Introduction. In orthopaedics, clinical outcome assessment (COA) is still mostly performed by questionnaires which suffer from subjectivity, a ceiling effect and pain dominance. Real life activity monitoring (AM) holds the promise to become the new standard in COA with small light weight and easy to use accelerometers. More and more activities can be identified by algorithms based on accelerometry. The identification of stair climbing for instance is important to assess the participation of patients in normal life after an orthopaedic procedure. In this study we validated a custom made algorithm to distinguish normal gait, ascending and descending stairs on a step by step basis. Methods. A small, lightweight 3D-accelerometer taped to the lateral side of the affected (patients) or non-dominant (healthy subjects) upper leg served as the activity monitor. 13 Subjects (9 patients, 4 healthy) walked a few steps before descending a flight stairs (20 steps with a 180o turn in the middle), walked some steps more, turned around and ascended the same stairs. Templates (up, down and level) were obtained by averaging and stretching the vertical acceleration in the 4 healthy subjects. Classification parameters (low pass (0.4 Hz) horizontal (front-back) acceleration and the Euclidian distance between the vertical acceleration and each template) were obtained for each step. Accuracy is given by the percentage of correctly classified steps. Results. In total the subjects took 537 (41+/-8 mean+/-std) steps, 525 of which were correctly identified as step. 12 Steps were not detected, and 2 steps were incorrectly identified as step. Per subject the accuracy of the classification algorithm ranged from 57% to 97%. In only 2 subjects the accuracy was less than 75%, giving an overall accuracy of 85%. Discussion. In literature algorithms able to identify walking the stairs and normal walking have been reported with an accuracy in the range of 80–95%1,2. Our algorithm falls well within this range, and can be even further improved. The low accuracy in two subjects can be explained by the fact that the sensor was placed more to the front of the leg, which influences the low-pass horizontal acceleration. Using a combination of front-back and left-right acceleration could possibly solve this problem. In the future we are confident to identify also other activities and even distinguish different types of stair climbing (i.e. taking a step with each leg versus only taking steps with the unaffected leg and ‘dragging’ the second leg) and obtain more specific activity profiles to be used in clinical outcome assessment

Introduction. In orthopaedics, clinical outcome assessment (COA) is mostly performed by questionnaires which suffer from subjectivity, a ceiling effect and pain dominance. Real life activity monitoring (AM) can objectively assess function and becomes now feasible as technology has become smaller, lighter, cheaper and easier to use. In this study we validated a custom made algorithm based on accelerometry using different orthopaedic patients with the aim to use AM in orthopaedic COA. Methods. A small, lightweight 3D-accelerometer taped to the lateral side of the affected upper leg served as the activity monitor. AM algorithms were programmed in Matlab to classify standing, sitting, and walking. For validation a common protocol was used; subjects were asked to perform several tasks for 5 or 10 seconds in a fixed order. An observer noted the starting time of each task using a stopwatch. Accuracy was calculated for the number of bouts per activity as well as total time per activity. 10 Subjects were chosen with different pathologies (e.g. post total knee/hip arthroplasty, osteoarthritis) since the difference in movement dynamics in each pathology poses a challenge to the algorithm. Results. In total the subjects performed 267 activities (99 standing, 80 sitting, 88 walking), 258 of which (99, 73, 87 resp.) were classified correctly by the algorithm, corresponding to a sensitivity of 97%. Sensor misplacement in 1 subject caused all missed instances in sitting, and exclusion of this subject increased sensitivity to 99.9%. 5 Instances of standing were incorrectly added by the algorithm, giving a specificity of 95% for standing. In total 80 sit-stand, and 78 stand-sit transitions were performed. Subjects were standing for 792 seconds, sitting for 764 s, and walking for 905 s. The algorithm found a total duration of 739, 583 and 1056 seconds for those activities respectively, and 83 seconds of lying (misclassification of sitting). Discussion. Sensor placement is an important factor to obtain reliable results. Even so sensitivity and specificity are comparable to values found in literature [85–99%]. The added instances of standing occurred when a subject did not immediately sit after a period of walking. It is doubtful if these instances should be considered false positives. The main difference in duration is also found in sitting, which is caused by the missed instances previously described, in combination with the fact that the duration of transitions are added to the walking period in the algorithm, whereas it is divided over sitting and walking by the observer. This corresponds to a difference of less than one second per transition. The algorithm produces reliable results when challenged with different movement patterns common with orthpaedic pathologies. The device may be used as as AM in objective assessment of clinical outcome after orthopaedic procedures