Abstract
OBJECTIVE: To evaluate and compare the predictive value of eight dementia risk scores for late-life cognitive function and cognitive decline; ANU-ADRI, CAIDE, CogDrisk, LIBRA, LIBRA2, UKBDRS(-APOE), and a Lancet commission-based risk score. METHODS: Using Norwegian Trøndelag Health Study (HUNT) data, we calculated risk scores from lifestyle and health data of 7221 dementia-free participants (mean age: 76.8 years, 54.1% female) collected in HUNT3 (2006-2008). Cognitive function was assessed using the Montreal Cognitive Assessment scale (MoCA) 11 years later in HUNT4 70+, and reassessed in 4716 participants 4 years thereafter. Associations between continuous risk scores or risk score tertiles, cognition and cognitive decline were examined using linear mixed-effects models. Logistic regression models were used to test associations between risk scores and a ≥ 3-point decline in MoCA scores. RESULTS: All risk scores were significantly associated with cognitive function and cognitive decline. Associations with cognitive function ranged from UKBDRS β (per 1SD)=-1.61(95%CI:-1.72,-1.51) to CAIDE (β=-0.74;95%CI:-0.82,-0.67), and with yearly cognitive decline from Lancet (β=-0.23;95%CI:-0.27,-0.18) to CAIDE (β=-0.04;95%CI:-0.07,-0.02). High-low risk group differences in cognitive function were largest for CogDrisk (β=-3.04;95%CI:-3.27,-2.81), LIBRA (β=-3.04;95%CI:-3.27,-2.80) and lowest for CAIDE (β=-1.65;95%CI:-1.86,-1.44). High-risk groups showed the steepest decline for UKBDRS-APOE (β=-0.43;95%CI:-0.52,-0.34), Lancet (β=-0.39;95%CI:-0.48,-0.30), and LIBRA (β=-0.38;95%CI:-0.47,-0.28). All scores predicted ≥3-point decline modestly: AUCs were highest for UKBDRS (AUC=0.61;95%CI:0.60,0.63), UKBDRS-APOE (0.61;95%CI:0.60,0.63), CogDrisk (0.60;95%CI:0.58,0.62), and Lancet (0.60;95%CI:0.58,0.61), but none outperformed a model including age and education alone (0.61;95%CI:0.60,0.63). CONCLUSION: Risk scores captured meaningful gradients in cognition and decline but offered limited discriminatory accuracy beyond demographics, supporting their use for prevention-oriented risk profiling rather than prediction.