ML to cure the world
-
Upload
xavier-amatriain -
Category
Technology
-
view
411 -
download
0
Transcript of ML to cure the world
Medicine is hard(er)
● Doctors have ~15 minutes to capture information* about a patient, diagnose, and recommend treatment
● *Information○ Patient’s history○ Patient’s symptoms○ Medical knowledge
■ Learned years ago■ Latest research findings■ Different demographics
● Data is growing over time, so is complexity● Very hard for doctors to “manually”
personalize their “recommendations”
Medical Diagnosis
● Diagnosis (R.A. Miller 1990):
○ Mapping from patient’s data (history, examination, lab exams…) to a possible condition.
○ It depends on ability to:■ Evoke history■ Surface symptoms and
findings■ Generate hypotheses that
suggest how to refine or pursue different hypothesis
○ In a compassionate, cost-effective manner
Cost of medical errors
● 400k deaths a year can be attributed to medical errors as well as 4M serious health events
○ This compares to 500k deaths from cancer or 40k from vehicle accidents
● Almost half of those events could be preventable
● 30% or $750B is wasted by the US Healthcare system every year
How to improve medical care?
● Automate processes through AI/ML
● Use of (big) data● More/better personalization● Improved user experience
both for patients and doctors
Does this sound familiar?
Medical Decision + Knowledge Bases
Medical Knowledge Bases encode years of Doctor Expertise
Doctor ExpertiseMedical Research
An example: Internist-1/QMR/Vddx
● Internist (1971) led by Jack Myers considered (one of) the best clinical diagnostic experts in the US
○ University of Pittsburgh, Chairman of the National Board of Medical Examiners, President of the American College of Physicians, and Chairman of the American Board of Internal Medicine
● Process for adding a disease requires 2-4 weeks of full-time effort and doctors reading 50 to 250 relevant publications
ML/AI Approaches to Diagnosis
● Early DDSS based on Bayesian reasoning (60s-70s)● Bayesian networks (80s-90s)● Neural networks (lately)
Ontologies
● Snomed Clinical Terms○ Computer processable collection of medical terms used in clinical
documentation and reporting.○ Clinical findings, symptoms, diagnoses, procedures, body
structures, organisms substances, pharmaceuticals, devices...
● ICD-10○ 10th revision of the International Statistical Classification of
Diseases and Related Health Problems (ICD)○ Codes for diseases, signs and symptoms, abnormal findings,
complaints, social circumstances, and external causes
● UMLS○ Compendium of many controlled vocabularies○ Mapping structure among vocabularies ○ Allows to translate among the various terminology systems
Electronic Health Records
● EHR/EMRs include digital information about patients encounters with doctors or the health system
NLP
Methods and algorithms to extract meaning and knowledge
from unstructured text
Patientunderstanding
The Language of Medicine
Doctor’sNotes
Medical researchpublications
ML/AIMedicalSystem
Clinical Decision Support +
Medical Knowledge Bases
Personalization
NLP
Multimodal input
Multimodal input
We will include many different signals besides direct patient
input
Speech interfaces
Image recognition
Sensors/lab data
ML/AIMedicalSystem
Clinical Decision Support +
Medical Knowledge Bases
NLP
Multimodal input
Personalization
Precision medicine
● Precision medicine (NIH):
"an emerging approach for disease treatment and prevention that takes into account individual variability in genes, environment, and lifestyle for each person."
● Term is relatively new, but concept has been around for many years.
○ E.g. blood transfusion is not given from a randomly selected donor
Personalization
The best and most relevant information “for you”
Patient profile & medical history
Biological markers & other lab data
Clinical Decision Support +
Medical Knowledge Bases
ML/AIMedicalSystem
Personalization
NLP
Multimodal input
What is different from other domains?
● Cost of errors● We care about causality● Implicit user signals not enough● Need of conversational approaches
○ Importance of eliciting information○ Importance of communicating outcomes
● Complex interactions between diseases and symptoms, including temporal sequences
What are we doing?
● Building an awesome team (Netflix, Quora, Facebook, Google, Microsoft, Uber, Stanford…)
● Combining AI/ML and best product/UX practices to build a service that revolutionizes healthcare by empowering patients to make their own decisions
● Leveraging pre-existing resources and state-of-the-art approaches
● We are stealth, too soon to say too much about what we have
Challenges
● Algorithmic: e.g. combining expert rule-based and ML● Data: quality, sparsity, and bias in data● UX: trustworthiness and engagement of the system,
incentives…● Legal● …
It’s about time we overcome all of these.
● “Probabilistic diagnosis using a reformulation of the INTERNIST-1/QMR knowledge base” . Shwe et al. 1991. ● “Computer-assisted diagnostic decision support: history, challenges, and possible paths forward” Miller. 2009.● “Mining Biomedical Ontologies and Data Using RDF Hypergraphs” Liu et al. 2013. ● “Health Recommender Systems: Concepts, Requirements, Technical Basics & Challenges”, Wiesner & Pfeifer, 2014. ● “A ‘Green Button’ For Using Aggregate Patient Data At The Point Of Care” Longhurst et al. 2014. ● “Building the graph of medicine from millions of clinical narratives” Finlayson et al. 2014. ● “Comparison of Physician and Computer Diagnostic Accuracy” Semigran et al. 2016. ● “Identifiable Phenotyping using Constrained Non-Negative Matrix Factorization”. Joshi et al. 2016. ● “Clinical Tagging with Joint Probabilistic Models” . Halpern et al. 2016. ● “Deep Patient: An Unsupervised Representation to Predict the Future of Patients from EHR”. Miotto et al. 2016. ● “Learning a Health Knowledge Graph from Electronic Medical Records” Rotmensch et al. 2017. ● “Clustering Patients with Tensor Decomposition”. Ruffini et al. 2017. ● “Patient Similarity Using Population Statistics and Multiple Kernel Learning”. Conroy et al. 2017. ● “Diagnostic Inferencing via Clinical Concept Extraction with Deep Reinforcement Learning”. Ling et al. 2017. ● “Generating Multi-label Discrete Patient Records using Generative Adversarial Networks” Choi et al. 2017● Suresh, H., Szolovits, P., & Ghassemi, M. (2017, March 20). The Use of Autoencoders for Discovering Patient
Phenotypes. arXiv.org.
References