What Single-Topic Summaries Miss in Hospital Reviews-Aspect-Level Evaluative Structure Using Generative Pretrained Transformer-Based Sentiment Analysis: Content Analysis.

Abstract

Health care service quality is inherently multidimensional; yet, the dominant practice in applied text analysis assigns each patient review to a single topic via Latent Dirichlet Allocation (LDA). This simplification may systematically compress evaluative information when patients discuss multiple service dimensions with varying sentiments within the same review. This study compared the dominant-topic operationalization commonly used in applied LDA research with Generative Pretrained Transformer (GPT)-based aspect-based sentiment analysis (ABSA) to examine (1) the extent to which patient reviews contain multiple service-quality aspects and how single-topic summaries represent or obscure this structure, (2) the prevalence and patterning of mixed-sentiment reviews, and (3) whether positive and negative reviews differ in aspect comention profiles, before and after adjusting for marginal aspect prevalence. We analyzed 5467 Google Reviews posted in 2024 from all 24 medical centers in Taiwan. LDA (K=7 topics) and GPT-based ABSA with structured prompts were applied to the same corpus, with the 7 ABSA categories aligned to the LDA topic labels for a controlled but information-asymmetric comparison. Two independent annotators achieved interrater reliability of Cohen κ=0.82; against the consensus gold standard, GPT-4o achieved an accuracy of 0.89, a weighted F₁ of 0.89, and a Cohen κ of 0.78. Mixed-sentiment reviews were identified as those containing both positive and negative aspect evaluations. Rating-stratified network analysis compared aspect comention patterns between positive and negative reviews using Jaccard similarity, with pointwise mutual information as a prevalence-adjusted sensitivity analysis. Aspect-bearing reviews discussed an average of 2.05 distinct aspects (95% bootstrap CI 2.02-2.08), yielding an illustrative 51.2% representational compression estimate under dominant-topic assignment (95% bootstrap CI 50.6%-51.9%). A soft-assignment LDA baseline reduced count-level compression to 1.7%, but semantic alignment with ABSA aspects remained limited (mean set Jaccard=0.33), and topic assignments carry no aspect-level sentiment polarity. Among multiaspect reviews, 11.0% exhibited cross-aspect mixed sentiment, with Technical-Functional Divergence-praising technical quality while criticizing functional quality-appearing in 61.6% of these cases. Clinical dimensions were more frequently comentioned in positive reviews and operational dimensions in negative reviews; however, pointwise mutual information analysis indicated that these differences were substantially confounded with marginal aspect prevalence rather than reflecting differential co-occurrence tendencies. In this corpus, dominant-topic assignment compressed multiaspect patient feedback; soft-assignment LDA recovered topic counts but did not restore semantic alignment or aspect-level sentiment polarity. A nontrivial subset of reviews exhibited cross-aspect mixed sentiment, most commonly praising clinical competence while criticizing functional service dimensions, and positive and negative reviews discussed different constellations of quality dimensions-differences that primarily reflect which aspects patients discuss rather than prevalence-independent associations. Aspect-level analysis that preserves both multidimensional structure and sentiment polarity may help organize patient feedback at a more diagnostically specific level than single-topic summaries.