Based on deep learning, we train our models to perform clinical findings extraction, normalization,diagnosis prediction on clinical notes from emergency department (ED) at National Taiwan University Hospital (NTUH),develop an algorithm to suggest crucial medical terms. These modules enable our interactive system to assist physicians during medical interview by showing the patient＇s clinical findings, differential diagnosis,most differentiating diagnostic terms. These components facilitate the exploration of possible diagnosisthe convergence to the final diagnosis.

We propose two novel tasks, namely incremental diagnosis predictiondynamic diagnostic term suggestion. Incremental diagnosis prediction aims to provide differential diagnosis successively based on currently available patient history, which is incrementally documented throughout medical interviews. Dynamic diagnostic term suggestion aims at suggesting differential diagnosis-related medical terms that can converge the diagnoses most. We design training schemesalgorithms to achieve the above goals,conduct experiments to prove their effectiveness.

The system can be utilized in the content search for electronical medical records, standardize medical terminologies,further to establish index as the basis for patients searchingcase studies. Meanwhile, through the hints of key wordsdiagnosis prediction, which can assist physicians in conducting treatment, increase the efficacy of inquiries, lesser the time for searching diagnostic codeslower the possibilities for misdiagnosis. Moreover, it can assist young physiciansnon-specialist physicians in conducting medical treatment.