acerisk of progression from global developmental delay to intellectual disability predicted by interpretable machine learning using clinical, neurophysiological and genetic data
Source article: Pathogenic Genetic Variants, Comorbid Autism and Adaptive Developmental Quotient as Independent Predictors of Intellectual Disability in Children With Global Developmental Delay: An Interpretable Machine Learning Model With Calibrated Risk Estimation
Abstract: Background Global developmental delay (GDD) frequently precedes intellectual disability (ID), but no validated multivariable prognostic tool exists to support individualised counselling during the initial diagnostic work-up. Existing risk indicators are typically considered in isolation, and their joint contribution within an interpretable predictive framework remains uncertain. Methods We retrospectively analysed 2453 children diagnosed with GDD between January 2014 and December 2023 at a provincial tertiary ch…
Contested: both sides are scored from claims and sources, not community votes.
"Mental Patients Against Stigma" by Mindy Fisher is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.
Researchers retrospectively analyzed 2453 children diagnosed with GDD between January 2014 and December 2023 at a provincial tertiary children's rehabilitation centre, followed to at least 60 months. Using 28 predictors across perinatal, developmental, neuroimaging, electrophysiological, genetic and comorbidity domains, they trained L2- and L1-regularized logistic regression, random forest, XGBoost and LightGBM, with Platt scaling for the L2 model, and evaluated on a held-out 30% test set.
By the August 2026 publication date, the calibrated L2 model showed moderate discrimination and good calibration for high-risk triage in this referral setting, with high positive predictive value but limited sensitivity and low negative predictive value. Because the cohort had 83% ID prevalence and came from a single tertiary centre, the authors note the tool is not ready for community use without refitting and external validation, and its impact on counselling decisions remains to be prospectively tested.
- Retrospective cohort of 2453 children diagnosed with GDD 2014-2023 at provincial tertiary children's rehabilitation centre, followed to minimum age 60 months; 2037 (83.0%) progressed to ID.
- 28 candidate predictors retained after multiple imputation and multicollinearity screening; five algorithms trained on 70% split with class-weight rebalancing, no synthetic oversampling; L2 logistic regression recalibrated by Platt scaling.
- At Youden-optimal threshold 0.85, sensitivity 66.1%, specificity 76.8%, PPV 93.3%, NPV 31.7%; parsimonious five-feature model AUC 0.762; discrimination robust to exclusion of post-baseline predictors (AUC 0.769).
Among 2453 tertiary-referred children with GDD, a Platt-calibrated L2 logistic regression integrating routine clinical, neurophysiological and genetic data stratified progression to ID with AUC 0.783 and near-ideal calibration, enabling high-PPV triage.
In the same tertiary GDD cohort, the Youden-optimal model missed 33.9% of children who progressed to ID and had NPV 31.7%, limiting safe rule-out and indicating poor transportability beyond high-prevalence referral settings.
The rundown
Training used stratified 70% partition with class-weight rebalancing and no synthetic minority over-sampling; evaluation on held-out 30% test set used 1000 bootstrap confidence intervals for discrimination, calibration slope and intercept after Platt scaling, Brier score and net benefit. Sensitivity analyses excluded post-baseline candidate predictors and benchmarked against one-, three- and five-feature regressions.
Strongest independent risk factors were pathogenic genetic variant pathogenicity, comorbid autism spectrum disorder and EEG epileptiform discharges; higher Gesell adaptive developmental quotient was strongest protective factor (OR 0.45 per standardised unit). Removal of early intensive intervention lowered AUC by only 0.012 (95% CI - 0.002 to 0.026), and complete-case analysis yielded AUC 0.839 and restriction to genetically tested children AUC 0.864.
Single-centre retrospective design with high ID prevalence reflecting tertiary referral; model not validated for community populations and requires refitting and external validation before broader use.
Sources
- Peer-reviewedJournal of Intellectual Disability Research2026-08-24
How should this claim be treated?
The debate