Building a Topic Framework for Hospital Reviews
How thousands of candidate names from reviews of U.S. hospitals became 29 topics, and why the boundaries between them matter.
Reviews of U.S. hospitals and other healthcare facilities can combine warm praise for nurses with complaints about reception, requests for assistance, or an unexpected bill. Putting everything under “patient experience” gives a team little guidance about which accounts need to be read together or who should investigate them.
A taxonomy gives those experiences a shared vocabulary: named topics, each with a definition. We developed one from a sample of 217,343 Google reviews, consolidating thousands of model-generated candidates into 29 topics grouped into seven categories.
The practical value lies in the distinctions it preserves. In the later labeling pass, clinical bedside manner received more favorable descriptions than front-desk conduct. Feedback about visible cleanliness and infection-control practices differed even more. Here we explain how the vocabulary was built, what those distinctions reveal, and how a hospital team could use them to choose a more focused follow-up question.
Start with the accounts the vocabulary needs to describe
The source sample covers 10,029 U.S. healthcare listings, including hospitals, emergency departments, medical clinics, and urgent-care centers. “Hospital reviews” is shorthand for this broader collection. A vocabulary built from it needs room for appointment access and billing as well as experiences inside a hospital ward.
About 2.09 million text reviews were eligible. The selected sample covers 2019 to early June 2026, with sampling across years within each listing and a cap of 200 reviews per listing. That cap limits the influence of heavily reviewed locations. It does not make the sample representative of patients or care delivered across the United States.
The visual overview examines the rating distribution, geographic coverage, and care-setting differences in detail. Its maps remain interactive. Here, the key implication is narrower: the sample shapes which experiences the discovery process can find. An experience absent from these accounts cannot be recovered by consolidating their topic names.
From 3,280 candidate names to 29 usable topics
Discovery worked in batches of up to 1,000 reviews. Rather than assign a fixed label to each review at this stage, the model proposed topics suggested by the batch. Across 216 usable outputs from 218 attempted batches, the run produced 3,280 topic candidates.
Those candidates were not 3,280 distinct kinds of experience. Separate batches can describe similar concerns with different names. Consolidation merged related candidates in two stages, first to 344 and then to 25 topics. A manual refinement pass brought the final vocabulary to 29.
The vocabulary is consolidated, then refined Each number counts topic candidates or definitions at that stage, rather than reviews. The steps are shown at equal size because they describe a process, not a scale.
3,280
Candidate topics
Proposed independently across discovery batches.
344
First consolidation
Related candidate names grouped together.
25
Second consolidation
A smaller set of common topic definitions.
29
Manual refinement
Selected distinctions restored before labeling.
Source: saved discovery report. The final increase reflects refinement, not an increase in the number of sampled reviews.
The staffing topic illustrates consolidation. Candidate names such as Staff shortages and understaffing and Understaffing and slow throughput became Staffing levels and operational capacity. Similar descriptions could now be counted together.
The final refinement made selected topics more specific. Twenty-one topics remained; four broad topics were replaced by eight narrower ones:
| Before refinement | Distinctions in the final framework |
|---|---|
| Staff interaction and bedside manner | Front-desk conduct; clinical bedside manner |
| Billing, insurance, and financial aspects | Billing and charges; insurance coverage and denials |
| Facility environment, comfort, and amenities | Facility condition and comfort; food, parking, and amenities |
| Medication and pain management | Medication and pharmacy; pain management |
A disputed charge and a coverage denial may require different records and expertise. A prescription delay raises a different question from pain not being taken seriously. Preserving those distinctions gives a team a more useful starting point than a broad “financial” or “medication” score.
Discovery counts alone cannot tell us how common any of these experiences are. A candidate can recur across batches or overlap another candidate. For that, we need the separate labeling pass that applies the final definitions to individual reviews.
What a single “staff” score would hide
Applying the framework makes the reason for these distinctions concrete. Clinical bedside manner appears in 41.7% of labeled reviews, front-desk conduct in 25.3%, and nursing responsiveness in 11.8%. Their tone differs too: about 21%, 37%, and 41% of their positive-or-negative statements are negative, respectively.
Praise for clinical interactions can therefore coexist with a less favorable picture of administrative contact or practical assistance. For a patient-experience team, that suggests separate questions: what do reviewers value in clinical interactions, what happens at reception, and what do accounts of requests for help describe?
Switch between staff feedback and cleanliness/protective practices to compare the selected topics on a common scale. These model-derived estimates describe sampled public feedback.
The topic boundary changes the picture Negative share of positive-or-negative statements, 2019 to early June 2026. Hover for each topic’s review reach and statement count.
Source: frozen labeling aggregates, independently checked against the original label file. Reach uses 196,183 labeled reviews; tone counts statements and excludes neutral ones. A review may mention several topics. Download these examples.
The second comparison shows why “the environment” is also too broad for some questions. Cleanliness statements are about 36% negative; infection-control statements about 78%. Cleanliness reaches 5.8% of labeled reviews, compared with 1.4% for infection control. Favorable accounts of tidy spaces do not answer what reviewers describe about masks, isolation, or protective equipment. Those accounts require different reading and verification; neither measure establishes whether an infection occurred.
Explore the seven categories and their 29 topics
The seven categories make the vocabulary easier to navigate. Their sizes describe how finely the taxonomy divides each area. Clinical Care & Safety has six topics; Financial & Administrative has two. That does not mean clinical concerns occur three times as often.
