Based on the development of an in-depth understanding of the four use case settings (medical goals, potential data sets to be obtained, level of acceptance of sensing and monitoring, etc.) in Reporting Period 1, in Reporting Period 2, first, we specified and detailed in a systematic manner the types of analytics and ML techniques to be used in each Touchpoint. Second, we started to implement, demonstrate and verify a supervised ML approach in Touchpoint 2 and an unsupervised ML approach in Touchpoint 3. In this context, we realized that, in each use case setting, we need to closely co-adapt the data analytics goal, the algorithms to be used, and the structure of the sensing system (sensed parameters, types of sensors, data structures, etc.), and demonstrate our ability to do so in particular in the context of the implementation of a Human Activity Recognition chain in Touchpoint 2. Third, in Touchpoint 4 we experimented with playful physical tests administered early in a naturalistic environment by the Playware tiles to supplement and replace traditional standardized performance tests such as Timed Up & Go (TUG), Chair Stand (CS) and Four Square Step Test (FSST) in order to detect critical trends of deterioration. Fourth, in order to cope practically with the relatively small datasets that we deal with in REACH, we developed and applied transfer learning techniques. In Reporting Period 3, we will extend demonstration and verification to Touchpoints 1 and 4, and transfer the methods and lessons learned.
Types of analytics and ML used in the Touchpoints
In line with the medical/geriatric goals specified per TP, we can specify the business understanding (in a data analytics sense and as per the CRISP-DM model) for the analytics task per TP as follows:
- Touchpoint 1: development of machine learning models to modulate and optimize the effects of preventative ADL training for older persons.
- Touchpoint 2: development of machine learning models to conduct Human Activity Recognition to a) increase motivation and elderly/patient engagement, b) optimize rehabilitative training and measures provided, c) optimize workflows and save cost.
- Touchpoint 3: development of machine learning models to a) predict future developments of activity levels of older persons, and b) predict if an intervention will work or not to personalize interventions and guide engagement and behavior change regimes.
- Touchpoint 4: development of machine learning models to a) accurately count steps for elderly users (typically having irregular gait) and cluster users into groups of “similar” walkers; and b) allow for an early on-device (Playware tiles) functional performance assessment of older persons.
Based on this business understanding (in a data analytics sense) and on the work conducted, we summarise and interpret the analytics types used and their complexity per Touchpoint and review/evaluate the status and ambition of the implementation of sensing-data-collection-machine learning chains per Touchpoint. Gartner categorized and defined four types of analytics (descriptive, diagnostic, predictive and prescriptive analytics) by plotting business value against the data analytics sophistication levels. We situate the REACH Touchpoints according to the nature of the analytic models used in them in Garter’s categorization graph.