Predicting the Placebo response in OA to Improve the Precision of the Treatment Effect Estimation


Scientific Poster presented on April 2021 at the Osteoarthritis Research Society International Meeting.


In osteoarthritis (OA) randomized clinical trials (RCTs), the magnitude and the variability of the placebo response have a negative influence when testing the statistical superiority of active compounds compared to placebo. Furthermore, the magnitude of this effect has tended to increase over time. The intrinsic complexity and the multifactorial aspect of this placebo response make it difficult to control in the statistical analyses.

To solve this issue, Tools4Patient has developed a machine learning-based model that predicts the placebo response in chronic pain diseases. This model, named Placebell©™, incorporates baseline patient features associated with the intensity of the disease, the demographics, and data from a questionnaire developed by Tools4Patient. This questionnaire assesses patients’ states and personality traits associated with a strong placebo response. Overall, this machine learning model provides a baseline covariate highly predictive of the placebo response. This baseline covariate can be used in the statistical analyses to adjust for each patient’s propensity to have a placebo response. As such, this adjustment provides a significant increase in the precision of the treatment effect estimation  and study power in chronic pain RCTs.

The purpose of this study was to assess the performance and applicability of the Placebell©™ placebo response model in a phase 2 trial conducted by UNITY Biotechnology testing the efficacy of single-dose intra-articular (IA) injection of UBX101 in patients suffering from painful osteoarthritis of the knee.


Study populations and designs.

This clinical trial was a randomized, double-blind, placebo-controlled, single-dose, parallel-group study to assess the efficacy, safety, and tolerability of a single-dose IA administration of UBX0101 in patients with moderate to severe painful knee OA. Approximately 180 patients were randomized (1:1:1:1) to one of four treatment groups (approximately 45 patients per group), all administered by IA route at Week 0. The four treatment groups were enrolled concurrently :

  • Group 1: Placebo
  • Group 2: UBX0101 0.5 mg
  • Group 3: UBX0101 2.0 mg
  • Group 4: UBX0101 4.0 mg

The study drug was administered as a single 8 mL IA injection at Week 0, Day 1. The primary endpoint analysis was the change from baseline to Week 12 of the WOMAC-Pain score in patients receiving a single 0.5, 2.0, or 4.0 mg dose of UBX0101 versus those receiving placebo.

Modeling the placebo response with ML

The placebo response model used in this study was built by meta-analysis including individual data from 211 placebo patients from several studies conducted by Tools4Patient. These chronic pain patients were suffering from peripheral neuropathic pain or painful osteoarthritis of the knee and hip. They received a blinded placebo (oral, BID) for a duration varying from 1 to 3 months. The primary endpoint was the change from baseline of the weekly mean of the daily average pain score (APS). The placebo response, as measured by the primary endpoint, was modeled using a gaussian process with a linear kernel.

In the current trial conducted by UNITY, the model predicted the placebo response for each patient and each endpoint using only data available at baseline.

Testing the performance of the model

The predictive performance of the model was tested both in the placebo group and the per-protocol population. The performance was evaluated with the Pearson’s correlation between the predicted placebo response and the actual observed response. The R-squared was also reported and represents the diminution of the variance of the estimated treatment effect. Higher R-squared values means better estimation of treatment effect and increased study power.

The model’s predictions were performed at baseline and could be considered as a baseline covariate under the EMA and FDA guidances for the use of baseline covariates. For the estimation of the treatment effect, we used this predicted baseline placebo covariate as any other typical covariate in an ANCOVA analysis. This adjusted analysis corrected for the range of placebo responsiveness in all trial patients.


Results of UNITY study

A strong placebo response was observed right after the injection that reached a plateau after 8 weeks. There was no statistically significant difference between any arm of UBX0101 and placebo at the 12-week endpoint for change from baseline in WOMAC-Pain (primary endpoint).

Results in prediction

Table 1 presents the Pearson’s correlations between the predicted placebo response and the observed responses in the placebo group. The predictions are highly statistically significant for all primary and secondary endpoints. The correlations were excellent and ranged between 45.8% and 59.7%. In particular, the model could explain 35.6% of the variance related to the placebo response of the primary endpoint (WOMAC-Pain).

Table 1 : Performance of the placebo model to predict the primary and secondary endpoints using baseline data on the per-protocol placebo population.

The model was also tested on all patients from the per-protocol population (N=173). The performance (Table 2) was also highly significant with all p-values below 0.001. The predictive placebo response model was able to explain 27.7% of the WOMAC-Pain response.

Table 2 : Performance of the placebo model to predict the primary and secondary endpoints using baseline data on PPP population.

Those results were consistent across time points with excellent results until end-of-treatment (24 weeks).

Placebell©™ adjusted estimation of the treatment effect

This predicted placebo covariate was used in the estimation of the treatment effect. We observed a 40% increased in the precision of the estimated treatment effect. These analyses adjusted for the predicted placebo response further confirmed the lack of efficacy of the UBX101 treatment in painful knee OA. In particular, we were able to show that the small differences observed between the treatment groups could, in part, be explained by slight imbalances in placebo response.


Tools4Patient has developed a machine learning model able to predict the placebo response in chronic pain RCTs. This model, named Placebell ©™, was used in phase II, randomised clinical trial conducted by UNITY Biotechnology investigating the effect of an intra-articular injection in approximately 180 patients suffering from painful osteoarthritis of the knee.

The model predictions were highly significant for all primary and secondary endpoint (p<0.001). In particular, the model explained 27.7% of the primary endpoint (WOMAC-Pain) variance in the per-protocol population. Similar performance was also observed for the secondary time points confirming the robustness of the Placebell©™ model. The results also demonstrate its applicability in RCTs with different modes of administration (oral vs. intra-articular injection).

We have also shown how this machine learning model could be used to improve the estimation of the treatment effect. Indeed, using the predictions as a covariate significantly reduced the variance resulting in a 40% increase in precision of the treatment effect. This gain in study power by increasing precision is exactly similar to that one would achieve by increasing sample size by 40% (here 72 additional patients).Overall, the current results further demonstrate the usability and performance of Tools4Patient’s placebo response model. Its excellent results across two indications, two modes of administration, and several endpoints raise the hope for wide applicability of this approach in pain indications.

Scientific Poster
Samuel Branders1; Jamie Dananberg2; Frederic Clermont1; Ben Xie2; Benjamin Hsu2; J. Visich2; Akbar Khan2; Lauren Masaki2; Alvaro Pereira1 1-Tools4Patient, Mont-Saint-Guibert, Belgium 2-Unity Biotechnology, Inc., South San Francisco, CA
April 2021
Osteoarthritis Research Society International (OARSI)