SUMMARYActigraphy can assist in the detection of periodic limb movements in sleep. Although several actigraphs have been previously reported to accurately detect periodic limb movements, many are no longer available; of the existing actigraphs, most sample too infrequently to accurately detect periodic limb movements. The purpose of this study was to use advanced signal analysis to validate a readily available actigraph that has the capability of sampling at relatively high frequencies. We simultaneously recorded polysomnography and bilateral ankle actigraphy in 96 consecutive patients presenting to our sleep laboratory. After pre-processing and conditioning, the bilateral ankle actigraphy signals were then analysed for 14 simple time, frequency and morphology-based features. These features reduced the signal dimensionality and aided in better representation of the periodic limb movement activity in the actigraph signals. These features were then processed by a Na€ ıve-Bayes binary classifier for distinguishing between normal and abnormal periodic limb movement indices. We trained the Na€ ıve-Bayes classifier using a training set, and subsequently tested its classification accuracy using a testing set. From our experiments, using a periodic limb movement index cut-off of 5, we found that the Na€ ıve-Bayes classifier had a correct classification rate of 78.9%, with a sensitivity of 80.3% and a specificity of 73.7%. The algorithm developed in this study has the potential of facilitating identification of periodic limb movements across a wide spectrum of patient populations via the use of bilateral ankle actigraphy.

No abstract

Study Objectives We propose a unique device-independent approach to analyze long-term actigraphy signals that can accurately quantify the severity of periodic limb movements in sleep (PLMS). Methods We analyzed 6–8 hr of bilateral ankle actigraphy data for 166 consecutively consenting patients who simultaneously underwent routine clinical polysomnography. Using the proposed algorithm, we extracted 14 time and frequency features to identify PLMS. These features were then used to train a Naïve–Bayes learning tool which permitted classification of mild vs. severe PLMS (i.e. periodic limb movements [PLM] index less than vs. greater than 15 per hr), as well as classification for four PLM severities (i.e. PLM index < 15, between 15 and 29.9, between 30 and 49.9, and ≥50 movements per hour). Results Using the proposed signal analysis technique, coupled with a leave-one-out cross-validation method, we obtained a classification accuracy of 89.6%, a sensitivity of 87.9%, and a specificity of 94.1% when classifying a PLM index less than vs. greater than 15 per hr. For the multiclass classification for the four PLM severities, we obtained a classification accuracy of 85.8%, with a sensitivity of 97.6%, and a specificity of 84.8%. Conclusions Our approach to analyzing long-term actigraphy data provides a method that can be used as a screening tool to detect PLMS using actigraphy devices from various manufacturers and will facilitate detection of PLMS in an ambulatory setting.