FasTag: Automatic text classification of unstructured medical narratives

Area of Science:

• Computational biology
• Bioinformatics
• Veterinary medicine

Background:

• Manual coding of unstructured clinical narratives in electronic health records is labor-intensive and prone to inaccuracies.
• Accurate and timely coding is crucial for billing and cohort identification.

Purpose of the Study:

• To automate the assignment of top-level International Classification of Diseases version 9 (ICD-9) codes to human and veterinary clinical records.
• To evaluate the performance of Long Short-Term Memory (LSTM) recurrent neural networks (RNNs) for this task.
• To assess model portability and compare performance against baseline models.

Main Methods:

• Trained LSTM recurrent neural networks (RNNs) on a large dataset of human (52,722) and veterinary (89,591) clinical records.
• Investigated separate-domain and combined-domain models, as well as model portability.
• Compared LSTM performance against Decision Trees (DT) and Random Forests (RF) models, with and without MetaMap Lite data transformation.

Main Results:

• LSTM-RNNs achieved average weighted macro F1 scores of 0.74 for veterinary and 0.68 for human records.
• The model trained on veterinary data showed high accuracy (F1=0.91) for 'neoplasia' in veterinary data and moderate accuracy (F1=0.70) in human data.
• LSTM models slightly outperformed DT and RF models.

Conclusions:

• Automated ICD-9 code assignment using LSTM-RNNs is a scalable and accurate method for both human and veterinary data.
• This approach facilitates rapid cohort identification, particularly for comparative oncology research.
• Enables human and veterinary health data to inform each other, advancing medical research.