[home] [Personal Program] [Help]

---

Session: Poster Session 1 (Even numbers)
Session starts: Thursday 26 January, 16:00
Presentation starts: 16:00

---

Hans van Gorp (Eindhoven University of Technology)
Merel van Gilst (Eindhoven University of Technology)
Pedro Fonseca (Philips Research)
Ruud van Sloun (Eindhoven University of Technology)
Sebastiaan Overeem (Eindhoven University of Technology)

Abstract:
Objectives: Sleep staging is a crucial but labour-intensive diagnostic tool, compelling the use of automated machine learning models. These models are trained on scorings made by humans, which have imperfect agreement, resulting in uncertainty about the correct scoring. Recent developments in automatic sleep staging, such as hypnodensity, try to capture this uncertainty on an epoch-by-epoch basis. However, these methods are inadequate to capture the uncertainty of overnight sleep statistics. Here, we propose U-Flow, which given a single input PSG, outputs a desired number of plausible hypnograms, mimicking the diversity seen in human scorings. Methods: The model was trained with 529 recordings of the Stanford Sleep Cohort (SSC; single scorer), all 80 recordings of the Dreem Open Datasets (5 scorers), and all 110 recordings of the Institute of Systems and Robotics dataset (2 scorers). The validation set consisted of the remaining 132 recordings of SSC (single scorer). The hold-out test set comprised 70 recordings of the Inter-Scorer Reliability Cohort (6 scorers). We compared U-Flow to a hypnodensity-based U-Net model of similar computational complexity. For each input PSG, both models output 100 hypnograms, which were compared to those of the 6 experts. Accuracy and kappa were calculated through majority voting. Additionally, 9 commonly-used summarizing sleep statistics were calculated from each hypnogram, including total sleep time, time spent in each stage, sleep onset latency, REM onset latency, and number of awakenings. Each sleep statistic for each recording was modelled as a normal distribution, resulting in a mean and a variance, where the latter expresses the uncertainty. The distance between the distributions of the experts and the model predictions was evaluated using the Kullback–Leibler divergence. Significance was tested through ANOVA. Results: U-Flow was found to better predict the distribution of overnight sleep statistics by a human panel for 8 out of 9 parameters (p<0.05), while not sacrificing accuracy and kappa. These were respectively 80.0% and 0.71 for U-Net compared to 82.7% and 0.74 for U-Flow. Conclusions: U-Flow outperforms a hypnodensity-based U-Net model of similar computational complexity, in its ability to capture the uncertainty of overnight sleep statistics, as well as accuracy and kappa.