Abstract:
The human body is the most complicated apparatus, comprising a plethora of mutually connected organs, performing diversely different sophisticated functions. These interconnected organs and processes communicate with each other in forms of signals, which are commonly termed as Biomedical Signals. In addition to playing a vital role in the proper operation of various physiological tasks, these signals can also be used as indications of whether the body is functioning appropriately or is suffering from diseases. For a long time, Doctors have been using ECG (Electro- cardiogram) signals to monitor the heart condition of the patients. Moreover, PPG (Photoplethysmogram) signals have been being used to infer blood pressure. Both have resulted in an abundance of academic research works, however, they seem to be disjoint in nature.
About half of the works in the literature revolve around Signal Processing based
methods, where after monumental analysis of the signal a rather simpler mathemat- ical model is developed. On the contrary, there are works following the approaches of Machine Learning, but they hardly consider the discernible patterns in the sig- nals during feature engineering. Fusing the ideas from both the schools of Signal Processing and Machine Learning, we, therefore, propose an improved algorithmic pipeline, VFPred, that can detect Ventricular Fibrillation from ECG signals, which catalyzes life-threatening cardiac arrests. VFPred extends upon traditional signal processing based feature extraction and subsequently utilizes a suitable machine learning based classifier to not only demonstrate an outstanding accuracy, but also a balance between sensitivity and specificity.
On the contrary, employing the potential of Deep Learning, we develop PPG2ABP,
that is capable of inferring the continuous arterial blood pressure waveform with minimal error from analyzing PPG signals. Use of deep learning emancipates PPG2ABP from the need for handcrafted features, which often restricts the input signals to follow an ideal shape. Furthermore, this enables us to surpass contempo- rary methods both in terms of reliability and versatility.
