Transparent AI for Orthopedic Surgery Prediction

CareFuse is the first company to build a counterfactual AI model for total knee arthroplasty (TKA), predicting individualized outcomes with and without surgery. With ~790k TKAs annually in the U.S., and 30% of patients gaining no benefit, our SHAP-explained platform empowers payers to identify the right moment for the right patient, reducing unnecessary surgeries and costs.

$9+ Billion
spent in the US per year
in knee replacement every year for the surgery alone (not including revisions or readmissions). [2] This number is predicted to increase to $16.4 billion by 2032. [3]
30%
of Total Knee Replacements
ultimately leave the patient dissatisfied [5]
93%
CareFuse Model AUC
Clinical prediction accuracy for MCID achievement at 12 months
CareFuse Solution

Clinical Decision Support

Predict whether a member will achieve MCID (ΔWOMAC ≥ 10) at 12 months with transparent, explainable AI built for UM workflows.[10]

Patient Profile

Feature Value
Age 74
Sex F
Body Mass Index (BMI) Obese
Baseline Pain & Function 15.6
Baseline Pain 4.0
Baseline Stiffness 1.0
Baseline Disability 10.6
Lifestyle Modification None

CareFuse is an FDA-exempt decision assistance tool that supports providers with data-driven insights. It is not intended to deny care but to assist clinicians in making informed decisions.

About TKA Surgery

Total knee arthroplasty (TKA) is an invasive surgery with risks and a lengthy recovery period. This software helps identify potentially ineffective surgeries before they occur to protect patient well-being.[11]

TKA Success

Prediction: Low Benefit Likelihood
Predicted Probability: 21%
Surgery benefit is unlikely
Summary: The model estimates a 21% chance of success, with a cutoff of 37%. Your prediction is driven up by Body Mass Index, but somewhat offset by Baseline Disability, Baseline Stiffness, Baseline Pain & Function. Recommend initiating intensive conservative care to mitigate factors that worsen knee osteoarthritis before revisiting surgical authorization. Support a 10% weight loss prior to surgical reconsideration.
Top Influencers
Baseline Disability 18% hurts
Baseline Stiffness 18% hurts
Baseline Pain & Function 15% hurts
Body Mass Index (BMI) 6% helps

Conservative Success

Predicted Probability: 60%
Conservative care benefit is likely
Summary: The model estimates a 60% chance of success, with a cutoff of 50%. Your prediction is driven up by Baseline Pain, Baseline Pain & Function, but somewhat offset by Lifestyle Modification. Recommend continuing conservative care practices such as weight loss and physical therapy to sustain improvement.
Top Influencers
Baseline Pain 23% helps
Baseline Pain & Function 18% helps
Baseline Disability 16% helps
Baseline Stiffness 12% helps

Counterfactual Analysis

Propensity Score Weighting: Our model generates counterfactual predictions for conservative care using propensity-score weighting, allowing direct comparison between surgical and non-surgical treatment outcomes for informed decision-making.

Download PDF
Back

Main contributing factors List of features that most influence the predicted outcome.

Decision Support

Not auto-denial—configurable thresholds aligned with plan policy. Human-in-the-loop review required for all decisions.

Workflow Integration

Built for InterQual/MCG workflows with HL7 FHIR PAS compatibility. Seamless integration with existing UM systems.

Complete Documentation

Predicted probability, risk narrative, literature snippets, and exportable FHIR bundles for comprehensive documentation.

Clinical Evidence

Validation & Performance

At CareFuse, we've built a calibrated, explainable binary classification model (OAI: demographics + WOMAC) that estimates a TKA patient's probability of achieving MCID and generates a counterfactual for conservative care using propensity-score weighting. Across 50 multi-seed holdout splits it achieves AUC ≈ 0.93, with transparent calibration/thresholding and SHAP explanations suited for clinical review.

How we generate a prediction

This tool can generate a personalized prediction using information the patient can readily provide (basic demographics and medical history) together with responses to an internationally validated arthritis questionnaire . No imaging or specialized tests are required for the demo.

Validation Highlights

Up to 97%
Accuracy
AUC 0.93
TKA pathway
AUC 0.87
Conservative pathway
Sensitivity
to non-MCID procedures

Internal validation. Performance depends on thresholds and population. For demonstration only; not a coverage determination.

Methods

Our binary classification model uses demographics and WOMAC scores from the OAI dataset. We completed a live pilot with Hapvida (15.7M policyholders) to demonstrate feasibility and workflow integration. In the U.S., our pilot would train a new model on your de-identified data using the same pipeline and deploy a secure, metered API with audit logging.

CareFuse is represented by Gunderson Dettmer, and our product is built under the FDA non-device CDS exemption.

We also follow machine learning best practices, including strict train/test data splits, holdout validation, and transparent metric computation to ensure trustworthy, reproducible results.

Next: Problem