Multi-label Classification for Depressive Disorder Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a lack of datasets labeled based on the DSM-5 diagnostic criteria for depressive disorder, hindering research and diagnosis in Korea, where depressive disorder prevalence is high among young people.
Innovation Solution
A multi-label classification method using transfer learning with a KoBERT model for initial labeling and a GRU model for prediction, generating labeled data based on DSM-5 criteria, and performing data augmentation to enhance the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If transfer learning with KoBERT model is performed for initial labeling, then labeling efficiency is improved, but manufacturing precision of labeled data quality may deteriorate
Solution Approach 1:
The patent introduces an intermediary manual review process between automated KoBERT labeling and final labeled dataset creation. Experts review and verify the labels generated by KoBERT, ensuring high quality while maintaining efficiency. This intermediary step resolves the contradiction by combining automated productivity with expert quality control.
Solution Approach 2:
The patent replaces manual expert labeling (mechanical human process) with automated KoBERT model labeling (algorithmic process) for initial labeling, then uses a hybrid approach where experts only review uncertain cases. This substitution improves productivity while maintaining precision through selective expert intervention.
2Measurement precision
If multi-label classification based on DSM-5 criteria is implemented, then diagnostic accuracy is improved, but device complexity increases
Solution Approach 1:
The patent segments the complex DSM-5 diagnostic criteria into multiple independent binary classification labels, each corresponding to a specific diagnostic criterion. The GRU model performs multi-label classification by predicting each criterion separately, which simplifies the overall system architecture while maintaining diagnostic accuracy. This segmentation approach allows the complex diagnostic process to be broken down into manageable classification tasks.
3Quantity of substance
If data augmentation is performed on residual data, then training dataset size is improved, but loss of time increases
Solution Approach 1:
The patent applies data augmentation only to residual data (data not already labeled by KoBERT) rather than the entire dataset. This partial action approach efficiently increases the training dataset size by focusing augmentation resources on the most valuable unlabeled portions, minimizing time loss while maximizing dataset quantity for model training.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A training method according to an embodiment may include: performing transfer learning a first artificial neural network model based on a consultation dataset sentence; inputting depressive disorder-related expression data into the first artificial neural network model, labeling the depressive disorder-related expression data according to depressive disorder diagnosis criteria, and generating labeled depressive disorder-related expression data; and training a second artificial neural network model based on the labeled depressive disorder-related expression data so that a second artificial neural network model may output the depressive disorder diagnosis criteria corresponding to input data.