Scholarlink

Normative modeling is a valuable tool for analyzing brain morphometry in patients. We used the isolation forest algorithm, trained on 93 Cognitively Normal (CN) subjects, to detect subcortical volume deviations from the norm in Frontotemporal (FTD) dementia subtypes (122 subjects). The SHapley Additive exPlanations (SHAP) method was used to interpret the models’ output. In the test set, our left and right subcortical models correctly identified 90.9% of the CN subjects as Control-like (20 out of 22), and 73.3% (left)/65.5% (right) of the FTD patients as Anomalous: 31-left, 34-right out of 52 behavioral variant FTD (bvFTD); 22-left, 16-right of out 33 progressive non-fluent aphasia (PNFA); 35-left, 28-right out of 35 semantic-variant (SV). FTD patients classified as Control-like showed lower subcortical atrophy than Anomalous. SHAP analysis highlighted distinct features relevant for each subtype, with the inferior lateral ventricle being the most significant. These findings suggest the models’ potential to detect different stages of FTD subtypes.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable Normative Modeling: Subcortical Changes in Frontotemporal Dementia Subtypes

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Helena Rico Pereira,
  • José Manuel Fonseca,
  • Hugo Alexandre Ferreira

摘要

Normative modeling is a valuable tool for analyzing brain morphometry in patients. We used the isolation forest algorithm, trained on 93 Cognitively Normal (CN) subjects, to detect subcortical volume deviations from the norm in Frontotemporal (FTD) dementia subtypes (122 subjects). The SHapley Additive exPlanations (SHAP) method was used to interpret the models’ output. In the test set, our left and right subcortical models correctly identified 90.9% of the CN subjects as Control-like (20 out of 22), and 73.3% (left)/65.5% (right) of the FTD patients as Anomalous: 31-left, 34-right out of 52 behavioral variant FTD (bvFTD); 22-left, 16-right of out 33 progressive non-fluent aphasia (PNFA); 35-left, 28-right out of 35 semantic-variant (SV). FTD patients classified as Control-like showed lower subcortical atrophy than Anomalous. SHAP analysis highlighted distinct features relevant for each subtype, with the inferior lateral ventricle being the most significant. These findings suggest the models’ potential to detect different stages of FTD subtypes.