Meta-analysis with over 1.3 million participants finds clear association between maternal polycystic ovary syndrome and ADHD in offspring

Disordered eating (a broad category of persistent, harmful patterns in eating or weight control) affects between 5% and 22% of children and adolescents worldwide, with similar rates seen in the United States. The consequences are far-reaching: these conditions are linked to bone fractures, anemia, malnutrition, dental erosion, obesity, diabetes, hypertension, and elevated cholesterol and triglycerides. They also carry one of the highest mortality rates of any psychiatric illness. 

Eating disorders rarely occur in isolation. They frequently arise alongside other psychiatric and neurological conditions. Yet, until now, no large-scale study had examined these co-occurrences in a nationally representative U.S. sample. A new study addresses that gap, focusing on children and adolescents aged 6–17 and the conditions most commonly associated with disordered eating, including ADHD. 

The Study: 

Researchers drew on data from the 2022–2023 National Survey of Children's Health (NSCH), a nationally representative, cross-sectional survey covering all 50 states and Washington, D.C. Households were selected using stratified, address-based sampling, and parents or guardians completed surveys about one randomly selected child per household. The final sample included 68,000 children and adolescents. 

Results: 

After accounting for factors including sex, age, race and ethnicity, household income, educational attainment, insurance status, and household language, children and adolescents with ADHD were 2.6 times more likely to have some form of disordered eating compared to their typically developing peers. 

The elevated risk appeared across a range of specific behaviors: 

• 60% more likely to over-exercise 

• Twice as likely to experience a fear of vomiting or choking 

• 2.4 times more likely to be extremely selective eaters, to skip meals, or to fast 

• 2.7 times more likely to purge food or vomit 

• 3 times more likely to show little interest in food 

• 3.2 times more likely to binge eat 

A greater tendency toward using diet pills, laxatives, or diuretics was also observed in the ADHD group, though this finding did not reach statistical significance. 

The Take-Away: 

These findings underscore a need to improve both prevention and treatment strategies for disordered eating, particularly in children and adolescents who have ADHD. Clinicians working with this population are advised to screen for a wide spectrum of disordered eating behaviors.

‍

‍

The Background:

For centuries, we’ve called the eyes the "windows to the soul," but for modern neurologists, they are quite literally a window into the brain. The retina and the central nervous system share the same embryonic origins, developing from the same neural tissue in the womb. Because of this deep biological connection, the back of your eye acts as a non-invasive map of your brain's health, displaying a complex web of nerves and blood vessels that can (theoretically!) mirror certain neurodevelopmental conditions. 

‍

Recently, a buzz rippled through the mental health community when a study published in partnership with Seoul National University Bundang Hospital claimed a massive breakthrough. Researchers developed an Artificial Intelligence (AI) model that could screen children for Attention-Deficit/Hyperactivity Disorder (ADHD) using nothing more than a simple retinal photograph. The study, which prospectively recruited children from Severance Hospital and Eunpyeong St. Mary’s Hospital, produced results that were staggering: the AI reportedly achieved an accuracy rate of  96.9%!

‍

In the world of medical testing, scientists use a metric called  AUROC  (Area Under the Receiver Operating Characteristic) to measure how well a test works.

• 0.5  means the test is no better than a coin flip (pure luck).
• 1.0  represents a perfect test with zero mistakes. 

‍

An AUROC of 96.9% is a near-perfect score, suggesting a tool is ready for immediate, real-world deployment. While headlines promised a revolution in mental health screening, a deeper look into this research and the study’s design has exposed that this 96.9% AUROC was more likely evidence of a flawed methodology rather than a biological reality.

‍

The Promise: How the AI "Sees" ADHD

To build their screening tool, researchers analyzed over 1,100 retinal images using a digital pipeline called AutoMorph and a machine-learning model known as XGBoost. The AI was trained to hunt for physical signals of the "Dopamine Connection." Dopamine is the primary neurotransmitter involved in ADHD, but it is also essential to the eye. It regulates synaptic formation, retinal blood flow, and vascular endothelial regulation. Because dopamine dysregulation influences how blood vessels grow and remodel, the study hypothesized that an ADHD brain would leave a unique "fingerprint" on the retinal vasculature, resulting in denser, thicker vessel structures.

On paper, the logic was sound: use AI to spot the subtle vascular remodeling caused by dopaminergic shifts. But a closer look at the investigation revealed that the AI wasn't just spotting ADHD; it was over-indexing on technical noise.

‍

Flaw #1: Batch Effects

The most significant "smoking gun" flagged by critics is a massive temporal mismatch. In other words, there was a severe disparity in the timeframes and conditions under which the retinal images for the two comparison groups were collected. For an AI to learn a biological condition, it must compare groups under identical technical conditions. Instead, this study created a time-traveling dataset:

• The ADHD Group:  323 children recruited prospectively in a tight 6-month window in  2022 .
• The Control Group:  323 children gathered retrospectively over a  17-year span  (2007 to 2024).This discrepancy triggers severe Batch Effects. This is a term scientists use to describe non-biological factors in an experiment that can cause inaccuracies in the data it produces. Fundus photography technology changed dramatically between 2007 and 2024. An investigation into the hardware uncovered shifts in camera models, lens optics, sensor degradation, and digital compression formats .Think of it this way: if you compare a selfie taken on the original 2007 iPhone with one from an iPhone 16, the AI doesn't need to look at your face to tell them apart; it just looks at the  2007 sensor noise  and pixel grain. The AI likely didn't learn to identify ADHD so much as it learned to distinguish between "old camera" and "new camera."

