Connectome Principal Component Drives Cross-Dataset Replication and Clinical Prediction in Symptom Lesion Network Mapping

Abstract

Lesion network mapping (LNM) describes a group of methods using normative functional connectivity data to map disparate brain lesions and stimulation sites onto common brain networks. Van den Heuvel and colleagues recently showed that these methods lack disease specificity, instead producing maps that converge toward intrinsic properties of the normative connectome dataset. Here, we investigate symptom LNM (sLNM), a recent advancement in the method which attempts to increase the robustness of results by incorporating symptom severity and incorporating replication across multiple datasets and prediction of clinical outcomes. Using clinical datasets of depression and Brocas aphasia, we show that sLNM maps from unrelated disorders nonetheless converge despite using null models which break the specific lesion-symptom structure in the datasets. Using simulated datasets with a known ground-truth disease network, we show that sLNM results are systematically biased towards the normative connectomes first principal component (PC1), which drives spurious convergence across unrelated datasets. We further show that the apparent clinical predictive capability of these maps are non-specific: network maps derived from unrelated disorders such as migraine and aphasia predict brain stimulation improvement in depression as well as -- or better than -- the cohorts own sLNM map. However, controlling for PC1 reduces spurious convergence across unrelated datasets and improves clinical prediction specificity, supporting the notion that disease-specific signal exists within sLNM but is confounded by the globally present PC1 signal in the normative connectome. These findings offer a practical correction applicable to existing and future sLNM studies.