NLP-Based Hospital Diagnosis Reporting Aid

Abstract

This thesis explores using NLP to automate hospital diagnosis reporting for Atrial Fibrillation patients following European Society of Cardiology guidelines. The two-phase study used English MIMIC-IV dataset (40,000+ records) and Dutch Jessa Hospital data (12,516 records) to develop AF classification and CHA2DS2-VASc score extraction models. For AF classification, XGBoost with TF-IDF achieved 95% accuracy on English cardiology data and 93% on Dutch data. Enhanced n-gram approaches improved Dutch performance from 0.71 to 0.78 F1-score by capturing negation patterns. For score extraction, fine-tuned MedRoBERTa.nl achieved 95% accuracy and 0.82 macro F1-score, excelling at identifying missing scores (0.967 F1-score). Quality analysis revealed documentation gaps: only 44.5% of 1,489 AF patients had documented CHA2DS2-VASc scores, with 26.7% having clinically significant scores requiring anticoagulation consideration. The research demonstrates feasibility of automated clinical text processing, combining traditional ML for classification with transformers for extraction. Results exceeded hypothesized thresholds (90% for classification, 85% for extraction) while highlighting cross-language processing challenges and class imbalance issues. Future work includes multi-label AF classification, improved embeddings, and continuing research on the automated reporting of other quality indicators for patients with AF.