ASSOCIATIONS BETWEEN PREDICTED AND ACTUAL STRUCTURAL PROGRESSION IN THE APPROACH COHORT

Osteoarthritis Imaging, vol. 2, pp. 100049

Abstract

In the APPROACH study, machine learning models were trained to predict structural and pain progression of individual participants. Predicted progression was differentiated into a structural (S) and pain (P) progression score, where the score (range 0-1) reflected the likelihood of a participant being a progressor. Structural progression was defined as a decrease in minimum JSW of ≥0.3mm/year. Only participants with the estimated highest 75% combined (P+S) score after screening were included. To evaluate the associations between the predicted and actual structural progression according to different parameters over two years. 297 Participants were included at five European centers. Radiographs and 1.5T or 3T MRIs were acquired at baseline and two years. Minimum and mean medial and lateral tibiofemoral (TF) JSW, subchondral bone density, and osteophyte size were measured from radiographs. Medial and lateral cartilage thickness was measured from 3D SPGR MRI scans using manual, quality-controlled segmentations. Semi-quantitative MOAKS cartilage damage, bone marrow lesions (BML) and osteophytes scores were assessed from triplanar PDW and T1W sequences, summarized to overall medial and lateral scores. Progression was defined as two-year change greater than the smallest detectable change (SDC) for continuous measures or a change in ≥1 full MOAKS score, in the most affected compartment (MAC). For minimum JSW, progressors as predefined in the study protocol were analyzed as well. The association of S scores with progression was analyzed using logistic regression. Data was available for 237 participants. TF progression was seen most often in MRI cartilage thickness (38%) and radiographic bone density (39%) and osteophyte size (35%). S scores were not significantly associated with progression of any TF structural parameter besides minimum JSW, both based on the predefined progression criterion of ≤-0.6mm and on the SDC of <-0.49mm over two years. However, baseline minimum JSW was used for calculation of S scores, and after adjustment for baseline minimum JSW, the S scores were no longer associated with minimum JSW progression (both p>0.38). Selecting only participants with radiographic OA (KL grade ≥2, 53%) resulted in a higher percentage of progressors for all parameters (e.g. 50% MRI cartilage thickness) but this did not change results for S score association. Despite the use of machine learning models in the APPROACH cohort to only include participants that were predicted to show fast pain and/or structural progression, only around 1 in 6 participants was a structural progressor based on the predefined criterion. It is possible this number would have been lower without the model-based inclusion. Still, the S scores were generally not associated with progression. The machine learning model was trained using a different progression criterion than most criteria used in this validation; machine learning models trained on the actual criteria used for this analysis may have been more effective and is now possible with the data from the APPROACH cohort. EU/EFPIA Innovative Medicines Initiative Joint Undertaking (grant n° 115770) WW: Chondrometrics GmbH; FWR: Boston Imaging Core Lab, LLC IMI-APPROACH (NCT03883568) participants and investigators CORRESPONDENCE ADDRESS: [email protected].

Authors 13

1. Mylène P. Jansen

   University Medical Center Utrecht

   Affiliation as printed

   University Medical Center Utrecht, Utrecht, Netherlands

2. W. Wirth

   Paracelsus Medical University

   Affiliation as printed

   Chondrometrics GmbH, Freilassing, Germany & Paracelsus Med. Univ., Salzburg, Austria

3. Frank W. Roemer

   Boston University · Universitätsklinikum Erlangen

   Affiliation as printed

   University Hospital & University of Erlangen, Erlangen, Germany & Boston University School of Medicine, Boston, MA, USA

4. Jaume Bacardit

   Newcastle University

   Affiliation as printed

   Newcastle University, Newcastle, UK

5. E.M. van Helvoort

   University Medical Center Utrecht

   Affiliation as printed

   University Medical Center Utrecht, Utrecht, Netherlands

6. A.C. Marijnissen

   University Medical Center Utrecht

   Affiliation as printed

   University Medical Center Utrecht, Utrecht, Netherlands

7. M. Kloppenburg Aachen

   Leiden University Medical Center

   Affiliation as printed

   Leiden University Medical Center, Leiden, Netherlands

8. Francisco Javier Blanco

   Universidade da Coruña · Instituto de Investigación Biomédica de A Coruña

   Affiliation as printed

   INIBIC-Universidade de A Coruña (UDC), A Coruña, Spain

9. Ida Kristin Haugen

   Diakonhjemmet Hospital

   Affiliation as printed

   Diakonhjemmet Hospital, Oslo, Norway

10. Françis Berenbaum

    Inserm · Sorbonne Université · Hôpital Saint-Antoine

    Affiliation as printed

    Sorbonne University, Inserm, APHP hôpital Saint-Antoine, Paris, France

11. Floris P J G Lafeber

    University Medical Center Utrecht

    Affiliation as printed

    University Medical Center Utrecht, Utrecht, Netherlands

12. Paco M J Welsing

    University Medical Center Utrecht

    Affiliation as printed

    University Medical Center Utrecht, Utrecht, Netherlands

13. S.C. Mastbergen

    University Medical Center Utrecht

    Affiliation as printed

    University Medical Center Utrecht, Utrecht, Netherlands