A balanced sleep/wakefulness classification method based on actigraphic data in adolescents

Several research groups have developed automated sleep-wakefulness classifiers for night wrist actigraphic (ACT) data. These classifiers tend to be unbalanced, with a tendency to overestimate the detection of sleep, at the expense of poorer detection of wakefulness. The reason for this is that the measure of success in previous works was the maximization of the overall accuracy, disregarding the balance between sensitivity and specificity. The databases were usually sleep recordings, hence the over-representation of sleep samples. In this work an Artificial Neural Network (ANN), sleep-wakefulness classifier is presented. ACT data was collected every minute. An 11-min moving window was used as observing frame for data analysis, as applied in previous sleep ACT studies. However, our feature set adds new variables such as the time of the day, the median and the median absolute deviation. Sleep and Wakefulness data were balanced to improve the system training. A comparison with previous studies can still be done, by choosing the point in the ROC curve associated with the corresponding data balance. Our results are compared with a polysomnogram-based hypnogram as golden standard, rendering an accuracy of 92.8%, a sensitivity of 97.6% and a specificity of 73.4%. Geometric mean between sensitivity and specificity is 84.9%.