EUSFLAT 2019: explainable neuro fuzzy recurrent neural network to predict colorectal cancer

Explainableneuro-fuzzy
recurrentneuralnetworkto
predictcolo-rectalcancerwith
differenttimeframedata
EUSFLAT 2019
Ostrava, Czech Republic
Servio Fernando Lima Reyna, PhD (c ) University of Fribourg, Switzerland

Agenda
 Why this research?
 Proposed neuro-fuzzy system
 Components of the system
 Future work
 Conclusions
 Q&A
2

Why this
research?
 In 2018, 1.8 Million cases worldwide, are attributed to CRC:
Colo rectal cancer (CRC) incidence worldwide
Source:World Health Organization-WHO
3

Why this
research?
Screening:
duplicates and
misclassification
removal
Eligibility: revisión
of title and
abstract
Inclusion: final
selection of
articles
Keywords:cancer AND neuro fuzzy AND colo AND rectal
between 2006 and 2019
ACM
2
Xplore
1
Elsevier
34
Springer
15
695 435 102 52
Too few papers are related to using neuro-fuzzy systems
for analyzing time sensitive EMR (electronic medical records) applied to CRC.
4

Why this
research?
 In past years, it has been seen that even though deep neural
networks (DNN) are accurate in its prediction, the medical
community hasn’t adopted it, because the lack of explainability.
 Moreover it has been shown that CRC prediction is benefit by
analyzing past patient data (e.g. EMR data) [2] .There are
algorithms such as LSTM (Long ShortTerm Memory) that can
take into account this data and find patterns in time. But LSTM still
remains difficult to explain due to its internal complexity.
 We propose to explain LSTM algorithm by means of Fuzzy Logic
which helps to humans:
 In representing the results of DNN in if-then format, easier to
interpret.
 In having traceability of the LSTM decisions
 In adding or modifying fuzzy rules (by humans) that will allow more
explainable predictions.
Rationale of this research
5

Proposed
neuro fuzzy
system
Explainable neuro fuzzy system for
colorectal cancer (ENF-CRC)
6

Data source
 Data was compiled on approximately 4,300 patients, including
patients between 18-49 years old diagnosed with CRC (200
patients), patients between 50-90 years old diagnosed with CRC
(3300 patients), and health controls with the same year of birth as
patients in Group 1 without CRC (800 patients).
 Dataset created Feb 12, 2018
 Dataset provided by
 NYU Health Sciences Library
 Time period covered
 Jan 1, 2011 - Dec 31, 2017
 Area covered
 NewYork
 Source: https://datacatalog.med.nyu.edu/dataset/10164
NYU Langone Health EMR Dataset
7

Components
of the system
 Our source of information are EMR.We are going to split the data in
two: 1. known predictors of CRC (as per Hippisley-Cox J, Coupland
C model, in blue) and 2. yet to know predictors (anything else).
EMR and data pre-processing
Type Purpose Output
Raw data
*Patient: age, gender
Consults: symptoms/diagnosis
Medication: medication type
Referral: specialism
Lab results: measurement type:value
*Diagnosis of specific disease:0/1
Lab value
contextualization
Contextualize raw lab values
to ground values
Lab value: low/normal/high
Lab value: increasing/decreasing/stable
Semantic
enrichment
Add existing knowledge to
compensate for
incompleteness of data
*Consults: comorbidities, diagnosis type
Medication: side effects
Events that co-occur or occur in succession
Table 1: EMR data and its preparation
*Hippisley-Cox J, Coupland C model uses specific data to predictCRC [3] 8

Components
of the system
EMR and data pre-processing: Hazard risk
Source: Hippisley-Cox J, Coupland C model [3]. CI: Confidence Interval. Bowel cancer includes colon and rectal cancer
 Thanks to the Hippisley-Cox J, CouplandC model, we know in
advance the impact that the variables in this model have over CRC.
It uses a score named Hazard Risk or HR.Therefore is up to us to
input them in LSTM.
9

Components
of the system
 Clinical variables are measured at different moments and are
irregularly sampled in time.
 LSTM may have a hard time in distinguishing lengthy and
incomplete sequences of events.
 Moreover, temporal pattern mining usually returns a large number of
temporal patterns, most of which may be irrelevant to the
classification task. To address all these problems we can use Batal [1]
pre-processing algorithm. 10