‍

Flaw #2: Control Group

A scientific study is only as reliable as its control group. The control in any experiment acts as a baseline against which the study group is compared. In this case, the control group should be composed of children without any neurodevelopmental disorders, or of “typically developing” children. 

In this study, the control group wasn't composed of healthy children from the community. Instead, they were patients visiting a tertiary ophthalmology clinic. Children visiting a specialist eye hospital are rarely "typical." They are there because they have symptomatic eye issues. This introduced a massive selection bias involving three major confounders:

• Refractive Errors (Myopia/Nearsightedness):  Severe myopia physically stretches the retina. This stretching alters vessel density and optic disc size, which were the exact markers the AI was examining.
• Strabismus:  Misaligned eyes.
• Ocular Anomalies:  Physical eye defects.Because these conditions directly alter retinal architecture, the AI likely learned to distinguish between "kids with ADHD" and "kids with severe eye problems," rather than "kids with ADHD" and "typical kids."

‍

Fatal Flaw #3: The "Mirror Image" Leakage

When training AI, you must never allow the "test questions" to leak into the "study material." The researchers, however, committed a fundamental violation of machine learning hygiene known as  Eye-to-Eye Data Leakage. The study split the data by the eye rather than by the participant. 

Human eyes are highly correlated; the left eye is a near-mirror of the right. If a child's left eye was used for training and their right eye was used for testing, the AI was effectively "cheating." Instead of learning the general traits of ADHD, the model was potentially memorizing individuals. This error artificially balloons accuracy metrics. 

‍

The True Test: Differential Diagnosis 

The true test of medical AI is diagnostic specificity, or differential diagnosis. This refers to the ability to tell one condition apart from another. While the model claimed 96.9% accuracy against a flawed control group, its performance collapsed when faced with real-world complexity.

When the researchers asked the AI to differentiate between ADHD and Autism Spectrum Disorder (ASD), the accuracy plummeted to a poor  63% AUROC. In real-world clinical settings, an accuracy of 63% is dangerously close to a 50% coin flip. Since ADHD frequently co-occurs with ASD, anxiety, or intellectual disabilities, an AI that cannot handle these "clinical differentials" is functionally useless in a doctor's office. The failure at this stage proves the model was likely detecting technical quirks of the dataset rather than a unique biological marker for ADHD.

‍

Conclusion:

To move from the lab to the clinic, we must establish a foundation built on rigor rather than high-speed data scraping. Moving forward, we must demand these 3 Pillars of Trusted Medical AI :

1. Prospective, Unified Hardware:  Data must be collected on identical camera systems with the same protocols to eliminate technical "batch effects."
2. Healthy, Community-Based Controls:  Comparisons must be made against truly "typically developing" children, not patients from eye clinics with their own retinal anomalies.
3. Rigorous External Validation:  AI models must be tested on independent datasets from entirely different hospital networks to ensure they aren't just "memorizing" one hospital's specific machinery.Artificial Intelligence holds immense potential, but we must demand detective-like scrutiny before these tools reach our children. In the search for the "window to the mind," we have to make sure we aren't just looking at a smudge on the glass.

‍

The dream of a quick eye scan to diagnose ADHD is not dead, but it must be rescued from "fast science" shortcuts and buzzy headlines. 

‍

Background:

One of the more persistent concerns among parents of children with ADHD is whether stimulant medications will stunt their child's growth. A large Israeli cohort study now offers some of the most rigorous reassurance to date, and its methodology sets it apart from earlier research. 

The question has long been complicated by a more fundamental uncertainty: do growth differences in children with ADHD stem from the condition itself, from stimulant treatment, or from factors present before any medication is ever prescribed? Without a clear answer, clinicians and families have faced a genuine dilemma when weighing the benefits of stimulant therapy against potential long-term physical costs. 

Most previous studies compounded this difficulty by comparing group-average heights, which ignores the crucial variable of genetic potential. A child who is short relative to the general population may simply have short parents. Failing to account for this introduces systematic bias and can make medications appear more harmful than they are. 

The Study:

The Israeli research team addressed this directly. Using health records from a nationwide provider, they assembled a retrospective cohort of children born between 1995 and 2003, following them through 2023. This amount of time was long enough for all participants to have reached adult stature (defined as 17 or older for females, 19 or older for males). Their sample included 5,671 children with untreated ADHD, 11,846 who received stimulant treatment, and 47,258 non-ADHD controls. Children who took stimulants for only one to two months, or who had chronic medical conditions requiring long-term medication, were excluded to avoid confounding the results. 

Crucially, adult height was evaluated not against population norms but against each individual's expected height, calculated from parental heights using the Tanner-Goldstein-Whitehouse method, a standard approach for estimating genetic height potential via mid-parental height. 

When the researchers compared adult heights across the three groups using analysis of variance (ANOVA), they did find statistically significant differences. But statistical significance, particularly in studies with tens of thousands of participants, does not automatically translate into clinical significance. The effect sizes were consistently very small, and the absolute differences were under one centimeter, which is a margin considered clinically negligible. 

Their conclusion is measured but clear: after accounting for genetic growth potential, neither an ADHD diagnosis nor stimulant treatment was associated with meaningful reductions in adult height. The findings, they argue, support prioritizing behavioral and functional outcomes when making treatment decisions, since the risk of clinically significant height loss appears to be minimal. 

The Take-Away:

For families navigating ADHD treatment, the practical implication is significant: concerns about permanent growth suppression, while understandable, should not be the primary driver of whether or how long a child receives stimulant therapy. 

‍