A Multi-Sensor Data Fusion Approach for Atrial Hypertrophy Disease Diagnosis Based on Characterized Support Vector Hyperspheres

Yungang Zhu, Dayou Liu, Radu Grosu, Xinhua Wang, Hongying Duan, Guodong Wang2017 · DOI

TL;DR

This paper introduces a novel multi-sensor data fusion method called Characterized Support Vector Hyperspheres (CSVH) for diagnosing atrial hypertrophy. CSVH addresses limitations of traditional methods like SVM by using hyperspheres to better capture data diversity and separation, achieving higher diagnostic accuracy on real patient data.

What Problem Does This Paper Address?

Cardiovascular diseases, including atrial hypertrophy, are a leading cause of death globally. Accurate and timely diagnosis is crucial, but current computer-aided diagnosis (CAD) methods for atrial hypertrophy face challenges. Specifically, a lack of sufficient labeled training data and the inherent complexity and diversity of patient medical records make it difficult for traditional machine learning algorithms, like Support Vector Machines (SVMs), to perform well. SVMs, which rely on hyperplanes for classification, struggle with overlapping classes and outliers, common issues in medical data. Existing approaches like Locally Discriminative SVM (LDSVM) attempt to address some of these issues but have their own limitations, such as biased metrics for overlapping data and a lack of explicit outlier handling.

Methods

The proposed method, Characterized Support Vector Hyperspheres (CSVH), departs from the hyperplane-based approach of SVMs. Instead, it constructs two "hyperspheres" to represent the patient and healthy classes.

  • Characterized Hyperspheres:
    • Patient Class Hypersphere: This hypersphere is designed with a "weighted schema" to account for the diversity within patient data, including outliers and overlapping-region data. This allows it to better capture the nuanced characteristics of patients.
  • Healthy Class Hypersphere: This hypersphere is constructed to be maximally separated from the patient class hypersphere, aiming for clear distinction.
  • Labeling Phase:
    • Membership Functions: Two membership functions are defined based on the distance of a query point to each hypersphere.
  • Decision Making: If a query's membership degrees to both classes are sufficiently different, it's classified based on the higher degree.
  • Outlier/Overlap Handling: If the membership degrees are close, indicating uncertainty or proximity to outliers/overlapping regions, CSVH analyzes the angle information between the query and these specific data points to make a refined decision.
  • Adaptive Parameterization: The method incorporates heuristics for data-adaptive parameter tuning, enhancing efficiency.

Datasets

The paper utilizes two real electrocardiogram (ECG) datasets for atrial hypertrophy diagnosis:

  • MGH/MF Dataset: Contains 250 patient instances.
  • Fantasia Dataset: Contains 40 healthy subject instances.

For validation and comparison, the authors also used two synthetic datasets:

  • Exclusive OR (XOR) Dataset
  • Crossing Line Dataset

Additionally, they evaluated their method on several benchmark datasets from the UCI Machine Learning Repository.

Key Insights and Ideas

  • Hyperspheres for Data Description: The core idea is that hyperspheres are more powerful than hyperplanes for describing complex data distributions, especially in the presence of diversity and overlap.
  • Weighted Hypersphere for Patient Data: Recognizing the inherent variability in patient data, a weighted hypersphere approach is crucial for capturing this diversity and handling outliers effectively.
  • Maximizing Separation: Explicitly pushing the healthy class hypersphere as far as possible from the patient class hypersphere is key to achieving robust classification.
  • Leveraging Outlier/Overlap Information: Instead of discarding or poorly handling outliers and overlapping data, CSVH actively uses their geometric relationships (angles) to refine classification decisions, a significant advantage over methods that struggle with these cases.
  • Adaptive Parameterization: The use of data-adaptive heuristics for parameter tuning is essential for making the method robust across different datasets and avoiding manual tuning.

Key Results

  • Atrial Hypertrophy Diagnosis: CSVH achieved the highest diagnosis accuracy on the real atrial hypertrophy datasets compared to other state-of-the-art methods, including SVM, LDSVM, TWSVM, MVSVM, and NN2.
  • Benchmark Datasets: On UCI benchmark datasets, CSVH performed competitively, often matching or exceeding TWSVM, demonstrating its generalizability.
  • Stability: CSVH exhibited more stable performance, indicated by narrower standard deviations in accuracy across different runs and datasets.
  • Imbalanced Data: CSVH and TWSVM were less affected by imbalanced training data compared to other methods, a critical factor in real-world medical scenarios.
  • Statistical Significance: Statistical tests (Friedman and Wilcoxon signed-ranks) confirmed that CSVH's performance improvements were statistically significant compared to its competitors.

Limitations and Future Work

  • Computational Cost: While CSVH offers performance benefits, its time complexity is noted as 2·O((N/2)³), similar to TWSVM and MVSVM, due to quadratic optimization problems. This is still better than SVM's O(N³).
  • Future Work:
    • Bayesian Networks Integration: Combining CSVH with Bayesian networks to incorporate prior medical expert knowledge.
  • Ensemble/Information Fusion: Developing techniques for ensemble methods or information fusion to further boost accuracy.
  • Visualization: Studying and visualizing the high-dimensional distribution of atrial hypertrophy data.
  • Multi-Classification: Extending CSVH to handle multi-classification problems.

Summary updated

BibTeX
@misc{zhu2017multi,
  author = {Yungang Zhu and Dayou Liu and Radu Grosu and Xinhua Wang and Hongying Duan and Guodong Wang},
  title = {{A Multi-Sensor Data Fusion Approach for Atrial Hypertrophy Disease Diagnosis Based on Characterized Support Vector Hyperspheres}},
  year = {2017},
  doi = {10.3390/s17092049},
}
Figure