This was a secondary analysis of longitudinal data from the Healthy Beginnings Trial, a home-based early childhood obesity prevention intervention involving first-time mothers and their children. Detailed information on the randomised controlled trial and results has been previously published [15,16,17]. Briefly, the trial consisted of two phases, a two-year intervention phase conducted from 2007 to 2010 and a three-year follow-up phase from 2011 to 2014. The intervention group received eight home visits from a trained early childhood nurse, consisting of one antenatal visit at 30–36 weeks gestation and seven postnatal visits starting from birth to 24 months [17, 18]. Nurses promoted healthy infant feeding, child and family nutrition and physical activity and social support; these visits were complemented with proactive telephone support. The control group received usual care.

Pregnant women who attended one of two metropolitan antenatal clinics in Western Sydney were invited to participate in the trial. Women were eligible to participate if they were at least 16 years old, were 24–34 weeks pregnant, they or their guardian could communicate in English, were local residents and could provide informed consent. Women were excluded if, after giving birth, their infant was diagnosed with a medical condition affecting physical activity, eating behaviours, weight or height/length. Ethics approval was obtained from the Sydney South West Area Health Service ethics review committee (RPAH Zone, Protocol No X10-0312 and HREC/10/RPAH/546) and the Human Ethics Advisory Group—Health at Deakin University (HEAG-H 194_2021). Informed consent was obtained from all participants. This study is reported according to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (Supplementary Table 1) [19].

Women’s dietary data were collected at baseline (24–34 weeks gestation) and at 1, 2 and 3.5 years postpartum. Maternal dietary intake was collected using a 15-item semi-quantitative food frequency questionnaire (FFQ) through face-to-face interviews with a trained nurse at all time points [19]. This questionnaire was based on dietary questions from the New South Wales Population Health Survey Australia [16, 21] and has been previously validated [22]. The FFQ collected information on frequency of intake of vegetables, fruit, bread, breakfast cereal, cooked grains/cereals, milk, processed meat, takeaway meals/snacks, potato products, sugar-sweetened beverages, fruit juice and water [16, 21]. Women were asked to answer how many times/serves per day, week or month they consumed the food and beverage items or whether they were rarely or not consumed (Supplementary Table 2).

Diet quality was calculated using modified versions of the DGI-2013 and the RDGI, which assess adherence to the 2013 Australian Dietary Guidelines [7, 10]. The DGI-2013 was chosen because it is a commonly used index with food-based components designed to score dietary intakes according to national recommendations [7] and can be applied to brief dietary assessment tools [8]. The RDGI was selected because the dietary questions from the questionnaire could be directly applied to the indicators of the index.

The DGI-2013 is a food-based score developed to assess adherence to age- and sex-specific recommendations of the 2013 Australian Dietary Guidelines [7] and has been shown to be a valid measure of diet quality in the Australian population [7, 9, 23, 24]. The original DGI-2013 comprises 13 components consisting of seven encouraged (i.e., diet variety, vegetables, fruit, grains/cereals, lean meat and alternatives, dairy and alternatives and fluid intake) and six discouraged (i.e., discretionary foods, saturated fat, unsaturated fat, added salt, added sugar and alcohol) components to give a total score ranging from 0 to 130 [7]. The DGI-2013 was adapted for use in the present study based on the dietary data available. The modified DGI-2013 consisted of eight components, six of which were scored out of 10 (i.e., vegetables, fruit, dairy and alternatives, fluid intake, discretionary foods and added sugar), and two of which were scored out of 5 (i.e., grains/cereals and saturated fat; Supplementary Table 3). Discouraged components were reverse scored. Scoring was proportional, with a maximum score indicating that the national recommendations had been met. Questions on diet variety, type of bread consumed, lean meat and alternatives intake, whether fat was trimmed from meat, salt intake and alcohol intake were excluded. The total score for the modified DGI-2013 ranged from 0 to 70, with a higher score indicating better diet quality.

The RDGI is a food-based score that assesses adherence to the 2013 Australian Dietary Guidelines recommendations in adults >18 years [10]. The original version of the RDGI contains 10 components comprising six encouraged (i.e., vegetables, fruit, grains/cereals, lean meats, dairy and alternatives and fluid intake) and four discouraged (i.e., saturated fat, added salt, added sugar and alcohol) components. The original RDGI score ranges between 0 and 100 [10]. Similarly, the RDGI was modified to suit the dietary intake data in the present study, resulting in seven components (Supplementary Table 4). This included three components (i.e., vegetables, fruit, fluid intake) scored out of 10, one component (i.e., grains/cereals) scored out of 7.5, one component (i.e., saturated fat) scored out of 6 and two components (i.e., dairy and alternatives and added sugar) scored out of 5. Discouraged components were reverse scored. Scoring was proportional, with a maximum score for a component suggesting that national recommendations had been met. The components in the original RDGI on intakes of lean meats, cheese, hot drinks (e.g., tea and coffee), salt, foods high in added sugars (e.g., biscuits, cakes and chocolate) and alcohol, as well as how often fat was trimmed from meat and the type of bread consumed, were excluded. The total score for the modified RDGI ranged between 0 and 53.5, with a higher score suggesting a better diet quality.

