Download Apply Deep Learning Technology to Implement AI Doctor

Apply Deep Learning Technology
to Implement AI Doctor
Kan Deng，Ph.D
[email protected]
139-1122-6337
11/09/2016
1 / 14
Computer can play go,
but how about practicing medicine?
In 1997, IBM DeepBlue
outperformed Kasparov.
In 2011, IBM Watson won the
human champion of Jeopardy!
In 2016, Google AlphaGo
outperformed Lee Sedol
2 / 14
What is the correct objective?
Not to replace human doctor, but assist?
8 Dan - 10 Dan:
Recommend personalized precision medicine，
Covers rare diseases.
4 Dan – 7 Dan:
Following clinical guideline,
• Intermediate level,
Recommend the most likely diagnosis results
and therapies
• No surgery,
Of common diseases.
• Focus on 400 common
diseases,
1 Dan – 3 Dan：
Looks up the medical documents smartly,
• Recommendation only.
Enhances human doctor’s efficiency.
3 / 14
How big is the market?
全国每日约 1.7
亿 人次，
在网上咨询医疗，基本是两类问题。
1. 知情权：
确认医院的诊断是否准确，
了解医院的诊断依据。
2. 行动方案：
找性价比好的药品，
找临床经验好的医生，
找治疗质量可靠的医院，
找性价比好的医疗保险。
四种行动方案都直接是赢利点。
4 / 14
Where to use AI Doctor?
Health-care Community
• Help patients to enjoy
quick and good medical
consulting service.
• Help hospitals to find
target patients.
• Help government to
construct public healthcare records.
• Help insurance company
to investigate credit and
audit claims.
5 / 14
How to use AI Doctor?
Diagnosis and pathway
• Given (limited) illness status
description, output likely diseases.
• Recommend the further lab tests
and examination, to diagnose the
disease more accurately.
• Instantly respond to the patients.
• Reduce hospital labor cost.
修改病情输
入
6 / 14
To-do list to implement AI doctor
• Construct a private cloud inside hospital intra-network.
• Collect data from all sources, to assemble the EHR data
warehouse.
• Construct a medical synonym dictionary. Using WordVector
algorithm to cluster the synonym candidates automatically, and
then verify by human doctors.
• Transform the free-text EHRs into structured vectors, using
NER / LSTM algorithm.
• Construct the medical knowledge graph, by approximating the
correlations among the variables.
• Train the models to predict the diagnosis results given
symptoms, lab tests and other disease status.
7 / 14
Construct a private cloud
HIS
Synonym
Dictionary
Medical
Knowledge-Base
EMR
Data
Cleaning
LIS
EHR
Structuration
Data
Indexing
Data
Mining
Model
Training
PACS
Gene
Scheduler
Original EHRs
Structured EHRs
EHR Indices
Medical Knowledge Graph
8 / 14
So far we have processed 160 million EHRs
Similar to conventional CDR, but our system is capable of data mining and clinical decision supporting.
3 weeks to launch the private cloud,
0 on-site engineers,
No need to collaborate with the original HIS, EMR, LIS, PACS system vendors.
9 / 14
Construct the medical synonym dictionary
Using WordVector algorithm to find
the synonym candidates.
Disease standard name number: 25,856
Disease synonym name number: 43,476
Symptom term number: 1591
Lab testing indicator standard name number: 4,495
Lab testing indicator synonym name number: 4,495
Drug standard name number: 18,189
Drug synonym name number: 16,443
10 / 14
Transform free-text into structured entities
Using Long Short Term Memory algorithm
to recognize the entities from free-text.
11 / 14
Construct the medical knowledge graph
Do statistics on the co-occurrence among
symptoms, lab testing indicators,
radiology marks, diseases, and drugs.
Each symptom, lab testing indicator,
radiology mark, disease, and drug
is regarded as a node,
the co-occurrence frequency between
two nodes is an edge.
12 / 14
Clinical decision support
Given a collection of symptoms, lab test
and radiology exam results,
remind doctors of the likely diseases and
recommend the guideline of further tests and exams.
13 / 14
http://www.iyoudoctor.com
14 / 14