Ensemble learning for improved sentiment analysis in doctor-patient communication

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning in Healthcare
• Computational Linguistics

Background:

• Accurate sentiment analysis of clinician-patient interactions is crucial for evaluating healthcare quality and patient experience.
• Existing research lacks comprehensive benchmarking of advanced machine learning models for this specific task.
• Sentiment classification in clinical dialogue presents unique challenges due to nuanced language and context.

Purpose of the Study:

• To benchmark deep learning, transformer, and ensemble models for three-class sentiment classification (low/medium/high) in doctor-patient consultations.
• To address the gap in standardized evaluation of sentiment analysis models within the clinical domain.
• To identify the most effective model architecture for analyzing sentiment in anonymized doctor-patient dialogues.

Main Methods:

• Utilized a publicly available dataset of 3325 anonymized doctor-patient consultations.
• Evaluated Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Convolutional Neural Networks (CNN), CNN-LSTM, and Bidirectional Encoder Representations from Transformers (BERT).
• An ensemble model (hard voting over Logistic Regression, Random Forest, and Support Vector Classifier) was also tested using stratified five-fold cross-validation.

Main Results:

• The ensemble model achieved the highest accuracy (75.5% ± 0.5%), outperforming individual models including BERT (66.98% ± 0.6%).
• The ensemble demonstrated robust detection of high-severity interactions (F1 score: 90.8% ± 1.3%), though low-severity classification remained challenging.
• BERT offered the highest precision for low-severity interactions (65.5% ± 1.0%), while the ensemble improved recall (58.7% ± 1.0%).

Conclusions:

• Ensemble learning provides the strongest and most balanced performance for three-class sentiment classification in clinician-patient dialogue.
• Transformer models like BERT offer valuable precision for challenging low-severity cases, complementing ensemble approaches.
• Interpretability analyses enhance transparency, supporting clinical application; future work should explore multimodal and privacy-preserving models.