PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
April 4, 2026Topics in Spinal Cord Injury Rehabilitation0 citations

The ASIA Data Science Challenge: Predicting Functional and Neurological Recovery From Acute ISNCSCI Scores

View Full Paper
JVJ. VillinesRSR. StirnimannLLL. P. Lukas

Key Points

  • This research aims to develop predictive models for functional and neurological recovery after spinal cord injury using acute assessment scores.
  • Utilized data from the Sygen clinical trial involving 797 participants.
  • Involved two prediction tracks: motor status and walking function.
  • Applied ensemble boosting models and TabPFN transformer for predictions.
  • Conducted interpretability analyses to identify key predictors.
  • The winning motor status model achieved RMSE of 1.0.
  • The successful walking function model recorded a Spearman's ρ of 0.85.
  • Both models showed competitive performance despite limitations in dataset size and missing data.

Abstract

Background: Recovery after spinal cord injury (SCI) is driven by a variety of factors, making SCI recovery prediction particularly suitable for data-driven modelling. The American Spinal Injury Association (ASIA) Engineering and Data Science Committee launched the first open SCI data science challenge focused on recovery prediction. Objectives: Hosted on Kaggle in 2025, the challenge addressed 2 tasks using data from the Sygen clinical trial (N = 797): (1) predicting motor status as assessed by the International Standards for Neurological Classification of SCI (ISNCSCI; score: RMSE root mean square error between true and predicted motor scores), and (2) predicting walking function on the 8-point modified Benzel scale (score: Spearman's ρ). Method: Eight teams participated in the motor status track and 14 in the walking function track. The winning solution (RMSE = 1.0) in the motor status track applied an ensemble of boosting models with multiple imputation and Bayesian hyperparameter optimization, whereas the winning solution in the walking function track (Spearman's ρ= 0.85) harnessed TabPFN, a pretrained transformer for tabular data, combined with feature engineering from early ISNCSCI assessments and auxiliary predictions of future motor scores. Results: Both approaches demonstrated competitive performance for SCI recovery prediction from acute phase assessments, despite modest dataset size and missing data. Both winning solutions further conducted an interpretability analysis, identifying key predictors consistent with clinical understanding of SCI recovery. Conclusion: This challenge highlights the value of publicly accessible formats, which encourage broader engagement of diverse audiences, for developing robust predictive models and provides a proof of concept for future initiatives leveraging larger, multicenter, multimodal datasets to improve personalized prognostication in SCI rehabilitation.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Villines et al. (2026) studied this question.

synapsesocial.com/papers/69d0af83659487ece0fa5762https://doi.org/10.46292/sci25-00137
Ask AI
Helpful
Bookmark
Share
View Full Paper