CS4486 AMP-PD checkpoint collection

Existing fitted models from the CS4486 Parkinson's disease progression project.

Results

SMAPE+1 is lower-is-better. The main comparisons use 248 patients, 1,068 clinical origins with proteins, and 12,578 observed labels.

Model Original split Mean ± sample SD over split seeds 42/142/242
Direct MLP baseline 70.611 Not run
Shared distribution MLP reference 64.026 64.120 ± 0.166
Selected AE + Ridge residual 63.685 63.845 ± 0.181
PCA + Ridge residual 63.537 63.813 ± 0.517
PLS + Ridge residual 63.379 63.900 ± 0.496

PCA has the lowest three-split mean. Selected AE improves its same-split reference in all three splits and has the smallest variation among the protein methods. These are local CV results, not hidden Kaggle leaderboard scores.

Contents

checkpoints/ preserves the local experiment hierarchy:

Directory Contents
retained/baseline_mlp/ Five original CV folds and an all-patient model; preprocessing + direct regression MLP
retained/pls/ Five original CV folds; shared reference ensemble + PLS/Ridge parameters
retained/selected_ae/ Five original CV folds and an all-patient model; AE + reference ensemble + selected residual heads
improvement_ablation_2026-11-16_v2/ Rebuilt PCA and fixed-parameter ablation heads; loads neural dependencies from retained/
three_split_seeds_2026-11-17/ Five refitted model bundles for each of the three split seeds
original_configs_2026-11-10/ Earlier regression/distribution MLP checkpoints and preprocessing
top_methods_2026-11-10/ History-enabled adaptations of leading solution ideas
isolated_inference_2026-11-10/ Earlier frozen and retrained inference-boundary models
top_nohistory_2026-11-10/ Current-visit adapters with frozen and retrained models
leakage_2026-11-10/ Saved preprocessing from early controls; Ridge coefficients were not exported in that run

History-enabled and INVALID_* diagnostic artifacts retain their original names. They are supplementary records, not main zero-history results. Old weights previously deleted locally are not recreated in this release.

checkpoint_manifest.json and .csv list every published parameter file, source-relative path, SHA-256 and any packaging transformation. Preprocessing arrays retain fitted statistics but omit dataset row-index arrays. Evaluation-probability dumps, raw clinical/protein/peptide data, participant-level OOF files and split membership tables are not included.

code/ampd/ contains the inference source snapshot needed to load the current packages, including modules that may not yet be present in the linked GitHub revision. load_models.py provides portable convenience loaders. Ablation references are relative to the downloaded checkpoint root. Legacy pickle preprocessors require the matching Python package environment; current main model loaders use numeric arrays and PyTorch state dictionaries.

Load a current model

Install requirements.txt and huggingface_hub in a separate environment, then:

import sys
from huggingface_hub import snapshot_download

snapshot = snapshot_download('fryabcfry/cs4486-amp-pd-checkpoints')
sys.path.insert(0, snapshot)
from load_models import load_model

model = load_model('pls', fold=0, root=snapshot)
# Replace protein_vector with current measurements in model.spec['proteins'] order.
events = [{'visit_id': 'current_sample', 'visit_month': 12,
           'proteins': protein_vector}]
prediction = model.predict(events)['pls']   # N x 16

Outputs are ordered by UPDRS component, with horizons 0, 6, 12, 24 inside each component. Missing protein values can be None or NaN.

Loader name Output key Final all-patient model
baseline_mlp month_protein fold='final'
selected_ae selected_latent_r fold='final'
pls pls CV folds only
pca pca_full CV folds only
ablations ae_full, pca_full, pls_full, and ablation names CV folds only

For repeated-split models:

model = load_model('selected_ae', split_seed=142, fold=0, root=snapshot)
outputs = model.predict(events)
# outputs: reference, selected_ae, pca, pls

The split-seed loader returns all three residual predictions around the same reference. Full-data models have no separate test score; the reported metrics come from held-out CV folds. The original data must be obtained from the Kaggle competition to retrain the models.

Related work

The report cites SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling (Lendering, Akdag and Bondarev, CVPR 2026, pp. 28557–28566) as a related frozen-feature/subspace design. SubspaceAD uses a frozen DINOv2 image encoder, PCA and reconstruction residuals. Our selected AE is trained on training-patient proteins, frozen for latent screening, and followed by supervised Ridge prediction of UPDRS residuals. This collection does not contain DINOv2 or SubspaceAD pretrained weights.

Paper/project: https://clendering.github.io/SubspaceAD/

Scope

This is a course research artifact for comparing input availability and protein representations. It is not a clinically validated prediction system.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support