CGA-KWP: Neuro-Symbolic Pediatric Diagnosis

CGA-KWP for Early Diagnosis of Pediatric NDDs A research-oriented PyTorch prototype for the Cognitive Graph Assembler (CGA) and the Knowledge Weaving Protocol (KWP) — models designed to support early, interpretable diagnosis of pediatric neurodevelopmental disorders (NDDs). This code is based on the paper: Exploring Image Classification for Early Diagnosis in Pediatric Neurodevelopmental DisordersAuthor: Xiao Li Background & Motivation Neurodevelopmental disorders (NDDs) such as autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD) are conditions that typically emerge in childhood. Early identification is crucial for effective interventions, but current machine learning models often act as black boxes, offering little interpretability to clinicians. :contentReference[oaicite:0]{index=0} The CGA-KWP framework aims to overcome these issues by: Combining temporal encoding with graph-based reasoning. Embedding clinical knowledge (ontologies, rules, symptom–condition relationships). Producing interpretable diagnostic graphs instead of opaque probability vectors. Key Components 1. Temporal Encoder Encodes longitudinal patient data with explicit modeling of time gaps (Δt). This helps differentiate between recent vs. outdated evidence. Implemented here with a GRU plus a time-gap embedding. :contentReference[oaicite:1]{index=1} 2. Hypothesis Generator Projects hidden states into candidate diagnoses (multi-label logits). Active hypotheses correspond to nodes in the graph. :contentReference[oaicite:2]{index=2} 3. Logic-Aware Edge Constructor Predicts edges between nodes while respecting a plausibility mask derived from taxonomy distances and co-occurrence priors. This ensures the generated graph adheres to known clinical relationships. :contentReference[oaicite:3]{index=3} 4. Knowledge Weaving Protocol (KWP) Performs dual-channel attention: Symptom channel: compares embeddings with symptom templates. Condition channel: compares embeddings with condition templates.This injects prior knowledge into the model, improving robustness under scarce data. :contentReference[oaicite:4]{index=4} 5. Graph Validation & Uncertainty A lightweight plausibility scorer evaluates the constructed graph, while MC-dropout estimates uncertainty — providing confidence levels for predictions. :contentReference[oaicite:5]{index=5}