"Continuous Digital Biomarkers from Wearable Devices" - Brandon Ballinger (Co-Founder, Cardiogram)
-
Upload
hyper-wellbeing -
Category
Technology
-
view
115 -
download
0
Transcript of "Continuous Digital Biomarkers from Wearable Devices" - Brandon Ballinger (Co-Founder, Cardiogram)
N OV E M B E R 1 4 -1 6
M O U N TA I N V I E W, C A
Continuous Digital Biomarkers from Wearable Devices Brandon Ballinger, @bballinger
@AppCardiogram
2
3
TRADITIONAL BIOMARKERS CONTINUOUS DIGITAL BIOMARKERS
Biomarkers are Becoming Continuous and Digital
Necessary but not fun Engaging
Point-in-time 100,000+ data points per person per year
Clinician-interpreted Hybrid of human and artificial intelligence
Well-understood science based on Framingham (1948) and other historic studies
Emerging field at the intersection of medicine, artificial intelligence, and mobile design.
Fitbit HuaweiApple
Fossil Motorola
What’s your ♡ telling you? Examples from Cardiogram for Apple Watch
5
Fitness StressRunning
Hockey
Driving in Rush Hour
Microsoft interview
IllnessAtrial Fibrillation
Atrial Flutter
SleepRestful Sleep
Sleep after Alcohol
Heart rate biomarkers known from the medical literature
Reverse commute(SF to Palo Alto in AM)
Regular commute(SF to Palo Alto in PM)
8:00AM 8:30AM 4:30PM4:00PM
60
100
140Heart rate
60
100
140
Heart Rate Variability
11:00AM
80
130
180Heart rate
12:00PM11:30AM
Intense workout StretchingRecovery
Heart Rate Recovery & Resting Heart Rate
HEART RATE RECOVERY
RESTING HR
What can heart rate metrics tell you about yourself?
Metric Associated with (not exhaustive)
Resting heart rate (RHR) (Physical fitness, stress)
• Mortality (1.3x in pts with coronary artery disease)
• Heart attack • Mortality (1.3x women)
Heart rate recovery (HRR) (Physical fitness, vagal activity)
• Mortality (4x) • Coronary artery disease
Heart rate variability (HRV) (Stress response, vagal activity)
• PTSD • Depression • Anxiety • Sudden death after heart attack
Heart rhythm abnormality e.g. atrial fibrillation (AF)
• Mortality (1.5x men, 1.9x women) • Stroke (5x)
40
50
60
70
80
90
100
110
0 200 400 600 800 1000 1200 1400
MovementvsRestingHeartRate
40
50
60
70
80
90
100
110
0 5 10 15 20
StandingvsRestingHeartRate
40
50
60
70
80
90
100
110
0 20 40 60 80 100 120
ExercisevsRestingHeartRate
Rest
ing
Hea
rt Ra
te (b
eats
per
min
ute)
Exercise (minutes per day) Stand (hours per day) Move (calories per day)
Exercise vs Resting Heart Rate Standing vs Resting Heart Rate Movement vs Resting Heart Rate
Do Apple Watch’s activity rings drive cardiovascular improvements?
Can we use artificial intelligence to design novel biomarkers?
Semi-Supervised Sequence Learning for Continuous Digital Biomarkers
5 12 5 28 30 30dtheart rate
steps
86 0 74 0 85 83
0 5 0 97 0 0
5
83
0
…
…
…
1 0 0 -1 1atrial fibrillation hypertensionsleep apnea
0 0 -1 0 0 1
0 0 1 0 0 -1
-1
0
0
…
…
…
…
Multi-channel, multi-timescale
sensor input
Multi-task Output0
time ▶
Recurrent Neural Network
Fully connected tanh
Long Short-Term Memory
1D Convolution
Fully connected ReLu
13
…
AF Not-AF
86 79 74 82 85 83 74Raw Heart Rate Measurements
1 0
Clinical Gold Standard
Continuous Digital Biomarkers Applying deep learning to health
Atrial Fibrillation Normal Heart Rhythm
Accuracy on Cardioversions
True
Pos
itive
Rat
e
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
False Positive Rate
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
What else can you detect?
Why does this matter?
17