Evaluating sentiment analysis models in healthcare: addressing bias and enhancing interpretability

Advancing trustworthy AI applications in healthcare necessitates systems that are not only high-performing but also capable of explaining decisions and addressing biases, particularly in critical tasks like sentiment analysis on clinical narratives and patient feedback. Conventional sentiment analysis methods, while effective in general applications, struggle with domain shift, linguistic variability, and ambiguous labeling in healthcare, limiting their interpretability and fairness in clinical contexts. To overcome these limitations, a novel sentiment analysis framework is proposed to improve both accuracy and interpretability.

This framework employs a formal probabilistic modeling approach that incorporates fine-grained sentiment granularity and domain-aware priors. Central to the framework is the Sentiment Modulated Encoding Network (SMEN), a transformer-based architecture featuring a gating mechanism that dynamically enhances sentiment-relevant features across network layers, enabling rich sentiment representation learning without external resources. Additionally, the Context Polarity Decoupling Scheme (CPDS) disentangles sentiment from domain-specific artifacts through a multi-stage adversarial and contrastive training process, accompanied by a polarity explanation module that provides token-level interpretability.

Together, SMEN and CPDS form a robust system capable of producing domain-invariant and explainable sentiment predictions. Experimental results on multiple healthcare datasets demonstrate superior generalization and more transparent model attributions compared to existing approaches. This research contributes to the development of explainable and bias-resistant AI tools for healthcare and highlights potential avenues for interdisciplinary exploration at the interface of affective computing and clinical informatics.