The Project
It appears from the question that emotional detection and response is the longer term goal of the project and that recognizing potential emotional manifestations in easily detectable physiological metrics is an initial R&D objective.
Mobile device applications are already available to do this, but biometric monitoring via a wristband, excellence in AI design, and marketing excellence could overtake these apps in the marketplace and provide emotional regulation to improve productivity and reduce health and wellness risks.
The (at least initial) goal seems to be basic binary classification, which simplifies the output, but not the detection, which seems to have the following two criteria.
The Biometric Device Challenge
There is definitely a challenge to classifying biometric trends as meaningful in this context.
What do you think about my reasoning of the problem? Do you have other perspectives on how to handle these problems? I have been told this could also be considered an anomaly detection problem, for example.
Novelty detection is not the first milestone in this challenge. Detection of useful features is the prerequisite. Novelty comes into play once changes in the organism can be characterized with some reliability and accuracy (few false positives and false negatives) and the particular user has provided subjective reporting to correspond with some change in the organism.
Biological systems have forms of stasis that can be indirectly sensed, including these.
- Blood pressure
- Metabolic rate
- Multi-ionic salinity
- Cognitive attention
- Blood sugar levels
These affect the dermis at the wrists through perturbations in externally detectable metrics, the inclusion of which depends on the capabilities of the wrist device.
- Circumference
- Coefficient of electrical resistance
- Surface temperature
- Multi-ionic salinity
- Surface moisture
With blood and urine lab tests, conditions can be controlled, however sampling of those fluids are usually infrequent, leading to the use of ranges to detect anomalies that may be indicative of disease, conditions, or other health risks. There are major challenges to using ranges of metrics at the wrist.
- Sensitivity to the general and ongoing physical condition of the individual
- Unique individual responses to emotional states
- Interference from changes in the temperature, pressure, air movement, and humidity of the environment
One inroad to detection of the internal metrics of the organisms is the fact that biological stasis is not the same thing as perfect regulation. Biological systems are chaotic. Normal signals from time series sampling of biological metrics are neither constant nor periodic. When changes in the attractors and spectra characterizing the chaotic fluctuation of direct metrics (like electrical resistance) or indirect metrics (like systolic blood pressure) occur, a request for subjective information may be asked from the wearer.
These are some chaotic principles to consider.
- Spectra can provide much information about the intensity of the force driving the chaotic fluctuation in the time series, which is likely to correlate with stress.
- The various ways to evaluate exponents, including the Lyapunov exponent, may be of significant value in correlating patterns in the external, indirect time series to emotional states.
- Autocorrelation at the primary frequencies involved may reveal changes indicative of significant mental transitions.
Once a correlation is established, then the novelty detection makes sense, except for an occasional re-acquisition of the correlation, since the wearer may become more internally connected to their physical and emotional states as they use the app, and the answers to inquiries may mature correspondingly.
Additional Questions Within the Question
These additional questions were also asked.
What algorithms would you use to separate normal from more meaningful physiological data? Support Vector Machines perhaps? Maybe some decision theory?
A recurrent neural network such as B-LSTM or GRU are commonly excellent at characterizing time series, but there may be a need for more than one network.
- One to extract the features of the chaotic patterns in the raw data
- Another to train to correlate changes in the features of the chaos with subjective reporting
This second training must be done during in situ field use of the first already trained network. Since training is not generally viable on a mobile device, there will need to be a client server arrangement where the training data and its resulting trained network can be communicated through secure RestFUL services so the training processes can be asynchronous and leverage GPUs or other hardware acceleration.
Do you know any books or papers on similar matters? Even if I have found some after hours and hours of research, you may be able to point me to something different than those I have.
The terms above in a scholarly search will bring up all kinds of study materials. Specifically these
- Lyapunov exponent
- Strange attractors
- Spectral analysis of biometrics
- Chaos autocorrelation
- Phase space
One good book for all but the middle item is Chaos Theory Tamed, 1997
by Garnett P. Williams.
It is worth noting that data collection is supposed to be done when no other factors are messing with signal readings, such as sport.
This is where higher level detection involving chaotic analysis is much more reliable than ranges. The chaotic features of athletic activity will be distinctly different from those of fear states or sexual arousal.
