Participants

This study is a part of the Nutrition Study within the larger DIPP study, which is an ongoing multidisciplinary prospective population-based birth cohort study screening for HLA-DQB1-conferred predisposition to type 1 diabetes using cord blood samples (ClinicalTrials.gov NCT03269084) [21]. Families with newborn infants born between September 1996 and September 2004 were recruited from the regions served by Oulu University Hospital and Tampere University Hospital in Finland. Children with high or moderate HLA-conferred risk (approximately 14% of the whole population) were eligible for the study and were invited to participate. The exclusion criteria were any severe systemic disease or congenital anomaly, or if the child’s parents were of non-European origin or did not speak Finnish, Swedish or English fluently. Altogether 6080 children were enrolled in this follow-up study. In the present study, the 6-year follow-up period was used. The inclusion criteria for the IA cohorts were the availability of at least one autoantibody assessment and at least one completed day in a 3-day food record at or before the assessment of autoantibodies. The inclusion criteria for the type 1 diabetes genetic risk cohort were available information on the type 1 diabetes status of the child and at least one completed day in a 3-day food record at or before the diagnosis. The inclusion criteria for the progression cohort were repeated positivity for at least one autoantibody and at least one completed day in a 3-day food record at or after the first assessment of seropositivity but before type 1 diabetes diagnosis. A total of 5626 children were included in the IA cohort, 5674 in the type 1 diabetes genetic risk cohort, and 506 in the progression cohort. The DIPP Nutrition Study flow chart is presented in Fig. 1.

Fig. 1

DIPP study participant flow chart. All participants in the IA cohort are within the type 1 diabetes genetic risk cohort, and all participants in the progression cohort are within the IA cohort

Outcome variables

ICA were screened at 3–12 month intervals as described previously [22]. If a participant tested positive for ICA, all available samples from that participant were analysed for IAA, GADA and IA-2A. ICA were quantified by a standard indirect immunofluorescence method, while IAA, GADA and IA-2A were quantified using specific radiobinding assays [23]. In the current study, two outcome definitions for IA were used: (1) repeated positivity for ICA and at least one biochemical autoantibody or a diagnosis of type 1 diabetes (ICA plus biochemical IA); and (2) repeated positivity for at least two biochemical autoantibodies or a diagnosis of type 1 diabetes (multiple biochemical IA). Both IA outcomes were assessed by February 2017. The type 1 diabetes outcome was defined by a diagnosis of type 1 diabetes obtained from the Finnish Paediatric Diabetes Register and the university hospitals by May 2017. The Finnish Paediatric Diabetes Register covers approximately 92% of children diagnosed with type 1 diabetes. Children who are not identified in the diabetes register as having type 1 diabetes were considered free from type 1 diabetes. Progression from islet autoantibody positivity to type 1 diabetes was assessed among children who were repeatedly positive for at least one of the assessed autoantibodies. In the analyses, the primary outcomes were: (1) ICA plus biochemical IA; and (2) type 1 diabetes. Secondary outcomes were multiple biochemical IA and progression from IA to type 1 diabetes.

Genetic methods

HLA-DQ genotyping using panels of sequence-specific oligonucleotide probes has been described previously [22]. The HLA-DQB1(*02/*0302) genotype represents high risk for type 1 diabetes, and the HLA-DQB1*0302/x genotype (where x indicates alleles other than DQB1 *02, *0301 or *0602/3) represents moderate risk. After April 1997, the DQB1*0602/3 probe recognising both DQB1*06:02 and DQB1*06:03 alleles was replaced by a probe specific for DQB1*06:02.

Ethical aspects

The DIPP study adheres to the Declaration of Helsinki, and the local ethics committees of Oulu and Tampere University Hospitals approved the study protocol (ETL 97193M). Families gave their written informed consent for the genetic testing of the newborn infant and for their participation in the follow-up study, and were able to discontinue the study whenever they want.

Assessment of diet

The intake of vitamins from food (including breastmilk and drinks) and dietary supplements was assessed by using 3-day food records containing two weekdays and one weekend day at the 3, 6 and 12 month visits and the 2, 3, 4, 5 and 6 year visits. The collection of food consumption data has been described in detail previously [24, 25]. The food and nutrient calculations were based on the constantly updated in-house nutrition calculation software Finessi (Finnish Institute for Health and Welfare, Finland), which utilises the Finnish National Food Composition Database (Fineli) [26]. Calculated dietary components used in the statistical analysis include energy (MJ), vitamin A, retinol, β-carotene, carotenoids, thiamine (B1), riboflavin (B2), niacin (B3), pyridoxine (B6), folate (B9), vitamin B12, vitamin C, vitamin D, vitamin E (α-tocopherol) and γ-tocopherol. All vitamins were measured as milligrams or micrograms. Vitamin A comprised retinol and carotenoids with vitamin A activity (retinol activity equivalents). Carotenoids comprised of α-, β- and γ-carotenes and β-cryptoxanthin. Information on pantothenic acid and biotin intake in our food composition data was not up to date, and therefore those B vitamins were not included in this study. For breastfed children, we estimated the total energy intake based on age, body weight and the expected energy requirements for growth and development [27]. The amount of ingested breastmilk was estimated on the basis of weight, growth rate and energy intake from other foods [28]. Vitamin values in the food composition database are based on chemical analysis of food samples, recipe calculation and adopting values from other sources, e.g. other databases, scientific literature and food labelling. Fineli recipe calculations are performed according to European Food Information Resource (EuroFIR) guidelines [29], and nutrient retention factors are based on a report from the National Food Administration, Sweden [30]. As children’s diets change and expand during growth, we analysed the food sources for vitamins in two age groups: up to 1 year of age (3, 6 and 12 months) and from 2–6 years of age. We reviewed and updated the values for fat-soluble vitamins such as vitamin A and E in the Fineli database in 2012–2016 by screening and correcting possible errors and by quality checking of the new database versions' which improved the accuracy of data on fat-soluble vitamins.

