What a model score is
A model score is for one doctor and one procedure. A model trained on public records predicts how the result for patients of that procedure, at the hospitals the doctor is affiliated with, is expected to compare with the national rate in the next measurement period. The score is shown as a percentage below or above the national rate. A lower score is better.
Each score includes a likely range and the records behind it. In the tests, a range built this way for a single hospital contained the hospital's later result between 78% and 90% of the time. The range for a doctor affiliated with several hospitals was not tested.
The models predict a standing against the national rate, not a rate. The national rate changes from one period to the next, and the models do not predict that change.
A label uses only the predictions.
- Lower predicted rate. Every hospital behind the score is in the range of the lowest fifth of hospitals and below the national rate.
- Near the national rate. This label applies to a doctor whose hospitals are in neither group or are in different groups.
- No label. When every hospital used for the score is in the highest fifth of hospitals and above the national rate, OpenHealth shows the score without a label and does not put the doctor on a list. Hospitals in that group had worse results later in every test. The one check of doctors' own results did not confirm that pattern for the doctors, so OpenHealth does not label them.
- No model score. Public records do not let the model predict a result for this doctor. This does not mean the doctor is a poor choice.
How the models are built
CMS publishes results for each hospital every year: deaths, complications, readmissions and infections after named kinds of care. For each result, a model learns from earlier years and predicts the next year.
The models train on nine yearly sets of CMS hospital results from 2019 to 2026, Medicare billing counts for each doctor from 2013 to 2024, and eight yearly lists of which doctors are affiliated with which hospitals. The models that score doctors today were fitted on 114,076 training rows. Each row pairs a hospital's earlier records with one later result, and one later result can appear in several rows.
A model starts with the hospital's own earlier rate, pulled toward the national rate when the hospital had few cases. It then adds about 100 other public facts that predict results: the number of cases, the hospital's results on 58 other measures, patient survey scores, the type of hospital, and the experience and volume of the doctors affiliated with it.
Three models are trained for each result, and their predictions are averaged: a ridge regression, gradient-boosted decision trees trained on that result alone, and gradient-boosted decision trees trained on every result at once. All models are retrained every night on the newest public files.
How well the models predict
Every test predicts a later period from records published before it. The model for a test is trained only on results that were already public on the day its inputs were read. Every test concerns hospitals because the results are published for hospitals.
The newest period of each result was scored last. It is not an untouched test. An earlier version of the models scored it first. The models were then corrected twice and extended once: an input that made one part of the model misjudge shorter periods was removed, an independent review led to further corrections, and a third model trained on every result at once was added after an experiment on every period, including the newest.
For knee replacement and hip replacement, the models predict complications after hip or knee replacement. The newest results cover April 2023 to March 2025. The test covers the 1,392 hospitals that had a result both then and in the earlier period, out of 1,703 with a result then. The error was 22.1 for the model, 22.8 for each hospital's own earlier rate adjusted for its number of cases, 26.4 for the earlier rate as published and 24.7 for the national rate. A lower error is a better prediction. The model's gain over the adjusted earlier rate is larger than chance explains. Hospitals the model placed in its lower group later had a result of 3.29%, hospitals in its higher group 4.41%, and hospitals in neither group 3.95%.
For coronary artery bypass grafting, the models predict death within 30 days of bypass surgery. The newest results cover July 2022 to June 2025. The test covers the 827 hospitals that had a result both then and in the earlier period, out of 861 with a result then. The error was 28.9 for the model, 29.7 for each hospital's own earlier rate adjusted for its number of cases, 31.5 for the earlier rate as published and 32.0 for the national rate. A lower error is a better prediction. The model's gain over the adjusted earlier rate is larger than chance explains. Hospitals the model placed in its lower group later had a result of 1.96%, hospitals in its higher group 2.94%, and hospitals in neither group 2.34%.
