Evaluating the value and limits of null-models for Lesion Network Mapping

Lesion Network Mapping (LNM) is a technique employed to map symptom-related circuits within the human brain. However, recent studies have raised doubts regarding the method and the distinctiveness of LNM-derived networks, as they frequently exhibit high spatial overlap among unrelated conditions. To address these limitations, null models have been introduced as statistical tools in network neuroscience.

These null models serve as reference procedures in LNM, using the normative connectome matrix C as the foundation for derived networks, even when employing null modelling techniques. It is essential to scrutinize the proposed null models closely and evaluate their practical application across various datasets.

Through a step-by-step examination and empirical testing of the null models, we have discovered that the proposed null-model correction effectively only replaces C with a transformed matrix C (e.g., C ~ C - mean(C)). This new matrix is derived directly from C and maintains a significant portion of its low-dimensional structure. As a result, LNM maps that have undergone null-model correction continue to be influenced by the same set of connectome-derived patterns.

Consequently, these maps still display repetition and lack specificity when comparing unrelated conditions.

In conclusion, null models can offer a statistical benchmark for LNM, but their application does not resolve the fundamental limitation of non-specificity in LNM networks.