How Problems Connect in Patient Feedback
What linked concerns in reviews of U.S. hospitals reveal about financial processes, delays and the context a hospital needs for follow-up.
Among low-rated Google reviews of U.S. hospitals and other healthcare facilities, almost half of those mentioning insurance also mention billing. A concern routed only as an insurance issue could therefore lose the charge-related context needed to understand it. Looking at linked topics helps a hospital decide what to investigate together and which people or records to involve.
The same principle leads to different questions elsewhere: waiting alongside reported capacity constraints calls for an operational check; behavioral health alongside perceived unfair treatment calls for close examination of the described interaction. Counting each topic separately can identify priorities, but it cannot show this shared context.
These findings come from 72,038 labeled one- and two-star reviews of U.S. healthcare listings, including hospitals, clinics and urgent-care services, from 2019 to early June 2026. They describe self-selected public feedback, not measured incident rates. The labels have passed automated checks. A connection means two topics occur in the same account, without establishing a single visit, event order or cause.
Financial concerns often need both charge and coverage context
Billing and insurance occur together in 908 low-rated reviews, about 4.6 times the shuffled comparison’s expected count of 199. More useful for follow-up is the asymmetry: 46% of insurance-related accounts also mention billing, while 12% of billing-related accounts mention insurance.
| Starting set of low-rated feedback | Share also mentioning the other topic |
|---|---|
| Insurance: 1,989 reviews | 908 also mention billing: 46% |
| Billing: 7,607 reviews | 908 also mention insurance: 12% |
The same 908 accounts sit behind both percentages. They answer different questions about context, not which problem came first. Coverage concerns often come with a description of charges; most charge-related feedback does not mention coverage. A hospital should therefore avoid treating every billing dispute as an insurance problem, while checking for financial context when examining coverage feedback.
This pairing largely reflects concerns about both subjects: 843 of the 908 accounts, or 93%, contain a negative statement about each. That check matters because a low overall rating does not make every topic mentioned a complaint.
The two categories also cover different issues. Billing includes unexpected or disputed charges, payment and collections; insurance includes coverage verification, denials, network and authorization. Reading the joint accounts can distinguish what was charged, what coverage decision was described, and what explanation was given. Those distinctions determine whether financial-services records, coverage or authorization records, or both are relevant.
For a hospital team, the practical lesson is to keep the charge and coverage descriptions together during follow-up. An accurate charge could still be poorly explained; a described coverage dispute needs different evidence from an alleged duplicate bill. These are questions to test against the full account and available records. The aggregate pairing does not identify which explanation applies or demonstrate a handoff failure.
Waiting feedback calls for different operational questions
Waiting and staffing/operational capacity appear together in 2,075 low-rated reviews, about 1.36 times the shuffled expectation. Waiting is mentioned in 47% of capacity-related feedback, but capacity appears in only 9% of waiting feedback. These shares use 4,398 capacity-related and 22,622 waiting-related reviews as their respective denominators.
The distinction changes the question a hospital should ask. Capacity concerns often come with delays, but most waiting feedback does not explicitly describe a capacity constraint. The “staffing” label in the charts covers beds, equipment and overcrowding as well as personnel. It measures what people describe, rather than actual staffing levels, and cannot justify a general recommendation to hire more staff.
In 2,042 of the 2,075 paired accounts, or 98%, both topics have a negative statement. This gives an operations team a focused set of jointly described concerns to examine. The useful next step is to locate the waiting stage and the reported constraint: arrival, testing, admission or another part of care. Available demand, resource and stage-specific waiting records can then help assess whether that description fits an operational bottleneck. Explanations given during the wait belong in the same reading, since a reported constraint and uncertainty about a delay require different follow-up.
Administrative experiences offer a useful contrast. Front-desk conduct and waiting appear together in 4,436 low-rated reviews, more than twice the capacity/waiting count. Yet this pairing is not unusually concentrated under the comparison model: its independently calculated lift is about 0.96. A large recurring set of concerns can deserve attention even when it is absent from the retained association map.
That is why volume, concentration and seriousness need separate judgments. Counts help locate a substantial body of feedback. Lift helps identify relationships that appear more often than the comparison predicts. The full account and corroborating evidence determine the significance of a particular concern. Neither a large count nor a high lift estimates the benefit of an intervention.
Strength of association and number of reviews tell different stories Open dots show expected counts under the shuffle; filled dots show observed counts. The ratio is printed with each pair. Expected counts are averages over shuffles, not additional observed reviews.
Source: saved pair counts and lifts. The staffing–waiting expected count is recovered from its observed count divided by the saved lift. These are selected examples from the 31 retained pairs.
Smaller connected concerns deserve specific follow-up
Behavioral health and discrimination/equitable treatment appear together in 212 low-rated reviews, roughly twice the shuffled expectation. 203 of these accounts, or 96%, contain a negative statement about both topics. The lower volume does not settle how significant the described experiences may be.
The value of the pairing is a narrower reading question: how are the treatment setting and the interaction perceived as unfair described together? Among these 212 accounts, 51 also mention bedside manner, 34 medication and 32 safety. These overlapping subsets suggest additional context to check in particular accounts; they do not show that everyone describes the same set of problems.
A service or patient-experience team could use this selection to examine the reported interaction, what treatment or explanation was described, and what would need verification. Relevant concerns can then enter the appropriate existing service, experience or quality process. The model’s categories do not establish a diagnosis, the author’s role, verified discrimination or a safety incident.
The learning is to preserve the specific concern when summarizing smaller categories. A broad “staff experience” summary could obscure the question of perceived fairness. A network group cannot determine seriousness or ownership on its own; those decisions need human reading and appropriate supporting evidence.
