Convolutional Neural Networks For Modeling Temporal Biomarkers And
Disease Predictions
Narges Razavian New York University Langone Medical Center
GTC 2017
In collaboration with: David SontagPhD, Saul BleckerMD, Ann-Marie SchmidtMD, Enrico BertiniPhD, Rahul Krishnan, YD Choi, Josua Krause, Somesh Nigam, Aaron Smith-McLallen, Ravi Chawla
Deep learning progress Healthcare world getting digital
Parallel Developments
EHR adoption by healthcare centers in the US
Error rate on Image-Net object recognition challenge
What is captured in the EHR?
Source: healthcare.gov
Healthcare has joined the data-rich world
Moving from Treatment to Prevention
Challenges: Each Individual has a different ‘healthy’ baseline.
- Temporal Patterns/Trends are predictive Each biomarker varies at a different speed in our bodies Measurements are sparse, asynchronous and correlated Many correlated outcomes are observed per patient
- Can we leverage this correlation?
Biomarkers and Outcomes
Biomarkers measurements
over time
Biomarkers and Outcomes
Biomarkers measurements
over time
Phenotype (diseases) over time
Biomarkers and Outcomes
Biomarkers measurements
over time
Phenotype (diseases) over time
Biomarkers and Outcomes
Biomarkers measurements
over time
Phenotype (diseases) over time
Step 1 Learn each biomarker from other biomarkers time-series
Kernel Regression
Observations
X
(Mea
sure
men
t Ti
me-
Ser
ies)
Time Not Observed Want to estimate
Kernel Regression
Observations
X
(Mea
sure
men
t Ti
me-
Ser
ies)
Time Not Observed Want to estimate
E[X(v)]= xP(x |∫ t = v,Xtrain )dx
E[X | t = v,Xtrain ]= x∫ P(x, t = v | Xtrain)P(t = v | Xtrain)
dx
E[X | t = v,Xtrain ]= x∫K(x − xi,v− ti )
xi ,ti
∑
K(v− ti )ti
∑dx
Kernel Regression
Observations
X
(Mea
sure
men
t Ti
me-
Ser
ies)
Time Not Observed Want to estimate
E[X(v)]= xP(x |∫ t = v,Xtrain )dx
E[X | t = v,Xtrain ]= x∫ P(x, t = v | Xtrain)P(t = v | Xtrain)
dx
E[X | t = v,Xtrain ]= x∫K(x − xi,v− ti )
xi ,ti
∑
K(v− ti )ti
∑dx
E[X | t = v,Xtrain ]=(K ⊗ Xtrain )(v)
(K ⊗ I(Xtrain :Observed))(v)
Use convolution framework to LEARN those kernels
We can learn the kernel (No need for parametric forms and cross validations) Easily extendible to multivariate!
Unsupervised: All needed is (asynchronous) sequence of observations. Fast to train. Fast to apply.
Benefits
Data:30KIndividualsfromtheoriginaltrainingset.Datasetsplitequallybetweentrain,testandvalidateset.Loss:MSE.Trainandevaluateonlyon(lab,person)withmorethan1observaGon.
Mul$variateKernelslearnedforeachinputdimension(total18)
More details in our ICLR paper
Narges Razavian, David Sontag Temporal Convolutional Neural Networks for Diagnosis from Lab Tests http://arxiv.org/abs/1511.07938 Open Source code available (torch/lua implementation): https://github.com/clinicalml/deepDiagnosis
Step 2 Predict 200+ correlated outcomes using multi-resolution convolutional neural networks and multi-task learning
Multi-Resolution Convolution Networks The Architecture - model (1)
Multi-Resolution Convolution Networks The Architecture - model (2)
Prediction AUCs on the held-out test set
More details in our JMLR paper
Narges Razavian, Jake Marcus, David Sontag, Multi-task Prediction of Disease Onsets from Longitudinal Lab Tests JMLR, 2016 http://arxiv.org/abs/1608.00647 Open Source code available (torch/lua implementation): https://github.com/clinicalml/deepDiagnosis
Following up in clinical world • Prediction models built and deployed for
– Nurse calls and home visits for 250,000+ NYUMC patients at high risk for a number of these outcomes
– Improved documentation in EHR • Automation of mandatory visits/screening/follow-ups • Best practice alerts • Reimbursement for intense lifestyle management programs
• Extending to broader outcomes and domains
New York University (i2b2) Database
New York University (i2b2) Database Nuclear Medicine Procedures
Magnetic Resonance Imaging
Conclusions • Applications of deep learning in healthcare are unlimited
• Unsupervised learning + back-propagation + deep learning can recover biomarker models from asynchronous high-dimensional time-series data
• Multi-task learning benefits prediction tasks with smaller datasets.
Thanks!
Questions/comments: [email protected]
Open Source Package: https://github.com/clinicalml/deepDiagnosis