a Life Science and AIOps Perspective
Towards Trustable AI for Complex Systems Research Fellow
Data Science Institute
Imperial College London
Xian Yang
Conclusion
Towards Trustable AI for Complex Systems
Make data trustable
Make a good understanding of
systems
Make AI algorithm trustable
Background
Overview
Ways to achieve
trustable AI
Complex system: a system of systems
Complex life systems
Source: Wikipedia, GREATOPS
The signal transduction pathway in a
cell
The deployment diagram of a large-scale IT systemcomplex networks of biologically relevant entities
all related computer hardware, software, firmware, and
data for the communication, transmission, processing,
manipulation, storage, or protection of information.
Complex IT systems
Connected
Communicated
Complicated
Medical AI AIOps
Source: Itgsopedia, Riverbed
Missions of AI
Understand systems
Diagnose systems
Control systems
use of complex algorithms and software to emulate
human cognition in the analysis of complicated medical
data.
automate and enhance IT operations by
1) analyse big data collected from various tools and
devices via analytics and machine learning;
2) automatically spot and react to issues in real time.
AI for complex life systems and IT systems
AI for complex life systems and IT systems
AI based diagnosis
Embedding mapping
layer
Categorical feature
Classification
Layer
......Title
σ σ tanh
x
σ
x
x
+
tanh
σ σ tanh
x
σ
x
x
+
tanh
σ σ tanh
x
σ
x
x
+
tanh
Description i-1 Description i Description i+1
Output:
Failure type
Input:
Failure description
Input: Engineers’ discussion
Input:
Failure’s characteristics
Disease diagnosis Failure diagnosis
Source: ReferralMD
Medical AI AIOps
Elements of AI in Complex systems
Surgical
robot,
AI CT scan
reader,
AI nurse
Pathologic
analysis,
Efficacy
analysis
Clinical
pathway
optimization,
Hospital bed
management,
Disease
prevention
AIOps
component
library,
Intelligent
prediction,
Chatbot
Anomaly
detection
& prediction,
Root Cause
Analysis
Performance
Optimization,
Defragmentati
on,
Cost analysis,
Capacity
management
Efficiency Quality Cost
Data
Standardizatio
n
Data
Acquisition
Data Channel
Data Cleaning,
ETL,
Meta Data
Management,
Offline
Computation
Realtime
Computation
Feature
Engineering
Efficiency Quality Cost
Medical AI AIOps
AI
Applications
Big Data
Platform
Regression
analysis
Time series
analysis
Causal
inference
Dynamical model
construction
Correlation
analysis
Differential
feature selection
Cluster
analysis
Component
analysis
AI algorithms
Categories of AI algorithms
Ease of experiments
Power of explanation
My focus: Trustable AI for complex systems
What's AI’s
holdup?
It is not technical
The barrier is the trust aspects.
If it is not trustable, then it is
not useful.
Working towards trustable AI.
How to get humans to use our
AI technology and rely on it?
My
focus
Ways to achieve
trustable AI
1.1. Deeper 1.2. Wider 1.3. Bigger
2.1. A holistic view: Simplification VS. Complication
3.1. Moving beyond correlation 3.2. Moving beyond AI black-box
Make data trustable
Make a good understanding of systems
Make AI algorithm trustable
1.
2.
3.
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
Deeper: Extracting detailed information from the data1.1
Make data trustable
Phenotype annotation
from electronic health record
EXAMPLE
Cohort selection
for precision
medicine
Use as the
inputs of the
AI model
extracted
features
Extract information from
raw data
Problem:
Disease code cannot fully represent medical information in electronic health records (EHRs).
Case study: two patients with the same ICD code have different level of severity.
Example: Phenotype annotation from electronic health record
Case 1 admission_id=198908, subject_id=28912
brief hospital course the patient was seen in the emergency room at the
request of neurology and the emergency room staff at doctor last name
family she received a dilantin load on arrival to hospital a ct was obtained
which showed subarachnoid blood the patient was neurologically intact
except for some mild confusion about her location stating she was still at
doctor last name family hospital her films were reviewed by the neurosurgery
staff and the decision was made to take her for cerebral angiogram the
angiogram showed an aneurysm at the vertebral basilar junction mm to mm
it was unable to be treated endovascularly an open repair was considered to
be complicated given the location and the patients age dr first name stitle
elected to transfer the patient to hospital to dr last name stitle for further
evaluation and possible intervention
Case 2 admission_id=188170, subject_id=56707
brief hospital course patient was admitted to neurosurgery on for further
management she underwent the above stated procedure please review
dictated operative report for details she had a negative angiogram and was
to intensive care unit in stable condition she had an uncomplicated intensive
care unit course and was transferred to floor in stable condition throughout
her hospital course she remained neurologically stable and intact she
complained of a mild headache that worsens when she sits up and walks
around now dod patient is vss and neurologically stable patient s pain is well
controlled and the patient is tolerating a good oral diet pt s incision is clean
dry and inctact without evidence of infection patient is ambulating without
issues she is set for discharge home in stable condition and will follow up in
month for mri a brain with dr first name stitle
ICD 430: Subarachnoid hemorrhage
Find EHRs that diagnosed as 430 ONLY, Phenotypes (including severity) in red
Phenotype (HPO) terms can better characterize patients by providing deeper information .
