ENDOCRINOPATHY OR EARLY PUBERYNY: NUTRITIONAL AND CHEMICAL ASSESSMENT OF PACKAGED FOOD PRODUCTS IN CHILDREN

ENDOCRINOPATHY OR EARLY PUBERYNY: NUTRITIONAL AND CHEMICAL ASSESSMENT OF PACKAGED FOOD PRODUCTS IN CHILDREN

Abstract

The global rise in early puberty in children is an important public health problem, which needs a multi-disciplinary, toxicological, nutritional and computational study. Packaged foods are most common foods consumed by children in today's diet and are a double-edged sword as hyper palatable foods containing excess amounts of sugar and caloric density, and simultaneously containing a hidden vector of exposure to endocrine disrupting chemicals (EDCs) via the synthetic packaging materials. This multi-faceted issue is addressed by a novel, comprehensive method that couples the use of AI-based dietary assessments with quantitative structure activity relationships (QSAR) toxicological modelling and an efficient Bayesian ordinal quantile regression model to dissect complex developmental endpoints. The framework allows for high fidelity exposure information as the packaging materials are identified by computer vision and machine learning pipelines and a chemical contact area component is estimated within the package, though due to nature of manual dietary logging. These exposure measures are flexibly linked with causal inferences models and augmented with multi-time window clinical trials designs for fine tuning causal bounds on real world cohorts of children. It is expanded to include a variety of other phenotypic features, including onset of puberty, metabolic syndrome, childhood obesity, and neurodevelopmental delay; the framework is expanded to a wider range of features in a categorical mathematical model that maps overlapping biological systems.

Finally, comprehensive lifecycle assessments and multi objective optimization strategies are taken into account to prevent unintended negative environmental consequences from lowering the level of toxic exposures. This provides the confidence that suggested biodegradable packaging have minimal human endocrine risk and environmental effects. To conclude, the proposed approach here is comprehensive and not only provides a link between theory and data, but also between computational toxicology and sustainable intervention, and thus offers an adequate model for computational toxicology, health policy in children and global environmental change.

 

Introduction

Synthetic Food Packaging comes with a Crisis.

Endocrine-disrupting chemicals (EDCs) are one of the largest environmental health problems of the modern era and occur in the synthetic food packaging industry. Chemicals, such as bisphenols, phthalates and perfluorinated chemicals or perfluorocarbons (PFAS) can be released into packaged foods and interfere with the body's hormone levels. Children are especially susceptible because of windows of rapid development, smaller body size, and underdeveloped metabolic detoxification systems.

The modern diet given to kids is made of processed food, containing two vectors working together: high sugar, macronutrient density and a "stealthy payload" of synthetic chemicals seeping from plasticized food packaging and metal can linings. Since the age of appearance of secondary sexual characteristics has been dropping worldwide, the knowledge of the interaction between nutritional biochemistry and environmental toxicology is an urgent need for paediatric endocrinology.

 The Methodological Gap

Despite the need for this crisis, current analytical and epidemiological tools are inadequate for a number of important reasons:

1)       Methodological Silos – Nutrition epidemiology (observational) and mechanistic toxicology (deterministic) are well established, with a long-standing division. This prevents having a full understanding of the effect of toxic packaging on a variety of pediatric health issues.

2)       Dietary tracking is traditionally done by self-reported diaries and these diaries have large reporting bias and cognitive fatigue. Furthermore, the traditional instruments do not take into account incidental ingestion of micro-plasticics and leachates.

3)      Statistical Limitations: Linear regression models are suitable only for simple linear relationships between variables and do not have good predictive power for the complex, non-linear relationships, or the different stages of the pubertal processes and outcomes (e.g., Tanner stages).

4)      4. Reactive Sustainability: Today’s generative AI materials discovery tools explicitly consider structural stability as a key goal, and environmental Life Cycle Assessment (LCA) is considered later, as a secondary goal; and it might result in the adoption of other materials that have higher carbon footprints.

 Proposed Unified Architecture

These gaps are filled by this paper which introduces a three-part, multidisciplinary architecture that integrates probabilistic AI-based nutritional models, deterministic toxicological simulations, and Bayesian causal networks.

