DRBN Models

Dynamic Relational Bayesian Inference

Dynamic Relational Bayesian Network (DRBN) models can infer important actions, behaviors, conditions and states from our embedded assessment sensor data, which has been obtained without resorting to the use of intrusive technologies like video cameras with face recognition software or the compulsory wearing of ID beacons on one’s body or tags in the clothing.

Relational Bayesian networks have been cited as one of the “10 emerging technologies that will change your world” (MIT Technology Review 2004). They are the result of the combination of relational databases and machine learning technologies. They allow for the presentation of relational information (for example, how people relate to sensors in a home setting at different times of day), at multiple scales and from multiple points of view.

Dynamic Relational Bayesian Network = schema + data + model

Examples:

  1. Identify the individual being monitored in a multi-person household, by inferring their location from a combination of room motion sensors and sedentary pressure sensors embedded in key furniture, including the bed and favorite chairs. These sensors will also provide data that may be used to draw inferences from changes in activity baselines, based on nighttime behaviors (e.g., number of trips to the bathroom per night, ratio of restful to fitful sleep time and hours slept per night) and detect wandering and sundowning behaviors, to inform caregivers and track behavioral changes over time.
  2. Discover what sense-able activities are useful determinants of a person’s state of wellness, in terms of the performance of ADLs, by tracking access to food, drink and medications and inferring changes in health status based on longitudinal analyses of sleeping patterns, pacing behaviors and usage frequencies of the kitchen and bathroom facilities.
  3. Infer whether a person is at home (and possibly incapacitated) or away, without burdening them to remember to log in and out, as required by home security systems. Occupancy status will be inferred from the whole-house motion sensor network, combined with direction-sensing differential motion sensors and door switches, to determine exit and entrance events. Conditions to be inferred include: whether a person has exited and not returned within a specified interval, whether the person is at home but has not gotten out of bed or emerged from the bathroom within a specified time interval.
  4. Develop intelligent reminding methods to close the feedback loop between automatic prompting and task fulfillment. It is easy to generate reminders, but difficult to remind well. Typical prompting systems often provide inappropriate reminders – whether they are required or not – adding to people’s frustrations. Context-aware methods can interpose intelligent reminders to support task completion within the supportive environment of one’s own smart home, protecting a person’s remaining competencies and safety by allowing them to do as much as possible for themselves before intervening.
  5. Identify and explicate multiple stakeholders’ information needs with respect to the data available from a sensored home and generate a model and probabilistic queries that capture the range of possible distinctly personalized content that can be presented to these stakeholders, including the resident, family, informal and formal caregivers and healthcare professionals.

Intelligent reminding leads to a more challenging set of assisted cognition applications that seek to infer the person’s intent and provide assistance with daily plan management, based on implicit sensor and explicit plan inputs. For example, medication compliance may be based on a set of implicit inputs including: the person’s current state, the time of day and the time when they last accessed their medicine cabinet. By using implicit inputs, context-aware applications will give people with cognitive challenge, optimized degrees of support, while minimizing their cognitive loads and their caregiver’s burdens.

The CleverSet Dynamic Relational Bayesian Modeling technology will allow many living behavior variables to be modeled simultaneously while discovering models from the data in real and near-real time.  DRBNs are an instance of a new family of approaches emerging in the machine learning community based on a relational data model. DRBNs are well-suited for this modeling activity because they can directly model relationships among people, places, objects, activities, and sensed phenomena. DRBN algorithms exploit the data model and meta-data from the schema to guide and frame relational queries about behavior and events. The DRBNs formulated can represent complex, dynamic, multi-scale processes involving multiple actors, as probability distributions over the elements, queries, and relationships in the DRBN model.