Federated Privacy-Preserving AI for Global Primary Care Screening in Autism Spectrum Disorder
Keywords:
— federated learning; differential privacy; autism spectrum disorder; primary care screening; global health; secure aggregation; edge computing; privacy–utility trade-off; low- and middle-income countriesAbstract
Background. An estimated 61.8 million people worldwide are autistic. The Global Burden of Disease Study 2021 puts age-standardised prevalence at 788.3 per 100,000, against 369.4 in the previous cycle; the revision followed a change in estimation method, not a change in prevalence. Reported figures also differ widely between regions, from 1,559.5 per 100,000 in High-income Asia Pacific to 614.5 in Tropical Latin America. A review of 30 studies found first parental concern at a mean of 23.6 months and diagnosis at 56.0 months, with no significant difference between high- and low-income countries. Machine learning could support earlier triage in primary care, but the training data would have to come from clinics that are often unable to transfer patient records.
Objective. To test whether a useful screening model can be trained across such clinics without moving any records, and to measure what formal privacy guarantees cost in diagnostic accuracy.
Methods. Each clinic trains on its own data using local hardware. Per-sample gradients are clipped and Gaussian noise is added, giving record-level differential privacy. Updates are masked before transmission, so the coordinating server can compute their sum without seeing any individual contribution. We used FedProx for aggregation to allow for differences between sites, and a Rényi accountant to track cumulative privacy loss. Nine privacy budgets were evaluated at two federation sizes in simulation, and the resulting curves were compared against published DP-SGD results.
Results. Reported costs of federation in the literature vary. A federated model for autism screening from serious-game data reached an AUC-ROC of 78.02%, against 79.59% for a centralised model; a second study using FedAvg on multi-site behavioural data reported 70.30% against 82.61% for local training. In our simulation the federated model reached 0.897 AUROC against 0.912 for pooled data. Differential privacy accounted for a larger share of the loss than federation did in every configuration we tested. With 20 clinics, AUROC was 0.781 at ε = 1, 0.868 at ε = 4 and 0.884 at ε = 8. An 8-clinic federation performed worse at every budget.
Conclusions. At ε ≤ 1, sensitivity fell to 0.47 at fixed specificity, which is too low to be useful for referral. Choosing the smallest available privacy budget is therefore not necessarily the safer option. We suggest a record-level budget of 4 to 8 combined with secure aggregation. Accuracy at a fixed budget also improved as more sites joined, so recruitment and privacy are not straightforwardly in tension. The model is intended to prioritise children for evaluation rather than to diagnose them.
References
1. World Health Organization. Autism. Fact sheet, 17 September 2025.
2. GBD 2021 Autism Spectrum Disorders Collaborators. The global epidemiology and health burden of the autism spectrum: findings from the Global Burden of Disease Study 2021. Lancet Psychiatry. 2025;12(2):111–121.
3. Liu et al. Global inequality and future burden of autism spectrum disorders: a frontier and projection analysis based on GBD 2021. Brain and Behavior. 2026;16:e71365.
4. Global, regional, and national burden of autism spectrum disorder: trends and decomposition analysis from 1990 to 2021, and projections for 2045. PMC12728558.
5. Costa et al. Autism spectrum disorder diagnoses: a comparison of countries with different income levels. Clinical Epidemiology. 2022;14:1001–1013.
6. Robins DL, Casagrande K, Barton M, Chen C-MA, Dumont-Mathieu T, Fein D. Validation of the Modified Checklist for Autism in Toddlers, Revised with Follow-Up (M-CHAT-R/F). Pediatrics. 2014;133(1):37–45.
7. Wieckowski AT, et al. Meta-analysis of the Modified Checklist for Autism in Toddlers, Revised/Follow-up for screening. Pediatrics. 2023;151(6):e2022059393.
8. Aishworiya R, et al. Meta-analysis of the M-CHAT-R/F: diagnostic accuracy across global samples. 2023.
9. Guo C, et al. Reliability and validity of the Chinese version of the M-CHAT-R/F. J Autism Dev Disord. 2019;49:185–196.
10. Zheng RM, Chan SP, Law EC, Chong SC, Aishworiya R. Validity and feasibility of using the M-CHAT-R/F in primary care clinics in Singapore. Autism. 2024;28(5).
