DTA

Corso Ordinario Secondo Livello

MARANI, EDOARDO

etd-05152026-111117

Data-Driven Healthcare Management: An LLM-Based Framework for Transforming Natural Language into Actionable Analytics

Classe Scienze Sociali

SCIENZE ECONOMICHE E MANAGERIALI - SCIENZE ECONOMICHE E MANAGERIALI

tutor Prof. FREY, MARCO
relatore Prof.ssa NUTI, SABINA

• Nessuna parola chiave trovata

15/06/2026;

completa

This dissertation proposes a Large Language Model-based framework to transform unstructured diagnostic queries from electronic prescriptions into structured administrative data for regional healthcare analytics. The work addresses the critical limitation that while diagnostic queries, free-text clinical justifications written by prescribers, contain valuable information about care pathways and appropriateness, their natural language format prevents quantitative analysis. Traditional rule-based approaches fail due to Italian linguistic flexibility, spelling variability, and semantic complexity in multi-pathology cases. The methodology follows a rigorous iterative pipeline with dual validation gates: models must achieve ≥90% accuracy against a manually-classified gold standard of 500 cardiological prescriptions, and residual unclassifiable cases must remain below 10%, indicating adequate taxonomy breadth. The author tested four LLM architectures (Gemma 2 27B, Llama 3.1 70B, Mistral Small 22B, Qwen 2.5 72B) across eight iterative prompt refinements, benchmarking performance using a WAMPI composite index that balances accuracy, latency, and operational scalability. Key findings show larger models (Qwen 2.5, Llama 3.1) achieve 85–87% accuracy but incur ~50% computational overhead, while meticulous prompt engineering yields greater performance gains than model selection alone. The framework addresses privacy constraints through EU-compliant cloud infrastructure and structural anonymization, demonstrating that clinical utility and data security are compatible. Although results fall slightly short of the 90% threshold, the author establishes that this gap is bridgeable through additional model testing and hardware improvements, while the modular pipeline design ensures scalability across any specialist branch without infrastructure overhaul.

Nome file	Dimensione
DIPLOMA_4.pdf	1.53 Mb
Contatta l'autore