Professional Abstract
Objectives
(1) To use nonrandomized data from patients with knee osteoarthritis to create models that predict patient-specific outcomes of different treatment options; (2) To use the models to develop software to help identify patients who may be eligible to enroll in randomized controlled trials (RCTs)
Study Design
| Design Elements |
Description |
| Design |
Empirical analysis (nonrandomized study) |
| Data Sources and Data Sets |
A consolidated database of 1,452 patient knees, half with total knee replacement and half with nonsurgical treatment, from 1,322 patients, combining data from 4 databases: Multicenter Osteoarthritis Study, Osteoarthritis Initiative, New England Baptist Hospital Orthopedic Registry, and Tufts Medical Center Orthopedic Surgery Registry |
| Analytic Approach |
Multivariable linear regression, “greedy” matching computer algorithm, multiple imputation |
| Outcomes |
Pain, based on observed or estimated Western Ontario and McMaster Universities Arthritis Index (WOMAC) score; functional status based on SF-12 Health Survey score |
When conducting RCTs to compare treatments, researchers must recruit only patients with clinical equipoise, that is, patients for whom insufficient evidence exists to favor one treatment over another. When limited prior RCT evidence is available to identify patients with clinical equipoise, researchers can apply mathematical models to clinical registries, electronic health records (EHRs), and other non-RCT data to predict patient-specific outcomes of the treatments under study. If predicted outcomes are similar across treatments, called mathematical equipoise, random treatment assignment may be appropriate, and patients may wish to consider participating in an RCT.
For patients with knee osteoarthritis, the choice between total knee replacement and nonsurgical treatment is an important clinical question for which there are few RCTs. Nonsurgical treatment may include medication and/or physical therapy. In this study, the researchers developed Knee Osteoarthritis Mathematical Equipoise Tool (KOMET) software, with accompanying clinician and patient web-based interfaces, for use in EHR systems to
- Help identify patients with mathematical equipoise who could consider enrolling in RCTs
- Support decision making by providing patients with individualized, predicted outcomes for treatment options
To develop KOMET, the researchers used nonrandomized data from four databases to match total-knee-replacement knees to similar nonsurgical-treatment knees. The researchers then developed models to predict one-year outcomes for knee pain and functional status, modeling each outcome separately. Analysis consisted of three rounds of predictive modeling based on estimation and testing using data from various combinations of the four available databases. After identifying optimal prediction models for each outcome, the researchers programmed associated algorithms into the KOMET software.
During the study, knee osteoarthritis researchers, patients, clinicians, and patient advocates provided input on study questions, modeling issues, outcomes, and user interface development.
Results
The final model used for predicting WOMAC pain scores included main effects for baseline WOMAC knee pain, treatment type, and an interaction of the two. The predictors accounted for 32% of the variance in pain scores.
The final model used for predicting SF-12 functional status scores included main effects for age, gender, baseline SF-12 mental and physical component scores, body mass index, and treatment type. The predictors accounted for 34% of the variance in functional status scores.
Limitations
Differences among the various databases made it challenging to match comparable patients, which could have affected the models’ accuracy. The research team defined functional status using a measure of overall physical functioning; a survey that measured knee-specific functioning may have better captured meaningful knee function improvement. Many of the patient variables under consideration were burdensome to collect or difficult to capture in a consistent manner. Thus, the team chose to develop the KOMET software using algorithms from only the models that were based on databases containing a more limited set of predictor variables.
Conclusions and Relevance
The results of this study demonstrate the use of mathematical modeling for identifying potential enrollees in RCTs, or when RCT enrollment is not appropriate, for informing treatment decisions based on predicted outcomes.
Future Research Needs
Future research could test this approach with other medical conditions requiring important treatment decisions. Researchers could also develop approaches to lessen the bias inherent in nonrandomized data. In addition, using a more knee-specific functional scale, researchers could work to develop improved models to predict patient outcomes, applying newer statistical procedures and validation and using larger databases.
