Statistical Methods for Wearable Technology in CNS Trials Andrew Potter, PhD Division of Biometrics 1, OB/OTS/CDER, FDA ISCTM 2018 Autumn Conference – Oct. 15, 2018 Marina del Rey, CA www.fda.gov
Disclaimer This presentation reflects the views of the author and should not be construed to represent FDA’s views or policies. 2 www.fda.gov
Outline • Data • Statistical Methods – Signal Processing – Feature Selection – Modeling of treatment effect evolution over time • Simulated case study in sleep medicine 3 www.fda.gov
Movement Data from Acceleration Sensors 8 min 0 min Gyllensten, IC, Physical Activity Recognition in Daily Life using a Triaxial Accelerometer, Master’s Thesis, 2010. 4
Converting Acceleration Sensor Data to Health Outcomes • Dense information on a person’s movement while the device is recording – At least 100 measurements per day – Days to weeks of data • What are the important features of the signal? – How does a feature relate to a disease state? – How do features change over time? – How to compare between people and groups? – How to define a drug effect? • How to identify features? – Have subject tag events in real time on the device? – Can machine learning automate this task? 5
An Important Feature: Circadian Variation in Sensor Data • Blood flow data from a ventricular assist device recorded every 15 min. • Circadian patterns present in multiple types of sensor data 6
Total Sleep Time Derived from Acceleration Sensor High device use Low device use Calendar Day 7
Weekday to Weekend Variability In Total Sleep Time Weekday mornings Weekend mornings 8
Extracting Features – Fourier Transform Circadian Cycle Feature • Focuses on periodic features in a signal – Represents the strength of a signal over a range of frequencies – Signals with circadian variation have a peak at 1 cycle/day – Spectral representation of EEG 9
Extracting Features – Smoothing Signal in Time Source: Wang et al. Journal of Circadian Rhythms 2011, 9:11 10 http://www.jcircadianrhythms.com/content/9/1/11
Feature Selection - LASSO • LASSO – least absolute shrinkage and selection operator • Extension of regression – Automatically selects covariates – Subset of all covariates most predictive of outcome – Shrinks covariate coefficients towards zero - regularization • See ‘The Elements of Statistical Learning’ by Hastie, Tibshirani, and Friedman for more details 11
Feature Selection – Neural Networks • Automatic techniques that selects a combination of features associated with a class membership – Creates features from the data – Applications - Digital Phenotypes, detection of cancer in radiology images, classification of sleep states in polysomnography Source: Nielsen, Neural Networks and Deep Learning, • Automated classification of PSG http://neuralnetworksanddeeplearning.com/chap6.html and sleep events 12
Feature Selection – Neural Networks • Multiple models using different feature of the PSG – Examples: • Supratak et al, DeepSleepNet: A Model for Automatic Sleep Stage Scoring Based on Raw Single-Channel EEG, IEEE Transactions on Neural Systems and Rehabilitation Engineering , 2017, 25 (11), pp. 1998-2008 https://ieeexplore.ieee.org/document/7961240/ • Chambon et al, A Deep Learning Architecture for Temporal Sleep Stage Classification Using Multivariate and Multimodal Time Series, IEEE Transactions on Neural Systems and Rehabilitation Engineering , 2018, 26 (4), pp. 758-769 https://ieeexplore.ieee.org/document/8307462/ • Olsen et al, Automatic, electrocardiographic-based detection of autonomic arousals and their association with cortical arousals, leg movements, and respiratory events in sleep, Sleep , Volume 41, Issue 3, 1 March 2018, zsy006, https://doi.org/10.1093/sleep/zsy006 Source: Chambon et al. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2018, 26 (4), pp. 758-769 https://ieeexplore.ieee.org/document/8307462/ 13
Modeling Features – Functional Linear Models AHI – Sleep Apnea Severity • Method for analyzing curves • Extends regression to curves • Multiple cases: – Cross-sectional observation of a single curve per patient on a single outcome measurement – Longitudinal observations of the same curve on a single outcome measurement within a patient – Cross-sectional and longitudinal of multiple curves on the same patient Source: Wang et al. Journal of Circadian Rhythms 2011, 9:11 14 http://www.jcircadianrhythms.com/content/9/1/11
Case Study - Sleep • Wearable sensors introduce two statistical challenges • Analysis of the within day data recorded densely • Analysis of the longitudinal evolution of the daily sensor • Illustrate an approach to analyzing longitudinal evolution using total sleep time (TST) as a summary measure of daily sensor data • Compare changes in TST between a new sleep medication to placebo over four weeks • Focus on modeling the linear trend in TST in both groups • Use all observed data • Calculation of TST at specific time points conducted after statistical modeling • Framework extends to multiple sleep parameters and functional models 15
Case Study - Sleep • Simulated data: • 300 patients • 30 minute improvement in TST by day 15 • Similar change in TST to several NDAs submitted to FDA • Measure treatment effect by: • Difference in TST at four weeks • Average TST trajectory in each group – focus on the linear trend • Use two statistical models • Linear mixed model with random slopes – strong assumption on covariance between days • Generalized estimating equation (GEE) model – robust to misspecification of covariance between days 16
Simulated Clinical Trial – The Data Example Subjects Subject Specific Change from Baseline in TST Triangles – Weekday Circles - Weekend 17
Population Average Total Sleep Time Trajectories 18
The Linear Mixed Model Results 95% Confidence Average TST Trajectories Estimate Interval Intercept 312.937 302.183 323.690 Day 1.339 0.844 1.834 Treatment 7.333 -7.741 22.406 Tue -1.522 -4.687 1.643 Wed 0.000 -3.175 3.174 Thurs -0.267 -3.444 2.911 Fri -0.740 -3.908 2.428 Sat 89.546 86.422 92.671 Sun 90.848 87.727 93.969 Treatment by Day 0.782 0.076 1.488 Week 4 Placebo 29.229 5.933 52.524 Subtracted Treatment Effect 19
The GEE Results 95% Confidence Average TST Trajectories Estimate Interval Intercept 314.989 304.139 325.840 Day 1.189 0.629 1.748 Treatment 1.922 -12.964 16.807 Tue -0.605 -3.576 2.366 Wed 0.397 -2.657 3.451 Thurs -0.125 -3.124 2.873 Fri -0.909 -3.835 2.017 Sat 89.420 86.623 92.217 Sun 90.874 88.044 93.704 Treatment by Day 0.781 0.009 1.552 Week 4 Placebo 23.776 0.519 47.030 Subtracted Treatment Effect 20
Final Thoughts • Rich new data source • Contains new information about neurology and psychiatry diseases • Existing statistical method provide starting point • Explore new methods to show population and individual drug effects 21
Recommend
More recommend