Carotid plaque detection in patients with type 2 diabetes: development, interpretability, and limited transportability of a clinical model

OBJECTIVE: To develop and evaluate a parsimonious model for detecting prevalent carotid plaque in type 2 diabetes mellitus (T2DM), and to assess its interpretability, robustness, and transportability. METHODS: This retrospective cross-sectional study enrolled 2,035 hospitalized patients with T2DM, randomly assigned 70:30 to training and test sets. Bidirectional stepwise regression screened 39 candidate variables; stability was assessed by bootstrap resampling (1,000 replicates) and penalized regression. Performance was compared with random forest and XGBoost, and TreeSHAP was used for interpretability. Interaction terms, a non-medication model, and high-sensitivity thresholds were evaluated. The model was applied to an independent cohort of 3,191 health-screening individuals, with updating assessed by split-half validation. RESULTS: Eight predictors were retained. The test-set AUC was 0.827 (95% CI 0.791-0.862), versus 0.804 for random forest and 0.816 for XGBoost (P = 0.023 versus random forest). Only age and statin use were selected in 100% of bootstrap resamples. Removing statin use and GLP-1RA lowered the AUC to 0.787 (ΔAUC - 0.040, P < 0.001), retaining a history of coronary heart disease. No interaction term was significant (joint P = 0.511). In the external cohort, the AUC was 0.666 with marked miscalibration; updating left discrimination unchanged, and at 85% sensitivity, specificity was 33.4%. CONCLUSION: An eight-variable logistic model using routinely available variables performed comparably to machine learning models for detecting prevalent carotid plaque in the development cohort. However, variable selection was unstable, and statin use was the strongest but indication-confounded correlate. External discrimination was moderate and did not improve after recalibration, indicating limited transportability. The model is best regarded as an exploratory candidate requiring further setting-specific validation, rather than one that may already aid in the detection of clinical cases.

Authors

Institutions

Publication Details

Journal
Endocrine
Published
2026-11-16
DOI
https://doi.org/10.1007/s12020-026-04828-9
Primary Topic
Artificial Intelligence in Healthcare
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Carotid plaque detection in patients with type 2 diabetes: development, interpretability, and limited transportability of a clinical model

Xun Liu, Xiao ju Wang, Gang Xu
1 links
Endocrine
Artificial Intelligence in Healthcare
article

Carotid plaque detection in patients with type 2 diabetes: development, interpretability, and limited transportability of a clinical model

Xun Liu, Xiao ju Wang, Gang Xu
article en

Abstract

OBJECTIVE: To develop and evaluate a parsimonious model for detecting prevalent carotid plaque in type 2 diabetes mellitus (T2DM), and to assess its interpretability, robustness, and transportability. METHODS: This retrospective cross-sectional study enrolled 2,035 hospitalized patients with T2DM, randomly assigned 70:30 to training and test sets. Bidirectional stepwise regression screened 39 candidate variables; stability was assessed by bootstrap resampling (1,000 replicates) and penalized regression. Performance was compared with random forest and XGBoost, and TreeSHAP was used for interpretability. Interaction terms, a non-medication model, and high-sensitivity thresholds were evaluated. The model was applied to an independent cohort of 3,191 health-screening individuals, with updating assessed by split-half validation. RESULTS: Eight predictors were retained. The test-set AUC was 0.827 (95% CI 0.791-0.862), versus 0.804 for random forest and 0.816 for XGBoost (P = 0.023 versus random forest). Only age and statin use were selected in 100% of bootstrap resamples. Removing statin use and GLP-1RA lowered the AUC to 0.787 (ΔAUC - 0.040, P < 0.001), retaining a history of coronary heart disease. No interaction term was significant (joint P = 0.511). In the external cohort, the AUC was 0.666 with marked miscalibration; updating left discrimination unchanged, and at 85% sensitivity, specificity was 33.4%. CONCLUSION: An eight-variable logistic model using routinely available variables performed comparably to machine learning models for detecting prevalent carotid plaque in the development cohort. However, variable selection was unstable, and statin use was the strongest but indication-confounded correlate. External discrimination was moderate and did not improve after recalibration, indicating limited transportability. The model is best regarded as an exploratory candidate requiring further setting-specific validation, rather than one that may already aid in the detection of clinical cases.

EndocrineVol. 91(1)
People's Hospital of Kaizhou District (CN), Chongqing Emergency Medical Center (CN), Chongqing Medical University (CN)
Good health and well-being
Artificial Intelligence in Healthcare
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon