AI Analytics

Technical writing

CMS Hospital Quality Data: Outcomes, Readmissions, and Star Ratings for 6,000 US Hospitals

· AI Analytics
Federal DataCMSHealthcare QualityHospitals

Every Medicare-certified hospital in the United States is required to report dozens of quality measures to the Centers for Medicare & Medicaid Services. CMS publishes those measures publicly through Care Compare (care.compare.cms.gov), formerly known as Hospital Compare. The result is one of the largest structured quality-measurement datasets in American healthcare: roughly 6,000 hospitals, hundreds of measure-hospital combinations per release, and annual refreshes going back to the early 2000s.

What Care Compare Is

Care Compare is CMS's unified consumer-facing portal for provider quality data, launched in 2021 by consolidating Hospital Compare, Nursing Home Compare, Physician Compare, and several other predecessor sites. The hospital quality data itself predates the portal — Hospital Compare launched publicly in 2005 — and the underlying datasets are released quarterly to annually depending on the measure group.

The data is not a survey or a sample. CMS collects it as a condition of Medicare participation: hospitals that do not submit required quality data face a payment reduction under the Inpatient Prospective Payment System (IPPS). Compliance rates are near-universal for acute-care hospitals. Critical access hospitals, psychiatric hospitals, long-term acute care hospitals, and children's hospitals follow separate reporting requirements and are sometimes excluded from specific measure groups.

Bulk downloads are available at data.cms.gov/provider-data as ZIP files containing flat CSVs. The Care Compare API offers JSON access to individual provider records but is less suitable for population-level analysis than the bulk downloads.

Dataset Key Fields

Every row in the CMS hospital quality CSVs shares a common spine of identifier and administrative fields:

The Five Main Data Categories

Process Measures

Process measures capture whether hospitals perform specific evidence-based clinical actions. They are expressed as percentage compliance: the share of eligible patients who received the recommended care. Classic examples include timely administration of aspirin for acute myocardial infarction (AMI), antibiotic prophylaxis within one hour of surgical incision, and venous thromboembolism (VTE) prophylaxis for surgical patients. Because these measures ask “did the hospital do X for patients who should have received X,” they are directly actionable for quality improvement teams and do not require risk adjustment. However, they can approach ceiling effects: well-managed hospitals routinely hit 99%+ on antibiotic prophylaxis, compressing the signal at the top of the distribution.

Outcome Measures

Outcome measures capture what actually happened to patients. The flagship outcomes in the CMS dataset are 30-day mortality and 30-day unplanned readmission rates, calculated for six clinical conditions: acute myocardial infarction (AMI), heart failure (HF), pneumonia, chronic obstructive pulmonary disease (COPD), coronary artery bypass graft (CABG), and elective hip and knee replacement. CMS also publishes a composite hospital-wide all-cause readmission rate.

These rates are the most analytically demanding piece of the dataset because they require risk adjustment. A hospital that treats sicker patients will mechanically show worse raw outcomes. CMS uses a hierarchical logistic regression model that adjusts for patient age, sex, principal diagnosis, and a standard set of comorbidities drawn from claims data in the 12 months preceding the admission. The output is a risk-standardized rate (RSR): what the mortality or readmission rate would have been if the hospital had treated the national average patient mix.

Risk adjustment matters enormously for interpretation. A safety-net hospital serving a high-comorbidity, low-income population will have a worse unadjusted readmission rate than a suburban community hospital, even if both provide equivalent care. The RSR partially corrects for this, though debate continues about whether CMS's model adequately adjusts for social determinants of health — housing instability, food insecurity, and lack of follow-up care are not captured in Medicare claims.

Patient Experience: HCAHPS

The Hospital Consumer Assessment of Healthcare Providers and Systems (HCAHPS) is a standardized, publicly reported patient experience survey administered to a random sample of adult inpatients after discharge. CMS requires all IPPS hospitals to administer HCAHPS as a condition of full payment updates.

