Language Model Training for Real-Time Depression Symptom Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current methods for detecting depression symptoms and intensity through social media analysis and Hamilton Depression Rating Scale surveys face challenges in providing real-time services and accurately classifying symptoms of Major Depressive Episode (MDD) using existing technologies.

Innovation Solution

A text-based emotion detection language model training method that involves receiving and processing user input text to generate clean data, detecting emotion data, and training a language model to predict depression intensity and classify symptoms using a combination of first and second language models, including a BERT-based model for intensity calculation and DSM-5-based classification.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If real-time conversation services like Chatbot are used to predict users vulnerable to depression, then real-time service capability is improved, but the requirement for user status information makes it difficult to implement

Engineering Contradiction:

Improvereal-time service capabilityVSAvoiduser status information requirement

Core Design Contradiction:

SpeedVSEase of operation

Solution Approach 1:

The system enables users to self-diagnose their depression symptoms by inputting their own text data, eliminating the need for external status information collection. The language model processes user-generated content directly to identify symptoms and predict depression vulnerability, making the service both real-time and easy to access without requiring additional user information.

Inventive Principle:

Principle #25Self-service

2Measurement precision

If Hamilton Depression Rating Scale surveys are used to determine depression degree, then depression level classification is improved, but the requirement for survey completion makes it difficult to apply in real-time services

Engineering Contradiction:

Improvedepression level classificationVSAvoidsurvey completion time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system extracts depression symptom information directly from user input text without requiring completion of structured survey forms. The language model identifies and extracts relevant symptom indicators from natural language input, enabling rapid depression assessment that maintains classification precision while eliminating the time burden of traditional survey completion.

Inventive Principle:

Principle #2Taking out (Extraction)

3Reliability

If previous depression classification studies are used, then depression detection capability is improved, but the nine symptoms for diagnosing depression are not classified making it difficult to determine specific symptoms

Engineering Contradiction:

Improvedepression detection capabilityVSAvoidspecific symptom classification

Core Design Contradiction:

ReliabilityVSLoss of information

Solution Approach 1:

The system segments depression detection into two distinct components: a first language model that predicts overall depression vulnerability and a second language model that classifies specific depression symptoms according to DSM-5 criteria. This segmentation preserves the reliability of depression detection while recovering detailed symptom-level information that was lost in previous classification studies.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20250021883A1Method for training language model, device for emotion diagnosis using pre-trained language model, and storage medium storing instructions to perform method for training language model

Publication Date: 2025.01.16 RES & BUSINESS FOUND SUNGKYUNKWAN UNIV

AI summary

There is provided a method for training text-based emotion detection language model. The method comprises receiving text data; generating clean data by removing errors included in the received text data; detecting emotion data related to a specific emotion from the clean data and generating a training dataset including the detected emotion data; and training the language model using the training dataset.