QUANTIFICATION OF POLITENESS: MULTI-CLASS CLASSIFICATION OF PATIENT INTERVIEWING PHRASES FOR IMPROVING CLINICIAN BEDSIDE MANNER

Abstract

This thesis introduces work on quantifying politeness for implementation into a classification machine learning model to improve the bedside manner of healthcare practitioners with patient-centric technology design. Using AI to predict politeness allows for virtual patient simulations to give live feedback to clinicians, as opposed to post-interviewing feedback in current education practices. The AI simulated interview mirrors the traditional interviewing experience, as polite and impolite stimuli are met with appropriate patient responses. To obtain the model’s training data, a crowdsourcing survey was released to analyze the disparity between patient-clinician perceptions of politeness on common patient interviewing phrases. The aggregated patient data was input into a Naïve Bayes multi-class classification model to construct an algorithm for predicting the politeness of phrases on a patient-derived baseline. The results of the survey, while not statistically significant, show an observable gap between patient and clinician perceptions, highlighting the biases of clinician self-evaluation, as well as key linguistic features for defining politeness. The model returns a perfect accuracy upon its test of the dataset and is bolstered by near-perfect average accuracy across the model’s cross-validation. However, acquiring a more comprehensive dataset is necessary for determining this task’s true performance, although the algorithm’s structure is stable for implementation. Overall success of the study proves the ability to operationalize qualitative variables and build concrete methods to teach subtle behaviors and emotions. Likewise, the study emphasizes the need for patient consideration in developing standards of care, as well as showing the potential for AI in healthcare education.