Multiple patterns of structural brain change trajectories in AD: MRI deep learning verification with enhanced sequential pattern analysis

Abstract

AbstractBackgroundOur recently proposed interpretable Ensemble 3DCNN deep learning (DL) technique has allowed the derivation of a P‐score based on brain structural magnetic resonance imaging (sMRI) modelling to predict patterned neurodegeneration in AD (https://doi.org/10.1002/advs.202204717). Sequential association rules with a support rate greater than 18% have been used to define a neurodegenerative progression pattern. To more prudently test the meaningfulness of the patterns found, we further calculated the lift values of the various sequential patterns underlying the suggested longitudinal sMRI changes.MethodThe lift value of an association rule is defined as the rule’s confidence rate divided by the support rate of the rule consequent, i.e., the support rate of the rule divided by the product of the support values of the rule consequent and the rule antecedent. For example, with {L.NAC}→{L.NAC, R.NAC} (https://doi.org/10.1002/advs.202204717), among the 167 AD subjects employed for analyzing neurodegenerative progression patterns, the number (N) of the subjects with at least an sMRI image containing both L.NAC and R.NAC as neurodegenerative regions at all time points examined was firstly calculated. Secondly, among the above‐mentioned 167 AD subjects, the number (M) of the subjects with at least an sMRI image containing only L.NAC as neurodegenerative regions at all time points examined was counted and the time point (P) corresponding to the selected (earlier/earliest) sMRI image was saved. Thirdly, among the above selected M subjects, the number (K) of the subjects with at least an sMRI image containing both L.NAC and R.NAC as neurodegenerative regions at all time points latter than the P was computed. Thus, lift value = (K/M)/(N/167).ResultThe lift values of the fifty sequential association rules reported in the recent paper (https://doi.org/10.1002/advs.202204717) ranged between 1.27 and 2.59(i.e., all > 1; Table 1), further confirming the longitudinal sMRI change patterns in AD.ConclusionThis study analyzing the lift values of the sequential association rules further verified the multiple patterns of longitudinal sMRI trajectories in AD brains, as reported recently. The research demonstrated featured and heterogeneous whole‐brain structural degeneration with AD progression.