Back

Development and comparison of natural language processing models for abdominal aortic aneurysm repair identification and classification using unstructured electronic health records

Thompson, D. C.; Mofidi, R.

2024-12-12 health informatics
10.1101/2024.12.11.24318852 medRxiv
Show abstract

BackgroundPatient identification for national registries often relies upon clinician recognition of cases or retrospective searches using potentially inaccurate clinical codes, potentially leading to incomplete data capture and inefficiencies. Natural Language Processing (NLP) offers a promising solution by automating analysis of electronic health records (EHRs). This study aimed to develop NLP models for identifying and classifying abdominal aortic aneurysm (AAA) repairs from unstructured EHRs, demonstrating proof-of-concept for automated patient identification in registries like the National Vascular Registry. MethodUsing the MIMIC-IV-Note dataset, a multi-tiered approach was developed to identify vascular patients (Task 1), AAA repairs (Task 2), and classify repairs as primary or revision (Task 3). Four NLP models were trained and evaluated using 4,870 annotated records: scispaCy, BERT-base, Bio-clinicalBERT, and a scispaCy/Bio-clinicalBERT ensemble. Models were compared using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve. ResultsThe scispaCy model demonstrated the fastest training (2 mins/epoch) and inference times (2.87 samples/sec). For Task 1, scispaCy and ensemble models achieved the highest accuracy (0.97). In Task 2, all models performed exceptionally well, with ensemble, scispaCy, and Bio-clinicalBERT models achieving 0.99 accuracy and 1.00 AUC. For Task 3, Bio-clinicalBERT and the ensemble model achieved an AUC of 1.00, with Bio-clinicalBERT displaying the best overall accuracy (0.98). ConclusionThis study demonstrates that NLP models can accurately identify and classify AAA repair cases from unstructured EHRs, suggesting significant potential for automating patient identification in vascular surgery and other medical registries, reducing administrative burden and improving data capture for audit and research.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.