1. Automated Data Acquisition (Computer Vision)

We suggest that the mobile-integrated computer vision (CV) solution is a revolutionary solution. These algorithms can be automated and if they take a picture of a meal with the original packaging, they can:

A.      Identify different types of materials (e.g., plastic (PS), plastic (PE) etc.).

Estimate area of contact between food and packaging (geometric)

Supply high fidelity exposure data for use in developing a chemical leaching rate estimation model, minimize burden on the user, and eliminate reporting bias.

B.       The Dual-Vector Risk Engine (QSAR & AI)

This model combines the AI-based nutritional profiling with the deep learning-based Quantitative Structure-Activity Relationship (QSAR). This allows for the estimation of synthetic EDC exposures, and the simultaneous measurement of macronutrient loads. The system features open-ended dietary patterns, and the unstructured data is fed into the powerful biochemical rules, made possible with the use of Large Language Models (LLMs).

C.      There is no direct relationship between the variables, so an indirect one must be used. There is, however, no direct relationship between variables and thus, we are required to use an indirect link. We use a partially collapsed Gibbs sampler version of Bayesian ordinal quantile regression to properly model these inputs to clinical outcomes. Unlike a linear model, this model:

D.      Relates multi-omics, dietary covariates to discrete and ordinal early puberty stages.

Describes the ‘complex morphisms' in a coherent manner between biological systems including overlapping pathologies such as child obesity, metabolic syndrome and neurodevelopmental delays.

E.        Sustainable Communities

The final piece brings clinical evaluations to open source sustainability modelling. We formulate an optimization approach as multi-objective constraint satisfaction problem that embeds the sustainability, manufacturability, and chemical safety requirements. This will make sure that recommended biodegradables are taken into account from several perspectives – carbon-footprint, land-use etc. – before they are recommended as alternatives.

2.      Technological Pipeline:

We suggest an automated pipeline for packaging identification and estimation of the chemical contact area based on CV.

3. Mathematical Innovation:

We provide a high-end statistical pipeline that integrates statistical exposure data into mathematical models of HPG axis disruption using statistical harmonisation, incorporating both statistical harmonisation, Bayesian causal integration and ordinal quantile regression.

4. Holistic Sustainability:

We will present a holistic integration module for LCA that will ensure that non-toxic packaging options can support global environmental health without taking unacceptable ecological trade-offs.

Conclusion:

Toward Clinically Actionable Models

The approach could be used in the future as a template for empirical studies, which also involves deterministic biochemical modeling and clinical validation. It is a multi-faceted approach that's necessary, not just because of the calorie count, but because the packaging is a chemically active substance that can produce widespread disruption of the hormones of young people, if it is indeed packaged food. This architecture is needed to unite the "shared inferential language" needed to mitigate risk in the current food system, and to protect development of futures generations.

II. Related Work

To give context to the framework proposed, the literature was broken down into four key domains:

1. Recent advances are showcased with the ability of identifying a predictable log-normal signature of macronutrient concentration variations for packaged foods (Bagler et al., 2026). The development of vision capable conversational agents and AI nutrition lenses have been made to dynamically infer the sugar content of the products from their images (Berengueres 2026). All such models, however, only consider the number of calories of the food itself, not the chemical materials in which it is stored, called "packaging materials.

2. EDCs can interfere with the extremely conserved hormonal signaling systems and hence are a threat to human health (Baker, 2021). The binding energy between the human estrogen receptor alpha and EDCs with strong electrostatic interaction of the ligand binding domain (LBD) has been estimated using molecular research via multilayer fragment molecular orbital (MFMO) calculations (Ugarte, 2019). To accelerate screening, deep neural network (DNN)–based QSAR models are used to predict the binding of a large number of molecular descriptors (MDs) to the estrogen receptor (ER) (Desai et al., 2026).

3. The HPG axis needs to be modeled beyond simple linear regression, which only explains conditional means, and is not well suited for the ordinal variables of puberty (pubic hair type, pubic hair number, and pubic hair volume) (Grabski et al., 2019). Bayesian ordinal quantile regression with a partially collapsed Gibbs sampler (BORPS) to decompose associations with early pubertal milestone items is a recent innovation that has been developed (Grabski et al., 2019). In addition, causal inference frameworks are being used to set credible causal limits on the effects of environmental covariates on medical outcomes (Zivich, 2025).

