JMIR Publications

ABSTRACT

Background:

Online reviews provide unsolicited accounts of health care experiences and may complement conventional patient-experience surveys. However, analyses based on overall sentiment can obscure the specific service domains discussed, while aspect-level approaches require distinction between detecting an entity and attributing sentiment to it.

Objective:

This study aimed to characterize service entities and lexical discourse patterns in online reviews of a regional public health care system, compare primary and hospital care, identify where negatively rated feedback was concentrated across facility types, and evaluate the lexical detection instrument.

Methods:

We conducted a retrospective observational analysis of 47,589 Google Maps reviews from 812 publicly funded health care facilities in Andalusia, Spain. Six service entities and seven complaint- or praise-related lexical markers were detected using a prespecified rule-based lexicon. Sentiment was assigned from review-level star ratings. Negative review burden was defined as the proportion of all reviews that both mentioned an entity and received a negative rating. Lexical-marker differences were estimated using facility-clustered logistic regression within sentiment strata. The final lexicon was validated against manual annotation of 400 reviews.

Results:

System/Management was the most frequently detected entity (42.6%) and had the highest negative review burden overall (37.3%, 95% CI 35.1%-39.5%) and across all nine facility types. Among negatively rated reviews, access-barrier formulations were more frequent in primary than hospital care (12.5% vs 2.2%; OR 6.47, 95% CI 4.57-9.15), whereas waiting-and-delay formulations were less frequent (5.6% vs 12.0%; OR 0.43, 95% CI 0.37-0.51). Among positively rated reviews, professionalism showed no statistically detectable difference (OR 0.99, 95% CI 0.86-1.13), while gratitude (OR 0.45, 95% CI 0.38-0.54) and humanistic-care formulations (OR 0.33, 95% CI 0.25-0.45) were less frequent in primary care. System/Management burden was highest in outpatient consultation centers (52.5%), health centers (46.0%), and local health posts (36.2%), forming a descriptive pattern of higher burden in ambulatory consultation settings. Five of six entities achieved precision between 0.91 and 1.00; Facilities showed insufficient validation performance and was excluded from substantive interpretation. Lexical markers generally showed high precision but limited recall.

Conclusions:

Online reviews revealed distinct patterns in organizational, access, waiting, and positive-care discourse across settings. Organizational processes showed the greatest concentration of negatively rated feedback, particularly in ambulatory consultation settings. These findings describe reviewer discourse rather than population-representative patient experience or entity-specific sentiment. A transparent, empirically evaluated lexical approach can provide a reproducible means of identifying domains of negatively rated feedback warranting closer organizational investigation.