Slide 1: Real-time Implementation of Sleep Staging using Interbeat Intervals
Gary Garcia-Molina and Jiewei Jiang (US) Real-time Implementation of Sleep Staging using Interbeat Intervals
Slide 2: Conflict of Interest Disclosure
CONFLICT OF INTEREST DISCLOSURE With respect to this CME activity, No, I (nor my spouse/partner) do not have a relevant financial relationship. X Yes, I (and/or my spouse/partner) do have a relevant financial relationship. Nature of Relevant Financial Relationship Name(s) of Company or Companies (ch
Slide 3: The activity of the autonomic nervous system changes according to behavioral state
The activity of the autonomic nervous system changes according to behavioral state AMY, amygdala; BaroR, baroreflex; BS, brainstem centers; ChemoR, chemoreflex; INS, insula; MCC, midcingulate cortex; REM, rapid eye movement; Resp, respiration. Modified image from: Chouchou F and Desseilles M. Front.
Slide 4: Sleep staging algorithms can be categorized into 2 groups
Sleep staging algorithms can be categorized into 2 groups Examples of cardiac signals Feature extraction Classic sleep Representation staging learning feature extraction features are and classification automatically discovereda Classifier • Deep • Light • REM • Wake BCG, ballistocardiography; ECG, e
Slide 5: Study aim
Study aim • To develop a sleep staging representation-type of algorithm (deep neural network) that: o utilizes a variety of cardiac signals including ECG, PPG, BCG o can operate in real-time with sufficient accuracy o has a small memory footprint that enables micro-processor embedding o has low comp
Slide 6: The ECG from 3 sleep data sets were used in this research
The ECG from 3 sleep data sets (D1, D2, D3) were used in this research Dataset Number of PSG recordings from unique subjects Diagnosis D1 994 Healthy and apnea ▪ ~50% for training ▪ ~20% for validation 3× cross-validation ▪ ~30% for testing D2-healthy 16 Healthy D2-INS 9 Insomnia D2 D2-RBD 22 REM be
Slide 7: Signal processing and deep neural architecture
Signal processing and deep neural architecture Convolution (filter) group 208 parameters D, deep; ECG, electrocardiogram; HP, high pass; IBI, interbeat interval; L, light; LSTM, long short-term memory; R/REM, rapid eye movement; W, wake.
Slide 8: Results show moderate agreement with manual scoring for the healthy datasets
Results show moderate agreement with manual scoring for the healthy datasets Four-sleep-stage accuracy Mean (SD) D1-test D2-healthy D2-INS D2-RBD D2-PLMD D3 Kappa 0.39 (0.16) 0.31 (0.08) 0.39 (0.13) 0.24 (0.09) 0.36 (0.12) 0.44 (0.09) Accuracy 0.62 (0.13) 0.52 (0.06) 0.58 (0.12) 0.47 (0.06) 0.57 (0.
Slide 9: Model size, latency, and accuracy for comparable approaches
Model size, latency, and accuracy for comparable approaches Model size Decision Latency Approach Kappa Accuracy Type of model (# of parameters) (time) Sridhar et al 20201 0.66 0.77 1.5 M DNN > 6 hours Radha et al 20192 0.61 0.77 260 K DNN > 6 hours Wei et al 20193 0.55 0.78 10 K Decision tree 5 minu
Slide 10: Qualitative analysis of the DNN's spectral response
Qualitative analysis of the DNN’s spectral response • Four clusters could be distinguished, each with 9 components • The cluster profiles show band-pass type of spectral responses • The peak at 0.25 Hz that is most prominently visible for clusters 2 and 4 reflects typical respiration frequency (15 b
Slide 11: Conclusions
Conclusions • Real-time interventions that compensate for sleep deficiencies1 or boost beneficial aspects of sleep2 require real-time sleep staging, preferably based on minimally obtrusive sensing • We developed an algorithm that is minimally obtrusive, relying on IBI signals that can be extracted t
Slide 12: Disclosures and acknowledgments
Disclosures and acknowledgments • This research was supported by Sleep Number Corporation • Gary Garcia-Molina, PhD and Jiewei Jiang are employees of Sleep Number Corporation, Minneapolis, MN, USA • Shawn Barr, Cristina Jocson, Anna Pastushko, and Farzad Siyahjani provided insightful recommendations
Slide 13: Supplemental Slides
Supplemental Slides
Slide 14: Overview of signal processing and DNN architecture
Overview of signal processing and DNN architecture • The training and validation subsets of D1 were used to train the DNN • Three-fold cross-validation was used • DNN output approximated the probabilities of a window belonging to light, deep, REM, or wake stages • Cohen’s Kappa, accuracy, and sensit
Slide 15: DNN training and validation
DNN training and validation • Our model included 6408 parameters • Training accuracy increased with the DNN- epoch, while the validation accuracy fluctuated more than the training accuracy • Both datasets showed an increasing trend in accuracy overall DNN, deep neural network.
Slide 16: Kappa optimization and correction of light sleep bias
Kappa optimization and correction of light sleep bias • Given that light sleep is significantly more prevalent than other sleep stages, the DNN can be biased to detect light sleep • To correct the bias, we analyzed the cumulative distribution functions of the probability ratio of light sleep versus
Slide 17: Sleep staging accuracy for all datasets using probability thresholds
Sleep staging accuracy for all datasets using probability thresholds Stage Sensitivity Specificity D2- D2- D1-test D2-INS D2-RBD D2-PLMD D3 D1-test D2-INS D2-RBD D2-PLMD D3 healthy healthy 0.37 0.57 0.51 0.25 0.22 0.46 0.96 0.76 0.93 0.86 0.89 0.94 Deep (0.33) (0.23) (0.36) (0.28)* (0.22) (0.33) (0.
Slide 18: The confusion matrices show a diagonal dominance
The confusion matrices show a diagonal dominance, with lower Kappa values associated with the D2 datasets Acc, accuracy; INS, insomnia; PLMD, periodic limb movement disorder; RBD, REM behavior disorder; REM, rapid eye movement sleep.