Private Ensembles, Public Confidence: A PATE-to-MedPrompt System for Autism Detection

Background/Objectives: Early autism screening needs to be both accurate and privacy-preserving, but single-source assessments can miss clinically important context. We therefore study a preliminary integrated framework that combines privacy-preserving questionnaire-based risk estimation with a second reasoning component based on a large language model (LLM) that evaluates symptom narratives. The objective is to test whether structured screening outputs can be translated into uncertainty-aware narrative reasoning within one privacy-conscious workflow. Methods: The proposed pipeline links a PATE-style AQ-10 screening stage to a MedPrompt-style consensus reasoning stage that operates on behavioral summaries and transcript-style inputs. Evaluation includes component-wise testing on AQ-10 data, an end-to-end controlled setting, synthetic stress testing, and transcript-only analysis on 26 examples. Results: In component-wise evaluation, the combined pipeline reaches ceiling performance on a controlled AQ-10 split, synthetic stress testing reduces accuracy to 97.2%, and transcript-only testing shows that contextual factors such as age substantially improve sensitivity. Conclusions: These findings support only a highly preliminary proof-of-concept under constrained evaluation conditions and should be interpreted as motivation for broader external validation rather than as evidence of practical decision-support readiness across settings.