Background data of women were collected at baseline. Women reported their age (years), country of birth (Australia versus other), annual gross household income (<AUD$40,000, $40,000–80,000 and >$80,000) [16, 20], current employment status (employed versus unemployed) [16, 20], highest completed qualification (high school or less versus trade certificate/diploma versus university degree or higher) [16, 20], marital status (married versus not married), household composition (two-parent household versus other), smoking status (non-smoker versus past/current smoker) and pre-pregnancy weight and height without shoes. Past and current smokers were grouped, as prior research indicates that former smokers have similar dietary behaviours to current smokers [25]. The World Health Organization’s classification of overweight or obesity was used to categorise women’s pre-pregnancy body mass index (BMI; kg/m2) as underweight/normal weight (BMI < 25.0 kg/m2) and overweight/obese (BMI ≥ 25.0 kg/m2) [26].

Group-based trajectory modelling (GBTM) was used to estimate diet quality trajectory groups from 24 to 34 weeks gestation to 3.5 years postpartum using the traj command in Stata version 17.0 [27]. To facilitate comparison between the two diet quality indices, the raw DGI-2013 and RDGI scores were standardised (z-scores) before identifying the diet quality trajectories. GBTM identifies heterogeneous groups of women following similar diet quality trajectories within a study population and estimates the proportion of women in each group and their probability of membership to that group over time [28, 29]. Women with two or more diet quality data over four time points (median = 3) were included in the analyses to identify the diet quality trajectories. Mean gestational age in months was used as the time point variables, with baseline (i.e., 24–34 weeks gestation) coded as 0 and 1, 2 and 3.5 years postpartum coded as 12, 24 and 42 months, respectively. Censored normal models specifying the cubic function of time (in months) as the independent variable and repeated measurements of diet quality z-scores as the outcome variable were performed. Models with 2–5 groups were conducted and compared using the following criteria: Bayesian information criterion (BIC), entropy (> 0.7), the minimum proportion of women (> 5% of total sample) assigned to each trajectory group and model parsimony [29, 30]. The final model was chosen based on higher entropy and a higher BIC (less negative), which indicate a better model fit, clinical interpretability and model parsimony, in which a simpler model with greater interpretability is preferred [29, 30]. Once the optimal number of groups was chosen, the optimal shape of each trajectory was tested using various polynomial functions (i.e., linear, quadratic and cubic) of the time points of diet quality assessment [29, 30]. Significant polynomial functions were retained [29, 30].

Descriptive analyses were conducted to summarise cohort characteristics by the identified diet quality trajectories. Analyses were conducted to assess differences in diet quality scores by intervention allocation (intervention versus control group). There were no differences by intervention allocation at all time points except for RDGI score at 2 years (mean difference −1.20; 95%CI −2.35, −0.06) (Supplementary Table 5). A 1.2-point difference in RDGI score was considered small given that the overall RDGI score is 53.5. Therefore, the intervention and control groups were pooled for the present analysis. Intervention allocation was included as a covariate in all analyses to account for potential differences between groups.

Multivariable logistic regression was used to assess the association between maternal factors (maternal age, country of birth, educational attainment, marital status, household composition and smoking status) and the identified diet quality trajectory groups. First, univariable logistic regression was conducted with each maternal factor as the exposure and the low diet quality trajectory groups as the outcome; the high diet quality trajectory group was the reference group in all analyses. All maternal factors were then included in the multivariable model. Pearson’s correlation was undertaken to test for multicollinearity among the maternal factors (Supplementary Table 6). Household income was removed from the analysis due to data missingness (n = 41, 8.5%) and high correlation with employment status (r = –0.51). Maternal pre-pregnancy BMI was included in sensitivity analyses due to missing data (n = 26, 5.4%) (Supplementary Fig. 1).

In addition, t-tests or χ2 tests for continuous and categorical variables, respectively, were also performed to compare cohort characteristics between included and excluded participants.

To examine the effect of missing maternal factors on the analysis, multiple imputations by chained equation with 10 datasets were conducted to impute the missing maternal factors (Supplementary Fig. 1). Linear and logistic regressions were used to predict the missing continuous and categorical maternal factors, respectively, and the ‘mi estimate’ Stata command was then used to pool estimates from the 10 datasets.

Additional sensitivity analyses were conducted using raw DGI-2013 and RDGI diet quality scores as the outcome to identify the diet quality trajectory groups and associated maternal factors. Further analyses were conducted to investigate associations between pre-pregnancy BMI, along with other maternal factors, and DGI-2013 and RDGI trajectories. All analyses were performed using Stata 17.0. Results were considered significant at p < 0.05. Results are reported as odds ratio (OR) and 95% confidence interval (95%CI).