Applying Speech Recognition and Language Processing Methods to Transcribe and Structure Physicians’ Audio Notes to a Standardized Clinical Report Format

Abstract

Clinical documentation is an audio recording of the clinical encounter by the specialist which is subsequently manually transcribed to be added to the patient’s medical record. The current clinical documentation process is tedious, error-prone and time-consuming, more so for specialists working in the emergency department given the rapid turnaround of high-acuity patients. In this thesis, we investigate methods to automate the clinical documentation processes for a pediatric emergency department, leading to the generation of a SOAP report of the clinical encounter. Our approach involves (a) speech recognition to transcribe the audio recording of the clinical encounter to a textual clinical encounter report; and (b) identifying and classifying the sentences within the textual report in terms of the standard SOAP format for clinical reports. For speech recognition, we worked with the DeepSpeech application and used recurrent neural network and n-gram based methods, augmented with medical terminologies and heuristics, to develop domain-specific acoustic and language models. Our approach resulted in a reduction of 49.02% of critical errors as compared to the baseline acoustic and language models provided by DeepSpeech. For generating a SOAP report from the clinical text, we extended an exemplar-based concept detection algorithm to learn a sentence classifier to identify and annotate the clinical sentences in terms of subjective, objective, assessment and plan. Our SOAP classifier achieves a precision of 0.957 (subjective), 0.919 (objective), 0.626 (assessment) and 0.82 (plan).