Predicting the Year of Total Knee Replacement: A Transformer-Based Multimodal Approach

Ozkan Cigdem, Refik Soyak, Kyunghyun Cho, Cem M Deniz

Keywords: Multimodal Learning, Year of TKR Prediction, Deep Learning, Knee Osteoarthritis

Abstract:

Accurate prediction of the year of total knee replacement (TKR) is challenging due to
the complex interplay of factors influencing the surgical decision. Current deep learning
models often rely on single-modality data, limiting their predictive power. Multimodal
approaches integrating imaging and patient data offer the potential to improve predictions
and support clinical decisions. This study presents an end-to-end trained, transformer-
based multimodal model that integrates MR imaging with tabular data, including clinical
variables and image readings, to predict the year of TKR for each subject. Our model lever-
ages cross-modal attention to fuse features from an image encoder with a self-supervised
pretrained tabular encoder, achieving the highest accuracy of 63.4% among tested mod-
els. We evaluated its performance against three unimodal models and four multimodal
fusion strategies, including simple concatenation, DAFT, and multimodal interaction. The
results demonstrate that our model’s cross-modal interaction approach with pretrained
TabNet not only outperformed all unimodal models but also showed improvements over
other multimodal fusion techniques, highlighting the effectiveness of cross-modal attention
fusion for integrating complex data modalities in TKR year prediction tasks. Source code
is available at https://github.com/denizlab/2025_MIDL_time2TKR.

Primary Subject Area: Integration of Imaging and Clinical Data

Secondary Subject Area: Application: Radiology

Paper Type: Both

Registration Requirement: Yes

Reproducibility: https://github.com/denizlab/2025\_MIDL\_time2TKR

Visa & Travel: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl25_NNN.zip file with midl25_NNN.tex, midl25_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2025}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., Appendices and supplementary material are included in the same PDF after references., Main paper does not exceed 9 pages; acknowledgements, references, and appendix start on page 10 or later.

Submission Number: 220