Deep Learning and Machine Learning Model Comparison for Diagnosis Detection from Medical Records

Abstract

Structured data is needed in hospitals as a means of exchanging information between doctors, nurses, pharmacy department, coder/medical record section, and administration section. Structured data improves interoperability and uniformity of interpretation between entities working in the hospital. One of the stages of the process to generate structured data such as ICD codes is to detect diagnoses from medical records written by doctors. The entity in charge of interpreting medical records and determining the relevant ICD code according to the doctor's diagnosis written in the medical record is called a coder. In determining the ICD code, the coder looks for the patient's diagnosis in the medical record. However, coders with minimal experience may find it challenging to find a patient's diagnosis. This will cause inaccuracy in determining the ICD code to diagnose the patient's disease. This research constructed a predictor in the diagnostic recommendation system. We developed a supervised deep learning model, which is an LSTM model, and a Stochastic Gradient Descent model as a baseline in this study. Compared to the Stochastic Gradient Descent model, it was discovered that the proposed LSTM model produced the best results, reaching up to 98% accuracy in 14 epochs.