AI studies report 87% sensitivity and 91% specificity for ADHD—but not one test

A new meta-analysis reports pooled sensitivity of 0.87 and specificity of 0.91 for AI approaches to ADHD diagnosis. Those numbers look like the performance of a single test. They are averages across 54 studies using different populations, data types, models, reference standards and validation designs—and the remaining heterogeneity exceeded 96%.[1]

The study’s most useful conclusion is therefore not that an 87/91 diagnostic product exists. It is that the literature contains promising signals alongside enough variation and bias risk to prevent a pooled average from being treated as a clinical operating point.

Sensitivity and specificity need a population

Sensitivity describes how often the evaluated system identifies ADHD among people classified as having it under the study reference. Specificity describes how often it rejects ADHD among those classified as not having it. Neither number says how many positive results will be correct in a clinic until prevalence and referral selection are known.

Consider an illustrative population of 1,000 people in which 10% truly meet the reference definition. Applying the pooled sensitivity and specificity mechanically would produce approximately 87 true positives, 13 false negatives, 81 false positives and 819 true negatives. Only about 52% of positive results would be true positives. This is arithmetic with an assumed prevalence, not a forecast for a real service.

High heterogeneity means the average may describe nobody

Studies can differ in age, sex, comorbidity, medication, recruitment source and the way ADHD was established. They also use MRI, EEG, behaviour, cognitive tests, digital traces or combinations. A model trained to distinguish clear research cases from healthy controls may perform very differently in a clinic where anxiety, autism, sleep problems and learning difficulties overlap.

The review found no statistically significant modality winner. That result should not be read as proof that every modality is equal. With heterogeneous studies and limited direct comparisons, absence of a detected difference can reflect uncertainty rather than equivalence.

External and prospective validation are the missing bridge

A clinically useful system should be frozen before evaluation and tested at a new site on consecutive or representative patients. It should report calibration, missing-data handling, subgroup performance and how it affects clinician decisions. The comparison should be against current diagnostic practice, not merely against another machine-learning model.

ADHD remains a clinical diagnosis informed by development, impairment across settings, history and exclusion of alternative explanations. A data-driven model may become a useful screening or decision-support component. The current literature does not support replacing that assessment with a pooled benchmark number.

The meta-analysis advances the field by replacing isolated impressive studies with a view of the full evidence base. Its message is cautiously positive: useful signal exists across modalities. Its equally important warning is that extreme heterogeneity, publication bias and incomplete external validation make readiness a study-level question, not a headline average.

Report an error or suggest a correction