Generative Multiple-Instance Learning Models For Quantitative Electromyography

Tameem Adel, Benn Smith, Ruth Urner, Daniel W. Stashuk, Daniel J. Lizotte

Introduction

Diagnosing neuromuscular disorders, such as myopathy or neuropathy, relies heavily on Quantitative Electromyography (QEMG). In clinical practice, a needle electrode is inserted into a muscle to record electrical activity. This raw signal is then decomposed into Motor Unit Potential Trains (MUPTs), which represent the firing patterns of individual motor units.

Overview of EMG signal acquisition and decomposition into Motor Unit Potential Trains (MUPTs).

The diagnostic challenge lies in the nature of the data. A diseased muscle (the "bag") does not consist exclusively of diseased motor units (the "instances"); rather, it contains a heterogeneous mixture of healthy and abnormal units. This structure defines a Multiple-Instance Learning (MIL) problem. Traditional MIL approaches often rely on a "trigger" rule—where a bag is labeled positive if at least one instance is positive. However, in clinical neurology, the relationship is generative: the underlying disease state of the muscle determines the statistical distribution of its constituent motor units.

The BIF Generative Framework

To better reflect this biological reality, the authors propose a generative framework for MIL. Unlike discriminative models that merely seek a decision boundary, generative models describe the joint distribution of the data, effectively acting as simulators of the disease process.

The primary contribution is the BIF (Bag $\to$ Instance $\to$ Feature) model. This Bayesian network structure assumes that the bag label (the clinical diagnosis) generates the hidden instance labels, which in turn generate the observed features of the MUPTs.

The BIF (Bag $\to$ Instance $\to$ Feature) model structure.

The authors compared this to the traditional FIB (Feature $\to$ Instance $\to$ Bag) view, where features determine instance labels that are then aggregated via a logical rule. To handle the complex, non-Gaussian distributions typical of EMG data, they explored several density estimation techniques:

  • Multivariate Gaussians: A computationally efficient baseline.
  • Kernel Density Estimation (KDE): A non-parametric approach to capture complex data shapes.
  • Gaussian Copulas with KDE Marginals: A hybrid method that uses KDE for individual feature distributions while using a Copula to model dependencies between features.

Since instance-level labels are typically unavailable in clinical settings, the models are trained using a Hard-EM (Expectation-Maximization) algorithm, iteratively estimating hidden instance labels and updating model parameters.

Evaluation and Datasets

The models were validated on two clinical MUPT datasets:

  1. Upper Leg: 88 bags, 1534 instances.
  2. Lower Leg: 70 bags, 1500 instances.

Expert-labeled instances served as a "gold standard" for validation, though these labels were withheld during training. Additionally, the MUSK1 benchmark dataset was used to test the model's generalizability to non-medical MIL problems.

Key Results and Insights

The BIF model demonstrated significant improvements over existing methods:

  • Classification Accuracy: On the lower leg dataset, the BIF model achieved 98.6% accuracy, a substantial increase from the previous state-of-the-art of 82.3%.
  • Instance Recovery: The model successfully identified individual motor unit states with up to 98.4% accuracy, despite never seeing instance labels during training.
  • Robustness: Even on the MUSK1 dataset, the BIF model achieved 87% accuracy, suggesting that the "Bag causes Instance" assumption is a powerful inductive bias for various real-world scenarios.

Conclusion

By shifting from discriminative to generative modeling, this research provides a more biologically plausible and accurate tool for QEMG diagnosis. The BIF framework not only improves classification but also offers interpretability, allowing clinicians to understand the distribution of healthy versus diseased units within a muscle. Future work aims to extend this framework to estimate continuous disease severity rather than providing simple binary classifications.

Summary updated

Abstract

We present a comprehensive study of the use of generative modeling approaches for Multiple-Instance Learning (MIL) problems. In MIL a learner receives training instances grouped together into bags with labels for the bags only (which might not be correct for the comprised instances). Our work was motivated by the task of facilitating the diagnosis of neuromuscular disorders using sets of motor unit potential trains (MUPTs) detected within a muscle which can be cast as a MIL problem. Our approach leads to a state-of-the-art solution to the problem of muscle classification. By introducing and analyzing generative models for MIL in a general framework and examining a variety of model structures and components, our work also serves as a methodological guide to modelling MIL tasks. We evaluate our proposed methods both on MUPT datasets and on the MUSK1 dataset, one of the most widely used benchmarks for MIL.
BibTeX
@inproceedings{adel2013generative,
  author = {Tameem Adel and Benn Smith and Ruth Urner and Daniel W. Stashuk and Daniel J. Lizotte},
  title = {{Generative Multiple-Instance Learning Models For Quantitative Electromyography}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2013},
  url = {https://openreview.net/forum?id=gBgr9DLmY4Y},
}
Figure