Automatic multilabel detection of ICD10 codes in Dutch cardiology discharge letters using neural networks

Abstract Standard reference terminology of diagnoses and risk factors is crucial for billing, epidemiological studies, inter/intranational comparisons diseases. The International Classification Disease (ICD) a standardized widely used method, but the manual classification an enormously time-consuming endeavor. Natural language processing together with machine learning allows automated structuring using ICD-10 codes, limited performance models, necessity gigantic datasets, poor reliability terminal parts these codes restricted clinical usability. We aimed to create high performing pipeline reliable in free medical text cardiology. focussed on frequently well-defined three- four-digit that still have enough granularity be clinically relevant such as atrial fibrillation (I48), acute myocardial infarction (I21), or dilated cardiomyopathy (I42.0). Our uses deep neural network known Bidirectional Gated Recurrent Unit Neural Network was trained tested 5548 discharge letters validated 5089 procedural letters. As practice may labeled more than one code, we assessed single- multilabel main cardiovascular factors. investigated both entire body only summary paragraph, supplemented by age sex. Given privacy-sensitive information included letters, added de-identification step. high, F1 scores 0.76–0.99 three-character 0.87–0.98 four-character best when complete Adding variables age/sex did not affect results. For model interpretability, word coefficients were provided qualitative assessment manually performed. Because its performance, this can useful decrease administrative burden classifying serve scaffold reimbursement research applications.