Risk stratification models in elderly patients: recalibrating or remodeling?

Area of Science:

• Cardiovascular Surgery
• Medical Statistics
• Geriatric Medicine

Background:

• Current risk stratification models for cardiac surgery demonstrate suboptimal performance in elderly patient populations.
• Accurate risk assessment is crucial for optimizing surgical decisions and patient outcomes in older adults undergoing cardiac procedures.

Purpose of the Study:

• To evaluate the performance of the Italian Coronary Artery Bypass Graft Outcome Project (ItCABG) and logistic European System for Cardiac Operative Risk Evaluation (LogEuroSCORE) models in predicting 30-day mortality across different age groups.
• To develop and assess a novel risk prediction model specifically tailored for elderly patients (≥80 years).

Main Methods:

• Analysis of a large cohort (n = 34,310) from the ItCABG and Mattone Outcome-BYPASS (MO-BYPASS) studies, stratified by age (<60, 60-69, 70-79, ≥80 years).
• Assessment of model accuracy (Brier score, pseudo-R2), calibration (Hosmer-Lemeshow test), and discrimination (AUC) for ItCABG and LogEuroSCORE.
• Development of a logistic regression model using data from 2,255 octogenarian patients to create an elderly-specific risk function.

Main Results:

• Both ItCABG and LogEuroSCORE models showed good discrimination in younger age groups but significantly poorer performance in patients aged ≥80 years (AUCs 0.64-0.65).
• Model calibration was poor for octogenarians, with significant p-values for the Hosmer-Lemeshow test (ItCABG: P=0.05; LogEuroSCORE: P<0.001).
• The newly developed elderly-specific model demonstrated improved calibration (P=0.86) but retained poor discrimination (AUC=0.66) in octogenarian patients.

Conclusions:

• The ItCABG and LogEuroSCORE models are inadequate for predicting mortality in octogenarian cardiac surgery patients.
• Improved risk stratification for patients aged 80 and older necessitates the identification and incorporation of elderly-specific risk factors into predictive models.