ASIA 2026 Data Science Challenge — Track 2

Longitudinal reconstruction of full ISNCSCI sensory examinations

• Challenge: ASIA Data Challenge
• Paper: in preparation; public link and DOI forthcoming
• Archived code release: Zenodo

This repository contains the SCAI Lab winning solution for Track 2 of the 2026 American Spinal Injury Association (ASIA) Data Science Challenge. Track 2 asked teams to reconstruct omitted follow-up sensory findings from a complete baseline International Standards for Neurological Classification of Spinal Cord Injury (ISNCSCI) examination and an expedited follow-up examination.

TL;DR

• What we did: Reconstructed the missing follow-up sensory grid using a target-wise TabPFN v2.5 classifier, while retaining all observed follow-up measurements unchanged.
• Why: Sensory grades are ordered clinical states, and the examination has meaningful longitudinal and anatomical structure that a general-purpose tabular model does not enforce by itself.
• What worked: Five-fold probability averaging provided the selected base model; small, auditable anatomical corrections then improved the final challenge submission.
• Main result: The selected pipeline achieved 0.38056 public RMSE and 0.40706 private RMSE, ranking first in Track 2. Lower RMSE is better.
• Takeaway: Cautiously injecting domain-specific structure may help contextualize broadly pretrained tabular models for clinical data, but this finding remains challenge-specific and requires prospective external validation.

Method at a glance

Complete baseline ISNCSCI
            +
Expedited follow-up ISNCSCI
            +
Clinical metadata
            │
            ▼
Target-wise leakage masking
            │
            ▼
Five-fold TabPFN classification
            │
            ▼
Probability averaging → expected sensory score
            │
            ├── observed follow-up value → copy unchanged
            │
            └── omitted value → conservative anatomical post-processing
                                  1. pairwise shrinkage
                                  2. anchor correction
                                  3. extended anchor correction

Sensory outcomes were represented as the ordered classes 0, 1, and 2; binary anal sensation was handled as a two-class target. For each sensory target, the corresponding follow-up feature was masked before fitting to prevent target leakage. Class probabilities were converted to a continuous expected score for RMSE evaluation.

The selected implementation used shuffled row-wise five-fold cross-validation with random_state=42. Each round used approximately 80% of development rows for training and 20% for validation, with every row served once for validation, and the five test-prediction sets were averaged.

Results

Final submissions

Pipeline	Description	Public RMSE	Private RMSE
Method 1 — selected	Five-fold discrete TabPFN ensemble plus sequential anatomical post-processing	0.38056	0.40706
Method 2	Five-seed probability ensemble without the Method 1 post-processing chain	0.38514	0.41142

Development-set model comparison

We reported the following performance across all sensory entries and on the subset of cells that required reconstruction:

Model formulation	Validation	All-entry RMSE	Reconstructed-cell RMSE
Direct TabPFN regression	3-fold	0.393 ± 0.010	0.422 ± 0.010
Single discrete TabPFN classifier	3-fold	0.377 ± 0.010	0.405 ± 0.011
Discrete TabPFN ensemble	5-fold	0.374 ± 0.015	0.402 ± 0.015

Values after ± are the sample standard deviation of fold-specific RMSE. Post-processing was deterministic and evaluated using fixed public/private leaderboard partitions.

Included methods

Method 1: selected Track 2 submission

Method 1 details · pipeline entry point

Method 1 combines the five-fold discrete TabPFN base with three conservative post-processing stages:

1. Pairwise hedge shrink softly reduces large disagreements between left/right or light-touch/pin-prick pairs when both values are missing.
2. Anchor correction moves a missing prediction toward a training-derived conditional mean associated with a related observed follow-up value.
3. Extended anchor correction applies a final, narrowly gated correction where a nearby observed value indicates that the current prediction may be too low.

All three stages operate only on unobserved target cells. Recorded follow-up measurements are copied through and are never overwritten.

Method 2: secondary submission

Method 2 details · script

Method 2 trains five full-development-set models with seeds 11, 22, 33, 44, 55, averages their class probabilities, and converts the result to expected sensory scores. The Method 1 post-processing chain was not applied to this submission.