The survey covers ten domains:

Each domain is reported as the percentage of respondents who chose the most positive response option (“Always” for frequency items, 9 or 10 for the overall rating). CMS also assigns an HCAHPS star rating (1–5) for the overall hospital rating domain and for each of the ten composites. HCAHPS scores are adjusted for patient mix (case-mix adjustment) to account for the fact that some patient populations, such as the elderly or those with lower health literacy, systematically rate hospitals differently independent of care quality.

Structural Measures

Structural measures assess hospital characteristics rather than care processes or patient outcomes. They answer questions like: Does this hospital have intensivist physicians managing ICU patients around the clock? Has the hospital adopted a certified electronic health record system? Does it use a safe surgery checklist? These measures are typically self-reported by hospitals and verified through attestation rather than claims data. They tend to appear as binary (yes/no) fields and serve more as organizational quality indicators than performance metrics suitable for benchmarking or financial incentive programs.

Efficiency and Spending: MSPB

The Medicare Spending Per Beneficiary (MSPB) measure captures the total Medicare payments for services provided from three days before through 30 days after a hospital inpatient stay. It is an episode-based spending measure — capturing not just the inpatient bill but the professional fees, post-acute care, and outpatient follow-up associated with a hospitalization. The MSPB ratio compares a hospital's risk-adjusted, price-standardized spending per episode to the national median. A ratio above 1.0 means the hospital is associated with higher Medicare spending per episode than the national benchmark.

MSPB is important because it enters the Value-Based Purchasing calculation (discussed below) and because post-acute spending variation is large: two hospitals with identical inpatient costs can have dramatically different MSPB ratios depending on whether their patients predominantly go home, to skilled nursing facilities, or to inpatient rehab.

The Overall Star Rating System

CMS assigns each hospital an Overall Quality Star Rating of 1–5 stars based on performance across seven measure groups: mortality, safety of care, readmission, patient experience, effectiveness of care, timeliness of care, and efficient use of medical imaging. The methodology uses a latent variable model that first standardizes each measure within its group, weights the groups, and then applies a k-means clustering algorithm to sort hospitals into five star-rating tiers.

The Star Rating has been persistently controversial since its 2016 launch. Critics point out that large academic medical centers and teaching hospitals cluster in the 1–3 star range despite handling the most complex cases in the country. Major academic centers, safety net hospitals, and their trade associations have repeatedly argued that the model inadequately adjusts for case complexity and social risk factors, effectively penalizing hospitals that serve sicker, lower-income, and more medically complex patients. CMS has revised the methodology several times in response to public comment but the fundamental tension — between consumer-facing simplicity and the statistical nuance required to fairly compare heterogeneous hospitals — remains unresolved.

Rural critical access hospitals face an additional disadvantage: many CAHs do not report enough cases in enough measure groups to receive a star rating at all, and those that do often have wide confidence intervals on their outcome rates. The binary question of whether a hospital qualifies for a star rating in the first place systematically disadvantages small rural hospitals in any public comparison tool.

Hospital-Acquired Condition Reduction Program

The Hospital-Acquired Condition (HAC) Reduction Program, authorized by the Affordable Care Act, imposes a 1% Medicare payment reduction on hospitals that rank in the worst-performing quartile (bottom 25%) on a composite of hospital-acquired infection and patient safety metrics. The HAC measure set includes:

HAC penalties apply automatically based on the total HAC score; hospitals have no ability to appeal out of the penalty once they fall into the bottom quartile. Because the penalty is cut-off based rather than continuous, hospitals near the quartile boundary face binary cliff effects on their Medicare revenue. The program applies to general acute-care hospitals; critical access hospitals, psychiatric hospitals, and certain specialty hospitals are exempt.

