A Comparative Study of Healthcare Emotion and Sentiment Analysis using Machine Learning approach

Healthcare-related communication is now avail-able in many forms, including patient reviews, clinical notes, telemedicine conversations, and online feedback. These texts can reveal how patients respond to their experiences, but sentiment labels alone may not capture the full meaning of what is written. For example, a negative statement about a hospital service may reflect dissatisfaction, whereas a negative statement about a diagnosis may express fear or uncertainty.

In this study, we compare different approaches for identifying sentiment and emotion in healthcare text. The comparison covers four conventional machine-learning classifiers, namely Naive Bayes, Logistic Regression, Support Vector Machine, and Random Forest, together with Bi-LSTM, BERT, and ClinicalBERT. Healthcare datasets such as MIMIC-III, CADEC, and MedDialog are considered to represent different types of clinical and patient-authored text.

The evaluation is planned using accuracy, precision, recall, macro-F1, and class-wise results, along with factors such as com-putational cost, calibration, interpretability, and robustness. We also consider class imbalance and the differences between clinical documentation and patient-written text. The experimental runs are not yet complete, so no numerical results are presented as final findings. Instead, this paper describes the evaluation procedure and the comparisons that will be carried out. The resulting framework is intended to provide a consistent basis for future studies of sentiment and emotion analysis in healthcare communication, where model outputs can be considered alongside human judgment.