Figure 4
Schematic representation of main features of a prediction using the proposed ‘high-dimensional subsymbolic biomarker’ (from top to bottom). (a) The biomarker creation starts from the analysis of the relevant single markers, gene sequences, symbolized as arbitrarily colored squares (for example, let this be nucleotides and their heterozygous or homozygous presences). In the present analysis, let these squares denote the variants in the opioid receptor genes. (b) Based on the clinical background, the complex markers were grouped into either high-opioid doses demanding patients (orange small squares at the right) or controls (blue small squares at the right). Subgroups within the groups are possible and should be addressed by clustering (not shown). (c) From these composed makers, ABC analysis identifies the relevant submarkers that are most informative for the prediction (noninformative submarkers grayed out). (d) A new patient is analyzed (markers with red margins). Based on distance measuring in the high-dimensional space, this patient will be assigned to the most similar group, that is, to its nearest neighbors in the high-dimensional space based on the chosen distance measure (here, the Jaccard distance was used for genetic markers being either 0, 1 or 2). Concomitantly a learning process of the biomarker starts. In the present example, it proves useful, in a new ABC analysis, to include a further marker that, in a previous ABC analysis, had been found uninformative (third marker from the left, green margins). The process is repeated (dotted arrow) with each prediction where the disease background is known, that is, the biomarker ‘learns’ in the sense of Artificial Intelligence, where valid information about a patient’s background is used to improve the marker. This will improve its predictive accuracy in cases where the background is not known.
