Application of Latent Semantic Analysis and Supervised Learning Methods to Automate Triage of Referral Letters for Spinal Surgery

Referral letters are the most common mean used by healthcare practitioners to exchange information relevant to patient care. However, their triage takes a significant amount of administrative resources, which may be amenable to automation. This study aims to evaluate a pilot study for automating the triage of referral letters sent to a spine surgery department in secondary care using Natural Language Processing (NLP) techniques and supervised machine learning methods. Text data in referral letters were represented using Term Frequency-Inverse Document Frequency (TF-IDF) scores and Latent Semantic Analysis (LSA) and fed into supervised learning models to classify it into urgent and non-urgent. Their outcomes have been evaluated using performance metrics (precision, recall, F1, area under the curve (AUC), and accuracy), and their clinical value is assessed based on decision curve analysis. The AdaBoost classifier obtained the maximum AUC score (61.8%) in the test set. The model also reached an F1 score of 53.1% with 48.6% and 58.6% of recall and precision scores, respectively. Comparing its net benefit against baseline clinical alternatives proves that the model has potential clinical value, and it might be a valuable tool for automating the referrals' triaging process. This study successfully demonstrated the potential for automating the triage of referrals and provides a foundation for further work.