Aims

To evaluate the prognostic value of QyPredict(R), a tunable machine learning model for clinical trial enrichment, in amyloid-positive (A+) and all-comer mild cognitive impairment (MCI) populations.

Methods

QyPredict® incorporates multiple inputs (QyScore® vMRI results, demographic, clinical, genetic and biological markers) and generates a score (0-1) representing the probability of an individual demonstrating a specific outcome (here, CDR-SOB increase of 0.5+ over 12 months). We used ADNI subjects with ages (55-85), MMSE (24-30), CDR=0.5 and available amyloid status.

We evaluated performance using probability cut-off scores of 0.1, 0.2, 0.3 and 0.4. For both all-comer and A+ populations we calculated the mean, standard deviation and Cohen’s d for change in CDR-SOB values over 12, 24 and 36 months, and Pearson’s correlation between the probability score and observed change in CDR-SOB.

Results

For the full sample without QyPredict(R) enrichment, change in CDR-SOB was 0.33 (±1.02) with d=0.32. Change in CDR-SOB increased with increasing cut-off values (0.46(±1.04) d=0.44 for cut-off 0.1, to 0.73 (±1.01) d=0.72 for cut-off 0.4).

For the A+ population without QyPredict(R) enrichment, change in CDR-SOB was 2.06 (±2.89) with d=0.71. Increasing cut-off values again led to increased CDR-SOB change (2.26 (±2.27) d=0.76 for cut-off 0.1, to 3.23 (±3.32) d=0.97 for cut-off 0.4).

Correlation values ranged from r=0.39-0.5. Higher cut-off values, longer follow-up time, and combining QyPredict(R) enrichment with amyloid positivity were all associated with higher correlations.

Conclusions

Subsets of MCI populations selected using QyPredict® exhibit faster cognitive decline, whether or not amyloid positivity is also enforced.