4. Although Sustainability and Materials Discovery Generative AI have revolutionized materials sciences, models are usually optimized for structural stability and Life Cycle Assessment (LCA) is seen as a downstream analysis.Though the field of materials discovery has been transformed by Generative AI, the optimization of the models for structural stability has resulted in LCA being perceived as a downstream analysis. This results in a "sustainability paradox," pointing to the need to be careful with biodegradable materials such as PBAT or chitosan-based nanomaterials throughout their entire life cycles from their extraction to their end-of-life stage to ensure that there are no "unintended" carbon footprints (Cosic et al., 2026).

III. This is a brief overview of Method: The Nutri-Chem Integration Framework (NCIF).

We present a multi-modal computational pipeline for the prediction of early puberty risk profiles in four integrated phases:

Phase 1: Automated Data Acquisition (CV & LLM) This phase involves using computer vision (CV) and Large Language Models (LLMs) to track diet through automation. Semantic segmentation is used to classify the materials from the image of the packaging material (e.g., polystyrene, PET) and estimate the geometric chemical contact area of the meals in the photo using a CV module. Meanwhile, a LLM agent is used to process unstructured diet records to obtain the amount of sugar and calorie density (Liu et al., 2025).

Based on the material type and contact area determined, the material is passed to a deep neural network QSAR model (Phase 2: Chemical Exposure and QSAR Modeling). This engine is also able to predict the estrogen receptor-binding affinity of known EDCs, and incorporates MFMO calculations to rank compounds with high electrostatic receptor interaction (Ugarte, 2019). This gives an individual "chemical load" indicator.

The outputs of nutritional and chemical modules are then used as covariates in the statistical synthesis (Phase 3) with BORPS. The structure turns out to be mathematically optimal for full conditional response distribution of ordered discrete Tanner stages (Grabski et al., 2019). We use a partial identification strategy to estimate upper and lower bounds of EDC effect on pubertal onset (Zivich, 2025).

In phase 4, Categorical Expansion and Sustainability Optimization, we use category theory (Spivak, 2014) to map the morphisms between the endocrine system and overlapping categories such as metabolic and neurodevelopmental systems. This will allow for outcomes like paediatric metabolic syndrome and obesity to be included. Finally, a multi-objective optimization algorithm is used to compare the packaging options to a library of open-source LCA data, with the goal of determining the packaging alternatives with the lowest human toxicity and environmental impact (Pareto optimization).

IV. Evaluation Plan

We suggest a two-step process for validating NCIF, a synthetic 10 year longitudinal cohort of 10,000 children with pediatrics:

The first step is the benchmarking that compares estimated contact area based on the CV with the ground truth that is measured using 3D-scanning.

Step 2: Testing the BORPS algorithm for discovering “hidden” causal truths in the space computed by BORPS.

Step 3: Simulating retrieval of alternatives that are non-toxic and biodegradable from LCA databases, with identified chemical vulnerabilities.

V. Discussion

Consumer AI: This is a model that could shape the future of consumer AI assistants, which would go beyond counting calories and become “endocrine risk” monitors. It provides the regulatory authorities with a pipeline for algorithmic, proactive and evidence-based limitations of classes of EDCs.

Limitations: The model is subject to 'cohort attrition' and 'unmeasured confounding' such as EDCs in dust or air etc. Furthermore, there are cross-market biases in the performance of AI vision models due to their inability to perform well on products not found in their training set (Berengueres, 2026).

Ethics: The tracking of children's development over time generates many privacy issues and should be made anonymous using cryptographic methods. Also, care must be taken not to "stigmatize" low-income households in food deserts by holding manufacturers accountable for the models, rather than parents.

 

Conclusion

 

The integrated field of interdisciplinary approach between computer science, molecular biology and public health is needed to combat the root cause of endocrine disruption in children. The strength of the NCIF is to recognise that the packaged product is one biologically active product, with the wrapper and macronutrients closely intertwined, thus providing a basis for protecting future generations development. This versatility is crucial in converting complex and contextual situations from the real world into formal representations, which then can be continually improved for greater impact on public health (Liu et al., 2025).