Related Experiment Video
Updated: Oct 20, 2025
09:47
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
1.4K
Data analysis with Shapley values for automatic subject selection in Alzheimer's disease data sets using
Louise Bloch1,2, Christoph M Friedrich3,4,
1Department of Computer Science, University of Applied Sciences and Arts Dortmund, Dortmund, 44227, Germany.
Alzheimer'S Research & Therapy
|September 16, 2021
Summary
Identifying informative subjects using Data Shapley values improved machine learning models for predicting Alzheimer's disease (AD). This method enhances early AD detection by focusing on key data points, aiding in subject recruitment for therapy studies.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Medical Imaging
Background:
- Predicting Alzheimer's disease (AD) progression in mild cognitive impairment (MCI) is crucial for therapy studies.
- Machine learning (ML) offers potential for early AD prediction, but data variability poses challenges like overfitting.
- Heterogeneous AD etiology and data acquisition variations necessitate robust methods for signal identification.
Purpose of the Study:
- To evaluate an automatic data valuation method based on Shapley values for identifying informative subjects.
- To enhance ML classification accuracy for predicting AD in MCI subjects.
- To improve the efficiency of subject recruitment and monitoring in AD clinical trials.
Main Methods:
- Developed an ML workflow using volumetric MRI, feature selection, and sample selection via Data Shapley.
- Trained and validated models on Alzheimer's Disease Neuroimaging Initiative (ADNI) and Australian Imaging, Biomarker and Lifestyle Flagship Study of Ageing (AIBL) cohorts.
- Employed Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) for classification, with Kernel SHAP for interpretation.
Main Results:
- RF models excluding subjects with low Data Shapley values improved accuracy by 5.76% on the ADNI test set (mean accuracy 62.64% to 68.40%).
- XGBoost models showed improved classification accuracy by 2.98% on the AIBL dataset after excluding subjects with the lowest RF Data Shapley values.
- Data Shapley effectively identified informative subjects, enhancing model performance on independent test sets.
Conclusions:
- The Data Shapley method successfully improved mean accuracies for AD prediction test sets.
- Informative subjects identified by Data Shapley were linked to ApolipoproteinE ε4 (ApoE ε4) alleles, cognitive test results, and MRI measurements.
- This approach aids in optimizing ML models for early AD detection and clinical trial subject selection.
