Explainable machine learning radiomics model for Primary Progressive Aphasia classification

Introduction: Primary Progressive Aphasia (PPA) is a neurodegenerative disease characterized by linguistic impairment. The two main clinical subtypes are semantic (svPPA) and non-fluent/agrammatic (nfvPPA) variants. Diagnosing and classifying PPA patients represents a complex challenge that requires the integration of multimodal information, including clinical, biological, and radiological features. Structural neuroimaging can play a crucial role in aiding the differential diagnosis of PPA and constructing diagnostic support systems. Methods: In this study, we conducted a white matter texture analysis on T1-weighted images, including 56 patients with PPA (31 svPPA and 25 nfvPPA), and 53 age- and sex-matched controls. We trained a tree-based algorithm over combined clinical/radiomics measures and used Shapley Additive Explanations (SHAP) model to extract the greater impactful measures in distinguishing svPPA and nfvPPA patients from controls and each other. Results: Radiomics-integrated classification models demonstrated an accuracy of 95% in distinguishing svPPA patients from controls and of 93.7% in distinguishing svPPA from nfvPPA. An accuracy of 93.7% was observed in differentiating nfvPPA patients from controls. Moreover, Shapley values showed the strong involvement of the white matter near left entorhinal cortex in patients classification models. Discussion: Our study provides new evidence for the usefulness of radiomics features in classifying patients with svPPA and nfvPPA, demonstrating the effectiveness of an explainable machine learning approach in extracting the most impactful features for assessing PPA.