60/100

Moderate evidence

Cite
More
EmailLinkedInX

Study of 400 joint replacement patients finds AI screening matched doctors' picks

Study title: A machine learning model for predicting hospitalist consult appropriateness among elective joint arthroplasty patients: A single-center experience with hospitalist comanagement screening

Evidence type
Observational study
Sample
n = 400
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: A machine learning model for predicting hospitalist consult appropriateness among elective joint arthroplasty patients: A single-center experience with hospitalist comanagement screening

Researchers analysed the post-operative courses of 400 total joint arthroplasty patients to work out who actually benefited from hospitalist comanagement, then trained a random forest model to predict that. Against the hospital's own screening, the model caught 97.0% of patients who warranted comanagement, compared with 91.8% for the current approach.

Read the study abstract

Hospitalist comanagement improves outcomes for total joint arthroplasty (TJA) patients, but is a limited resource, and data on which patients are most likely to benefit are sparse. A total of 400 unique TJA patients and their post-operative trajectories were analyzed retrospectively to determine which patients warranted comanagement, defined by hospitalists making meaningful contributions to post-operative care. From this data, a random forest machine learning (ML) model was developed to predict hospitalist cons…

A machine learning model for predicting hospitalist consult appropriateness among elective joint arthroplasty patients: A single-center experience with hospitalist comanagement screening
"Medical Care" by DarkoStojanovic (Pixabay), CC0.

The gain

In a retrospective analysis of 400 joint replacement patients at one hospital, a random forest model identified 97.0% of patients who warranted hospitalist comanagement versus 91.8% for current physician-led screening, with specificity of 46.3% versus 37.8%.

In brief

Hospitalist comanagement is a limited resource, and 303 of the 400 patients received it even though only 134 were judged to have warranted it. A tool that flags fewer unnecessary consults could ease that load, but the authors describe the model's performance as comparable to physician screening, not clearly better.

Functional status, albumin, medical optimization and cardiac comorbidities were among the features that drove the model's predictions.

Main points

  1. The study analysed 400 unique joint replacement patients and their post-operative trajectories retrospectively.
  2. 303 of the 400 patients received hospitalist comanagement, but only 134 were judged to have warranted it.
  3. Current screening, which relies on physician discretion and a risk stratification tool, had sensitivity of 0.918 and specificity of 0.378.
  4. The random forest model had sensitivity of 0.970, specificity of 0.463 and an area under the curve of 0.783.
  5. Only one missed opportunity for a consult was identified among the 400 cases under current screening.

The full story

The team defined comanagement as warranted when hospitalists made meaningful contributions to post-operative care. Applying that definition to 400 patients, they found that 134 truly needed it, although 303 received it.

The comparison baseline was the hospital's current process, which combines physician discretion with a set of criteria called the risk stratification tool. That process reached sensitivity of 0.918 and specificity of 0.378. The model reached 0.970 and 0.463, with an area under the curve of 0.783 that the authors call good discriminatory capacity.

The authors conclude the model performed comparably to manual screening and that patients deemed low risk can proceed safely without comanagement under both ML and physician screening. Functional status, albumin, medical optimization and cardiac comorbidities were important features.

Sources

1 on file
  1. Journal of Hospital Medicinedoi:10.1002/jhm.70554
    Peer-reviewed2026-11-17

Talk it over

Three questions to start on, two more if you have time
The random forest model caught 97.0% of patients who warranted hospitalist comanagement against 91.8% for current screening, but specificity stayed at 46.3% versus 37.8%. How would you weigh the extra catches against the many patients still flagged without needing it?

Start with a question

Study of 400 joint replacement patients finds AI screening matched doctors' pickshttps://truvace.com/item/a-machine-learning-model-for-predicting-hospitalist-consult-appropriateness-amon

  1. The random forest model caught 97.0% of patients who warranted hospitalist comanagement against 91.8% for current screening, but specificity stayed at 46.3% versus 37.8%. How would you weigh the extra catches against the many patients still flagged without needing it?

    Answer in the thread
  2. The study looked back at 400 patients at one hospital, and the researchers judged afterwards who warranted comanagement. How much should that single-hospital, after-the-fact judgment shape what we make of the model's numbers?

    Answer in the thread
  3. 303 of the 400 patients received hospitalist comanagement, yet 134 were judged to have warranted it. If you were a hospitalist with a limited number of consult slots, what would you want from a screening tool before relying on it?

    Answer in the thread
If you have more time
  • Which results from a second hospital, with different patients and different screening habits, would make you trust the model's 0.783 area under the curve more or less?

    Answer in the thread
  • Functional status, albumin, medical optimization and cardiac comorbidities drove the predictions. How might a surgical team check whether those features hold up for patients in a hospital you know?

    Answer in the thread

What readers say