Team, Visitors, External Collaborators
Overall Objectives
Research Program
Highlights of the Year
New Software and Platforms
New Results
Bilateral Contracts and Grants with Industry
Partnerships and Cooperations
XML PDF e-pub
PDF e-Pub

Section: New Results

Probabilistic Activity Modeling

Participants : Elisabetta de Maria, Sabine Moisan, Jean-Paul Rigault, Thibaud L'Yvonnet.

Serious games constitute a domain in which real-time activity recognition is particularly relevant: the expected behavior is well identified and it is possible to rely on different sensors (biometric and external) while playing the game. We focus on games to help in diagnosis and treatment of patients.

We developed a formal approach to model such activities, taking into account possible variations in human behavior. All the scenarios of an activity are not equivalent: some are typical (thus frequent) while others seldom happen. We propose to quantify the likelihood of these variations by associating probabilities with the key actions of the activity description. We rely on a formal model based on probabilistic discrete-time Markov chains (DTMCs). We used the PRISM framework and its model checking facilities to express and check interesting temporal logic properties (PCTL).

As a use case, we considered a serious game to analyze the behavior of Alzheimer patients. We encoded this game as a DTMC in PRISM and we defined several meaningful PCTL properties that are then automatically tested thanks to the PRISM model checker. Two kinds of properties may be defined: those to verify the model and those oriented toward the medical domain. The latter may give indications to a practitioner regarding a patient's behavior. These properties include the use of PRISM “rewards” to quantify the performance of patients.

We expect that such a modeling approach could provide doctors with new indications for interpreting patients' performance and we identified three medically interesting outcomes for this approach. First, to evaluate a new patient before the first diagnosis of doctors, we can compare her game performance to a reference model representing a "healthy" behavior. Second, to monitor known patients, a customized model can be created according to their first results, and, over time, their health improvement or deterioration could be monitored. Finally, to pre-select a cohort of patients, we can use a reference model to determine, in a fast way, whether a new group of patients belongs to this specific category.

This year we first addressed the model definition and its suitability to check behavioral properties of interest [24]. Indeed, this is mandatory before envisioning any clinical study.

The next step will be to validate our approach as well as to test its scalability on three other serious games selected with the help of clinicians. We wrote a medical protocol to be submitted to CERNI proposing clinical experimentations with patients. This protocol will be a collaboration with the ICP institute, member of the CoBTEX laboratory. The new games will be modeled in PRISM and different configurations (for example for Mild, Moderate or Severe Alzheimer) will be set up with the participation of clinicians. Then, several groups of patients will play these games and their results will be recorded to calibrate our initial models.