Value-Based Purchasing Program

The Hospital Value-Based Purchasing (VBP) Program adjusts Medicare payments up or down based on a Total Performance Score (TPS) calculated across four domains: clinical outcomes, person and community engagement (HCAHPS), safety, and efficiency and cost reduction (MSPB). The adjustment is budget-neutral at the program level: CMS withholds 2% of each hospital's base operating payments, then redistributes the pooled funds to hospitals proportionally to their TPS. Hospitals above the median receive net payments greater than their withhold; hospitals below the median receive less.

In practice, VBP adjustments range roughly from –2% to +4% of Medicare payments for most hospitals, with outliers at both ends. For a hospital with $200 million in annual Medicare revenue, a 3 percentage-point swing represents $6 million — material for operating margin management. The program creates direct financial incentives aligned with the quality measures in the Care Compare dataset, linking the public reporting infrastructure to real payment consequences.

Suppression Rules and Small-Hospital Limitations

Measures with fewer than 25 eligible patients over the reporting period are suppressed and appear as “Not Available” in the public dataset. The suppression threshold exists to protect patient privacy (a small enough sample could be re-identified) and to avoid publishing statistically unreliable rates. However, it has significant analytical consequences.

Rural hospitals and critical access hospitals frequently fall below the 25-case threshold for individual condition-specific outcome measures. A small rural hospital may admit only 10 to 15 AMI patients per year — not enough to generate a reportable 30-day mortality rate. This means that the hospitals serving the most geographically isolated populations are, by design, among the least visible in the quality measurement system. Any analysis that treats suppressed scores as missing at random will systematically underrepresent rural facilities and overcount the proportion of low-volume hospitals that appear high-performing.

Notable Patterns in the Data

Several empirical patterns emerge consistently across Care Compare releases:

Ownership type and outcomes. For-profit hospitals tend to show modestly higher 30-day readmission rates for heart failure and pneumonia relative to non-profit hospitals after risk adjustment, a pattern documented across multiple academic analyses of CMS data. The direction of causality is contested: selection effects (for-profits concentrate in certain markets and patient populations), staffing ratios, and post-discharge follow-up infrastructure are all candidate mechanisms. Government-owned hospitals, which include many urban safety-net institutions, show the widest performance spread in the data.

Teaching hospital process compliance. Large teaching hospitals consistently report high process measure compliance rates — aspirin for AMI, antibiotic prophylaxis — reflecting both the administrative capacity to track and report measures and the clinical protocol infrastructure of academic medicine. Where teaching hospitals underperform is on HCAHPS scores: academic centers often score below community hospitals on nurse responsiveness and quietness domains, reflecting the complexity of house-staff shift handoffs and the physical environment of large inpatient towers.

Rural critical access hospital star rating disadvantage. CAHs that do receive star ratings are concentrated in the 3-star tier. The combination of suppressed outcome measures, smaller HCAHPS sample sizes, and the absence of data on efficiency measures (many CAHs are IPPS-exempt and do not have MSPB calculated) leaves too few qualifying measures for the clustering algorithm to reliably differentiate top performers.

Analyzing 30-Day Heart Failure Readmissions in Python

The following script downloads the CMS readmissions ZIP, filters to the heart failure 30-day readmission measure (READM-30-HF), computes the distribution by hospital ownership type, and identifies outliers — hospitals whose risk-standardized readmission rate exceeds the national mean by more than two standard deviations.

import requests, zipfile, io
import pandas as pd
import numpy as np

# CMS Provider Data Catalog — Hospital Readmissions Reduction Program
# Unplanned Hospital Visits — General Information (30-day readmissions)
# Download URL: https://data.cms.gov/provider-data/dataset/unplanned-hospital-visits
ZIP_URL = (
    "https://data.cms.gov/provider-data/sites/default/files/resources/"
    "unplanned_hospital_visits_general_information.zip"
)

print("Downloading CMS readmissions ZIP...")
resp = requests.get(ZIP_URL, timeout=120)
resp.raise_for_status()

with zipfile.ZipFile(io.BytesIO(resp.content)) as z:
    # Identify the readmissions CSV (name varies by release)
    csv_name = next(n for n in z.namelist() if n.endswith(".csv"))
    with z.open(csv_name) as f:
        df = pd.read_csv(f, dtype=str, low_memory=False)

print(f"Loaded {len(df):,} rows, columns: {list(df.columns)[:8]} ...")

