Ensemble learning for improved sentiment analysis in doctor–patient communication

Abstract

Objective: To fill the benchmarking gap in clinician–patient sentiment analysis, we compare deep learning, transformer, and ensemble models for three-class (low/medium/high) sentiment classification in doctor–patient consultations. Methods: We used a publicly available dataset of 3325 anonymized doctor–patient consultations from the Hugging Face repository (mahfoos/Patient-Doctor-Conversation) labeled as low, medium, or high severity. Preprocessing included text cleaning, tokenization, and padding; class balancing was applied only within the training split of each fold. Models evaluated were long short-term memory (LSTM), bidirectional LSTM (BiLSTM), convolutional neural networks (CNN), CNN–LSTM, and bidirectional encoder representations from transformers (BERT); an ensemble (hard voting over Logistic Regression, Random Forest, and Support Vector Classifier (SVC)) was also tested. Evaluation used stratified five-fold cross-validation, with metrics reported as mean ± SD across outer test folds (accuracy; macro-averaged precision/recall/F1). Interpretability was examined via BERT attention and feature attributions. Results: The ensemble achieved the highest accuracy (75.5 ± 0.5), outperforming BERT (66.98 ± 0.6), CNN–LSTM (65.68 ± 0.9), CNN (64.17 ± 0.8), BiLSTM (64.82 ± 0.7), and LSTM (58.66 ± 0.19). Class-wise analysis showed robust detection of high-severity interactions (e.g. ensemble F1 = 90.8 ± 1.3), while low-severity remained most challenging; the ensemble improved class 0 recall (58.7 ± 1.0), and BERT provided the highest class 0 precision (65.5 ± 1.0). Conclusion: Under stratified five-fold cross-validation, ensemble learning delivered the strongest and most balanced performance for three-class sentiment classification of clinician–patient dialogue, while transformers offered complementary precision on difficult cases. Attention- and feature-attribution analyses improved transparency, supporting clinical interpretability. Future work should scale to larger, multimodal (text/audio/vision) and multilingual datasets, and develop privacy-preserving, lightweight models for real-time deployment in clinical settings.