Explainable NLP for ADHD Anamnesis

Overview

Building an explainable NLP system for German primary school reports to support retrospective ADHD assessment in collaboration with clinicians.

Role and scope

• Full-stack ownership: Python backend (models, REST API) and React/Node.js frontend.
• Designed evidence highlighting and clinical metrics (AUC/ROC, sensitivity, specificity) for decision support.

Technical highlights

• Compared encoder and decoder models plus LLM APIs for quality, explainability, and cost.
• Supervised fine-tuning and reinforcement learning of Qwen3-4B for 22 criterion-level indicators in primary school reports.
• Published encoder experiments used focal loss and inverse-frequency class weighting, with macro F1 as the primary metric for imbalanced labels.
• Implemented RAG-style evidence retrieval and token-level attribution for UI highlighting.
• Iterative evaluation with clinician feedback on evidence review and workflow integration.

Status

• The collaboration system is published in CHI 2026 Extended Abstracts. The paper reports encoder and LLM comparisons; it does not document the separate Qwen3-4B SFT/RL recipe.

Publication

A Human-AI Collaboration System for ADHD Assessment from Primary School Reports · Open-access paper