Probabilistic Scoring of Validated Insights for Personal Health Services Aki H¨ arm¨ a and Rim Helaoui Philips Research Eindhoven, The Netherlands Abstract—In connected health services automatic discovery of recurring patterns and correlations, or insights, provides many interesting opportunities for the personalization of the services. In this paper the focus is on insight mining for a health coaching service. The basic idea in the proposed method is to generate a large number of insight candidates which have been pre-validated with domain experts and to score them using the data. The dynamic performance of the scoring is studied with a collection of lifestyle sensor data from volunteers. The proposed method is compared to a conventional data mining approach based on the Apriori algorithm. We demonstrate that the proposed method gives significantly more variability among the subjects and types of insights it finds which may reflect better the underlying statistics of individual lifestyle patterns of the different subjects. I. I NTRODUCTION Personal health coaching services typically focus on guiding the user to adopt a healthier lifestyle, for example, by being physically more active, sleeping and eating better. In the case of a conventional human health coach the opportunities for change are identified in a dialogue between the coach and the coachee. In an automated coaching machine based on a web service and an app, for example, it would be necessary to find those opportunities, or insights, automatically from the data [1]. Automatic generation of insights is a central topic in data mining literature. Conventional association mining is based on algorithms that find co-occurrences of sets of discrete data items [2], for example, particular books or food items in a marketing application, office behavioral data [3] or health data in clinical databases [4], [5]. In the case of health sensors with continuous data values, such associations cannot be uniquely defined but require a discretized and probabilistic framework for the description of insights [6], [7]. Let us call the proposed method the Probabilistic Scoring of Validated Insights, PSVI. The insights are found by computing probabilistic confidence scores for a large number of insight candidates which have been pre-validated by domain experts in the design phase. The pre-validation is necessary in a health coaching application to exclude potentially harmful insights. For example, the data may suggest that the user has a lower blood pressure on days when the user has slept less in the previous night. This insight may be interpreted by the user as an advice to sleep less while it most likely only refers to a correlation and not to a causal relation. The proposed method can be seen as a modification of conventional association mining algorithms such as Apriori [8] or CHARM [9] but it has also interesting similarities with various machine learning algorithms. The proposed method also resembles recommendation systems [10], [11] but the problem is different and the same methodology is generally not applicable here. In Sections II-III we give an overview of the PSVI algo- rithm and a use case in health programs. In Section IV the performance of the method is then studied using a collection of lifestyle sensor data from a group of volunteers, and the final results are discussed in Section V. II. PSVI ALGORITHM Conventional association rule mining is based on counting co-occurrences of discrete items {I k ,I j } [12], [2]. In case that the items are continuous measurements such as a step count and heart rate the algorithm can also be applied after discretization [7]. This is often performed by dividing the continuous measurement range into a small number of bins and using for example fuzzy membership functions to describe the associations [13]. The presented PSVI method uses a relative discretization where one measurement in a context is either smaller or larger than another measurement. In the Apriori algorithm the discovery of insights would be then based on occurrences of the these cases. This basic method is developed and tested further in this document and it is shown that it is not necessarily efficient for dynamic selection of interesting insights. The interestingness score of an associative rule can be characterized in many different ways, see, e.g., [14]. Due to the probabilistic nature of the insights discussed in this paper let us call this score a confidence value of an insight. For the purposes of this paper a high confidence should be related to a detection of an opportunity, which is typically a context or condition that somehow stands out from the data. For example, an insight may state that “a user walks less on Mondays than on Tuesdays”. The confidence value of this statement should be based on (1) the statistics of walking on those weekdays, e.g., based on data from an activity bracelet, and (2) the observation that it differs from some other context. In an insight mining application we might be interested in finding the highest scoring insight out of a collection of statements of the following form “in context a you walk less than in context b”, where a and b could be, for example, two