admission_id=188170, subject_id=56707
"brief hospital course patient was admitted to neurosurgery on for further management she underwent the above stated procedure please review dictated
operative report for details she had a negative angiogram and was to intensive care unit in stable condition she had an uncomplicated intensive care unit course
and was transferred to floor in stable condition throughout her hospital course she remained neurologically stable and intact she complained of a mild
headache that worsens when she sits up and walks around now dod patient is vss and neurologically stable patient s pain is well controlled and the patient
is tolerating a good oral diet pt s incision is clean dry and inctact without evidence of infection patient is ambulating without issues she is set for discharge
home in stable condition and will follow up in month for mri a brain with dr first name stitle"
Example: Phenotype annotation from electronic health record
• Problem:
HPO terms cannot be fully found using the keyword search method: synonyms and
implicit information
• Solution:
apply AI to do automatic phenotype annotation
unsupervised
learning with no
labelled data
ICD 430: Subarachnoid hemorrhage
Synonyms: subarachnoid blood == subarachnoid hemorrhage
Synonyms: vertebral basilar == vertebrobasilar
Implicit information: terms in blue
Example: Phenotype annotation from electronic health record
J. Zhang, X. Zhang, K. Sun, X. Yang, C. Dai, and Y. Guo, “Unsupervised Annotation of Phenotypic Abnormalities via Semantic Latent Representations on Electronic Health Records”, 2019 IEEE
International Conference on Bioinformatics and Biomedicine (IEEE BIBM), 2019
There are two types of data sources.
: a collection of EHRs and each EHR consists of textual notes written by clinicians.
: a standardized general category of human phenotypic abnormalities provided by HPO.
The HPO also provides additional subclasses
Example: Phenotype annotation from electronic health record
J. Zhang, X. Zhang, K. Sun, X. Yang, C. Dai, and Y. Guo, “Unsupervised Annotation of Phenotypic Abnormalities via Semantic Latent Representations on Electronic Health Records”, 2019 IEEE
International Conference on Bioinformatics and Biomedicine (IEEE BIBM), 2019
Assumptions
The semantics of a general phenotype is represented by a prior distribution. The prior distribution of each phenotype should be ‘distinct’ enough from each other.
The semantics of EHR is a composition of the semantics of phenotypes.
Represented by
Generated by
Sampled from
phenotype vector prior
‘Distinct’ enough other priors
Sampled from
Represented by
Generated by
Example: Phenotype annotation from electronic health record
J. Zhang, X. Zhang, K. Sun, X. Yang, C. Dai, and Y. Guo, “Unsupervised Annotation of Phenotypic Abnormalities via Semantic Latent Representations on Electronic Health Records”, 2019 IEEE
International Conference on Bioinformatics and Biomedicine (IEEE BIBM), 2019
Loss 1: text reconstruction of EHRs. Loss 2: text reconstruction of the general phenotypic
abnormalities.
Loss 3: text reconstruction of the phenotype subclasses.
Loss 4: the latent vectors sampled from different priors can be
classified to different classes, then the priors are thought to be
‘distinct’ enough.
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
Wider: Integrating multi-modal data1.2
Make data trustable
Pan-cancer Classification
based on Multi-Omics analysis
EXAMPLE
…
Combining
Multimodal
Combine Data from different modalities
Provide a
comprehensive
view of patients
Make
more accurate
clinical decision
Example: Pan-cancer
Classification based on
Multi-Omics analysis
Our method:
We combine the
variational
autoencoder with a
classification network
to achieve task-
oriented feature
extraction and multi-
class classification.
Inputs
Outputs
X. Zhang, J. Zhang, K. Sun, X. Yang, C. Dai, and Y. Guo, “Integrated Multi-omics Analysis Using Variational Autoencoders: Application to Pan-cancer Classification”, (short paper) 2019 IEEE International
Conference on Bioinformatics and Biomedicine (IEEE BIBM), 2019
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
Bigger: Augmenting data1.3
Make data trustable
Augmented
Data
Augment data for limited samples
Increase the volume
of training samples
• Rare disease study
• System failure study
Imbalance
d Data
EXAMPLE
Synthetic medical image
augmentation
Example: Synthetic medical image augmentation
Frid-Adar, Maayan, et al. "GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification." Neurocomputing 321 (2018): 321-331.
