Biomedical signal analysis has traditionally been helpful in diagnosing various health issues for example. EEG for seizure and epilepsy detection and ECG for various heart ailments. The recent developments in signal processing and machine learning have led to two important developments:
1. The diagnostic capabilities of machine based algorithms have improved a lot, with deep learning algorithms detecting arrhythmia as well as human diagnostician. and better cancer detection using convolutional neural networks.
2. Making use of wearable devices and the various bio and activity signals collected from them to predict various psycho-physiological constructs like stress, anxiety, mental workload etc. Such predictions can further be improved using other social-interaction and phone usage information. Currently, such projects are being undertaken in academia on a large scale.
My research is focused on the marriage between commercially available wearable devices and pattern recognition algorithms by using signal processing and physiological insights. Currently, a large number of data is being collected using smart wearables such as fitbits, apple watches and other devices. However, without proper signal processing techniques to assess the usability of such data (referred to as Signal Quality Assessment) and extract relevant information from them (Feature Engineering), these signals may only yield sub-optimal results which aligns well with the old adage in machine learning "garbage in- garbage out". Extracting relevant features requires knowledge of several domains including:
signal transformations/representations
physiological behavior and insights
signal decomposition
Once proper signal processing steps have been performed on the input signals, the extracted physiological information needs to be fed into pattern recognition algorithms. These algorithm can be simple, broadly referred to as machine learning, or much more complex, called deep learning that include millions of parameters. Several considerations go into development of robust pattern recognition systems. These include:
handling missing data
imbalance in dataset labels
data dimensionality
feature redundancy
classifier explainability
By combining knowledge from physiological signal processing and machine learning it is possible to develop robust diagnostic, prognostic and prediction systems. Such models can be individualized on completely generalizable to new subjects.