QyPredict prognostic model enriches selection for faster decliners in mild cognitive impairment.

Background: The suboptimal selection of patients is a key challenge for disease‐modifying trials in mild cognitive impairment (MCI) and Alzheimer's disease. Improved selection strategies are urgently needed to better power trials. Recent advances in AI predictive modeling, such as the QyPredict® algorithm, are promising tools to improve the selection of patients likely to clinically progress during the clinical trial timeframe. Method: QyPredict® was applied to 519 MCI patients from ADNI: age: (71.8(±7.1), 303 Male, MMSE range 24‐30, CDR = 0.5, with available amyloid and APOE‐ε4 + status. QyPredict®, a tunable machine learning model, incorporated baseline QyScore® volumetric MRI results, demographic (age and sex) and clinical (MMSE and CDR) inputs. A QyPredict® score (0‐1) was produced for each individual, representing the probability they will demonstrate a modelled outcome (here an increase in CDR‐SOB of 0.5+ over 24 months). A score close to zero indicating very low probability of change. Predictive performance was evaluated using positive predictive value (PPV), sensitivity and specificity. Mean and standard deviation for change in CDR‐SOB over 24 months were calculated and compared using Mann‐Whitney test at increasing QyPredict® cut‐offs (0.1, 0.2, 0.3, 0.4 and 0.5). Actual change versus predicted change in CDR‐SB scores was further investigates for 'Stable' (QyPredict®<0.5) versus 'Decliners' (QyPredict®>0.5) (Figure 1). Sample sizes to detect a 30% treatment effect (reduction in change in CDR‐SOB) were calculated for the full sample and a cohort enriched with only 'Decliners'. Result: For the full sample without QyPredict® enrichment, change in CDR‐SOB at 24 months was 2.3 (±2.11), which significantly increase (p<0.001) from a QyPredict®>0.2, reaching a change of 3.4 (±2.30) points for MCI patients with QyPredict®>0.5. PPV, sensitivity and specificity were: 0.76, 0.70 and 0.75 (the full cohort); 0.83, 0.77 and 0.70 (Amyloid positive patients) and 0.82, 0.75 and 0.74 (APOE‐ε4 + patients). Sample sizes for a putative trial where 304 (full sample) compared with 110 (enriched cohort) per treatment arm. Conclusion: Using baseline QyScore® metrics, basic demographic, and typical clinical data available at screening visits, QyPredict® accurately modelled the likelihood of future clinical decline, resulting in substantially reduced patient cohorts to detect treatment effects when enriched with 'decliners'. [ABSTRACT FROM AUTHOR]