PredictNeuropathy / Research artifacts

PNS 2026 — Downloads

Figures, de-identified ML dataset, trained model, and reports backing the PredictNeuropathy abstract accepted at the Peripheral Nerve Society Annual Meeting 2026 (Maastricht, 13–16 June). All files are released for academic and educational use.

Back to calculator

One-click bundle

All figures, data, model and reports — single ZIP archive.

Download full bundle (.zip)

Abstract ID10303

Cohort408 labeled exams, 5 centers

ModelExtraTrees, 11 features

Internal CVAUROC 0.890, BalAcc 0.812

Cohort summary

Task: binary classification, hereditary (CMT spectrum) vs acquired demyelinating (CIDP / MMN / POEMS) neuropathy.
Total exams used: 408 labeled records (227 hereditary, 181 inflammatory) from HCFMRP-USP, legacy HCRP, Humanitas Milano, UFU and USP-SP.
Features (11): age, sex, sensory absence flag, median & ulnar distal latency / NCV / CMAP amplitude, plus median-to-ulnar CMAP ratio and NCV difference greater than 10 m/s.
Validation: stratified 5-fold CV with permutation, center-only baseline and leave-one-center-out group CV.

Figures

High-resolution PNG and PDF artifacts used for the poster and slide deck.

2026_neuropathy_class_distribution.pngF1, class balance
2026_neuropathy_roc_curve.pngF2, ROC AUROC 0.890
2026_neuropathy_precision_recall.pngF3, PR AP 0.846
2026_neuropathy_confusion_matrix.pngF4, confusion matrix
2026_neuropathy_feature_importance.pngF5, feature importance
2026_neuropathy_groupcv_by_center.pngF6, leave-one-center-out
2026_neuropathy_ml_summary.pdfsummary PDF
FIGURE_INDEX_2026-01-03.mdfigure index
FIGURE_LEGENDS.mdnarrative legends

De-identified ML dataset

Record-level data ready for re-analysis. No patient names; record_id is a UUID and source_center is the only provenance kept.

ml_dataset_hereditary_vs_inflammatory.csv408 x 17
ml_dataset_hereditary_vs_inflammatory.jsonlJSONL mirror
label_counts_by_center.csvclass x center counts
ml_feature_missing_report.csvmissingness per feature
feature_correlations_pearson.csvPearson
feature_correlations_spearman.csvSpearman
model_metrics_folds.csvper-fold, 7 models
model_metrics_summary.csvaggregated
ml_run_manifest.jsonmanifest + SHA-256

Trained model

ExtraTrees classifier (scikit-learn) used by the live calculator.

extra_trees_model.joblib~14 MB
final_model_manifest.jsonfeatures + CV metrics
model_manifest.jsonruntime manifest
group_cv_by_center.csvLOCO results

Reports and abstract metadata

Methodology summaries, ML comparison report (13 runs), and PNS 2026 metadata.

PNS_2026_ABSTRACT_METADATA_2026-01-02.mdauthors, affiliations
SESSION_SUMMARY_AND_PNS_ABSTRACT_2026-01-02.mdfinal abstract draft
ML_COMPARISON_REPORT_2026-01-03.mdall 13 runs compared
ML_PROCESSING_2026-01-02.mdpipeline + leakage audit
ARTIGO_UPDATE_2026-01-03.mdmanuscript update notes
NEUROPATHY_CALCULATOR_2026-01-02.mddeployment notes

Released for research and educational purposes. The dataset has been de-identified and aggregated under multi-center IRB oversight (Athos USP and partner committees). Re-use should preserve the cohort scope and acknowledge the original abstract (PNS 2026, ID 10303). The model is a triage aid and does not replace clinical evaluation.