Abstract
Alzheimer's disease (AD) is one of the most common progressive neurodegenerative diseases, and the number of AD patients has increased year after year with the global aging trend. The onset of AD has a long preclinical stage. If doctors can make an initial diagnosis in the mild cognitive impairment (MCI) stage, it is possible to identify and screen those at a high-risk of developing full-blown AD, and thus the number of new AD patients can be reduced. However, there are problems with the medical datasets including AD data, such as insufficient number of samples and different data distributions. Transfer learning, which can effectively solve the problem of distribution discrepancy between training and test data and an insufficient number of target samples, has attracted increasing attention over recent years. In this paper, we propose a multi-source ensemble transfer learning (METL) approach by introducing ensemble learning and our tri-transfer model that uses Tri-Training, which ensures the transferability of source data by the tri-transfer model and high performance through ensemble learning. The experimental results on the benchmark and AD datasets demonstrate that our proposed approach has effective transferability, robustness, and feasibility, and is superior to existing algorithms. Based on METL, we propose an auxiliary diagnosis system for the initial diagnosis of AD, which helps doctors identify patients in the MCI stage as quickly as possible and with high accuracy so that measures can be taken to prevent or delay the occurrence of AD.
Introduction
Alzheimer’s Disease (AD) represents a significant challenge for global healthcare systems, particularly as populations age. The most critical window for intervention is the Mild Cognitive Impairment (MCI) stage, where early diagnosis can lead to treatments that delay disease progression. However, training effective machine learning models for early AD detection is complicated by two major issues: data scarcity and distribution shift. Medical datasets are often small due to the high cost of expert labeling, and data collected from different clinical settings or demographics often follow different statistical distributions, leading to poor generalization.
While Transfer Learning (TL) offers a solution by leveraging knowledge from related "source" domains to a "target" domain, it frequently encounters "negative transfer." This occurs when irrelevant or noisy information from the source domain degrades the performance of the target model. To overcome these hurdles, this paper introduces Multi-Source Ensemble Transfer Learning (METL), a framework designed to filter source data and intelligently weight multiple sources to maximize diagnostic accuracy.
Methodology: The METL Framework
The METL framework is structured into two distinct phases to ensure that only high-confidence, relevant information is transferred to the target model.
Phase 1: Single-Source Tri-Transfer Learning
In the first phase, the system filters source data to align it with the target domain's distribution. This is achieved through a "Tri-Training" approach using three heterogeneous classifiers: Softmax, Support Vector Machine (SVM), and a Deep Neural Network (DNN). By using different types of models, the system creates a robust "agreement filter." A source sample is only deemed reliable and added to the training set if all three models agree on its classification relative to the target domain. This iterative process minimizes the risk of negative transfer by excluding noisy or irrelevant source samples.
Phase 2: MI-Based Multi-Source Ensemble
Not all source datasets are equally useful for a specific target population. METL employs Mutual Information (MI)—a metric from information theory—to mathematically quantify the similarity between each source domain and the target domain. In the final ensemble stage, classifiers trained on sources that are more similar to the target are assigned higher weights. This ensures that the final diagnostic model is primarily influenced by the most relevant available data.
Experimental Results
The researchers validated METL using both standard medical benchmarks and real-world Alzheimer's data.
- UCI Medical Benchmarks: Tested across 12 datasets (including diabetes and heart disease), METL achieved an average accuracy of 87.16% when using three source domains, significantly outperforming the single-source baseline of 79.87%.
- ADNI Dataset: Using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI3), METL was tested in scenarios with very few labeled target samples. At just 3% labeled target data, METL significantly outperformed existing algorithms like TrAdaBoost and MTrA, demonstrating its effectiveness in "cold start" clinical scenarios.
Conclusion
METL provides a robust framework for clinical auxiliary diagnosis by effectively managing the trade-offs in multi-source transfer learning. By combining an agreement-based data filter with information-theoretic weighting, the system achieves high accuracy even when target data is scarce. Future work will explore Heterogeneous Transfer Learning to allow for knowledge transfer between datasets with different feature sets, further expanding the system's utility in diverse medical environments.
Summary updated