Dementia Risk Scores Show Uneven Predictive Power in Landmark Meta-Analysis

A new systematic review and meta-analysis published in GeroScience has delivered the most comprehensive head-to-head assessment yet of the tools clinicians use to forecast dementia, and the verdict is a nuanced one. The international research team, led by Tímea Lázár and András Attila Horváth, examined ten established dementia risk assessment scores across observational cohort studies and found that their ability to predict who will actually develop the condition varies dramatically depending on which score is used, how old the person is when assessed, and how far into the future the prediction extends. The findings arrive at a critical moment: the World Health Organization estimates that 57 million people currently live with dementia worldwide, a figure projected to soar to 139 million by 2050, placing mounting pressure on health systems that are already struggling to cope.

The stakes behind this statistical exercise are considerable. Roughly 45 percent of dementia cases are estimated to be linked to potentially modifiable risk factors such as hypertension, low education, obesity and physical inactivity, although modifying these factors does not guarantee prevention. Risk scores promise a simple, inexpensive and rapid way to flag high-risk individuals in routine clinical practice, particularly in primary care, so that risk-reduction strategies can begin years before any symptom appears. Yet until now, there has been a striking lack of comparative evidence on how well the various existing scales actually perform, leaving clinicians without a clear guide on which tool to trust for which patient.

To fill that gap, the researchers conducted their review in strict adherence to the Cochrane Handbook and reported it under the PRISMA 2020 guideline, registering the protocol on PROSPERO in December 2023 without deviating from it. They searched MEDLINE, Embase and the Cochrane Library, identifying 9,383 records, from which 29 papers ultimately met the inclusion criteria. The eligible studies enrolled adults free from dementia or cognitive impairment at baseline, with or without other health conditions such as diabetes or hypertension, and followed them for incident dementia diagnosed through clinical evaluation, standardised criteria, neuropsychological testing, neuroimaging and biomarker analysis. Thirteen of the 29 articles contained data comparable enough for quantitative pooling, and those covered just three of the ten prespecified scores: CAIDE, LIBRA and ANU-ADRI.

The statistical machinery behind the analysis was correspondingly rigorous. The team used random-effects meta-analyses with restricted maximum likelihood estimation, and because several studies reported multiple area-under-the-curve results while some reused the same cohorts, they applied a hierarchical random-effects structure with robust correction methods to account for correlated estimates. Follow-up durations were categorised as short (under five years), medium (five to ten years) and long (more than ten years), while study populations were split into middle-aged (baseline age under 65) and older (65 and above) groups. The headline metric was the area under the receiver operating characteristic curve, or AUC, a value where 0.5 indicates prediction no better than chance and 1.0 indicates perfect discrimination.

The standout performer was the Cardiovascular Risk Factors, Ageing and Dementia score, or CAIDE, which was originally designed to identify middle-aged individuals at risk over long horizons. In the long follow-up, middle-aged subgroup, CAIDE achieved a pooled AUC of 0.74 (95 percent CI 0.73 to 0.76), and when the APO-E genetic allele was added to the score, the pooled AUC in that same subgroup rose slightly to 0.76 (95 percent CI 0.74 to 0.77). Combining both versions of the score, the long follow-up, middle-aged subgroup yielded a pooled AUC of 0.75 (95 percent CI 0.73 to 0.76). By contrast, performance collapsed in other settings: in medium follow-up, older populations the pooled AUC was just 0.55, and in short follow-up, middle-aged groups it was 0.52, barely distinguishable from a coin flip.

Notably, adding the APO-E allele did not meaningfully improve the model’s performance overall. In the medium follow-up, middle-aged subgroup, the version with APO-E reached a pooled AUC of 0.70 (95 percent CI 0.69 to 0.71), but the authors caution that all three estimates feeding into that figure came from the UK Biobank, limiting generalisability. The researchers suggest that whether integrating genetic information into routine clinical practice is worthwhile should be carefully considered, given the added complexity and cost it introduces without a clear gain in discrimination.

