S.Y.TAY1, W.Y.LIM1, M.CHEN2, C.H.LEE1, K.JANARDANAN3, Y.H.G.TAY4
Tan Tock Seng Hospital1, National Centre for Infectious Diseases2, MOH Holdings Pte Ltd (MOHH)3, Lee Kong Chian School of Medicine4
Emergency Department free text notes (EDFTN) may contain additional diagnoses that have not been formally coded in electronic medical record (EMR) systems. It may be possible to extract these diagnoses to aid in early detection of some infectious disease syndromes (IDS) such as acute respiratory infections, gastroenteritis, fever with rash (eg for measles or dengue), and severe community-acquired pneumonias. This study uses an easily-available, off-the-shelf open-source transformer deep neural network (TfDNN) to perform natural-language processing (NLP) multi-label classification on EDFTN for the above four IDS’s.
15955 EDFTN were classified by a team of human coders, working in pairs, into one or more of five classes: “Acute Respiratory Infection” (ARI), “Gastroenteritis” (GE), “Acute Febrile Breathlessness” (Acute SOB), “Fever with Rash” (Fever/Rash) and “Others” (Others). 70% of the dataset was then used to train a RoBERTa-based TfDNN available from the NLP SpaCy package, 20% for validation, and 10% reserved as a test set. The performance of the TfDNN was measured against human coders.
TfDNN achieved macro-averaged area under ROC curve of 0.973 (ARI), 0.968 (GE), 0.918 (Acute SOB), 0.959 (Fever/Rash) and 0.934 (Others) respectively
TfDNN use should be considered for surveilling ARI, GE and Fever with Rash.