Workshop in Biostatistics
|DATE:||May 19, 2016|
|TIME:||1:30 - 3:00 pm|
|LOCATION:||Medical School Office Building, Rm x303|
|TITLE:||Analytics of sleep in large samples|
Sleep is core to mental health and has large medical, psychological and societal impact. We all experience sleep; yet why we sleep and how the brain generates sleep is one of the last remaining mysteries in biology. A sleep study, or nocturnal polysomonography (PSG), is comprised of multiple digital signals (EEG, EGG, EOG, chin and leg EMG, breathing, oxygen saturation, etc.) recorded simultaneously throughout the night to provide physiological measures of human activity and behavior during sleep. Millions of sleep studies are carried out yearly in the US alone to diagnose sleep disorders, most of which have unknown pathophysiology. Data analysis is currently performed using trained technicians who recognized PSG patterns to determine sleep stages and the presence of sleep apnea, insomnia or periodic leg movements.
In the last few years, engineers, sleep specialists and geneticists have teamed up in our lab to analyze large PSG data sets that are associated with sleep symptoms or diagnosed sleep disorders. PSG signals are first transformed using methods such as Fourier, Wavelet Transform, Hilbert-Huang Transform and Holo-Hilbert Spectral Analysis to provide us with more detailed and precise spectral information about the signals. Those parameters were previously analyzed using rule based processing, pattern recognition, but more recently machine learning algorithms that use hierarchical neural networks or convolution networks have been explored and provide even better results. We are also now exploring unsupervised learning algorithms.
Our machine learning algorithms are now as or more reliable than human scorers, and offer the possibility of discovering new phenotypes that have clinical correlates. For example, we are now creating a detector that can diagnose Type 1 narcolepsy (caused by the loss of hypocretin neurons) based on a simple PSG and working at detecting microarousals and autonomic arousals without EEG correlates. Correlations with clinical symptoms can be performed using the Stanford Sleep Clinic cohort, a 10,000 PSG database of patients with various sleep disorders, or/and the Wisconsin sleep cohort a 20-year longitudinal study of volunteers that have undergone PSG every four years and are deeply phenotyped for sleep, as well as psychological, metabolic and cardiovascular consequences. We are also conducting Genome Wide Association studies on these phenotypes and are planning a large cohort study that will include 40,000 subjects with GWAS, PSG, on line sleep questionnaire, actigraphy, facial picture analysis (for sleep apnea) and neuropsychological testing. As sleep is genetically controlled, genetics in combination with these algorithms is likely a powerful approach to revealing its underlying biological basis. We aim at being a catalyst for changing the sleep field.
Carmel Armon, William Nowack, Asim Roy, et al. Polysomnography (2016).
Campbell IG. EEG recording and analysis for sleep research. Curr Protoc Neurosci. 2009 Oct;Chapter 10:Unit10.2. doi: 10.1002/0471142301.ns1002s49.
LeCun Yann, Bengio Yoshua, Hinton Geoggrey. Deep learning. Nature, 2015 May 28;521(7553):436-44. doi: 10.1038/nature14539.
Warby SC, Wendt SL, Welinder P, Munk EG, Carrillo O, Sorensen HB, Jennum P, Peppard PE, Perona P, Mignot E. Sleep-spindle detection: crowdsourcing and evaluating performance of experts, non-experts and automated methods. Nat Methods. 2014 Apr;11(4):385-92. doi: 10.1038/nmeth.2855. Epub 2014 Feb 23.
Christensen JA, Carrillo O, Leary EB, Peppard PE, Young T, Sorensen HB, Jennum P, Mignot E. Sleep-stage transitions during polysomnographic recordings as diagnostic features of type 1 narcolepsy. Sleep Med. 2015 Dec;16(12):1558-66. doi: 10.1016/j.sleep.2015.06.007. Epub 2015 Jul 7.
Sehgal A, Mignot E. Genetics of sleep and sleep disorders. Cell. 2011 Jul 22;146(2):194-207. doi: 10.1016/j.cell.2011.07.004.