The ANU-ADRI score, developed by the Australian National University, showed a steadier but more moderate profile. Six studies contributed to its meta-analysis, with a pooled AUC of 0.71 (95 percent CI 0.65 to 0.77) at short follow-up and 0.65 (95 percent CI 0.61 to 0.69) at medium follow-up. Unlike CAIDE, its performance did not differ clearly across baseline age groups, suggesting it may also be useful for risk prediction in older populations. The LIBRA score told a more troubled story: the original version showed the lowest discrimination of the three, with individual AUCs ranging from 0.51 to 0.60 across short and medium follow-up, while a modified version performed better, reaching AUCs of 0.65 to 0.67 at short follow-up and a pooled AUC of 0.70 (95 percent CI 0.66 to 0.74) at medium follow-up.

The qualitative review surrounding the meta-analysis added important context on where these tools may and may not travel. In low- and middle-income Latin American and Asian populations comprising 11,143 participants, CAIDE showed weak discrimination (c-statistics of 0.52 to 0.63), as did the AgeCoDe score (0.57 to 0.74), whereas ANU-ADRI, the Brief Dementia Screening Indicator and the Rotterdam Study Basic Dementia Risk Model generally showed higher, though still variable, discrimination (0.62 to 0.78). The authors read these findings as a warning that dementia risk models developed in high-income countries should be applied cautiously elsewhere. Separately, a creative study used a deep learning algorithm applied to fundus photographs to estimate CAIDE scores in two large Chinese datasets, achieving AUCs of 0.944 and 0.926, though those values refer to predicting a high CAIDE score rather than incident dementia itself and are not directly comparable.

Other findings from the systematic review underscored the importance of timing and context. Evidence suggests that optimal cardiovascular and behavioural health in midlife is associated with lower late-life dementia risk, while metrics assessed in late life alone show no consistent association, highlighting that when risk is measured matters as much as what is measured. LIBRA research showed that lower wealth and education were associated with higher scores and increased dementia risk, and that the score’s utility appears limited in very old populations above age 85. Cardiovascular risk scores such as ASCVDRS and SCORE2-OP were also associated with dementia and cognitive decline risk in older adults, hinting that tools already in cardiology clinics could double as cognitive risk flags.

The authors are careful about what prediction does and does not imply. A score’s ability to stratify risk does not necessarily mean that modifying individual risk factors will reduce dementia incidence: randomised studies suggest blood pressure lowering may reduce dementia incidence, while multidomain cardiovascular interventions in older adults have not consistently shown a significant effect on dementia incidence, though they have benefited cognitive function. The team also acknowledges limitations, including high between-study heterogeneity in some analyses, insufficient data to pool all subgroups and scores, and the inability to meta-analyse sensitivity, specificity or effect measures such as hazard ratios. Egger’s test for publication bias could not be conducted given the small number of studies per analysis. Their bottom line is measured: dementia risk scores demonstrate moderate predictive ability and may be clinically useful for stratification, particularly in primary care, but the choice of score should match the patient’s age and the intended prediction horizon, and further prospective validation across different populations, ethnic and socioeconomic groups, and dementia subtypes is needed before widespread clinical use.

Subject of Research: Predictive accuracy of dementia risk assessment scores in cohort studies

Article Title: How well do dementia risk scores predict dementia? A systematic review and meta-analysis

Article References: Lázár, T., Radics, E., Engh, M. A., Turkel, C., Tóth, R., Kói, T., Hegyi, P., Kéri, S., & Horváth, A. A. (2026). How well do dementia risk scores predict dementia? A systematic review and meta-analysis. GeroScience. https://doi.org/10.1007/s11357-026-02615-y

Image Credits: AI Generated

DOI: 10.1007/s11357-026-02615-y

Keywords: dementia, risk scores, CAIDE, LIBRA, ANU-ADRI, meta-analysis, systematic review, predictive accuracy, AUC, primary care, risk factors, GeroScience