Research‎ > ‎

Model-Based Sampling Optimization


Faculty: Anind Dey, Deepak Ganesan
Post-docs: Jin-Hyuk Hong, Katarzyna Wac


The overall goal of the project is to develop a FieldStream system that enables obtaining scientifically valid physiological and behavioral data from human subjects in their natural environments in an accurate and timely manner. These data is to be obtained with use of Body Area Wireless Sensor Network (BAWSN) systems worn on the body of the subject. A BAWSN can sample physiological data as ECG, respiration, temperature, skin conductance and photoplethysmogram (PPG) with different frequency and resolution. The aim is that the BAWSN system automatically and accurately discriminates between the expected and unexpected behaviors, based on the models of behaviors learned from historical physiological data obtained from the subject in the past. Example behaviors include panic attack, stress or addiction withdrawal behavior.  

As the data is obtained from the subjects in their natural environments, the BAWSN needs also capture subject’s context and this to enable the FieldStream to deal with noise, motion artifacts, and existence of other uncontrollable confounding factors influencing the quality of data obtained from the subject. For this purpose, subject' s activity, ambient temperature, light and humidity conditions conditions are sampled.

Our research addresses the BAWSN system requirement for deriving accurate behavior models and for the system personalization to account for wide between person differences in physiological measurements. Secondly, our research aims at optimization of the sampling frequency of physiological  data for an accurate of behavior recognition based on the learned behavior models. The approach we take will be necessarily an experimental one. Namely, we will firstly attempt to develop personalized models for subjects’ behavior based on the data obtained in laboratory experiments. These models will not account for subject’s context. Secondly, we will aim to obtain data for subjects being in particular sets of contexts, such that models can be updated to account for these contexts. And finally we will derive optimization function for sampling frequency  and resolution of physiological measurements based on the required speed and accuracy of subject's behavior recognition.


We structure our research along the following activities:
  • Sampling - various data including physiological signals and user context is collected using BAWSN; parameters of physiological sensors that affect the performance of sampling and behavior recognition are sampling frequency and resolution
  • Feature extraction - we extract informative features to the target behaviors from the sampled data; the extracted features include statistical and structural features as well as domain-specific ones; we exploit various machine-learning techniques such neural networks, decision trees, Bayesian networks, and kNN
  • Multi-module decomposition - multiple modules are developed and decomposed to adjust the performance of behavior recognition in terms of its speed and accuracy
  • Behavior recognition - we dynamically organize multiple modules and construct behavior recognition models; we providing feedback to the sampling module upon a required frequency and resolution of sampling

Latest Work

Currently we are focusing on state-of-the-art research on personalized modeling of behaviors based on physiological data and user context data. We have formulated of the following research questions.

Research questions related to the personalized behavior modeling:

  • Developing personalized models that adapt with changes in context
    • What models we need to investigate for each physiological sensor?
    • What models we need to investigate for groups of physiological sensors?
    • What models we need to investigate for each behavior?
    • Which context and personalization factors we need to use?
    • How to detect changes in context and personalization factors?
    • How to set frequency and resolution of sampling based on the required accuracy and speed of behavior recognition?
  • Online refinement of the models
    • What physiological data streaming techniques can we use?
    • How to use population data, given wide between subject differences?
    • How to achieve all the above in the system constrained on its battery lifetime as well as its communication, storage and processing resources?

Research questions related to behavior recognition:

  • Definition of the protocol and scope of behavior recognition
    • What kind of information can be used for recognizing behaviors?
    • What kind of behaviors should be recognized?
    • How to log the physiological and context data via BAWSN?
    • How many multi-modules are required for behavior modeling?
    • What is the relationship between speed and accuracy of behavior models?
  • Development of behavior recognition module
    • Feature extraction: statistical, structural, domain-specific features
    • Feature selection based on the relationship with behaviors: manual / automatic
    • Behavior classification / identification by using machine-learning techniques: decision trees, Bayesian networks, neural networks, support vector machine or kNN
    • Multi-modules decomposition & integration: ensemble learning strategies (boosting, bagging) / fusion strategies (static, dynamic)
  • Performance evaluation
    • Local behavior classification / recognition: computational speed, accuracy
    • Global prediction performance: speed and accuracy based on multiple modules
    • Construction of a set of classifiers having various candidates for speed and accuracy
    • Overall system performance with relation to BASWN battery lifetime 

The research is ongoing. Please return to this page for updates on our progress.

Related Work

[1] J. Kim & E. Andre, "Emotion recognition based on physiological changes in music listening," IEEE Trans. on Pattern Analysis and Machine Intelligence, vol. 30, no. 2, pp. 2067-2083, 2008.

[2] Q. Ji, P. Lan, and C. Looney, "A probabilistic framework for modeling and real-time monitoring human fatigue," IEEE Trans. on Systems, Man, and Cybernetics-Part A: Systems and Humans, vol. 36, no. 5, pp. 862-875, 2006.

[3] J. Healey and R. Picard, " Detecting stress during real-world driving tasks using physiological sensors," IEEE Trans. on Intelligent Transportation Systems, vol. 6, no. 2, pp. 156-166, 2006.