Graph-Based Machine Learning for Automated Health Care Services

Copyright © 2019, Oracle and/or its affiliates. All rights reserved. |
Graph-Based Machine Learning for
Automated Health Care Services
Rhicheek Patra, Oracle Labs Zürich
Email: rhicheek.patra@oracle.com
Twitter: @rhpatra, LinkedIn: @rhicheek-patra
Swiss Conference on Data Science 2019

Safe Harbor Statement
The following is intended to outline our general product direction. It is intended for
information purposes only, and may not be incorporated into any contract. It is not a
commitment to deliver any material, code, or functionality, and should not be relied upon
in making purchasing decisions. The development, release, and timing of any features or
functionality described for Oracle’s products remains at the sole discretion of Oracle.
2

Program Agenda
Benefits of graph models
HealthCare: Background & Problem Statement
KG-RNN: Overview
Empirical Validation
Concluding Remarks
1
2
3
4
5
3

Program Agenda
KG-RNN: Overview
Concluding Remarks
1
2
3
4
5
4

Your Data is a Graph!
• Represent your dataset as a
graph
– Entities are vertices
– Relationships are edges
• Annotate your graph
– Labels identify vertices and edges
– Properties describe vertices and
edges
Patient
name = ‘Jerald’
Admission
id = 1234
venue = ‘Triemli’
was_admitted
date = 01-23
Diagnosis
Procedure
diagnosed _with

Benefits of Graph Models
• Some of graph benefits
– Intuitive data model
– Aggregate information over heterogeneous data sources
– Fast query over multi-hop relationships
– Data visualization and interactive exploration

• A fast, parallel, in-memory graph analytics framework
• Offers 35+ built-in, native graph analytics algorithms
• Provides a graph-specific query language (PGQL)
7
SELECT v, e, v2
FROM graph
MATCH (v)-[e]->(v2)
WHERE v.first_name ='Jerald’
AND v.last_name = 'Hilpert'
GROUP BY … ORDER BY … LIMIT …
Parallel Graph AnalytiX (PGX)

Graph in Various Domains
Health Science Financial Services
Patient
Admissions
Diagnosis
Procedures
Customer
Account
Address
Alert
Institution
ICIJ
Worldcheck

Program Agenda
KG-RNN: Overview
Concluding Remarks
1
2
3
4
5
9

Knowledge Graphs
10
• Represent information as entities, properties and
relations between entities
• Allows to naturally combine information from
different sources
• Lot of activity and research around leveraging
knowledge graphs in Machine Learning

Sample: Healthcare Knowledge Graph
11

Sample: Retail Knowledge Graph
12

Features of the HealthCare Knowledge Graph
• New relations updated in the Knowledge Graph (KG) over time
– Due to unseen disease-symptom analysis over new medical studies
– Due to incoming Electronic Medical Records (EMRs) for new medical admissions
• Dynamic properties of entities (a.k.a. “admissions” in the KG)
– Patient Admissions vary over long duration (few hours to few weeks)
– During the admission, patient goes through lots of different
• Medications: e.g., Aspirin (500mg/day)
• Lab tests: e.g., pH=7.4 at 15:35
• How do we construct such KGs?
– Internal information (e.g., admissions) aggregated over hospital databases
– External information (e.g., disease-symptoms) collected from public resources
13

What is the Learning Objective?
• Given an admission with all the dynamic medical inputs (e.g., medications,
lab results), how can we predict the diagnoses related to this admission?
14
1. Hypertension
2. Heart disease

Program Agenda
KG-RNN: Overview
Concluding Remarks
1
2
3
4
5
15

How do we achieve this Learning objective?
• Model the problem as a temporal sequence one (split by N-hour intervals)
16
Now we can train a Recurrent Neural Network for this objective!

Recurrent Neural Network and how do we employ it?
• Used in multiple sequence-based applications
– Stock Market Prediction
– Machine Translation
• Recurrent Neural Network requires the temporal states of the input x
(which is admission in our case)
17

Input to the RNN: Admission Timeline Chunks
18

Admission Training (Admission Encoder Module)
19
AggregatorAggregator
Chunk 1 Chunk N

Where can the Knowledge Graph benefit in our objective?
• To feed the relations between previous “relevant” admissions (based on
some admission properties, e.g., Pre-diagnosis) and their diagnoses.
20

Relevant Admission Extraction
• Nearest Neighbor problem
– Given an input Admission, find the top-K nearest Admission in the healthcare
knowledge graph
• Approach: Weighted Personalized PageRank
– Weights
• Disease – Symptom: Importance of Symptom for specific disease
• Admission – Symptom/Disease: Cosine Similarity w.r.t. Pre-diagnosis and Symptom/Disease
– Take top-K admissions as the most relevant one w.r.t. WPPR score
21

Relevant Admission Extraction
22

How do we employ the neighbors information? – KG-RNN
• Encoding for Input Admission
– Obtain admission encoding for the input admission (similar to the baseline)
• Aggregation of neighbors diagnoses information
– Aggregate the diagnoses information of the neighbors and concatenate with the
encoding of the input admission
• Diagnosis Prediction
– Employ the concatenated vector for the diagnoses prediction
23

KG-RNN Prediction Module
24
Input Admission
Final Encoding
Neighbor 1 Final
Encoding
Neighbor M Final
Encoding
Prediction Module
Diagnoses

Program Agenda
KG-RNN: Overview
Concluding Remarks
1
2
3
4
5
25

Dataset: MIMIC-III
• Available at https://mimic.physionet.org/
• MIMIC is a relational database containing data of patients who stayed
within the intensive care units at Beth Israel Deaconess Medical Center
• Input: Admissions
• Output: Multi-label multi-class classification task
– We focus on top-50 most frequent ones
26
Type Patients Admissions Input events Output events Lab tests Prescriptions
Count 46,520 58,976 21,000,000 4,500,000 28,000,000 4,000,000

Results
• Incorporating Neigbors information boosts quality of KG-RNN on MIMIC-III
27
Metric Average Model Score
AUROC
Macro
Baseline 85.24%
KG-RNN 86.29% (+1.05%)
Micro
Baseline 90.55%
KG-RNN 91.03% (+0.48%)
Properties
3 hours per chunk
200 chunks per admission
25 events per type and per chunk
10 neighbors sampled per input admission

Sample example of Neighbor-boosting
28
• Baseline does not detect any
diagnosis
• Neighbor helped pushing
the confidence of KG-RNN
upward for 2nd diagnosis
• Reveals how much KG-RNN
relies on its neighbors

Prediction Timeline
29
• Our approach provides real-time
diagnosis probabilities during an on-
going diagnosis leading to better
medications and treatments.

Program Agenda
KG-RNN: Overview
Concluding Remarks
1
2
3
4
5
30

Summary
• Novel approach for processing evolving Knowledge Graphs with GraphML
• Multiple Applications in HealthCare, Retail, Finance, Security
• Online articles for detailed read
– https://blogs.oracle.com/ai/introduction-to-knowledge-graphs-in-healthcare
– https://blogs.oracle.com/ai/graph-machine-learning-for-enhanced-healthcare-
services
• Tools
– PGX: https://docs.oracle.com/cd/E56133_01/latest/index.html
– PGQL: http://pgql-lang.org/
– Github repository: https://github.com/oracle/pgx-samples
31

Graph-Based Machine Learning for Automated Health Care Services

More Related Content

Similar to Graph-Based Machine Learning for Automated Health Care Services (20)

More from Thomas Teske (8)

Recently uploaded (20)

Graph-Based Machine Learning for Automated Health Care Services