Interpretable ALS Diagnosis via Causal Path Retrieval

This repository contains the implementation of our study titled "Interpretable ALS Diagnosis via Causal Path Retrieval".

Authors

• Ali Salman - Department of Medical Biotechnologies, University of Siena, 53100 Siena, Italy.
• Alessio Rotelli - Department of Medical Biotechnologies, University of Siena, 53100 Siena, Italy.
• Matteo Leoncini - Department of Medical Biotechnologies, University of Siena, 53100 Siena, Italy.
• Leya El Halabi - Department of Medical Biotechnologies, University of Siena, 53100 Siena, Italy.
• Forough Izadi - Department of Medical Biotechnologies, University of Siena, 53100 Siena, Italy.
• Elena Niccolai - Department of Clinical and Experimental Medicine, University of Florence, Florence, Italy.
• Jessica Mandrioli - Department of Biomedical, Metabolic and Neural Sciences, University of Modena and Reggio Emilia, Modena, Italy.
• Amedeo Amedei - Department of Clinical and Experimental Medicine, University of Florence, Florence, Italy.
• Ernesto Iadanza - Department of Medical Biotechnologies, University of Siena, 53100 Siena, Italy.

Abstract

Amyotrophic lateral sclerosis (ALS) remains a diagnostic challenge because of its clinical heterogeneity and the lack of a definitive molecular test. We present a fully transparent diagnostic framework that replaces black-box classifiers with an explicit causal graph and deterministic path-based inference. A multi-omic panel of 34 blood and fecal biomarkers plus demographics was obtained from 118 subjects (97 ALS, 21 healthy controls). After a stratified 50/50 split, constraint-based causal discovery (PC algorithm, alpha=0.1) and path quantification were performed exclusively on the training half. The resulting directed graph, enriched with biological narratives, was indexed in a vector store. At inference time each held-out patient is scored by accumulating the signed contributions of the train-derived paths, yielding an auditable diagnostic logit. On the unseen test set of 59 patients, the system achieved 100 % accuracy, sensitivity and specificity. The strongest recovered edges, suppression of IL-2 and of propionic acid, are consistent with established immune and gut–brain mechanisms in ALS. Because every prediction is an explicit linear combination of path weights learned without access to the test data, the framework simultaneously delivers high predictive performance and complete clinical transparency.

Key Features

• Train-only Causal Discovery – Constraint-based structure learning (PC algorithm, α = 0.1) and path quantification are performed exclusively on a stratified training split; the held-out test set is never used for graph construction or coefficient estimation.
• Deterministic Path-Based Inference – Patient risk is computed by retrieving train-derived path weights from a ChromaDB vector store and accumulating signed contributions into an auditable diagnostic logit. No black-box classifier or language-model generation is involved.
• Fully Transparent Explanations – Every prediction is accompanied by a patient-specific balance sheet of risk and protective pathway impacts, enabling complete clinical auditability.
• Leakage-Free Preprocessing – Robust scaling parameters (median + IQR) are fitted only on the training half and then applied to both halves.
• Open & Reproducible Pipeline – End-to-end code built on causal-learn, ChromaDB, and modular Python scripts, released for seamless replication.

Citation

If you use this work, please cite it as:

@misc{Salman2026,
  author       = {Ali Salman and Alessio Rotelli and Matteo Leoncini and Leya El Halabi and Forough Izadi and Elena Niccolai and Jessica Mandrioli and Amedeo Amedei and Ernesto Iadanza},
  title        = {Interpretable ALS Diagnosis via Causal Path Retrieval},
  year         = {2026},
  institution  = {Department of Medical Biotechnologies - University of Siena, Siena, Italy},
  url          = {https://github.com/alexsalman/als/}
  urldate      = {2026-08-03}
}

⸻
© 2026 Ali Salman · MIT License