Researchers Aryadi Aryadi, Harris Gadih Pratomo, Kanoksak Wattanachote and colleagues published a 2026 Frontiers in Oral Health study on automated patient emotion recognition in restorative dentistry consultations. The team built an emotion-labeled dental consultation corpus because emotional communication matters in dentist-patient interactions, while domain-specific resources for automated emotion recognition in dental consultations remain limited.

The researchers transcribed audio recordings from 12 restorative dentistry consultations and segmented them into patient utterances. The final dataset contained 2,160 utterances, which three trained annotators independently labeled with a standardized framework across six emotion categories: Neutral, Fear, Sad, Happy, Angry, and Disgust. Annotation agreement was measured with Cohen's kappa (agreement beyond chance), and the corpus reached high reliability with κ = 0.89.

Neutral was the most common category at 23.15%, followed by Fear at 19.44% and Disgust at 18.98%. Fear was also the most difficult category to classify, even though it was one of the most frequent emotions in the corpus. The team trained a Bidirectional Long Short-Term Memory, or BiLSTM, classifier and compared it with transformer-based baseline models (models using attention over text), including BERT and RoBERTa.

The baseline BiLSTM model achieved a macro-F1 score of 0.87, while BERT reached 0.89 and RoBERTa reached 0.91. Hybrid frameworks used LLM-based confidence re-ranking on low-confidence BiLSTM predictions, with three open-source LLMs evaluated: Llama 3.1, DeepSeek-R1, and Qwen 2.5. BiLSTM + Qwen 2.5 produced the highest macro-F1 score of 0.92, but the performance gains were not statistically significant, with all p > 0.05. The authors described the findings as preliminary evidence for selective LLM-based re-ranking in emotionally ambiguous patient utterances.

Researchers Aryadi Aryadi, Harris Gadih Pratomo, Kanoksak Wattanachote and colleagues published a 2026 Frontiers in Oral Health study on automated patient emotion recognition in restorative dentistry consultations. The team built an emotion-labeled dental consultation corpus because emotional communication matters in dentist-patient interactions, while domain-specific resources for automated emotion recognition in dental consultations remain limited.

The researchers transcribed audio recordings from 12 restorative dentistry consultations and segmented them into patient utterances. The final dataset contained 2,160 utterances, which three trained annotators independently labeled with a standardized framework across six emotion categories: Neutral, Fear, Sad, Happy, Angry, and Disgust. Annotation agreement was measured with Cohen's kappa (agreement beyond chance), and the corpus reached high reliability with κ = 0.89.

Neutral was the most common category at 23.15%, followed by Fear at 19.44% and Disgust at 18.98%. Fear was also the most difficult category to classify, even though it was one of the most frequent emotions in the corpus. The team trained a Bidirectional Long Short-Term Memory, or BiLSTM, classifier and compared it with transformer-based baseline models (models using attention over text), including BERT and RoBERTa.

The baseline BiLSTM model achieved a macro-F1 score of 0.87, while BERT reached 0.89 and RoBERTa reached 0.91. Hybrid frameworks used LLM-based confidence re-ranking on low-confidence BiLSTM predictions, with three open-source LLMs evaluated: Llama 3.1, DeepSeek-R1, and Qwen 2.5. BiLSTM + Qwen 2.5 produced the highest macro-F1 score of 0.92, but the performance gains were not statistically significant, with all p > 0.05. The authors described the findings as preliminary evidence for selective LLM-based re-ranking in emotionally ambiguous patient utterances.