02 May 2024
14:00 Master's Defense IC3 Auditorium
Theme
Visual Representations for Classifying Sleep Stages
Student
Rebeca Padovani Ederli
Advisor / Teacher
Zanoni Dias - Co-supervisor: Anderson de Rezende Rocha
Brief summary
The presence of anomalies during sleep, originally classified by experts, is used to diagnose sleep disorders. Polysomnography is the gold standard used in clinical settings for monitoring and classifying sleep stages, however, this method requires expensive and uncomfortable equipment. Recently, sleep analysis using wearable devices such as smartwatches has been investigated for the automatic classification of sleep stages, aiming to combine the comfort of smartwatches with the accuracy of polysomnography. However, the challenges faced include the correct classification of the more complex stages and the imbalance of classes in the datasets. Using sensor data from smartwatches, the literature explores the performance of sleep stage classification through different data representations, such as raw data (sensor time series) and features extracted from them (such as mean, standard deviation, etc.). In this context, the representation of time series through images can produce informative and noise-robust characteristics. Thus, this study innovates by exploring different visual representations of time series derived from accelerometer and heart rate data, evaluating their effectiveness in classifying sleep stages. Based on the experiments carried out, although Gramian Angular Field (GAF) demonstrated the most promising results, representations by Recurrence Graphs (RP) and Markov Transition Field (MTF) also proved to be promising techniques for classification. Furthermore, the strategy of dividing images into patches and employing ensemble techniques was investigated with the aim of improving results in classifying sleep stages. Experiments revealed that this approach contributes substantially to improving classification. Additionally, the research evaluated the potential of data from accelerometer and heart rate sensors in smartwatches, exploring the combination of these two sets of data through the application of ensembles. The results indicate that, depending on the specific classification context (whether two- or three-stage), accelerometer or heart rate data may play a more significant role. Finally, the comparison with other representations commonly used in the literature highlighted the superiority of visual approaches in this context. Evaluation using different metrics confirmed the effectiveness of the proposed techniques, indicating a competitive approach for classifying sleep stages based on data from wearable devices.
Examination Board
Headlines:
| Zanoni Dias | IC / UNICAMP |
| Marcos Medeiros Raimundo | IC / UNICAMP |
| George Darmiton da Cunha Cavalcanti | CIn / UFPE |
Substitutes:
| Ricardo Ribeiro Gudwin | FEEC / UNICAMP |
| Celmar Guimaraes da Silva | FT / UNICAMP |