# Filter to heart-failure 30-day readmission measure (READM-30-HF)
hf = df[df["Measure ID"] == "READM-30-HF"].copy()
print(f"Heart-failure readmission rows: {len(hf):,}")

# Coerce score to numeric (suppressed rows are marked "Not Available")
hf["score"] = pd.to_numeric(hf["Score"], errors="coerce")
hf = hf.dropna(subset=["score"])
print(f"Rows with numeric scores: {len(hf):,}")

# Normalise ownership type to three buckets
def bucket_ownership(s: str) -> str:
    s = str(s).lower()
    if "government" in s:
        return "Government"
    if "non-profit" in s or "non profit" in s or "voluntary" in s:
        return "Non-profit"
    if "proprietary" in s or "for profit" in s:
        return "For-profit"
    return "Other"

hf["ownership_bucket"] = hf["Hospital Ownership"].apply(bucket_ownership)

# Distribution by ownership type
summary = (
    hf.groupby("ownership_bucket")["score"]
    .agg(
        hospitals="count",
        mean_rate=lambda x: round(x.mean(), 2),
        median_rate=lambda x: round(x.median(), 2),
        p10=lambda x: round(np.percentile(x, 10), 2),
        p90=lambda x: round(np.percentile(x, 90), 2),
    )
    .reset_index()
    .sort_values("mean_rate")
)
print("\n30-day HF readmission rate by ownership type:")
print(summary.to_string(index=False))

# Identify outliers: hospitals more than 2 SD above national mean
national_mean = hf["score"].mean()
national_std  = hf["score"].std()
cutoff = national_mean + 2 * national_std

outliers = hf[hf["score"] > cutoff][
    ["Facility Name", "State", "Hospital Ownership", "score"]
].sort_values("score", ascending=False)

print(
    "\nOutlier hospitals (score > mean + 2 SD = "
    + str(round(cutoff, 2))
    + "):"
)
print(outliers.head(20).to_string(index=False))

The script uses string concatenation for the outlier cutoff display to avoid nested template literal syntax. The ownership_bucket function normalizes CMS's verbose ownership strings (“Voluntary non-profit — Church”, “Proprietary”, “Government — Hospital District or Authority”) into three analytical buckets. After removing suppressed rows, expect roughly 2,500–3,200 hospitals with numeric HF readmission scores depending on the release year.

Typical output shows non-profit hospitals at a mean 30-day HF readmission rate of around 21–22%, for-profit hospitals at 22–23%, and government hospitals with the widest distribution. Outlier hospitals above the two-standard-deviation cutoff are predominantly small facilities with volatile rates due to low case volume — a reminder that high-apparent-outlier status is often a sample size artifact rather than a true quality signal.

Cross-References

Hospital quality data intersects with several other federal datasets. The HHS OIG Exclusions (LEIE) list identifies providers and entities barred from Medicare participation; crosswalking CCNs against the LEIE surfaced cases where hospitals or their clinical staff were billing despite active exclusions. CMS Open Payments data captures financial relationships between hospitals, physicians, and pharmaceutical and medical device manufacturers — useful for examining whether physician payment patterns correlate with prescribing, utilization, or outcome measures at the facility level. For behavioral health facilities, the SAMHSA Treatment Locator data covers substance use disorder treatment facilities that are not captured in the CMS acute-care quality framework but serve overlapping patient populations with high readmission risk.

Related: HHS OIG Exclusions — the federal healthcare fraud blacklist

Related: CMS Open Payments — pharmaceutical and device manufacturer payments to physicians

Related: SAMHSA Treatment Locator — substance use disorder facility data

Related: CMS Doctors and Clinicians, CMS Provider Ownership (private equity in healthcare), and CMS Post-Acute Care Utilization