Nonlinear biomarker interactions in conversion from Mild Cognitive Impairment to Alzheimer's disease

Abstract

The multi-faceted nature of Alzheimer's disease means that multiple biomarkers (e.g., amyloid-beta, tau, brain atrophy) can contribute to the prediction of clinical outcomes. Machine learning methods are a powerful way to identify the best approach to this prediction. However, it has been difficult previously to model nonlinear interactions between biomarkers in the context of predictive models. This is important as the mechanisms relating these biomarkers to the disease are inter-related and nonlinear interactions occur. Here, we used Gaussian Processes to model nonlinear interactions when combining biomarkers to predict Alzheimer's disease conversion in 48 mild cognitive impairment participants who progressed to Alzheimer's disease and 158 stable (over three years) people with mild cognitive impairment. Measures included: demographics, APOE4 genotype, CSF (amyloid-beta-42, total tau, phosphorylated tau), neuroimaging markers of amyloid-beta deposition ([18F]florbetapir) or neurodegeneration (hippocampal volume, brain-age). We examined: (i) the independent value each biomarker has in predicting conversion; and (ii) whether modelling nonlinear interactions between biomarkers improved prediction performance. Despite relatively high correlations between different biomarkers, our results showed that each measured added complementary information when predicting conversion to Alzheimer's disease. A linear model predicting MCI group (stable versus progressive) explained over half the variance (R2 = 0.51, P < 0.001); the strongest independently-contributing biomarker was hippocampal volume (R2 = 0.13). Next, we compared the sensitivity of different models to progressive MCI: independent biomarker models, additive models (with no interaction terms), nonlinear interaction models. We observed a significant improvement (P < 0.001) for various two-way interaction models, with the best performing model including an interaction between amyloid-beta-PET and P-tau, while accounting for hippocampal volume (sensitivity = 0.77). Our results showed that closely-related biomarkers still contribute uniquely to the prediction of conversion, supporting the continued use of comprehensive biological assessments. A number of interactions between biomarkers were implicated in the prediction of Alzheimer's disease conversion. For example, the interaction between hippocampal atrophy and amyloid-deposition influences progression to Alzheimer's disease over and above their independent contributions. Importantly, nonlinear interaction modelling shows that although for some patients adding additional biomarkers may add little value (i.e., when hippocampal volume is high), but for others (i.e., with low hippocampal volume) further invasive and expensive testing is warranted. Our Gaussian Processes framework enables visual examination of these nonlinear interactions, allowing projection of individual patients into biomarker 'space', providing a way to make personalised healthcare decisions or stratify subsets of patients for recruitment into trials of neuroprotective interventions.