MSSDA: Multi-Sub-Source Domain Adaptation for Diabetic Foot Neuropathy Recognition

Yan Zhong, Zhixin Yan, Yi Xie, Shibin Wu, Huaidong Zhang, Lin Shu, Peiru ZhouAAAI 2025 · DOI

TL;DR

This paper introduces MSSDA, a novel multi-source domain adaptation framework for recognizing diabetic foot neuropathy (DFN) using plantar pressure data. By intelligently selecting and aligning sub-source domains, MSSDA effectively bridges domain gaps, leading to improved DFN recognition accuracy, especially in challenging cross-subject scenarios.

What Problem Does This Paper Address?

Diabetic foot neuropathy (DFN) is a serious complication of diabetes that can lead to foot ulcers and amputations. Early detection is crucial, and plantar pressure data from wearable sensors offers a promising avenue for monitoring. However, existing DFN datasets are limited, and traditional machine learning methods struggle with domain shift – the performance drop when models trained on one set of data (source domain) are applied to another (target domain), especially when data comes from different individuals. This paper tackles the challenge of recognizing DFN from plantar pressure data by developing a domain adaptation technique that can effectively leverage multiple, diverse data sources (multi-source) while mitigating the negative impacts of domain shift.

Methods

The core of this paper is the proposed MSSDA (Multi-Sub-Source Domain Adaptation) framework, a three-stage approach designed to handle domain shift in plantar pressure data for DFN recognition.

  • Stage 1: Contrastive Learning: This stage uses contrastive learning not as a pre-training step, but to enhance the feature extractor's ability to create a more discriminative feature space. It aims to push samples further apart, improving the subsequent clustering step.
  • Stage 2: Clustering and Pseudo Domain Labels: Here, the framework leverages convolutional feature statistics (mean and standard deviation) from the feature extractor's output. These statistics are used to cluster the data, and the clusters are assigned "pseudo domain labels." This allows the model to identify distinct sub-domains within the source data without relying on individual patient labels. The number of clusters is determined using the Bayesian Information Criterion (BIC).
  • Stage 3: Sub-Source Domain Selection and Alignment: This is a key innovation. Instead of using all source domains, MSSDA selects a subset of "proper" sub-source domains that are closest to the target domain. This selection process is crucial to avoid "negative transfer," where irrelevant or dissimilar source data can harm performance. The selected sub-source domains are then aligned with the target domain in multiple feature spaces using a domain-adversarial approach, similar to DANN but applied in parallel for each selected sub-source domain.

Datasets

  • DFN-DS (Diabetic Foot Neuropathy Dataset): A novel dataset introduced by the authors, comprising continuous plantar pressure data from 94 DM patients with DFN and 41 DM patients without DFN. This dataset is collected from wearable shoes and captures foot-specific information.
  • Fall Risk Assessment (FRA) Dataset: An existing plantar pressure dataset used for evaluation, containing data from 48 subjects (23 high-risk, 25 low-risk).

Key Insights and Ideas

  • Intelligent Sub-Source Domain Selection is Crucial: The paper strongly advocates for selecting relevant sub-source domains rather than using all available data. This is a significant departure from some existing multi-source domain adaptation methods and is shown to prevent negative transfer, leading to better performance.
  • Clustering by Feature Statistics, Not Just Individuals: Instead of treating each patient as a separate domain, MSSDA uses convolutional feature statistics to discover latent domains. This approach is more robust to variations within individuals and can better capture underlying data structures.
  • Contrastive Learning for Enhanced Feature Separation: Using contrastive learning within the framework, rather than solely for pre-training, helps to create a more distinct feature space that aids in subsequent clustering and domain alignment.
  • DFN-DS: A Step Towards Realistic DFN Monitoring: The creation of a new, continuous plantar pressure dataset is a valuable contribution, addressing the lack of foot-specific, long-term data in existing DFN research.

Key Results

  • Superior Performance on DFN-DS: MSSDA achieved state-of-the-art results on the newly introduced DFN-DS dataset, outperforming various established domain adaptation methods across multiple metrics (Precision, Recall, Accuracy, F1-score). For instance, it achieved an accuracy of 0.889, significantly higher than other methods.
  • Strong Performance on FRA Dataset: The method also demonstrated competitive performance on the FRA dataset, indicating its generalizability to other plantar pressure-based biomedical signal processing tasks.
  • Ablation Studies Validate Components: Ablation studies confirmed the effectiveness of each stage of MSSDA, showing that combining sub-source selection with multi-feature space alignment yields the best results.

Limitations and Future Work

  • Threshold Sensitivity: The paper notes that MSSDA's accuracy drops more significantly at higher thresholds compared to some other methods. This suggests that while MSSDA excels at transferability, it might be overly sensitive to the exact decision boundary, which could be a focus for future work.
  • Parameter Tuning: The selection of the number of sub-source domains (M) is user-determined, and finding the optimal M might require further investigation.
  • Computational Cost: While not explicitly detailed as a limitation, the multi-stage alignment process, especially with multiple sub-source domains, can be computationally intensive.

The authors suggest that future work could explore more sophisticated methods for threshold selection and potentially investigate adaptive strategies for determining the optimal number of sub-source domains.

Summary updated

BibTeX
@inproceedings{zhong2025mssda,
  author = {Yan Zhong and Zhixin Yan and Yi Xie and Shibin Wu and Huaidong Zhang and Lin Shu and Peiru Zhou},
  title = {{MSSDA: Multi-Sub-Source Domain Adaptation for Diabetic Foot Neuropathy Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i21.34451},
}
Figure