Integrating BERT Representations and Psycholinguistic Features for Emotion Recognition in Clinical Texts

In clinical texts, recognizing emotions is crucial for monitoring mental health, though it is still a tough task because of the way language is used and the particular terms in this field. The hybrid framework suggested in this research uses ClinicalBERT for context and LIWC and the NRC Emotion Lexicon for psycholinguistic features to help improve multi-label emotion classification in clinical narratives. The data has been de-identified and annotated with anger, anxiety, sadness, joy, fear and neutral emotions and there is good agreement between annotators (Cohen’s κ = 0.81). Three approaches were studied: using Random Forest with psycholinguistic features, ClinicalBERT-based Multilayer Perceptron (MLP) and a hybrid MLP that combines both sets of features. The hybrid model was better than the baselines, achieving mean scores of 0.884 (±0.011) accuracy, 0.854 (±0.012) Micro-F1, 0.814 (±0.013) Macro-F1 and 0.924 (±0.011) AUC which were statistically significant (ANOVA p < 0.005; Cohen’s d = 1.24–2.89). The SHAP analysis found that ClinicalBERT contributed more than two-thirds of the predictive ability, while psycholinguistic features contributed the rest, making the model easier to understand. This method works to solve main problems in healthcare AI by ensuring the accuracy of predictions and making the results easy to understand. It backs up trustworthy use in clinics by giving clear and reliable emotion predictions that can support decisions, monitor risks and be used in digital mental health services. The results suggest that using deep learning together with existing psychological tools improves emotional detection in healthcare.