Clinical Natural Language Processing (NLP) integrates large language models (LLMs) to extract biomedical insights from unstructured clinical text. Most named entity recognition (NER) and relation extraction (RE) datasets rely on manual annotation, which is costly and difficult to scale. Many biomedical knowledge graphs (KG) suffer from underspecified relations, conflate causal and correlational claims, and edges lack evidence for reasoning. This dissertation presents a semantic stability framework for constructing explainable KGs, highlighting stable extraction as fundamental for scalable NER and RE, and essential for graph structure. We applied this to Substance Use Disorders (SUD) and Social Determinants of Health (SDOH) from PubMed corpus and NER and RE annotation guide. Multiple LLMs perform extraction under shared semantic constraints, with disagreements resolved through Human-in-the-Loop (HITL) validation. We define semantic stability through NER and RE metrics, using stabilized gold data for model training and evaluation. We then develop a claim-centered KG, where edges represent evidence, provenance, relation type, directionality, polarity, and stability indicators. This benchmark and pipeline supports multi-hop reasoning, triadic SUD–SDOH–SUD mediation patterns, and feedback loop analysis. This will advance etiological inquiries and data-driven health policy analysis.

This paper describes a system for the CLPsych 2026 shared task that uses retrieval-augmented in-context learning with frozen LLMs and no fine-tuning. The core contribution is a five-agent agentic pipeline for Task 3.1 sequence summarisation: two rule-based agents detect change type (Switch/Escalation) and direction (improvement/deterioration), an LLM-based DynamicsExtractor produces structured ABCD analysis, a SummaryWriter composes prose grounded in retrieved gold exemplars, and a Validator enforces structural constraints. This pipeline is iteratively refined across three submissions via NLI-based candidate reranking and per-sentence contradiction reduction. For Tasks 1.1 and 1.2, a single LLM call combines static and RAG-retrieved examples; for Task 2, an auto-tuned prompt detects moments of change. The system ranked 1st on Task 1.2 (RMSE 0.917) and Task 3.1 (score rank average 4.00), 3rd on Task 1.1 (F1 0.420), and 8th on Task 2 (F1 0.466).