ICD-10 Neoplasm Location Using Text Classification Models in Spanish Electronic Health Records.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Oncology Research

Background:

• Spanish healthcare systems store vast clinical data as unstructured text in electronic health records (EHRs).
• Automatic extraction of critical information, such as primary neoplasm location, from EHRs is essential for clinical analysis and Real-World Evidence (RWE) studies.
• Neoplasm location data is vital for oncology research and is often coded within the International Classification of Diseases, 10th Revision (ICD-10).

Purpose of the Study:

• To explore the classification of Spanish medical documents for extracting primary neoplasm locations.
• To evaluate the performance of various Natural Language Processing (NLP) methodologies for this task.

Main Methods:

• Utilized a private corpus of 23,704 real clinical EHRs.
• Developed four NLP methodologies: traditional machine learning (ML), ensemble ML, recurrent neural networks (RNNs), and Transformer-based models.
• Classified neoplasm locations into 12 primary organ groupings and 29 specific locations.

Main Results:

• Traditional ML models (XGBoost, SVMs) outperformed RNNs and Transformers, achieving an F1-score of 0.938 for 12-class and 0.838 for 29-class classification.
• The pre-trained RoBERTa-Base-Biomed model achieved an F1-score of 0.808 for the 29-location problem.
• Transformer models demonstrated superior generalization capacity on an external corpus.

Conclusions:

• Traditional ML models are highly effective for extracting primary neoplasm locations from Spanish EHRs.
• NLP techniques, particularly traditional ML, offer a robust solution for unlocking valuable oncological data within unstructured clinical text.
• Further research with Transformer models may enhance generalization for diverse clinical datasets.