Background: We sought to define a cutoff for β-amyloid 1-42 in cerebrospinal fluid (CSF), a key marker for Alzheimer’s disease (AD), with data-driven Gaussian mixture modeling in a memory clinic population. Methods: We performed a combined cross-sectional and prospective cohort study. We selected 2462 subjects with subjective cognitive decline, mild cognitive impairment, AD-type dementia, and dementia other than AD from the Amsterdam Dementia Cohort. We defined CSF β-amyloid 1-42 cutoffs by data-driven Gaussian mixture modeling in the total population and in subgroups based on clinical diagnosis, age, and apolipoprotein E (APOE) genotype. We investigated whether abnormal β-amyloid 1-42 as defined by the data-driven cutoff could better predict progression to AD-type dementia than abnormal β-amyloid 1-42 defined by a clinical diagnosis-based cutoff using Cox proportional hazards regression. Results: In the total group of patients, we found a cutoff for abnormal CSF β-amyloid 1-42 of 680 pg/ml (95% CI 660-705 pg/ml). Similar cutoffs were found within diagnostic and APOE genotype subgroups. The cutoff was higher in elderly subjects than in younger subjects. The data-driven cutoff was higher than our clinical diagnosis-based cutoff and had a better predictive accuracy for progression to AD-type dementia in nondemented subjects (HR 7.6 versus 5.2, p < 0.01). Conclusions: Mixture modeling is a robust method to determine cutoffs for CSF β-amyloid 1-42. It might better capture biological changes that are related to AD than cutoffs based on clinical diagnosis.