Analyzing Sleep Behavior Using BERT-BiLSTM and Fine-Tuned GPT-2 Sentiment Classification: Comparison Study

Area of Science:

• Computational linguistics
• Clinical informatics
• Sleep medicine

Background:

• Sleep disorder diagnosis is complex, often showing a gap between objective clinical data and subjective patient experiences.
• Individual perception of sleep quality and latency can vary significantly.

Purpose of the Study:

• To investigate the alignment between subjective patient experiences and objective measurements in sleep disorder assessment.
• To explore how clinical narratives can provide insights into sleepiness perception.

Main Methods:

• Developed an aspect-based sentiment analysis method using large language models (Falcon 40B, Mixtral 8X7B) to analyze clinical narratives.
• Identified sleep behavior aspects (day sleepiness, sleep quality, fatigue) and assigned sentiment scores (0-1) using BERT-BiLSTM (78% accuracy) and GPT-2 (87% accuracy).

Main Results:

• Approximately 15% of 100 patients showed discrepancies between subjective (Karolinska Sleepiness Scale) and objective (Multiple Sleep Latency Test) daytime sleepiness assessments.
• Sentiment analysis of clinical narratives revealed statistically significant divergence in sleepiness perception (P=.047), outperforming standardized measures.
• Narrative free text analysis highlighted the importance of subjective sources in assessing fatigue and sleepiness.

Conclusions:

• The developed sentiment analysis method can reveal critical insights into patient self-perception versus clinical evaluations.
• This approach aids clinicians in identifying patients who may require objective verification of self-reported sleep symptoms.
• Integrating narrative free text analysis enhances the comprehensive assessment of sleep disorders.