Wrist actigraphic scoring for sleep laboratory patients: algorithm development

Abstract

Wrist actigraphy is employed increasingly in sleep research and clinical sleep medicine. Critical evaluation of the performance of new actigraphs and software is needed. Actigraphic sleep-wake estimation was compared with polysomnographic (PSG) scoring as the standard in a clinical sleep laboratory. A convenience sample of 116 patients undergoing clinical sleep recordings volunteered to participate. Actiwatch-L recordings were obtained from 98 participants, along with 18 recordings using the newer Spectrum model (Philips Electronics), but some of the actigraphic recordings could not be adequately aligned with the simultaneous PSGs. Of satisfactory alignments, 40 Actiwatch recordings were used as a training set to empirically develop a new Scripps Clinic algorithm for sleep-wake scoring. The Scripps Clinic algorithm was then prospectively evaluated in 39 Actiwatch recordings and 16 Spectrum recordings, producing epoch-by-epoch sleep-wake agreements of 85-87% and kappa statistics averaging 0.52 (indicating moderate agreement). Wake was underestimated by the scoring algorithm. The correlations of PSG versus actigraphic wake percentage estimates were r = 0.6690 for the Actiwatch and r = 0.2197 for the Spectrum. In general, using a different weighting of activity counts from previous and subsequent epochs, the Scripps Clinic algorithm discriminated sleep-wake more successfully than the manufacturer's Actiware algorithms. Neither algorithm had fully satisfactory agreement with PSG. Further evaluations of algorithms for these actigraphs are needed, along with controlled comparisons of different actigraphic designs and software.