Clinical Dialogue Extraction Using Context-Aware Utterance Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for extracting information from dialogues in healthcare settings, such as patient-clinician interactions, are inefficient, require significant clinician effort, and fail to capture contextual information, leading to suboptimal EMR data quality and clinician burnout.
Innovation Solution
A machine learning-based system that uses a two-layer bidirectional GRU neural network for utterance classification, temporal tagging, and attribute classification to extract relevant entities and attributes from transcribed dialogues, incorporating linguistic context and time information to generate accurate EMR data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are used to automatically extract information from dialogues, then productivity and data quality improve, but device complexity increases
Solution Approach 1:
The information extraction system is divided into multiple specialized machine learning models: an utterance classification model that categorizes dialogue turns, an entity recognition model that identifies medical concepts, and an attribute classification model that extracts clinical features. This segmentation allows each model to specialize in a specific task, improving overall extraction accuracy while enabling modular deployment and maintenance.
Solution Approach 2:
The patent introduces intermediate processing layers including preprocessing modules that clean and normalize dialogue text, and postprocessing modules that structure the extracted information into standardized EMR formats. These intermediary components bridge the gap between raw dialogue data and structured clinical records, simplifying the integration of complex ML models into the healthcare workflow.
2Measurement precision
If comprehensive dialogue analysis is performed to capture contextual information, then measurement precision improves, but loss of time increases
Solution Approach 1:
The system performs preliminary utterance classification on each dialogue turn before entity recognition. The classification model pre-tags utterances with types such as patient symptoms, clinician questions, or administrative information, allowing the entity recognition model to focus computational resources on relevant segments. This preliminary action preserves contextual accuracy while reducing overall processing time.
Solution Approach 2:
The patent applies different levels of analysis intensity to different dialogue segments based on their classification. High-priority utterances containing clinical symptoms or diagnoses receive full contextual analysis with multiple ML models, while routine administrative utterances receive minimal processing. This selective application of analysis depth maintains measurement precision for critical information while reducing time loss across the entire dialogue.
Data Source
AI summary
Described herein are systems and methods of extracting information from a dialogue, the dialogue having transcription data associated therewith. In an embodiment, the method including: receiving the transcription data associated with the dialogue; classifying utterances in the transcription data using a trained classification machine learning model, the classification machine learning model trained using one or more corpora of historical data comprising previous dialogues labelled with utterance types; identifying entities in the transcription data; classifying attributes in the transcription data using a trained attribute machine learning model, the attribute machine learning model trained using one or more corpora of historical data comprising previous dialogues labelled with attributes; and outputting at least one of the utterances, the entities, and the attributes.