# Identifying metabolic parameters as key indicators of hyperuricemia and ischemic stroke comorbidity via interpretable Clinlabomics models

## TL;DR

This study identifies key metabolic indicators for hyperuricemia and ischemic stroke comorbidity using interpretable machine learning models.

## Contribution

The study introduces an interpretable Clinlabomics model to assess risk and identifies specific metabolic parameters linked to the comorbidity.

## Key findings

- AIP was the strongest risk factor for HUA-IS comorbidity (OR = 2.74).
- The rpart-based model achieved high performance with an AUC of 0.986 in testing.
- SHAP analysis identified UA_admission, UA_3d, TyG, TG, AIP, and LCI as key indicators.

## Abstract

Ischemic stroke (IS) with hyperuricemia (HUA) correlates with poor outcomes, yet the shared pathophysiological traits remain unclear. This study examined metabolic parameters in HUA-IS comorbidity and developed an optimal interpretable Clinlabomics model for risk assessment.

A total of 2,164 IS patients and 2,459 healthy controls (HCs) were retrospectively enrolled. Participants were divided into four groups: HUA-IS (comorbidity, n=1,082), non-HUA IS (n=1,082), HUA HCs (n=1,314), non-HUA HCs (n=1,145); the latter three were defined as the non-comorbidity group. After 1:1 propensity score matching (PSM), 1,031 cases were matched in each group. Ten metabolic parameters were analyzed: serum uric acid at admission (SUA_admission), SUA on the third day of hospitalization (SUA_3d), triglyceride-glucose index (TyG), triglyceride (TG), high-density lipoprotein cholesterol (HDL−C), atherogenic index of plasma (AIP), atherogenic coefficient (AC), lipoprotein combine index (LCI), Castelli’s risk index I (CRI-I), and Castelli’s risk index II (CRI-II). Univariate/multivariate logistic regression, quartile-based logistic regression, and restricted cubic spline (RCS) analysis were used to explore parameters - comorbidity associations. Post-PSM data were split 7:3 into training/testing sets, least absolute shrinkage and selection operator (LASSO) regression selected features, and 11 machine learning algorithms developed Clinlabomics models. Additionally, the optimal model was validated in the testing set and an independent validation set.

After PSM, multivariate logistic regression identified AIP as the strongest risk factor (OR = 2.74, 95%CI: 1.80-4.19). The Q4 of TyG, TG, AIP, and LCI elevated comorbidity risk (P < 0.05). Besides, RCS showed nonlinear association of LCI with comorbidity (P < 0.05). The Recursive Partitioning and Regression Trees (rpart)-based Clinlabomics model exhibited favorable performance with F1-score, accuracy (ACC), and area under the curve (AUC) of 0.960, 0.960, and 0.986. At optimal hyperparameter (cp=0.0017), the model achieved AUCs of 0.987 (95%CI: 0.982-0.993), 0.955 (95%CI: 0.939-0.972), and 0.957 (95%CI: 0.915-0.999) in the training, testing, and validation datasets, respectively, correctly identifying 87.7% non-comorbidity and 98.0% comorbidity patients in validation. SHapley Additive exPlanations (SHAP) analysis identified UA_admission, UA_3d, TyG, TG, AIP and LCI as key metabolic indicators.

TyG, TG, AIP, and LCI were critical metabolic parameters for HUA-IS comorbidity, which warrant heightened attention in future comorbidity research.

Participants from Suining Central Hospital were categorized into ischemic stroke (IS) patients and healthy controls (HC) based on CT/MRI imaging findings. The diagnostic criteria for hyperuricemia (HUA) were defined as a serum uric acid (SUA) level of ≥ 6.0 mg/dL in females and ≥ 7.0 mg/dL in males. Subsequently, all participants were stratified into four groups: IS+HUA (comorbid group), IS+non-HUA, HC+HUA, and HC+non-HUA, with the latter three groups merged into a non-comorbid group. The metabolic parameters and their associations with the comorbidity were analyzed. Following feature selection via LASSO regression, 11 Clinlabomics models were constructed. The optimal model was identified, and the "black-box" characteristics of this optimal model were interpreted using SHapley Additive exPlanations (SHAP).