how can data science revolutionize healthcare? nips 2016 notes
TRANSCRIPT
How can Data Science revolutionize healthcare? NIPS 2016 Notes
Tetiana Kodliuk, Data Scientist
In 2016, University of Tokyo’s
Institute of Medical Science
reported about the first case,
when Artificial Intelligence
saved the person’s life.
● continuous scientific research● powerful computing resources● government programs for health care
Preconditions
Carlson - is the first case of using drone in pedagogy
Trends in 2017 by Forbes
Artificial Intelligence
Big Data
Machine Learning
Agenda
Example of ML algorithms for
problems solution
AI use cases in
Medicine
Main directions in ML for Health
What do you expect from AI in Medicine?
Main directions
I. Personal medicine
Medical Information Mart for Intensive Care, Leo Seli
● Records analysis for millions of patients
● Compute disease risk from clinical report history
● Recommend treatment options● Recommend multiple drugs
based on drug interactions
Main directions
II. Decision support systems
Medical Information Mart for Intensive Care, Leo Seli
● Mouse-and-click for clinicians● 1-minute summary of patient
pathology● Continuously learn from
similar cases● Up-to-date medical
recommendations
Main directions
III. Remote control
Medical Information Mart for Intensive Care, Leo Seli
● Heart attack prediction● Diet control● Doctor or clinic
recommendations● Drug-interactions analysis● Queues avoiding
Key challenges for Healthcare system
1. Better care at lower cost- e. g. Fee-for-service -> value based
payment
2. Precision medicine- Reducing “number needed to treat”
3. Person-centred care- Treating people not organs
Fundamental theorem
>But not:
>
Use cases in Medicine
How can Data Science revolutionize his life?
Digital human
Electronic clinical records
1. Medical records analysis
Information extraction from unstructured data
Lab tests
Treatment
Symptoms
Diagnosis
Natural language processing
- Name- Time- Measurement- Diagnosis- Treatment- Drug- Analysis
Neural Networks
Entities recognitionPerson Treatment
Diagnosis
I know everything about you!
HEALTH DATA
Clinical data:● Structured● Unstructured● Medical Images
Behavioral data:● Social Network data● Mobile sensor data● Demographics
BioMed data:● Genomic● Proteomic● Drug responses
2. General data analysis
Heart rhythms from tracker sensors
What can Social networks tell us?
Convolutional neural network for images processing
What will happen to me tomorrow?
Health predicting
Goal: predict a person’s health, stress and happiness tomorrow
● self-reported health is strongly related to all-cause mortality
● stress increases susceptibility to infection and illness
● happiness is so strongly associated with greater longevity that the effect size is as cigarette smoking
● mood has historically been a difficult task, with typical classification accuracies ranging from 55-80%
Multi-task Learning for Predicting Health, Stress, and Happiness, Natasha Jaques
Daily features
Location Physiology: EDA Accelerometer Skin temperature
Behavioral surveys Weather Smartphone logs
What is good for one person is bad for another
Deep Learning
Happiness Stress Health
Recognize me by the image!
3. Clinical Images analysis
● Medical Images and Video parsing● Pathology detection from biopsy slides● Pathology detection from ultrasound imagery● Brain diseases detection from MRT● Pathology detection from endoscopic images
Mammography
● Mammography is an “obvious” application for Deep learning
● Annotated datasets are small
● Augmentation is necessary
Augmentation analysis
Augmentation methods
Gaussian blur
Rotate
Shear
Scale
Flips
Jitter
Powers
Gaussian noise
Original dataset Augmentation Training Evaluation
Accuracy
Blur 0.885
Rotate 0.881
Shear 0.875
Scale 0.87
Flip 0.84
Jitter 0.81
Power 0.73
Noise 0.66
Categorizing Anomalies in Retinal Imaging Data
● Anomaly detection in Optical Coherence Tomography images is a difficult and unsolved task, though many patients are affected by retinal diseases that cause vision loss
● Philipp Seeböck proposes to identify abnormal regions which can serve as potential biomarkers in retinal spectral-domain OCT images, using a Deep Convolutional Autoencoder
OCT dataset processing
Healthy OCT
One-Class SVM model
Unhealthy OCT
K-means
Anatomy Classification Through Visualization
● The design of the CNN model does not necessarily need to be a trial-and-error process, solely focused on optimizing the test set accuracy.
● Through visualization we managed to incorporate domain knowledge and overall managed to achieve a much more informed decision process.
MRI analysis
Tell me my diagnosis
● by patient histories ● by medical literature
● by outside information ● by similar patients searching
Diagnostic Prediction Using Discomfort Drawings
A discomfort drawing is an intuitive way for patients to express discomfort and pain related symptoms.
LDA
Demographical Priors for Health Conditions Diagnosis
Input data
● Age● Race● Gender● Income level
Clusters
https://flowingdata.com/2016/01/19/how-you-will-die/
How will you die?
Patient similarity matching
Idea: to determine the similarity of two ICD code sequences, one can compare their primary diagnosis
Word2Vec
Matching algorithm
Drug, treatment similarity
Diet2Vec: personal diet
Diet2Vec: personal diet
Despite the vectors being generated based on meal co-occurrence, our resulting diet vectors still have intuitive interpretations in terms of macro ratios.
5. Security of the medical data
Anomalies in doctor’s behaviors
Can you improve lung cancer detection?
Useful links:http://www.nipsml4hc.ws/
https://arxiv.org/pdf/1612.01356v1.pdf
https://arxiv.org/pdf/1612.00686v1.pdf
https://arxiv.org/pdf/1611.06284v2.pdf
https://arxiv.org/pdf/1612.02460v2.pdf
https://arxiv.org/pdf/1612.00388v1.pdf
http://www.sciencedirect.com/science/journal/09333657