Source-linked AI summary
Deep ensemble learning for Alzheimers disease classification
Ning An, Huitong Ding, Jiaoyun Yang, Rhoda Au, Ting Fang Alvin Ang
TL;DR
AD classification requires better use of limited, multisource clinical information and expert opinions, while existing methods face classifier dependence and cost-sensitive errors. DELearning uses deep learning across voting, stacking, and optimizing layers to integrate diverse classifiers and physician-like predictions. On NACC clinical measures, it outperformed six representative ensemble methods in AD prediction accuracy and was presented as a data-driven aid for primary care where expertise is limited.
Problem
AD prediction methods need to better integrate complementary multisource data and reduce dependence among classifiers while supporting diagnosis where specialist access is limited.
Method
DELearning uses sparse autoencoders for feature fusion and diversification, a deep-belief-network stacking layer for weighted classifier integration, and neural networks with cost-sensitive optimization.
Results
DELearning outperformed six representative ensemble learning methods in AD prediction accuracy on clinical measures from the NACC UDC.
Takeaways & Limitations
The framework provides a data-driven solution to aid AD primary care, particularly where access to AD expertise is limited.
Takeaways & Limitations
The optimizing layer assumes ordered outcome costs and uses over-sampling because under-sampling can discard useful training samples and degrade classifier performance.
Abstract
from arXiv · showhide
Ensemble learning use multiple algorithms to obtain better predictive performance than any single one of its constituent algorithms could. With growing popularity of deep learning, researchers have started to ensemble them for various purposes. Few if any, however, has used the deep learning approach as a means to ensemble algorithms. This paper presents a deep ensemble learning framework which aims to harness deep learning algorithms to integrate multisource data and tap the wisdom of experts. At the voting layer, a sparse autoencoder is trained for feature learning to reduce the correlation of attributes and diversify the base classifiers ultimately. At the stacking layer, a nonlinear feature-weighted method based on deep belief networks is proposed to rank the base classifiers which may violate the conditional independence. Neural network is used as meta classifier. At the optimizing layer, under-sampling and threshold-moving are used to cope with cost-sensitive problem. Optimized predictions are obtained based on ensemble of probabilistic predictions by similarity calculation. The proposed deep ensemble learning framework is used for Alzheimers disease classification. Experiments with the clinical dataset from national Alzheimers coordinating center demonstrate that the classification accuracy of our proposed framework is 4% better than 6 well-known ensemble approaches as well as the standard stacking algorithm. Adequate coverage of more accurate diagnostic services can be provided by utilizing the wisdom of averaged physicians. This paper points out a new way to boost the primary care of Alzheimers disease from the view of machine learning.
1 INTRODUCTION
Alzheimer’s disease classification faces limited specialist access, biased clinical judgment, and incomplete use of complementary data sources. The paper proposes DELearning to diversify and integrate classifiers, reduce their correlation, and leverage averaged physician-like expertise for primary care.
- AD accounts for more than 60% of dementia cases, while diagnosis services in many middle-income countries remain concentrated in tertiary hospitals.
- Existing AD prediction methods may lack the generalization capability needed for medical practitioners because of insufficient data and physicians’ biased clinical judgment.
- Most classification methods use one biomarker modality, although clinical, neuropsychological, family-history, neuroimaging, biological, and genetic measures can provide complementary information.
- DELearing uses sparse autoencoders to fuse multisource data and reduce feature correlation, then builds diverse classifiers from different algorithms and feature spaces.
- A deep-belief-network stacking layer addresses dependence among base classifiers, while neural networks and cost-sensitive methods support diagnosis under unequal misclassification consequences.
- The framework treats classifiers as physician surrogates whose predictions can be integrated to reduce reliance on a single potentially wrong classifier or physician.
2 METHODS
The method frames AD classification as combining predictions from multiple physicians or base classifiers over AD and non-demented-control outcomes. DELearning uses three ensemble layers to combine these sources.
- The task distinguishes probable or possible AD from non-demented control using independently identically distributed participant samples.
- Each of d physicians or base classifiers produces predictions for every participant, allowing the framework to combine weaker or less expert information sources.
- DELearning is organized as three ensemble layers that process classifier predictions to produce the final AD classification.
