Attention-Based Deep Entropy Active Learning Using Lexical Algorithm for Mental Health Treatment.

Area of Science:

• Computational linguistics
• Mental health informatics
• Machine learning for healthcare

Background:

• Internet-Delivered Psychological Treatment (IDPT) is crucial for mental health but complicated by overlapping emotions and the need for large, labeled datasets.
• Creating emotion-aware labeled datasets for IDPT is time-consuming and faces challenges like vocabulary size and data source variability.
• Personalized mental health interventions require efficient methods for extracting linguistic properties and segmenting emotions from text.

Purpose of the Study:

• To enhance trainable instances for IDPT applications by developing a semantic clustering mechanism.
• To improve the accuracy and efficiency of Natural Language Processing (NLP) models in identifying mental health conditions, specifically depression symptoms.
• To leverage unlabeled online forum data for training more robust personalized mental health interventions.

Main Methods:

• Proposed a method combining Natural Language Processing (NLP) with attention-based in-depth entropy active learning.
• Utilized synonym expansion via semantic vectors for clustering unlabeled text based on contextual semantic information.
• Implemented a cyclical active learning process where selected unlabeled text is iteratively added to the training set to update the model.

Main Results:

• The proposed synonym expansion semantic vectors enhanced training accuracy without negatively impacting performance.
• A bidirectional Long Short-Term Memory (LSTM) network with an attention mechanism achieved a 0.85 Receiver Operating Characteristic (ROC) curve on a blind test set.
• The method improved the detection rate of depression symptoms from online forum text by effectively utilizing unlabeled data.

Conclusions:

• The developed semantic clustering and synonym expansion method is effective in increasing trainable instances for NLP in mental health.
• Attention-based LSTM models demonstrate strong performance in classifying mental health symptoms using learned embeddings.
• This approach offers a viable solution for improving personalized mental health interventions by efficiently processing large volumes of unlabeled online data.