Automatic sleep scoring in normals and in individuals with neurodegenerative disorders according to new international sleep scoring criteria

Introduction: Reliable polysomnographic classification is the basis for evaluation of sleep disorders in neurological diseases. Aim: To develop a fully automatic sleep scoring algorithm on the basis of a reproduction of new international sleep scoring criteria from the American Academy of Sleep Medicine (AASM). Methods: A biomedical signal processing algorithm was developed, allowing for automatic sleep depth quantification of routine polysomnographic (PSG) recordings through feature extraction, supervised probabilistic Bayesian classification, and heuristic rule-based smoothing. The performance of the algorithm was tested using 28 manually classified day-night PSGs from 18 normal subjects and 10 patients with Parkinson's disease (PD) or multiple system atrophy (MSA). This led to quantification of automaticversus- manual epoch-by-epoch agreement rates for both normal and abnormal recordings. Results: Resulting average agreement rates were 87.7% (Cohen’s Kappa: 0.79) and 68.2%(Cohen’s Kappa: 0.26) in the normal and abnormal group, respectively. Based on an observed reliability of the manual scorer of 92.5% (Cohen’s Kappa: 0.87) in the normal group and 85.3% (Cohen’s Kappa: 0.73) in the abnormal group. Conclusion: The developed algorithm was capable of scoring normal sleep with an accuracy around the manual inter-scorer reliability, it failed in accurately scoring abnormal sleep as encountered for the PD/MSA patients, which is due to the abnormal micro- and macrostructure pattern in these patients.