Speaker: Valentina Vadori, Date: 13th of November 2015, Time: 13:30, Room: 201 (DEI/A)
Abstract. Recent advances in wearable devices hold the promise to enhance the efficiency and applicability of telemedicine solutions. These devices allow seamless, non-invasive and inexpensive gathering of biomedical signals such as electrocardiogram (ECG), photoplethysmogram (PPG), heart rate, blood pressure, blood oxygen saturation, and respiration (RESP). They can be integrated into wireless body sensor networks (WBSN) to update medical records via the Internet, improving prevention, care’s personalization and quality and can be used within personalized training applications. Since wearables are required to be small and lightweight, they are often resource constrained, and as such they need dedicated algorithms to optimally manage energy and memory. In this work, we design SAM, an original Subject-Adaptive (lossy) coMpression technique for physiological quasi-periodic signals such as ECG, PPG and RESP. The data volume reduction that we achieve allows efficient storage and transmission, and thus helps extend the devices’ lifetime. SAM is based upon a subject-adaptive dictionary which is constructed and adapted at runtime without requiring any prior information on the signal itself. This is achieved utilizing the time-adaptive self-organizing map (TASOM) unsupervised learning algorithm. The time adaptivity of the conceived TASOM architecture makes it possible to learn and refine the dictionary as the statistics governing the input data undergoes major changes, tailoring it to the subject that wears the device. Quantitative results show the superiority of our algorithm against state-of-the-art techniques: compression ratios of up to 35-, 70- and 180-fold are generally achievable respectively for PPG, ECG and RESP signals, while reconstruction errors (RMSE) remain within 2% and 7% and the input signal morphology is preserved.