Deep learning for scoring sleep based on cardiorespiratory signals as compared to auto and multiple manual sleep scorings based on neurological signals

Abstract

Background: Typically, no neurological signals are recorded in home sleep apnea testing (HSAT) and thus standard sleep scoring is not applicable. Previous attempts to estimate sleep based on cardiorespiratory signals achieved Cohen’s kappa coefficients up to 0.50. Aims and objectives: To evaluate if artificial intelligence approaches can improve automatic sleep scoring performance. Methods: Supervised deep learning for scoring sleep was trained with 472 and tested in 116 PSGs, scored independently by 2 experts and by a consensus scorer. Training was performed separately for neurological and cardiorespiratory features. The resulting recurrent neural networks (RNN) were integrated in the Somnolyzer system and validated in 97 PSGs of OSAS patients scored independently by 4 human experts. Epoch-by-epoch Cohen’s kappa agreement for 4 stages (W, L: N1+N2, D: N3, R) was determined as compared to a consensus scoring without the assessed scorer included in the consensus. Results: Cohen’s kappa for the 4 manual expert scorings were 0.80 (W:0.83, L:0.76, D:0.58, R: 0.91), 0.81 (W:0.85, L:0.77, D:0.51, R: 0.91), 0.75 (W:0.85, L:0.70, D:0.43, R: 0.89), and 0.84 (W:0.88, L:0.81, D:0.63, R: 0.91), for neurological autoscoring 0.86 (W:0.90, L:0.83, D:0.71, R: 0.95) and for cardiorespiratory autoscoring 0.65 (W:0.69, L:0.59, D:0.54, R: 0.78). Conclusions: Autoscoring based on neurological signals outperforms all human expert scorers. With a kappa of 0.65, the cardiorespiratory-based RNN classifier is far above previously published values and reflects a substantial agreement with the manual consensus scoring in patients with sleep-disordered breathing.