Please use this identifier to cite or link to this item:
https://hdl.handle.net/10216/175659| Author(s): | João Pedro Antão Sobral |
| Title: | Enhancing Clinical Summarization through Knowledge Graph Grounding and Timeline Visualization |
| Issue Date: | 2026-07-15 |
| Abstract: | The digitalization of healthcare has led to an overwhelming volume of unstructured Electronic Health Records (EHRs), creating a phenomenon known as Note Bloat. While Large Language Models (LLMs) offer promising summarization capabilities, their clinical adoption is hindered by the critical risk of "hallucinations", fluent but factually incorrect medical data. This dissertation proposes the Graph-Anchored Text Extraction and Summarization (GATES) architecture to enforce absolute factual precision. GATES uses a deterministic module to extract verified medical entities into a structured Knowledge Graph, strictly constraining the LLM to synthesize summaries exclusively from this validated structure. Equipped with an interactive traceability dashboard, GATES eradicates generative hallucinations, achieves an extreme compression ratio (reducing records to 2.84%), and maintains a domain-specific semantic preservation score of 88.3%. This research provides a scalable, factually grounded tool that converts chaotic clinical narratives into safe, actionable intelligence. |
| Description: | Context: The digitalization of healthcare has led to an overwhelming volume of unstructured Electronic Health Records (EHRs), creating a phenomenon known as Note Bloat. Problem: While Generative AI and Large Language Models (LLMs) offer promising summarization capabilities to alleviate physician burnout, their clinical adoption is severely hindered by the critical risk of "hallucinations", the generation of fluent but factually incorrect medical data. Proposed Solution: The dissertation proposes the GATES (Graph-Anchored Text Extraction and Summarization) architecture, a neuro-symbolic framework designed to enforce absolute factual precision without sacrificing narrative fluency. Methodology: The pipeline completely isolates the generative LLM from raw clinical text. It uses a deterministic Named Entity Recognition (NER) module to extract verified medical entities and map them into a structured Knowledge Graph. The LLM is strictly constrained to synthesize the summary solely from this validated graph. Additionally, an interactive dashboard provides bidirectional data provenance to trace generated claims back to the source. Results: Empirical validation demonstrates that the framework eradicates generative hallucinations and achieves an extreme compression ratio, reducing records to 2.84% of their original volume. The system maintains a domain-specific semantic preservation score of 88.3% (BiomedBERT F1). Impact: It provides a scalable, privacy-compliant, and factually grounded tool that converts chaotic clinical narratives into safe, actionable intelligence, directly supporting medical decision-making. |
| Subject: | Engenharia electrotécnica, electrónica e informática Electrical engineering, Electronic engineering, Information engineering |
| Scientific areas: | Ciências da engenharia e tecnologias::Engenharia electrotécnica, electrónica e informática Engineering and technology::Electrical engineering, Electronic engineering, Information engineering |
| URI: | https://hdl.handle.net/10216/175659 |
| Document Type: | Dissertação |
| Rights: | embargoedAccess |
| Embargo End Date: | 2029-07-14 |
| Appears in Collections: | FEUP - Dissertação |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 786596.pdf Restricted Access | Enhancing Clinical Summarization through Knowledge Graph Grounding and Timeline Visualization | 2.04 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.