How the 29 topics are organized Number of topic definitions in each category. These are design choices in the vocabulary, not counts of reviews, statements, or complaints.
Source: the 29 definitions in the saved taxonomy; category counts sum to 29.
Select a category below to read every topic and its short description, preserved from the original report. The full frozen catalog contains the longer definitions used by the analysis. Some categories bring together several kinds of distinction: Rights, Equity & Specialty Care, for example, includes both privacy and particular care settings. These are useful browsing groups, not seven mutually exclusive types of patient experience.
Explore the full vocabulary
Source: saved topic names and short descriptions, unchanged. Download the explorer data or the full topic definitions.
Clear boundaries matter more than a neat topic count
Consolidation is not just a naming exercise. Merge too much and “communication” could absorb every problem involving an interaction. Split too much and closely related accounts could be scattered across near-duplicate topics. The useful test is whether the definitions preserve a distinction a reader can understand and a labeler can apply.
The examples below illustrate where that judgment enters. They are invented examples for explaining the definitions, not quotations or additional findings. The suggested topics are a reading of the current definitions, not audited model outputs.
Where would you draw the boundary?
These examples also expose a limit of short definitions. Real accounts can sit between topics: a sentence about delayed test results may involve both diagnostic services and communication. The wording, the surrounding account, and consistent labeling rules matter. A human audit needs to examine those cases, not just confirm that the chosen topic exists in the list.
Some boundaries remain imperfect. The full facility-condition definition explicitly overlaps cleanliness, although the short explorer description makes that less apparent. A usable framework needs rules for such cases and checks of how labels are applied.
Different follow-up needs can also justify a distinction when the tone is similar. Medication/pharmacy and pain-management statements are about 73% and 74% negative, respectively. A refill problem and an account of inadequate pain relief still raise different questions. A large sentiment difference is one way to demonstrate a topic boundary’s value; it is not the only test.
From a label to a follow-up question
For a hospital team, the framework is most useful when a topic leads to a specific question and the appropriate people to examine it. These are possible uses, to adapt to the hospital’s organization:
| Feedback distinction | A concrete follow-up |
|---|---|
| Front-desk conduct vs clinical bedside manner | With registration and patient-experience teams, read whether reception accounts concern courtesy, unanswered questions, or difficulty completing an administrative task. Retain what reviewers value in clinical interactions. |
| Clinical manner vs nursing responsiveness | With the relevant nursing or service lead, inspect accounts of requests for practical help separately from descriptions of warmth and respect. Check service context and available response records. |
| Billing vs insurance | With billing and coverage/authorization teams, distinguish disputed charges, coverage decisions, and explanation gaps. Read connected accounts together before deciding which process to investigate. |
| Cleanliness vs protective practices | With housekeeping and infection-prevention leads, distinguish the condition of a space from a described practice. Verify the accounts through the hospital’s existing processes. |
Check the accounts, identify the process, and compare the feedback with the hospital’s own evidence. These follow-ups are proposed applications of the framework, rather than interventions tested by this study. The analysis cannot tell us which change would improve care or how many stars it would gain.
The hospital overview shows how frequently topics appear and how their tone differs. The later articles examine repeat reviewers, topics worth reading together, and associations with star ratings. The shared vocabulary gives those analyses a way to move from a broad rating to a focused question. Its usefulness depends on preserving relevant distinctions and applying them reliably; 29 is the result of this process, rather than a uniquely correct number of topics.
Method details, source material, and preserved visualizations
Sampling. The saved report records a minimum of 20 reviews per eligible source listing, a maximum of 200 selected reviews per listing, year-stratified sampling, and random seed 42. The 20-review minimum is an eligibility rule, not a guarantee of 20 sampled reviews per listing. The coverage sample includes 217,343 reviews; 217,135 are matched to a U.S. state and 208 are not. Collection ends in early June 2026.
Discovery and consolidation. The report lists GPT-5.4 nano for discovery, batches of up to 1,000 reviews, and DeepSeek V4 Pro for consolidation. Of 218 attempted discovery batches, 216 produced usable outputs. The summary does not establish which experiences might have been lost from the other two. Manual refinement follows the automated 25-topic result; it is distinct from auditing the labels later assigned to reviews.
Source trail. Discovery counts, topic names, short descriptions, and category membership come from the retained report. Comparing the saved 25-topic consolidation with the 29-topic final vocabulary establishes the four refinements shown above. The staffing-name example is documented in the source project’s notebooks/hospitals_topic_labeling/build_notebook.py. The four boundary comparisons are newly written illustrations. Download the explorer data or the full frozen catalog; the existing extraction script preserves the explorer’s descriptions without rerunning a model.
Measured examples. The later labeling pass retains 425,667 statements across 196,183 distinct listing-and-review keys after checks of topic name, topic ID, quoted phrase, review resolution, and known sentiment. Reach counts each review once per topic. Negative share is negative ÷ (negative + positive) statements. The new comparison reuses the overview’s saved aggregates, independently recomputed from the original label file for all 29 topics. Medication and pain figures come from the same checked aggregates. These checks confirm the calculations and mechanical validity, not human agreement with every label. Download the five chart examples.
All original visuals remain available. The original exploratory report retains the interactive coverage map, ratings and annual sample counts, listing mix and category breadth, topic explorer, and discovery examples. The overview carries the coverage map into the site’s current design. Interpretations here supersede the earlier report’s stronger claims about what the taxonomy establishes; its shortened snippets should not be treated as a new set of verified quotations.