Prediction of amyloid pathology in cognitively unimpaired individuals using structural MRI

Background: Structural MRI measurements can contribute to the prediction of amyloid pathology in cognitively unimpaired (CU) individuals. In this work, we aimed at studying the predictive capacity, robustness, and generalizability of ML techniques to predict amyloid-ß pathology in CU individuals, as...

Descripción completa

Detalles Bibliográficos
Autores: Cumplido Mayoral, Irene, Ingala, Silvia, Lorenzini, Luigi, Wink, Alle Meije, Haller, Sven, Molinuevo Guix, Jose Luis, Wolz, Robin, Palombit, Alessandro, Schwarz, Adam J., Vilaplana Besler, Verónica|||0000-0001-6924-9961
Tipo de recurso: artículo
Fecha de publicación:2021
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/363798
Acceso en línea:https://hdl.handle.net/2117/363798
https://dx.doi.org/10.1002/alz.053661
Access Level:acceso abierto
Palabra clave:Alzheimer's disease
Brain -- Magnetic resonance imaging
Machine learning
Alzheimer, Malaltia d'
Cervell -- Imatgeria per ressonància magnètica
Aprenentatge automàtic
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo
Àrees temàtiques de la UPC::Informàtica::Aplicacions de la informàtica::Bioinformàtica
Descripción
Sumario:Background: Structural MRI measurements can contribute to the prediction of amyloid pathology in cognitively unimpaired (CU) individuals. In this work, we aimed at studying the predictive capacity, robustness, and generalizability of ML techniques to predict amyloid-ß pathology in CU individuals, as well as identifying key brain regions contributing to this prediction. Method: We included 653 and 250 CU individuals from the EPAD and ADNI studies, respectively, with available T1w MRI and CSF Aß42 measurements. Subjects were categorized as amyloid-ß positive (Aß+) or negative (Aß-) according to established cut-offs (CSF Aß42<1000pg/mL for EPAD and <880pg/mL for ADNI). Volumes and cortical thickness in regions of the Desikan-Kiliany atlas were obtained with Freesurfer 6.0, as well as the Total Intracranial Volume (TIV). We trained XGBoost classifiers to predict amyloid-ß positivity using age, sex, education, MMSE, APOE-¿4 status, volumes/TIV, and cortical thickness measurements. To study the generalizability of the classifier, we performed within-cohorts classification (train and test within the same cohort); and between-cohorts classification (train with one cohort and test with another). We performed the latter classification, both with the original samples, and following a bootstrapping approach to force balanced data in the training. With the classification results, we conducted a ROC analysis and, additionally, calculated SHAP values to determine the most important brain regions used to predict Aß+. Result: Similar classification performance was achieved when training/testing within ADNI (ROC-AUC=0.72 (0.59-0.84)) and within EPAD (ROC AUC=0. 68 (0.60-0.76)) (Figure 1A). In the between cohort’s analysis, ROC-AUC decreased slightly in both cases (Figure 1B) but was consistent over subsampling of individuals. Train with ADNI and test with EPAD gave ROC-AUC=0.61 (0.58-0.64) and train with EPAD and test with ADNI gave ROC-AUC=0.59 (0.54-0.66). The most important features used for prediction were similar for both ADNI and EPAD (Figure 2). The increased volume and/or thickness of some brain regions, such as superior frontal areas, predicted Aß+. The decrease in cingulate cortex was also seen. APOE-¿4 status also had a high impact on classification performance. Conclusion: The use of structural MRI measurements in combination with Machine Learning can provide additional capacity to predict amyloid-ß status in CU individuals.