11. Validation of the Modified Checklist for Autism in Toddlers (M-CHAT): a replication study of diagnostic accuracy in a rural health system. J Autism Dev Disord. 2026.
12. Comparative screening accuracy of the M-CHAT-R/F for early autism spectrum disorder: a systematic review. 2026. PMID 42264543.
13. Cross-cultural similarities and differences in reporting autistic symptoms in toddlers: synthesizing M-CHAT(-R) data from ten countries. Res Autism Spectr Disord. 2022.
14. Marlow M, Servili C, Tomlinson M. A review of screening tools for identification of ASD and developmental delay in infants and young children: recommendations for use in low- and middle-income countries. Autism Research. 2019;12(2):176–199.
15. Early autism intervention components deliverable by non-specialists in low- and middle-income countries: a scoping review. PMC9277121.
16. Use of machine learning to investigate the Quantitative Checklist for Autism in Toddlers (Q-CHAT) towards early autism screening. Diagnostics. 2021;11(3):574.
17. McMahan HB, Moore E, Ramage D, Hampson S, Agüera y Arcas B. Communication-efficient learning of deep networks from decentralized data. AISTATS. 2017.
18. Li T, Sahu AK, Zaheer M, Sanjabi M, Talwalkar A, Smith V. Federated optimization in heterogeneous networks (FedProx). MLSys. 2020.
19. Arivazhagan MG, et al. Federated learning with personalization layers. arXiv:1912.00818. 2019.
20. Li T, Sanjabi M, Beirami A, Smith V. Fair resource allocation in federated learning (q-FedAvg). ICLR. 2020.
21. Unlocking multi-site clinical data: a federated approach to privacy-first child autism behavior analysis. arXiv:2604.02616. 2026.
22. Federated anomaly detection for early-stage diagnosis of autism spectrum disorders using serious game data. arXiv:2410.20003. 2024.
23. Privacy-sensitive speech analysis using federated learning to assess depression. arXiv:2205.00111. 2022.
24. Federated learning for fair autism spectrum disorder screening across age-heterogeneous populations. Frontiers in Digital Health. 2026. doi:10.3389/fdgth.2026.1760849
25. Alshammari H, et al. Explainable federated learning for enhanced privacy in autism prediction using deep learning. Journal of Disability Research. 2024. doi:10.57197/JDR-2024-0081
26. Identifying autism spectrum disorder from multi-modal data with privacy-preserving. npj Mental Health Research. 2024. doi:10.1038/s44184-023-00050-x
27. Das A, Pattnaik PK, Bandyopadhyay A, et al. Privacy preserving federated framework for autism spectrum disorder screening with ensemble stacking and deep learning approaches. Discover Artificial Intelligence. 2026;6:426.
28. Kaissis G, et al. End-to-end privacy preserving deep learning on multi-institutional medical imaging (PriMIA). Nature Machine Intelligence. 2021;3:473–484.
29. Bonawitz K, et al. Practical secure aggregation for privacy-preserving machine learning. ACM CCS. 2017.
30. Abadi M, Chu A, Goodfellow I, McMahan HB, Mironov I, Talwar K, Zhang L. Deep learning with differential privacy. ACM CCS. 2016:308–318.
31. Dwork C, Roth A. The algorithmic foundations of differential privacy. Found Trends Theor Comput Sci. 2014;9(3–4):211–407.
32. Mironov I. Rényi differential privacy. IEEE CSF. 2017:263–275.
33. Differential privacy for medical deep learning: methods, tradeoffs, and deployment implications. arXiv:2506.00660. 2025.
34. Dynamic differential privacy technique for deep learning models. Scientific Reports. 2025. doi:10.1038/s41598-025-27708-0
35. Tayebi Arasteh S, et al. Preserving privacy in domain-specific medical deep learning: multi-institutional evaluation of privacy–utility–fairness tradeoffs. 2024.
36. Federated learning in smart healthcare: a comprehensive review on privacy, security, and predictive analytics with IoT integration. PMC11728217.
37. Blanchard P, El Mhamdi EM, Guerraoui R, Stainer J. Machine learning with adversaries: Byzantine tolerant gradient descent (Krum). NeurIPS. 2017.
38. World Health Organization & UNICEF. Global report on children with developmental disabilities: from the margins to the mainstream. Geneva: WHO; 2023.
