumney, RNC, CCRC, Manuel Porto, MD, Valerie Pham, RDMS; University of Pennsylvania – Samuel Parry, MD, Jack Ludmir, MD, Michal Elovitz, MD, Mary Peters, BA, MPH, Brittany Araujo, BS; University of Utah – Robert M. Silver, M.D., M. Sean Esplin, MD, Kelly Vorwaller, RN, Julie Postma, RN, Valerie Morby, RN, Melanie Williams, RN, Linda Meadows, RN; RTI International – Corette B. Parker, DrPH, Matthew A. Koch, MD, PhD, Deborah W. McFadden, MBA, Barbara V. Alexander, MSPH, Venkat Yetukuri, MS, Shannon Hunter, MS, Tommy E. Holder, Jr, BS, Holly L. Franklin, MPH, Martha J. DeCain, BS, Christopher Griggs, BS; Eunice Kennedy Shriver National Institute of Child Health and Human Development – Uma M. Reddy, MD, MPH, Marian Willinger, PhD, Maurice Davis, DHA, MPA, MHSA; University of Texas Medical Branch at Galveston – George R. Saade, MD. This study was supported by grant funding from the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD): R01 HD102313 to Bodnar LM and Naimi AI as well as U10 HD063036 to RTI International, U10 HD063072 to Case Western Reserve University, U10 HD063047 to Columbia University, U10 HD063037 to Indiana University, U10 HD063041 to University of Pittsburgh, U10 HD063020 to Northwestern University, U10 HD063046 to University of California Irvine, U10 HD063048 to University of Pennsylvania and U10 HD063053 to University of Utah. Support was also provided by respective Clinical and Translational Science Institutes to Indiana University (UL1TR001108) and University of California Irvine (UL1TR000153). The sponsors had no role in the design, analysis or writing of this article and stipulated no restrictions regarding publication. L. M. B., A. I. N. and J. M. P. conceptualised the research. J. M. P. led the investigation, including the development of the methodology, performed the analysis, wrote the original draft of the paper, made revisions based on coauthor input and took primary responsibility for the final content. L. M. P. and A. I. N. acquired the financial support for the project and supervised the work, with A. I. N. specifically supervising the statistical analysis. L. M. P. and S. I. K. provided expertise in the perinatal and nutritional methodology and relevant resources. All authors reviewed and approved the final submitted version of the manuscript. The authors declare no conflicts of interest. The authors declare there was no use of generative AI and AI-assisted technologies in the writing process. The nuMoM2b data described in the manuscript, as well as the code book, are publicly and freely available without restriction via the Eunice Kennedy Shriver National Institute of Child Health and Human Development Data and Specimen Hub at https://dash.nichd.nih.gov/study/226675. The study investigators and the Data Coordinating and Analysis Center have copies of the entire database but cannot release that version to outside investigators due to the permissions granted by the participant during the consent process and participant confidentiality. The R code used in the analysis is available at: https://github.com/JulieMOPetersen/HEI_SuperLearner The main nuMoM2b study was conducted according to the guidelines laid down in the Declaration of Helsinki, and all procedures involving human subjects/patients were approved individually by each study site’s Institutional Review Board. Written informed consent was obtained from all participants. The University of Pittsburgh Institutional Review Board deemed this secondary analysis of the de-identified data to be exempt (protocol number STUDY19100034). Long description The table presents the distribution of Healthy Eating Index-2015 (HEI-2015) component scores across various pregnancy outcomes, including gestational diabetes, pre-eclampsia, preterm delivery, and small for gestational age (SGA) birth. The table has 15 rows and 12 columns, with each column representing a different pregnancy outcome category. The rows detail the total score and various adequacy and moderation components such as total fruit, whole fruit, total vegetables, greens and beans, whole grains, dairy products, total protein foods, seafood and plant proteins, fatty acids, refined grains, sodium, added sugars, and saturated fats. The median HEI-2015 total score is 65.4 with an interquartile range of 56.7 to 73.3. Notable trends include lower HEI-2015 scores among participants with adverse pregnancy outcomes compared to those without. Specifically, meaningful differences based on an effect size of 0.5 are observed for the Greens and Beans component among participants with pre-eclampsia, preterm delivery, or SGA birth compared to those without these adverse outcomes. The data highlights the variations in dietary quality as measured by the HEI-2015 across different pregnancy outcomes. Navigate back to Table 1.. Long description The bar graph compares the importance of thirteen component scores of the Healthy Eating Index-2015 for gestational diabetes outcomes. The x-axis lists the components: Sodium, Total Vegetables, Total Dairy, Added Sugars, Saturated Fats, Total Fruit, Whole Fruit, Fatty