Voting layer
The voting layer fuses multisource clinical data through sparse autoencoder feature learning, then trains diverse base classifiers on transformed feature spaces to improve generalization.
- Base classifiers: Different learning algorithms and feature subsets generate multiple diverse diagnosis hypotheses.The voting layer includes classifiers such as logistic regression, tree methods, Bayesian models, and ensemble algorithms.
- Feature learning: SAE fuses multisource sparse data into three feature spaces with reduced attribute correlation.The transformed spaces are intended to support more diverse base classifiers.
- Base classifiers: The voting layer uses SAE feature learning to reduce correlation and increase base-classifier diversity.The approach addresses sparse clinical measures and uses hidden-node activations to define different feature spaces.
- Feature learning: Features are learned by propagating groups of clinical measures through the SAE hidden layer.The framework extracts features from clinical-measure groups for each participant.
- Stacking layer: Stacking receives base-classifier predictions and uses a DBN-based transformation to rank classifiers before neural-network prediction.The method is designed to address dependence among base classifiers.
- Optimizing layer: Cost-sensitive learning addresses fewer AD patients than healthy participants and the greater cost of missed diagnosis.The framework uses resampling, threshold moving, and soft ensemble methods for neural-network training.
3 RESULTS AND DISCUSSION
The results show that SAE-derived feature spaces improve separability while reducing feature dependence, and that DELearning combines diverse, ranked classifiers with cost-sensitive optimization. On the NACC dataset, DELearning achieved higher prediction accuracy than six representative ensemble methods.
- Feature learning: SAE-transformed features made AD and NDC participants easier to distinguish, with a relatively clearer boundary than in the original space.The transformed spaces were reduced and judged more effective for representing the two outcome groups.
- Feature learning: The transformed feature spaces reduced attribute dependence and tended toward low or no collinearity.The authors state that this helps improve base-classifier accuracy and speed.
- Construction and evaluation of classifiers: Base-classifier performance in the 20- and 30-dimensional SAE spaces was comparable with the original feature space while retaining useful information.Different learning algorithms and feature data were used to maximize classifier or physician diversity.
- Construction and evaluation of classifiers: The Q statistic for logistic regression and REPTree was 0.86<1, indicating that the two classifiers tended to recognize the same individuals correctly.The comparison used classifiers operating in the 20-dimensional feature space.
- Stacking by deep learning: Classifier ranking identified base classifiers No. 3 and 5 as having the highest scores and concurrently higher prediction accuracy.The ranking method was used in the stacking layer to evaluate classifiers or physicians automatically.
- Stacking by deep learning: A DBN with 6 hidden units had the lowest error and was selected as the best model for initializing the artificial neural network.The authors compared hidden-layer sizes from 3 to 9 units.
- Optimization for the final decision: The selected cost matrix was based on comparisons of nine matrices, with the highest geometric mean observed for cost 2.The cost matrices used integer values between 1.0 and 10.0 and included at least one non-diagonal element equal to 1.
- Optimization for the final decision: More than 4% higher prediction accuracy was achieved by DELearning than by Logiboost, Bagging, Random forest, AdaBoostM1, Stacking, and Vote on the NACC dataset.The comparison used four algorithms to generate individual committees: ZeroR, decision table, Naïve Bayes, and multilayer perception.
4 CONCLUSIONS
The paper proposes DELearning, a three-layer deep ensemble framework for Alzheimer’s disease classification that integrates multisource data and expert-like classifier opinions. On NACC UDC clinical measures, it outperformed six representative ensemble methods and is presented as a data-driven aid for primary care where Alzheimer’s expertise is limited.
- Conclusions: DELearning uses deep learning to ensemble at each of three layers for Alzheimer’s disease classification.The framework is intended to integrate multisource data and leverage the wisdom of experts.
- Conclusions: DELearning outperformed six representative ensemble learning methods in AD prediction accuracy using clinical measures from NACC UDC.The conclusion reports the comparative outcome without specifying an additional metric.
- Conclusions: The framework provides a data-driven solution to aid AD primary care where access to AD expertise is limited.
- Conclusions: DELearning may also support medical image tagging by aggregating subjective labels from multiple experts or annotators when objective labels are infeasible or too expensive.The authors state that the framework can find meaningful yet hidden labels in such settings.