Traditional image augmentation methods:
• Flip: flip images horizontally and vertically
• Rotation: rotate images by angles
• Scale: scale images outward or inward
• Crop: randomly sample a section from the
original image
• Translation: move the image along the X or Y
direction
• Gaussian Noise: add noise to enhance the
learning capability
Advanced image augmentation method:
• GAN
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
A holistic view: Investigate a complex system as a whole2
Make a good understanding of systems
Biological
system
System of systems
Cloud
System
Investigate system in
a holistic view
Study all
anomaly/failure signals
across the whole
system
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
Simplification VS. Complication: Keep balance in between2
Make a good understanding of systems
Inferring model for
large-scale biological network
EXAMPLEComplicatio
n
Simplification
Model networks of
complex systems for
better understanding
Effective modeling of entities
and inter-connections in
large scale systems
Example: Inferring large-scale biological network
Simplification:
Easy to model
Hard to mimic the real
behavior of system
Complication:
Hard to model
Good to mimic the real
behavior of system
Problem: How to make a balance between simplification and complication?
A. Holehouse, X. Yang, I. Adcock, and Y. Guo, “Developing a novel integrated model of p38
MAPK and glucocorticoid signalling pathways”, Computational Intelligence and Bioinformatics
and Computational Biology (CIBCB), 2012.
Example: Inferring large-scale biological network
Key observations of large-scale
networks
• Separation of timescales
Sparsity of variations
• Cross-reactivity
Combined-measurement
Prerequisites of sparse signal
recovery
• The signal is sparse in some domain
• A measurement is a weighted linear
combination of several points of the signal
My suggestion:
• We can study the complex network under different timescales.
• Within each time scale, only some entities have dynamic changes.
• Thus, we can apply sparse learning to infer a model under each timescale.
• Then, we combine all models obtained from all timescales together.
L. Nie, X. Yang, I. Adcock, Z. Xu, and Y. Guo, “Inferring cell-scale signalling
networks via compressive sensing”, PLoS One, vol. 9, no. 4, 2014.
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
Moving beyond correlation: Explore towards causation3.1
Make AI algorithm trustable
A signal’s predictive power does not necessarily imply that the
signal is actually related to or explains the phenomena being
predicted.
Moving from correlation to causation is especially important for
understanding :
what are the conditions under which it may fail?
how long we can expect it to be predictive?
how widely applicable it may be?
For an AI model…
• Deriving functional connectivity from
Brain fMRI data
• Root cause diagnosis for system
failures
APPLICATIONS
Correlation Causation
the covariance of the two
variables divided by the
product of their standard
deviations.
Fast computation of minimum
partial correlation
The choice of a hyperparameter,
the significant threshold, greatly
influences the results.
The minimum of all absolute values
of partial correlations by controlling
on all possible subsets of other
nodes
Remove indirect relationship
L. Nie, X. Yang, P. M. Matthews, Z. Xu, and Y. Guo, “Inferring functional connectivity in fMRI using minimum
partial correlation”, International Journal of Automation and Computing, 2017.
Causation
Correlation
measures the degree
of association between
two random variables, with the
effect of a set of controlling
random variables removed.
Partial correlation
Pearson correlation
PC algorithm
Minimum partial correlation
Automatically increase the significant threshold
within a given time limit to maximally
approach the minimum partial correlation.
Avoid repeating partial correlation done
previously with lower significant threshold.
Elastic PC algorithm
1. 2. 3.Make data trustable Make a good understanding of systems Make AI algorithm trustable
Moving beyond AI black-box: Explore towards explainable AI3.2
Make AI algorithm trustable
Explanability is the process of giving explanations to Human
Why we need Explainable AI?
Demands from industry and society
Desires of human brain
My future research direction: towards AI algorithm explainability
Black-box Explainable
Towards AI algorithm explainability
AUTO FEATURE ENGINEERING
COMBINATION OF
PHYSICS/MATH/TRADITIONAL ML
MODELS WITH ADVANCED DL
LEARNING LAWS FROM DATA
RATHER THAN CURVE FITTING
Automatically generate
explainable features in the
model construction process
Features: height, weight -> Label:
health []
Features: height, weight, BMI =
w/h2 -> Label: health []
Conclusion
There is still a long way to go before fully trustable AI.
We must work on it now!
The future of AI for complex systems depends on trust.
To achieve this, we need to work beyond the algorithm
aspect and work on
1) trustable data,
2) good understanding of systems
3) trustable AI algorithms.
Moving towards causation is crucial for making AI
algorithms trustable.
Source:
Wikipedia
https://en.wikipedia.org/wiki/File:Signal_transduction_pathways.png#filelinks
GREATOPS
http://www.gaowei.vip/m/Library/detail?no=94991143