Acids, Greens and Beans, Refined Grains, Whole Grains, Seafood and Plant Proteins, and Total Protein Foods. The left y-axis measures the absolute difference in predicted probabilities, represented by grey bars, while the right y-axis measures the change in negative log likelihood, represented by black dots. The components are ordered by importance based on the change in negative log likelihood. Notable trends include higher values for Total Protein Foods and Seafood and Plant Proteins, indicating greater importance. All values are approximated. Navigate back to Figure 1.. Long description The bar graph compares the importance of thirteen component scores of the Healthy Eating Index-2015 for pre-eclampsia outcomes. The x-axis lists the components: Total Fruit, Whole Fruit, Sodium, Whole Grains, Added Sugars, Total Vegetables, Refined Grains, Total Dairy, Total Protein Foods, Fatty Acids, Seafood and Plant Proteins, Greens and Beans, and Saturated Fats. The left y-axis measures the absolute difference in predicted probabilities, represented by grey bars, while the right y-axis measures the change in negative log likelihood, represented by black dots. The components are ordered by importance based on the change in negative log likelihood. The graph highlights that Saturated FatsGreens and Beans, Seafood and Plant Proteins, Fatty Acids, Total Protein Foods, and Added Sugars had higher importance according to both or one metric. All values are approximated. Navigate back to Figure 2.. Figure 2 long description. Variable importance of the thirteen component scores of the Healthy Eating Index-2015 for pre-eclampsia outcomes. The components are listed in order according to importance based on the change in the negative log likelihood. The right axis and dots provide the values for the change in the negative log likelihood. The left axis and grey bars provide values for the absolute difference in the predicted probabilities (ADPP). A value (change in negative log likelihood or ADPP) for a given component that is relatively higher than those for the other components indicates greater importance of that component according to that metric. The actual variable importance values are provided in online Supplementary Table 5. Long description The bar graph compares the variable importance of the thirteen component scores of the Healthy Eating Index-2015 for preterm delivery outcomes. The x-axis lists the components: Whole Grains, Refined Grains, Total Protein Foods, Seafood and Plant Proteins, Total Vegetables, Fatty Acids, Dairy, Total Fruit, Sodium, Greens and Beans, Saturated Fats, Added Sugars, and Whole Fruit. The left y-axis measures the absolute difference in predicted probabilities, represented by grey bars, while the right y-axis measures the change in negative log likelihood, represented by black dots. The components are listed in order according to importance based on the change in the negative log likelihood. The graph highlights that Whole Fruit, Added Sugars, Saturated Fats, Greens and Beans, and Sodium are important according to both or one metric. All values are approximated. Navigate back to Figure 3.. Figure 3 long description. Variable importance of the thirteen component scores of the Healthy Eating Index-2015 for preterm delivery outcomes. The components are listed in order according to importance based on the change in the negative log likelihood. The right axis and dots provide the values for the change in the negative log likelihood. The left axis and grey bars provide values for the absolute difference in the predicted probabilities (ADPP). A value (change in negative log likelihood or ADPP) for a given component that is relatively higher than those for the other components indicates greater importance of that component according to that metric. The actual variable importance values are provided in online Supplementary Table 5. Long description The bar graph compares the importance of thirteen component scores of the Healthy Eating Index-2015 for small-for-gestational age birth outcomes. The x-axis lists the components: Whole Fruit, Total Fruit, Refined Grains, Added Sugars, Seafood and Plant Proteins, Whole Grains, Sodium, Total Protein Foods, Greens and Beans, Saturated Fats, Fatty Acids, Total Dairy, and Total Vegetables. The left y-axis measures the absolute difference in predicted probabilities, represented by grey bars, while the right y-axis measures the change in negative log likelihood, represented by black dots. The components are ordered by importance based on the change in the negative log likelihood. The graph highlights that Total Vegetables, Total Dairy, Fatty Acids, Saturated Fats, Greens and Beans, Total Protein Foods, and Whole Grains were important according to both or one metric. All values are approximated. Navigate back to Figure 4.. Supplementary material For supplementary material/s referred to in this article, please visit https://doi.org/10.1017/S0007114526107223 References Associated Data Supplementary Materials Petersen et al. supplementary material Petersen et al. supplementary material