Repository structure

asia-challenge-2026/
├── method_1/
│   ├── README.md
│   ├── scripts/                 # Selected five-fold model and post-processing
│   └── data/submissions/
│       ├── base_model_submission.csv
│       ├── pairwise_shrinkage_submission.csv
│       ├── anchor_corrected_submission.csv
│       └── final_submission.csv
├── method_2/
│   ├── README.md
│   ├── scripts/                 # Five-seed probability ensemble
│   └── data/submissions/
│       └── seed_ensemble_submission.csv
├── utils/                       # Data loading, metrics, and shared utilities
├── requirements.txt             # Recorded Python environment
└── README.md

The repository preserves the final submission CSVs. Run summaries, fold-level metrics, out-of-fold predictions, and stage summaries are generated when the pipelines are rerun under runs/; historical run directories are not included in the current repository snapshot.

Data access and expected files

The de-identified challenge data are not redistributed in this repository. Access is governed by the challenge data-use conditions. Place the authorized Track 2 files in data/:

data/
├── features_train_2.csv
├── features_test_2.csv
├── labels_train_2.csv
├── labels_test_2_dummy.csv
├── metadata_train_2.csv
└── metadata_test_2.csv

The scripts merge features and metadata by ID.

Environment

The recorded environment used:

• Python 3.11
• CUDA-enabled PyTorch
• tabpfn==6.4.1 with pretrained TabPFN v2.5 classifier weights
• numpy, pandas, scikit-learn, and huggingface_hub
• NVIDIA GeForce RTX 4090 for the primary runs

The pinned package snapshot is in requirements.txt. Install a PyTorch build compatible with the local CUDA driver, then install the remaining recorded dependencies in an isolated environment.

This release specifically used TabPFN v2.5. The v2.5 weights are not redistributed in this repository. They are obtained separately from Prior Labs on Hugging Face and are subject to their own gated, non-commercial terms. Access may require approval and acceptance of those terms.

Approximate recorded runtimes were:

Hardware	Method 1	Method 2
NVIDIA RTX 4090	~1 hour	~3 hours
NVIDIA RTX 2080	~2 hours	~6 hours

Running the code

Run commands from the repository root.

Selected Method 1 pipeline

python method_1/scripts/run_pipeline.py \
  --data-root data \
  --run-root runs

The pipeline executes the five-fold base model and all three post-processing stages. Its final submission is written to:

runs/<pipeline_run_id>/predictions_test.csv

Method 2

python method_2/scripts/train_seed_ensemble.py \
  --data-root data \
  --run-root runs \
  --do-cv 1 \
  --n-splits 5

When cross-validation is enabled, the run directory also contains cv_metrics.json, weighted_oof.json, and oof_predictions_train.npz.

Clinical use and limitations

This code is for research and benchmarking. Reconstructed values are estimates, not recorded neurological findings, and should not replace a clinically indicated full ISNCSCI examination. The challenge used one clinical-trial dataset and an evaluation-specific missingness pattern. Generalization across centres, injury groups, examination time points, and routine-care missingness has not yet been established.

Any downstream use should:

• clearly label imputed values and preserve their provenance
• retain observed clinical measurements unchanged
• assess clinically consequential errors, not only aggregate RMSE, and
• undergo external and prospective validation before clinical deployment.

Publication status

An ASIA Data Science Challenge manuscript is in preparation for Topics in Spinal Cord Injury Rehabilitation, with publication targeted for the ASIA 2027 cycle. The manuscript has not yet been assigned a public DOI. Frozen software releases are preserved on Zenodo; cite the version-specific DOI associated with the exact release used in the manuscript.

Citation

Citation metadata, including the author ORCID, are provided in CITATION.cff. For re-use and re-distribution, cite the version-specific Zenodo release rather than the mutable repository branch.

Purkayastha, Partha Sarathi. ASIA 2026 Track 2: Longitudinal ISNCSCI
Sensory Reconstruction. Zenodo, 2026.
ORCID: https://orcid.org/0009-0007-8879-7622

Licensing

The repository code is released under the MIT License:

Copyright (c) 2026 Partha Sarathi Purkayastha

The license permits use, modification, and redistribution while requiring the copyright and permission notice to be retained. The separately obtained TabPFN v2.5 model weights remain subject to the Prior Labs terms linked above.

Acknowledgments

We thank the American Spinal Injury Association, the ASIA Engineering and Data Science Committee, the challenge organizers, the Sygen data contributors, and the individuals whose de-identified examinations made this benchmark possible.