Complex biomarker discovery in neuroimaging data: Finding a needle in a haystack

Abstract

Neuropsychiatric disorders such as schizophrenia, bipolar disorder and Alzheimer's disease are major public health problems. However, despite decades of research, we currently have no validated prognostic or diagnostic tests that can be applied at an individual patient level. Many neuropsychiatric diseases are due to a combination of alterations that occur in a human brain rather than the result of localized lesions. While there is hope that newer imaging technologies such as functional and anatomic connectivity MRI or molecular imaging may offer breakthroughs, the single biomarkers that are discovered using these datasets are limited by their inability to capture the heterogeneity and complexity of most multifactorial brain disorders. Recently, complex biomarkers have been explored to address this limitation using neuroimaging data. In this manuscript we consider the nature of complex biomarkers being investigated in the recent literature and present techniques to find such biomarkers that have been developed in related areas of data mining, statistics, machine learning and bioinformatics.

Fig. 1

Illustration of linear biomarker discovery: (a) matrix representation by treating edges in the brain as features, (b) linear regression setup where X represents the features (edges in brain networks or volumetric information) for all subjects, β represents the weights for features, and Y represents the phenotype value for each subject, and (c) resultant β from linear regression and LASSO.

Fig. 2

Illustration of combinatorial biomarker discovery: (a and b) X is a hypothetical data matrix where columns represent features derived from neuroimaging data and rows represent subjects. The subjects belong to two groups Healthy and schizophrenia (SZ) as indicated by the column vector Y. In matrix X, an element (row, column) with black color indicates that the feature is present for a given subject. A, B, C, and D are interesting submatrices in X that have information about Y. The columns representing these submatrices in (a) are individually associated with Y, but those in (b) are not associated. (c) Efficient search space pruning: The Apriori principles allows pruning of supersets when a set is not interesting.

Fig. 3

Illustration of a ‘pathway’ based biomarker discovery approach. The features (often edges in the brain networks) are evaluated individually and then the functional groups (resting state networks) are evaluated for enrichment with highly significant features (edges).

Fig. 4

Illustration of a subgraph discriminating between three healthy subjects and three disease subjects. The figure shows 6 networks from 3 healthy and 3 disease subjects. The shaded region in these networks covers nodes that are densely connected in healthy subjects and sparsely connected in disease subjects. Discovering such novel sets of nodes or subnetworks is essential.