چگونگی آنالیز کلان داده ها توسط آمارگیران How do statisticians analyse big data—Our story

1. Introduction A big thanks to the editors for editing this special issue to discuss the role of statistics in the era of big data. Instead of ‘the role of statistics’, a common concern for statisticians, when they are faced with large data sets of many gigabytes or even terabytes, is on ‘where to start’. This was exactly the first question we (a team of three statisticians) asked ourselves when we started our joint project Limbs alive – monitoring of upper limb rehabilitation and recovery after stroke through gaming with two other teams (one in neuroscience and the other in computing science) in 2012. Given the issues that we faced when first starting out, we would like to share our story with readers who may have less experience on analysing big data, and hopefully to shed some light on the process. The overall aim of the project was to develop a home-based rehabilitation system using action video games for stroke patients, and the main task of our statistical team was to develop an automatic system to assess upper limb function. Assessing upper limb functions is a difficult task. The current commonly used clinical measure is the CAHAI (Barreca et al., 2005). A patient is asked to complete 9 tasks, e.g. ‘open a jar’ and ‘dial 999’. A therapist gives a score for each task ranging from 1 to 7 based on how well the patient completed the task. The CAHAI is the sum of the 9 scores. Although it is a validated measure, it is certainly expensive (each assessment lasts from 20 to 30 min) and subjective. We designed an assessment game including 38 simple movements (see the details in Serradilla et al., 2014), for example a forward roll. We recorded the 3D position data and 4D directional data of the trajectory for each movement using three wireless controllers, one in each hand and the other attached to the middle of the body; see Fig. 1. The size of the data is quite big, up to several dozen gigabytes. The aim of the system is to use the data to calculate the level of impairment of upper limbs after stroke. This is a typical modelling problem, but many obstacles have to be overcome before we can consider any models. Our experience shows that, to achieve the target, we need to treat it as a comprehensive scientific project rather than a statistical project. This may be the attitude we should have in analysing any big data problems.