Abstract
BACKGROUND: The recent approval of monoclonal antibodies for the treatment of Alzheimer's disease (AD) in several countries has accelerated the need for affordable, simple and scalable methods to identify patients who are eligible for treatment with the new disease-modifying therapies (DMT). Blood-based biomarkers offer less invasive alternatives to established gold standards. We have clinically validated a predictive model combining plasma Aβ42/Aβ40, apolipoprotein E (APOE) genotype and age, in two independent real-world cohorts to identify brain amyloid deposition. METHODS: We conducted a clinical validation study involving 450 patients with mild cognitive impairment (MCI) from two real-world cohorts (HCSC, Madrid, Spain and HUSM, Lleida, Spain). Plasma Aβ42/Aβ40 was measured by ABtest-MS, an antibody-free liquid chromatography-mass spectrometry method. CSF Aβ42/Aβ40 and p-tau181/Aβ42 (gold standards) were quantified with the Lumipulse(®) platform. The model was trained in the HCSC cohort and validated in the HUSM cohort. Finally, an overall analysis in the combined population was performed. A dual cutoff approach was used to classify the patients as positive or negative. Statistical analysis included bootstrap resampling and model calibration. RESULTS: In the HCSC, HUSM, external validation and combined analysis, AUCs were 0.89 (95% confidence intervals-CI: 0.84-0.93), 0.88 (0.84-0.93), 0.88 (0.83-0.92) and 0.88 (0.84-0.91), with corresponding accuracies of 82.3, 81.6, 82.3, and 81.1%, respectively. After the combined analysis, positive and negative predictive values (PPV and NPV) were established at 87.5%, resulting in cutoff values of 0.30 and 0.67 for the likelihood of amyloid negativity and positivity, respectively, for a prevalence of 62%. Probability values lower than 0.30 indicate low probability of brain amyloid deposition, while values greater than 0.67 indicate high probability. Less than 28% of the participants fell into the intermediate zone. Additional cutoffs were derived for different prevalence values. Predictive model calibration showed excellent agreement with observed data, confirming accurate predictions (slope = 0.98, intercept = -0.01). CONCLUSION: This predictive model has demonstrated high accuracy for the identification of brain amyloid deposition in patients with MCI. Derived cutoffs enabled over 70% reduction in invasive testing, supporting efficient and cost-effective identification of candidates for DMTs.