Performance, Generalizability, and Fairness of a Peripheral Artery Disease Detection Model Across Patient Phenotypes and Health Systems

Author affiliations

University of California San Diego
University of California Irvine
University of California San Francisco
University of California Davis
Stanford University

In plain English

The study pairs case-control prediction with phenotype clustering, showing that a stable overall score can hide weaker detection for a clinically distinct, lightly documented group.

How the study worked

A plain-language walk through the work behind the result.

Matched 33,739 peripheral-artery-disease cases with 33,739 controls across five health systems.
Trained LightGBM on 14,023 EHR-derived features and assessed demographic and phenotype subgroups.

What they found

Reported AUROC and AUPRC ranged from 0.76 to 0.79 across institutions.
Sensitivity ranged from 0.87 in one phenotype cluster to 0.40 in the sparsely documented cluster.

Why it matters

Phenotype-level audits can reveal deployment failure modes that pooled fairness and performance summaries miss.

The catch

The manuscript is a preprint.
The private EHR data are unavailable publicly, limiting independent reproduction.

Methods at a glance

Study type: Retrospective multicenter EHR modeling preprint
Biological system: Peripheral artery disease
Organism: Human
Dataset: University of California Health Data Warehouse across five systems
Sample size: 33,739 cases and 33,739 matched controls
Model architecture: LightGBM using 14,023 features
Training objective: Peripheral artery disease case detection
Code: No public code repository is linked from the preprint record.
Data: The authors state that the EHR data are not public because of privacy protections and data-use agreements.

Tasks and metrics

Tasks: Case detection; Calibration; Demographic fairness; Phenotype-stratified sensitivity

Metrics: AUROC 0.76-0.79; AUPRC 0.76-0.79; True-positive rate

Baselines: Cross-institution and subgroup comparisons of the same classifier

Principal findings

Sensitivity was 0.87 in cluster 2 and 0.40 in cluster 1.

What remains unresolved

External validation: Evaluation covered five University of California health systems.

The work is a preprint.
Private EHR data limit independent reproduction and transportability testing outside the UC system.