Exploring Necessary Conditions for High and Low Patient Ratings in Online Healthcare Consultation: An LLM-Based Weak Supervision Approach

Abstract

The prevalence of the "positive rating plateau" in Online Healthcare Consultation (OHC) hinders a clear understanding of the critical constraints required for service improvement. This study aims to identify the necessary conditions for high and low patient ratings. We utilized a dataset of 517,026 patient narrative reviews and adopted an LLM-based weak supervision approach to quantify 12 emotion categories. Subsequently, Necessary Condition Analysis (NCA) was conducted on patient reviews for 625 physicians, to identify necessary emotion categories and their bottleneck levels for high (5-star) and low (1-3 star) ratings. The results revealed that positive emotions are not necessary for high ratings. In contrast, Disappointment and Distrust (d=0.390, p<0.001) emerged as the core necessary condition for low ratings. Additionally, Price Complaints (d=0.196, p<0.001) and Poor Communication (d=0.180, p<0.001) exhibited distinct bottleneck characteristics across varying low-rating rates. Beyond informing a "stepwise improvement strategy" for physicians to place greater emphasis on caring practices and optimizing bilateral service design for platforms, this study establishes a low-cost and high-precision analytical paradigm fusing LLM-based weak supervision with NCA to address the semantic complexity and data sparsity inherent in unstructured medical text.

Keywords: Online Healthcare Consultation (OHC), Patient Ratings, Large Language Model (LLM), Necessary Condition Analysis (NCA), Weak Supervision

DOI: 10.54941/ahfe1007737