Microprocessor prostheses all have embedded sensors. The data recorded by these sensors could be used in classification algorithms for real life activity recognition. Activity recognition can be used for the adaptation and the evaluation of the rehabilitation by assessing the performance progression of the patient. It can also serve for a better prosthesis adaptation to the terrain. Recently, a knee ankle prosthesis system has been described which adapts to slopes and stairs situations. These prosthesis sensors were used to setup an activity recognition algorithm.

A single subject, equipped with the knee ankle system, walked continuously in four situations (stair descent, slope ascent and descent, and level walking) for four to five minutes. Sensors data were recorded and synchronized with a video. This protocol was repeated on three different days. Sensors signals were segmented according to the gait cycles and labelled according to the video as one of the four situations. One hundred and forty-nine features were extracted from extrema, thresholds and ranges of the signal sensors and used to feed a machine-learning algorithm. A decision tree algorithm was chosen for its simplicity and understandability. Data from the two first days were used to train the algorithm and set the tree complexity. The last day was used to evaluate the tree error with a confusion matrix, the specificity and the sensitivity of each situation versus all the others.

The optimal tree complexity led to four final branches, which means that only three different features had been used for the activity recognition. A 98% mean specificity and sensitivity across all situations was found.

The decision tree has the advantage to give an intelligible final model that can be explained to a patient. But decision tree are known to be prone to overfitting. This overfitting can be avoided by a precise methodology followed in this study. The results of this study are encouraging with a very low error rate. The next step of this work is to train a generic decision tree on several patients and to implement it on the knee ankle prosthesis to give an activity report to the user or the practitioner.