Synthetic Modeling of the Impact of Artificial Intelligence-Assisted Early Detection of Autism Spectrum Disorder on Lifetime Suicide Risk: A Study Protocol with Methodological Triangulation and Quantitative Validation
Authors/Creators
Description
Background: Autism Spectrum Disorder (ASD) affects approximately 1.5% of the global population, with suicide risk four to nine times higher than in the general population. Late diagnosis—particularly common among individuals without associated intellectual disability and among women—correlates with greater accumulated psychosocial suffering. Prospective longitudinal studies capable of estimating the population-level impact of early detection would require decades of follow-up and prohibitive resources.
Objective: To present the protocol for a synthetic study employing methodological triangulation—integrating causal inference (g-formula with target trial emulation), dynamic modeling (Agent-Based Model), and machine learning with synthetic data (TVAE)—to estimate the population-level effect of artificial intelligence-assisted early ASD detection on indicators of psychological distress and suicide risk across development. The study includes rigorous quantitative validation through simulation, establishing sample sizes, confidence intervals, and statistical power for estimates.
Methods: The study will be structured across three complementary layers: (1) an Agent-Based Model (ABM) to simulate individual trajectories with complete heterogeneity, calibrated using longitudinal cohort data; (2) a causal inference model via g-formula to estimate population-level counterfactual effects of the intervention; (3) a predictive model with synthetic data generated by Tabular Variational Autoencoder (TVAE) for data enrichment and increased statistical power. Quantitative validation will include statistical power analysis (N=2,000,000; power >99% to detect reductions >10%), Bayesian calibration, parametric and structural sensitivity analysis, and E-value calculation for unmeasured confounders. Estimate stability will be verified through Monte Carlo simulations with 500 replications.
Expected Results: Validation study results—based on empirical screening data from autistic adolescents (N=1,441)—indicate that 13.39% of autistic adolescents present positive suicide risk, compared to 6.76% in non-autistic individuals [OR=2.13; 95%CI:1.82-2.49], with 9.99% presenting moderate-to-severe depression versus 5.57% in non-autistic individuals (p<0.01). Monte Carlo simulation of the g-formula demonstrates that for a rare outcome (prevalence ~1%), the logistic model yields effect estimates centered precisely around the true effect (mean difference <0.001), with a credibility interval for the mediated proportion of [3.2%-8.4%]. TVAE application to health data achieved validation with similarity distribution of 60.4%, preserved correlation of 93.5%, geometric similarity of 84.4%, and MLU score of 90.7%, with overall quality of 80.2%. ABM models demonstrated that at least 1,000 replications are necessary to ensure high confidence in mean estimates. Uncertainty analysis indicates that the g-formula can be estimated with precision in samples of N=384,000, with mean difference between estimates and truth <0.001 and 2.5th and 97.5th percentiles of -0.000004 and 0.000005 for prediction error.
Conclusion: The study will establish a new standard of rigor for synthetic research in preventive psychiatry, providing initial population estimates to inform public policy and the design of future empirical studies. Quantitative validation demonstrates that the proposed methodology produces precise and stable estimates, with narrow confidence intervals and sufficient statistical power to detect clinically meaningful effects. Results will not replace evidence from randomized controlled trials but will offer the methodological framework and justification for their conduct.
Files
Files
(28.3 kB)
| Name | Size | |
|---|---|---|
|
md5:c5243e289524bea6005611fa6c823b3a
|
28.3 kB | Download |