Elisa Daffré, Marco Alifano, Bastien Rance
These findings highlight the potential of NLP to improve clinical data extraction by increasing efficiency, reducing human error and providing more comprehensive data sets. While traditional NLP methods using dictionaries and regular expressions provide interpretability and reliability, more advanced large language models (LLMs) could further enhance adaptability and data processing capabilities. A hybrid approach integrating both methods could provide a balance between accuracy, efficiency and scalability. Our study highlights the importance of NLP in clinical research and its potential to refine data-driven insights in oncology and beyond.
BACKGROUND: Lung cancer remains one of the leading causes of cancer-related mortality worldwide, with poor survival rates largely due to late diagnosis and the aggressive nature of the disease. Accurate and efficient extraction of clinical data is critical to improving research and patient outcomes. Traditional manual data collection methods, while reliable, are time consuming and prone to error. In this study, we evaluate the effectiveness of natural language processing (NLP) techniques in extracting relevant clinical data from unstructured medical records and compare them to manual data entry.
METHODS: We analyzed a cohort of 70 patients undergoing lung cancer surgery, using both manual data extraction using REDCap and NLP-based automated extraction from discharge summaries. Our NLP pipeline included sentence tokenisation, negation detection, dictionary-based entity recognition, fuzzy matching and regular expression-based named entity recognition (NER) to capture key clinical variables. The extracted data was validated against a senior surgeon's manual review to assess precision, sensitivity, specificity and overall performance.
RESULTS: Our results showed that NLP outperformed manual data entry for most of the metrics evaluated, achieving higher precision [100% for body mass index (BMI), Tumor, Node, Metastasis (TNM) classification, surgical approach and patient weight], higher sensitivity (100% for BMI and TNM classification) and superior F1 scores, particularly in complex data fields such as spirometry and tumors localization. Manual data entry showed lower sensitivity and specificity, indicating a higher rate of false negatives and false positives.
CONCLUSIONS: These findings highlight the potential of NLP to improve clinical data extraction by increasing efficiency, reducing human error and providing more comprehensive data sets. While traditional NLP methods using dictionaries and regular expressions provide interpretability and reliability, more advanced large language models (LLMs) could further enhance adaptability and data processing capabilities. A hybrid approach integrating both methods could provide a balance between accuracy, efficiency and scalability. Our study highlights the importance of NLP in clinical research and its potential to refine data-driven insights in oncology and beyond.