{
  "abstract": "Introduction We present a low-cost method for automatic screening of obstructive sleep apnoea (OSA) using overnight recordings captured from a smartphone. OSA detection performance is improved by using multi-task learning (MTL), 1 a machine learning approach that jointly trains the model on OSA detection and a secondary task – estimating oxygen desaturation from SpO2 data. Only the data for the primary task (audio recordings) are required during deployment.Methods MTL was evaluated against single-task learning (STL), using a data set of 1094 hours of smartphone audio and data from home sleep apnoea test (157 nights, 103 participants). OSA severity was annotated by a trained sleep technologist. The primary task was classification of OSA using acoustic features; the secondary task was prediction of the time spent below individual SpO2 baselines. We compared STL and MTL performance, including the effect of class weighting in STL models (see figure 1).Results Class weighting benefitted the STL model, but MTL generally outperformed STL, with the greatest improvement in detecting moderate-to-severe OSA (AHI>=25). The best-performing MTL variant (so-called hard parameter sharing) achieved a sensitivity of 0.84 and a specificity of 0.93 (see table 1).Discussion Acoustic analysis provides a promising, non-invasive method for accessible OSA screening. MTL improves the robustness of machine learning models by exploiting related physiological signals during training, even if those signals are not available during deployment. Future work will investigate the integration of other physiological signals that can be obtained from low-cost sensors (e.g. accelerometry).Reference Ruder S. 2017. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098.Abstract P53 Table 1OSA screening results using STL and MTL models under different AHI cut off points. STL-CW: Single-Task Learning with Class Weights; MTL-HS: Multi-Task Learning Hard Sharing; MTL-SS: Multi- Task Learning Soft Sharing; MTL-HR: Multi-Task Learning Hierarchical Sharing; MTL-SP: Multi-Task Learning Shared-Private. See details in figure 1Abstract P53 Figure 1Overview of single-task learning (STL) and multi-task learning (MTL) architec- tures. In STL, audio waveforms are transformed into mel spectrograms and passed through a convolutional neural network (CNN) followed by fully connected layers for binary OSA classi- fication. In MTL with Hard Sharing (MTL-HS), shared convolutional layers extract common features, which branch into two heads for classification and regression. The Soft Sharing variant (MTL-SS) uses two task-specific CNN branches and dynamically exchanges information through a gated mechanism. Hierarchical Sharing (MTL-HR) applies shared lower-level layers, with the regression task diverging earlier, while the classification path continues through higher-level processing. The Shared-Private model (MTL-SP) incorporates both shared and task-specific private branches, with outputs from each concatenated before task-specific predictions",
  "authors": [
    {
      "affiliations": [
        "School of Computer Science, University of Sheffield, Regent Court, 211 Portobello, Sheffield S1 4DP, United Kingdom"
      ],
      "name": "Chaoyue Niu"
    },
    {
      "affiliations": [
        "School of Computer Science, University of Sheffield, Regent Court, 211 Portobello, Sheffield S1 4DP, United Kingdom"
      ],
      "name": "Xiaolei Xu"
    },
    {
      "affiliations": [
        "School of Computer Science, University of Sheffield, Regent Court, 211 Portobello, Sheffield S1 4DP, United Kingdom"
      ],
      "name": "Guy J Brown"
    },
    {
      "affiliations": [
        "School of Computer Science, University of Sheffield, Regent Court, 211 Portobello, Sheffield S1 4DP, United Kingdom"
      ],
      "name": "Ning Ma"
    }
  ],
  "title": "P53 Multi-task learning for acoustic analysis of obstructive sleep apnea",
  "uid": "28ed0276-dbf4-51b0-989e-3afeeb5156e2"
}