For coronary angioplasty and stenting, the models predict death within 30 days of a heart attack. The newest results cover July 2022 to June 2025. The test covers the 1,844 hospitals that had a result both then and in the earlier period, out of 1,933 with a result then. The error was 9.7 for the model, 10.0 for each hospital's own earlier rate adjusted for its number of cases, 11.9 for the earlier rate as published and 10.9 for the national rate. A lower error is a better prediction. The model's gain over the adjusted earlier rate is larger than chance explains. Hospitals the model placed in its lower group later had a result of 11.00%, hospitals in its higher group 12.54%, and hospitals in neither group 11.85%.
For hysterectomy, the models predict surgical site infection after abdominal hysterectomy. The newest results cover January 2025 to December 2025. The test covers the 502 hospitals that had a result both then and in the earlier period, out of 540 with a result then. The error was 84.3 for the model, 84.5 for each hospital's own earlier rate adjusted for its number of cases, 100.9 for the earlier rate as published and 89.6 for the national rate. A lower error is a better prediction. The model's error cannot be distinguished from the error of the adjusted earlier rate. The model's lower and higher groups did not both separate from other hospitals in every test, so no doctor gets a lower or higher label for this procedure.
On every result, the models predict better than the national rate and better than a hospital's own last published rate. Against the earlier rate adjusted for the number of cases, the models show a gain larger than chance for knee replacement and hip replacement, coronary artery bypass grafting and coronary angioplasty and stenting, and that gain is between 3% and 4% of the error. For coronary artery bypass grafting, the gain is larger than chance only since the third model was added, and that model was chosen after this newest period had been scored. The earlier versions of the model could not be distinguished from the adjusted earlier rate there. For hysterectomy, the model predicts no better than the adjusted earlier rate.
Most predictable variation comes from the hospital's own record. The models received the doctor's own volume and experience, but those inputs did not improve the predictions. Two doctors affiliated with the same hospitals therefore get the same score.
No public file reports results for each doctor nationally, so the labels, ranges and list order were tested on hospitals. One check on doctors was possible. New York and California publish death rates for each heart surgeon and stent operator from 2017 to 2019. In each report, doctors at hospitals with a better earlier result tended to have better results of their own, but most of those differences are within chance. The reports do not support labelling a doctor as having a higher predicted rate, so OpenHealth does not use that label. The check uses each hospital's earlier rate, not the model, and covers two states and heart care only.
How a doctor gets a score
Only the operating clinician is scored. OpenHealth counts a procedure for a doctor only when Medicare paid that doctor the full fee as the operating clinician, and the doctor has such a record in the latest year of billing files. A clinician who billed only as an assistant or for aftercare is not scored.
The score uses the hospitals CMS lists the doctor as affiliated with. An affiliation does not show where the doctor operates, and CMS does not publish how the doctor's cases divide between hospitals. Each hospital's prediction is weighted by that hospital's number of cases. An affiliated hospital with no published result for the procedure is left out. When the hospitals differ, the range is wider.
A doctor on the federal exclusion list or with a revoked Medicare enrollment is never listed. OpenHealth still shows the score.
The list for a procedure and a place
OpenHealth does not recommend doctors and does not name a best surgeon. For one procedure and one place, it lists up to five doctors whose affiliated hospitals have the lowest predicted rate, lowest first. Doctors with the same predicted result are shown together as a tie, not numbered. Many doctors share the same hospitals, so ties are common. When more doctors share a result than the list can hold, the doctors with more procedures as the operating clinician in Medicare records are listed first, then the doctors with a nearer office.
OpenHealth never lists a doctor whose hospitals are all in the highest group or who has no model score. OpenHealth measures the distance to the nearest office address that the Census geocoder matched exactly.
Model version and changes
The scores on this site come from model version ml-4, dated 14 November 2026. It trains on CMS hospital results from 2019 to 2026. In the tests, the likely range for a single hospital contained the hospital's later result 78% to 90% of the time.
- 14 November 2026: Trained models replace the earlier rule tiers (rules-3). A doctor's score is the models' prediction for the doctor's affiliated hospitals.
