Leveraging Language Embeddings from EMA Surveys to Predict Perceived Social Isolation among Stroke Survivors

Area of Science:

• Neuroscience
• Psychology
• Computer Science

Background:

• Perceived social isolation (PSI) negatively impacts stroke survivors' emotional well-being.
• Ecological Momentary Assessment (EMA) is used to identify PSI precursors.
• Current prediction methods using handcrafted features lack semantic depth.

Purpose of the Study:

• To develop a novel approach for predicting PSI in stroke survivors.
• To leverage language embeddings for processing structured EMA data.
• To improve real-time monitoring of psychosocial outcomes.

Main Methods:

• Collected 11,802 EMA surveys from 218 stroke survivors.
• Extracted language embeddings from EMA surveys using a pre-trained model.
• Trained an autoencoder for latent representation and PSI prediction.

Main Results:

• Achieved accurate PSI prediction with a weighted F1 score of 0.84 and AUPRC of 0.92.
• Outperformed traditional handcrafted features in prediction accuracy.
• Demonstrated optimized validity and usability using only three selected questions.

Conclusions:

• The proposed language embedding approach offers an efficient method for real-time PSI monitoring in stroke survivors.
• This method enhances early intervention and personalized care strategies.
• The study highlights the potential of NLP in understanding and managing psychosocial health post-stroke.