XGBoost Model Outperforms LASSO in Predicting 3-Year Mortality Risk

A machine learning model using the XGBoost algorithm has demonstrated higher accuracy than the traditional LASSO method in predicting three-year mortality rates for hospitalized patients with cardiovascular-renal-metabolic (CKM) syndrome.

XGBoost performance in mortality prediction

The study compared the discriminative ability of the two models across three distinct cohorts. The XGBoost model consistently outperformed LASSO, recording ROC-AUC values of 0.831 for derivation, 0.826 for internal validation, and 0.813 for cross-center validation. In comparison, the LASSO model registered scores of 0.807, 0.807, and 0.799, respectively. These performance differences were statistically significant.

The research team observed that the XGBoost model showed particularly strong performance in specific patient subgroups, including those aged 60 or younger, women, and patients classified in stage 4 of the syndrome. Both models maintained acceptable calibration and provided net clinical benefits during decision curve analysis.

Double-threshold system classifies patient risk categories

The researchers established a double-threshold system to classify patients into three risk categories: low, moderate, and high. Using probability thresholds of 0.223 and 0.855, the study categorized 45.0% of the sample as low risk, 52.4% as moderate risk, and 2.6% as high risk. The observed mortality incidence rates per 1,000 person-years were 5.9, 62.9, and 254.7 for each respective group.

Based on these findings, the authors suggest that patients identified as low risk could potentially be managed within primary care settings through structured follow-up programs. Conversely, those categorized as high risk may benefit from referral to tertiary hospitals for specialized, multidisciplinary evaluations. An online risk calculator has been developed to assist clinicians, available at https://ckm-mortality-predict.shinyapps.io/CKM-3-year-Mortality-Calculator.

Chinese hospital data requires external validation

The authors identified several constraints regarding the study’s scope and methodology. Because the data was gathered from tertiary hospitals in China, the researchers emphasized the need for external validation using independent databases and cohorts from other international settings. The requirement for complete biomarker measurements may have created a selection bias by favoring patients with more advanced disease stages.

Other noted limitations include the absence of coronary catheterization and tomography data, which may have led to classification errors in disease staging. The study authors also cautioned that the dual-threshold strategy prioritizes precision over sensitivity for the high-risk category, meaning some high-risk patients might be incorrectly classified into lower-risk groups. Consequently, they advised that patients in the moderate-risk category should not be overlooked.

Future development efforts are expected to focus on integrating imaging variables and metabolomic biomarkers from blood or urine samples to enhance the model’s predictive accuracy. This report is based on an observational retrospective study and does not constitute medical advice or a substitute for individual clinical assessment.