Turn a connection into a better investigation
The practical output is a more specific question and the context needed to answer it. Start by validating the labels and reading each selected account in full. Identify the uncertainty before choosing participants or requesting records. The examples suggest different follow-up needs:
| Linked concerns | Context to preserve | Question and evidence for follow-up |
|---|---|---|
| Billing with insurance | Charges, coverage decisions and the explanation described | What was disputed or unclear? Which financial and coverage records, and which teams, can assess it? |
| Waiting with capacity | Waiting stage, reported resource constraint and communication | Do available process, demand and capacity records support the described bottleneck, or point to a different question? |
| Behavioral health with equitable treatment | Treatment setting, interaction and perceived unfairness | What needs human verification, and which existing service, experience or quality process should receive it? |
Keep the linked context attached if the concern moves between teams. If an investigation leads to a process change, assess that change using independent process and experience evidence. A shift in the number of public co-mentions alone cannot establish improvement: participation, listing mix and the language people use may also change.
The connections help choose what to examine together. They do not establish escalation over time or an extra combined rating penalty. The supporting checks below explain those limits and why topic-level sentiment must be checked even in highly rated feedback.
Why a connected pair does not establish an extra rating penalty
The separate rating-association model examines five selected pairs of negative topic mentions, accounting for other topic mentions, review length, year, reviewer group, total topic count and listing effects. In held-out listings, all five interaction estimates are positive: the combined rating association is less negative than adding the two separate associations would predict. This model therefore does not support a beyond-additive negative rating claim for these pairs.
The follow-up model did not find an additional negative rating interaction Five selected pairs in held-out listings. Dots are interaction estimates; intervals are 95% confidence intervals clustered by listing. A negative estimate would indicate an additional penalty beyond the model’s separate associations.
Source: saved holdout estimates from the rating-association analysis. This is a separate model of negative topic mentions across the labeled review sample, not a re-estimate of the low-rating co-mention screen.
A positive interaction does not mean two problems are good or less serious. It concerns an additive rating model on a scale bounded below by one star; that bound may contribute to the pattern, but its role is not established here. Neither this model nor the pair map evaluates the effect of fixing a problem.
Why high ratings do not automatically mean positive experiences
The saved comparison finds 16 qualifying pairs in four- and five-star feedback, versus 31 in one- and two-star feedback; 11 appear in both sets. It uses 299 shuffles for the high-rating analysis and 499 for the low-rating analysis.
Billing and insurance also appear together in 58 high-rated accounts. Of those, 21 contain a negative statement about both topics and 21 contain a positive statement about both; sentiment flags can overlap. A high overall rating does not establish satisfaction with every aspect of the experience. Check the text and topic-level sentiment before selecting examples of successful handling for learning.
The 20 pairs qualifying only in low-rated feedback are candidates for closer reading, rather than complaint-exclusive patterns. Behavioral health and equitable treatment, for example, also occur together in 26 high-rated accounts, below the 30-co-mention testing threshold. Absence from the qualifying list cannot establish absence from the feedback. These comparisons do not formally test whether associations differ between rating bands.
Supporting plots, methods and aggregate data
Scope and labels. The year- and listing-stratified coverage sample contains 217,343 reviews from 2019 to early June 2026. Of these, 196,183 have at least one label passing automated topic, quoted-phrase, review-ID and sentiment checks. Within the 72,038 labeled one- and two-star reviews, 47,405 mention at least two distinct topics. Reviews are keyed by listing plus review ID and counted once per topic. Healthcare listings are not necessarily distinct hospitals. Topic labels describe reported experiences rather than verified events.
Comparison. Topic columns are independently shuffled 499 times within original topic-count strata (2, 3, 4, 5+), skipping strata below 50 reviews. This preserves each topic’s stratum-level frequency, but not exact word length or every shuffled row’s topic count. Pairs with at least 30 observed co-mentions are tested; retention requires a Benjamini–Hochberg q-value below 0.05 and lift above 1.15. Greedy modularity groups the network using log₂ lift as the weight. Group names are editorial; stability across alternative clustering algorithms is not established.
New descriptive checks. Conditional shares divide co-mentions by all labeled reviews in the rating band mentioning the starting topic. A topic has a negative or positive flag when any validity-gated statement has that sentiment; flags can overlap. Co-mentions in the original screen can have any sentiment. The focal sentiment checks support interpreting these particular low-rated pairs as joint concerns, rather than establishing that every pair is jointly negative. Exact shuffle expectations reconcile the saved retained-pair lifts to within 1.28%; the charts retain their original finite-shuffle estimates. No statistical screen or interaction model is refitted.
Preserved analyses. The earlier exploratory report retains the unadjusted severity comparison, facility-category checks and monthly and pandemic-period series. Read its unadjusted ratings alongside the adjusted model above; causal or compound-penalty interpretations are not supported by these findings. The executed notebook reports mean correlations of monthly attention changes of 0.119 within groups versus −0.001 across groups; corresponding tone correlations are 0.105 versus 0.091. These exploratory summaries do not independently validate the grouping or establish an early-warning system. Listing mix and the single pandemic period complicate interpretation.
Downloads and reproducibility. Download the original aggregate chart data and the additional counts, conditional shares and sentiment checks. The chart generator extracts saved report aggregates and imports the later model’s holdout estimates. The additional evidence is recomputed from the frozen coverage and label parquet files by scripts/generate_hospital_pair_insights.py. These exports contain no raw review text or reviewer identifiers.
For the companion sensitivity analysis, see Whose Voices Shape Hospital Ratings.