Sociodemographic characteristics

Information on diabetes status (all types of diabetes) in first-degree relatives (yes, no), the child’s sex (male, female) and maternal education (none, vocational, secondary vocational, university studies, or degree) were collected from parents using a structured questionnaire. We were not allowed to register the children’s ethnicity due to study regulations. However, the selected participants were of European origin, as the HLA-DQ genotypes of interest may have lower predictive value in non-European populations [22]. Weight was assessed at each study visit and weight-for-age z scores were calculated separately for both sexes based on WHO criteria [31].

Statistical methods

Joint models that combine longitudinal and survival data into a single model [32] were used to analyse the association between the intake of each vitamin (between 3 months to 6 years) and the development of IA and type 1 diabetes in children by the age of 6 years. The exposure was modelled using a linear mixed-effects model, and a Cox proportional hazards regression model was used to build the time-to-event submodel. Joint models are especially useful with longitudinally collected exposure data, enabling reconstruction of a complete dietary intake profile for each participant even if a series of repeated measurements is incomplete due to dropout or missed diet records. A Bayesian framework was used to estimate the posterior distribution of model parameters, providing credible intervals (CrI) that reflect the probability of the parameter lying within a given range, unlike confidence intervals that rely on repeated sampling assumptions. Model fitting was performed using Markov chain Monte Carlo (MCMC) sampling, allowing flexible evaluation of spline-based hazard functions. Three chains were run, and convergence was confirmed (Gelman–Rubin statistic <1.1). The event time for the IA outcomes was defined as the midpoint between the first repeatedly positive test and the preceding sample. The event time for the type 1 diabetes outcome was the diagnosis date. A current-value association structure was used, and thus the HR at a given point in time t is provided for a 1 unit (0.1, 1, 10 or 100 mg or µg/MJ) increase in the longitudinal value of the vitamin intake at the same time point t. Vitamin intake from 3 months to 6 years of age was modelled using piecewise natural cubic spline functions with three knots in the linear mixed-effects submodels. The locations of knots were defined using an algorithm that selected the best suitable combination of knots by fitting all relevant combinations and selecting the best-fitting model based on the Bayesian information criterion. All analyses were adjusted for variables that had previously been observed to be potential confounders: total energy intake, sex, HLA genotype (high or moderate risk) and family history of diabetes of any type [33,34,35]. For energy adjustment, the multivariate nutrient density method [33] was used, in which absolute vitamin intake is divided by the total energy intake in MJ, and the energy-adjusted intake as well as the total energy intake are included in the model as longitudinal covariates. We controlled for multiple testing by using the Benjamini–Hochberg procedure to calculate the false discovery rate with a significance level of 0.05 [36]. The progression analysis was performed from the time of the first seroconversion until diagnosis of type 1 diabetes or the age of 6 years, and the analyses were adjusted for the age at first seroconversion in addition to other potential confounders. The use of the joint model has been described in more detail elsewhere [37].

We performed additional analyses on the risk of ICA plus biochemical IA and type 1 diabetes outcomes adjusted for maternal education and breastfeeding status at the age of 6 months. As these adjustments did not change the results, we did not include them in the model presented. Furthermore, we performed outcome analysis adjusting for weight-for-age z score to assess whether growth/overweight confounds our results. As the children were primarily breastfed at 3 months of age, with limited other dietary sources, we performed a sensitivity analysis on the associations between vitamin intakes and the risk of ICA plus biochemical IA excluding the 3-month age point. To study whether the association between the vitamin intakes and the risk of ICA plus biochemical IA varied over time, the interaction term between the vitamin intake and a natural cubic spline of age with one knot at 3 years (the midpoint of food consumption assessment) was added to the model. The model was compared with the original one using the deviance and Watanabe–Akaike’s information criteria, and log pseudo-marginal likelihood [38]. If at least two of the criteria suggested the existence of a time-varying coefficient, the association was visually inspected. As the number of children with type 1 diabetes was low (n=94), we did not assess the interaction of time with the type 1 diabetes outcome. We also performed an interaction analysis to test whether sex or HLA genotypes modify the associations between vitamin intakes and the risks of ICA plus biochemical IA and type 1 diabetes. To assess whether food avoidance influenced the association between vitamin intake and outcomes, we performed a sensitivity analysis excluding children who avoided fruits, vegetables, dairy or cereals. We also tested whether vitamin supplement use (yes vs no) was associated with the risk of ICA plus biochemical IA using Cox regression. Differences between energy-adjusted vitamin intakes by background variables at 6 months and 2 years of age were assessed using one-factor ANOVA and unpaired t test. The analyses were performed using the joint model function from the JMbayes2 package in R version 4.2.1 (https://www.r-project.org/) and IBM SPSS Statistics version 29.0 (IBM).