- 14 November 2026: After an independent review, the score shows the tested standing against the national rate, and a label requires agreement from every hospital behind the score.
- 14 November 2026: Version 4 (ml-4) adds a third learner, gradient-boosted trees trained on every outcome at once.
- 14 November 2026: No doctor is labelled as having a higher predicted rate.
- 16 November 2026: Lists show doctors with the same predicted result together as a tie, without numbers, and every score is shown with its likely range.
What is scored today
OpenHealth scores doctors on 5 of the 27 procedures it lists. A procedure can be scored when a public file measures a result for that procedure's own patients at a place that a public file ties to the doctor.
- Knee replacement. The score measures complications after hip or knee replacement. 5,538 doctors have a score.
- Hip replacement. The score measures complications after hip or knee replacement. 3,909 doctors have a score.
- Coronary artery bypass grafting. The score measures death within 30 days of bypass surgery. 1,377 doctors have a score.
- Coronary angioplasty and stenting. The score measures death within 30 days of a heart attack. 4,910 doctors have a score.
- Hysterectomy. The score measures surgical site infection after abdominal hysterectomy. 546 doctors have a score.
- Not scored: CABG, valve or combined surgery. Only New York publishes results for this group of operations by surgeon. Its newest report covers 2017 to 2019. No national file measures an outcome for these patients.
- Not scored: Colonoscopy. A model predicts unplanned hospital visits after colonoscopy for hospital outpatient departments. CMS measures surgery centers separately. No public file says where a doctor performs colonoscopies, so OpenHealth cannot tie a hospital's prediction to a doctor.
- Not scored: Cataract surgery. CMS publishes no hospital or surgery center outcome for cataract surgery. One doctor-level measure exists. Doctors choose whether to report it, and few do.
- Not scored: Spinal fusion. No public file measures an outcome for this procedure's patients. CMS publishes hospital safety results that cover every patient in the hospital, not the patients of this procedure.
- Not scored: Laminectomy or laminotomy, Lower limb arthroscopy, Upper limb arthroscopy, Pacemaker insertion or repair, Upper GI endoscopy, Leg revascularization, Varicose vein removal, Minimally invasive hernia repair, Open groin hernia repair, Mastectomy, Prostate resection. No public file measures an outcome for these procedures' patients. CMS measures unplanned visits after outpatient surgery for each hospital as a whole, across every kind of surgery. No public file says whether a doctor operates at the hospital or at a surgery center.
- Not scored: Melanoma excision, Skin biopsy, Skin cancer removal, Skin growth removal or shaving, Skin lesion removal, Nail tissue removal, Tissue removal. These procedures are done mostly in a doctor's office. No public file measures an outcome for their patients.
Limits
- No public file reports results for each doctor nationally. The score predicts a result for the hospitals the doctor is affiliated with, not for the doctor alone.
- A score describes a group of patients. It does not predict the result of any one operation.
- The ranges are wide. Many doctors near the national rate cannot be distinguished.
- The tests cover hospitals that had a published result in two periods in a row. Hospitals that lost their published result, which are mostly small, are not in the tests.
- The billing files that show who operates end in 2024, and the affiliation list is from 2026. Some affiliations are newer than the doctor's last billing record.
- The results the models learn from cover Medicare patients, most of whom are aged 65 and over.
- The score for coronary stent procedures predicts deaths after a heart attack among all heart attack patients at the hospital, not only patients who had a stent.
- The hysterectomy score covers infections after abdominal hysterectomy, and only doctors who bill those operations receive that score.
- The hip and knee result combines both operations and covers hospital stays only.
- About one in ten listable doctors has no office address that matched exactly, so those doctors cannot appear on a list.
- A model score is OpenHealth's reading of public records. It is not medical advice.
Report a problem
Each score shows the records behind it. If a record is wrong or belongs to another doctor, report it, and OpenHealth will check the public file.