CONTINUUM-AI: A Continuous Learning Federated Neuro-Symbolic Framework for Lifespan Autism Diagnostics and Personalized Intervention

Abstract

Autism Spectrum Disorder (ASD) affects approximately 1 in 36 children globally, yet existing
artificial intelligence systems focus almost exclusively on initial binary diagnosis, neglecting
the critical need for continuous monitoring, personalized intervention adaptation, and long-term
developmental trajectory prediction across the lifespan. This paper presents CONTINUUM-AI,
a novel continuous learning federated neuro-symbolic framework that addresses this fundamental gap through five integrated innovations. First, we introduce a lifelong continuous learning
protocol that enables model updating from longitudinal follow-up data across 18 international
sites (N=847 participants, 2,541 timepoints) without catastrophic forgetting, achieving diagnostic accuracy of 0.974 AUC-ROC at initial assessment and maintaining 0.969 AUC-ROC after
five years of continuous updates. Second, we develop a differentiable neuro-symbolic reasoning engine encoding DSM-5-TR and ICD-11 criteria that provides interpretable diagnostic rationales while simultaneously predicting individualized developmental trajectories across three
domains (adaptive behavior, language, symptom severity) with 0.892 accuracy at 24-month
horizons. Third, we implement a privacy-preserving federated continuous learning protocol
with zero-concentrated differential privacy (ε = 1.4, δ = 10−5
) that enables collaborative
learning from longitudinal data across international boundaries. Fourth, we introduce a personalized intervention optimization module that dynamically adapts treatment recommendations
based on predicted trajectories and observed responses, achieving 31.6% improvement in 24-
month outcomes compared to static intervention protocols. Fifth, we demonstrate ultra-edge
deployment via neural architecture search, producing a 0.12M parameter student model (96.7%
reduction) that maintains 94.2% of teacher accuracy with 6ms inference on ARM CortexM4 processors, enabling continuous monitoring in naturalistic settings. Extensive validation
demonstrates that CONTINUUM-AI achieves superior performance across all dimensions: diagnostic accuracy (AUC-ROC: 0.974, +10.4% improvement), trajectory prediction (24-month
accuracy: 0.892), intervention optimization (31.6% outcome improvement), privacy (AUCROC degradation: 0.974 → 0.967 under ε = 1.4), and fairness (demographic parity difference:
0.156 → 0.027, 82.7% reduction). Our findings establish that continuous learning frameworks
represent the next frontier in computational psychiatry, moving beyond static snapshots to dynamic, personalized, lifespan-spanning support for individuals with autism and their families.