Integration of Multivariate Beta-based Hidden Markov Models and Support Vector Machines with Medical Applications

Narges Manouchehri, Nizar BouguilaFlairs 2022 · DOI

Introduction

In medical artificial intelligence, two primary challenges often hinder the development of robust diagnostic tools: rigid statistical assumptions and the inherent trade-off between generative and discriminative models. Traditional Hidden Markov Models (HMMs) typically rely on Gaussian (Normal) distributions, which fail to accurately represent the skewed or asymmetric nature of real-world biological data. Furthermore, while generative models like HMMs excel at learning from small datasets, they often lack the classification accuracy of discriminative models like Support Vector Machines (SVMs). Conversely, SVMs struggle with variable-length sequences common in clinical observations.

Methodology: The MB-HMM-SVM Framework

The researchers propose a novel hybrid architecture, MB-HMM-SVM, designed to leverage the strengths of both paradigms. The model replaces standard Gaussian emissions with a Multivariate Beta distribution, offering significantly greater flexibility to model complex biological shapes and densities.

Visualization of flexible Multivariate Beta distributions used in the MB-HMM to capture non-Gaussian data patterns.

To bridge the gap between the generative HMM and the discriminative SVM, the authors utilize Fisher Kernels. By extracting "Fisher Scores"—the gradient of the log-likelihood—the model transforms variable-length sequences into fixed-length feature vectors. These vectors are then processed by an SVM to define high-performance decision boundaries. The team employed the Forward-Backward algorithm to efficiently estimate the complex parameters of the Multivariate Beta distributions.

Experimental Setup and Datasets

The model was evaluated using features extracted via a pre-trained VGG-16 architecture from three medical datasets:

  • Lung Cancer Histopathology: 3,000 samples categorized into benign, adenocarcinoma, and squamous cell carcinoma.
  • Colonoscopy Images: 1,000 images consisting of normal samples and polypoid lesions.
  • Colon Histopathology: 2,000 samples of benign and malignant colon tissues.

Key Results

The MB-HMM-SVM consistently outperformed baseline models, including standard HMMs and standalone SVMs, across all evaluated metrics. In lung tissue analysis, the hybrid model achieved an accuracy of 87.44% and an F1-score of 87.79%, significantly higher than the 79.36% accuracy of a standard linear SVM. The performance gain was even more pronounced in colonoscopy analysis, where the model reached 93.81% accuracy, compared to 81.74% for GMM-HMM and 77.19% for a standard SVM.

Performance comparison of the proposed MB-HMM-SVM against baseline models across multiple metrics.

Conclusion

By integrating flexible statistical distributions with powerful discriminative classifiers through the Fisher bridge, the MB-HMM-SVM provides a robust, interpretable framework for medical image analysis. Unlike "black box" deep learning models, this hybrid approach relies on interpretable statistical distributions. Future work will focus on optimizing the computational efficiency of gradient calculations and extending the model to temporal clinical data, such as longitudinal patient vitals.

Summary updated

BibTeX
@inproceedings{manouchehri2022integration,
  author = {Narges Manouchehri and Nizar Bouguila},
  title = {{Integration of Multivariate Beta-based Hidden Markov Models and Support Vector Machines with Medical Applications}},
  booktitle = {Flairs},
  year = {2022},
  doi = {10.32473/flairs.v35i.130667},
}
Figure