Components
of the system
 Despite all the previous efforts in data pre-processing, we still
may end up with a big number of input features.
 In this situation we are going to use Mann-WhitneyWilcoxon
rank-sum test to reduce the number of input features
considered.This test determine the difference in occurrences of
a feature among CRC and non-CRC cases.We are going to pick up
the features with the lowest p-values or probability of observed
result arising by chance.
Sympton A frequency (EMR data)
Patient 1 (non CRC)
Patient 2 (non CRC)
Patient 3 (CRC)
Patient 4 (CRC)
Time
11

Components
of the system
 Once we finish the data pre-processing, we can feed the LSTM
algorithm.
 LSTM has the ability to capture long term relationships in data. It
uses forget gate=1 to remember cell state and f=0 to forget.
LSTM for analyzing EMR
 It has been shown that CRC benefits from analyzing past and present
data of the patient up to six months [2], which is the most
symptomatic phase for CRC.
 But LSTM decisions for CRC are hard to explain because it considers
multiple input variables, each changing in frequency and time.
LSTM network and formulas: it contains: input, output and forget gates
Input
Forget
Output
12

Components
of the system
 We need a method that can assess the impact of each variable
individually (e.g. frequency of occurance of symptoms, lab value) and
its change in time.
 The method that achieves this is IMV-LSTM [6]
Explaining LSTM decisions
13

Components
of the system
 After finding the most relevant variables by IMV-LSTM and considering
that we already have the Hippisley-Cox J, CouplandC variables, we can
use them as input for a fuzzy rule based system (FRBS), that will
generate fuzzy rules automatically.
Explaining LSTM decisions: fuzzy rules
14

Components
of the system
GUAJE is a modeling software tool especially aimed for designing
interpretable fuzzy systems. GLMP is granular linguistic model of
phenomenon and it is usually represented as a hierarchical network of
computational perceptions. rLDCP is a software tool that is aimed at
producing GLMP. GUAJE and rLDCP can be connected in order to build
rule bases that can be embedded into some parts of a GLMP.
Explaining LSTM decisions: GUAJE, GLMP, rLDCP
NATURAL
LANGUAGE:
The patient has CRC
cancer with high
probability
because is white, is a
heavy smoker, very
heavy drinker, has
family history of bowel
cancer and has shown
pattern X in the lastY
months.
GLMP representation for CRC 15

Future work
 Implement the system.
 Change to other predictive models for CRC such as Marshall et
al. [4].
 Incorporate images: fMRI and radiography and use CNN
(Convolutional neural networks) for reading those images.
 Substitute LSTM with Bio-BERT (Bidirectional Encoder
Representations fromTransformers)[5], a pre- trained biomedical
language representation model for biomedical text mining that
has access to PubMed (4.5 billion words) and PMC (13.5 billion
words)
 Add an evaluation mechanism of generated explanations, such
as the Psychological model of explanation proposed by the
Defense Advanced Research Projects Agency (DARPA)
16

Conclusions
 Cancer analysis can be greatly improved fusing deep learning and
fuzzy logic, making the overall system more explainable to
medical doctors.
 A system that allows traceability of decisions and human
modification in case of any bias of the decisions, is a system that
is more prone to be accepted for the medical community for
helping in attending CRC patients.
17

Thank you
Servio Fernando Lima Reina, PhD student
Human-IST institute, University of Fribourg, Switzerland
servio.lima@unifr.ch

Bibliography
[1] Iyad Batal, HamedValizadegan, Gregory F Cooper, and Milos
Hauskrecht. A temporal pattern mining approach for classifying
electronic health record data. ACMTransactions on Intelligent Systems
andTechnology (TIST), 4(4):63, 2013.
[2] Amirkhan R. et al. Using Recurrent Neural Networks to Predict
Colorectal Cancer among Patients. IEEE Symposium Series on
Computational Intelligence (SSCI). Honolulu, HI, USA 2017
[3] Hippisley-Cox J, Coupland C (2015) Development and validation of
risk prediction algorithms to estimate future risk of common cancers in
men and women: prospective cohort study. BMJ Open 5: e007825.
[4] MarshallT, Lancashire R, Sharp D, PetersTJ, Cheng KK, HamiltonW
(2011)The diagnostic performance of scoring systems to identify
symptomatic colorectal cancer compared to current referral guidance.
Gut 60: 1242–1248.
[5] Lee J. et al. BioBERT: a pre-trained biomedical language
representation model for biomedical text mining. Oxford Press. 2019
[6] GuoT. et al, Exploring Interpretable LSTM Neural Networks over
Multi-Variable Data. Proceedings of the 36 th International Conference
on Machine Learning, Long Beach, California, PMLR 97, 2019.
20

EUSFLAT 2019: explainable neuro fuzzy recurrent neural network to predict colorectal cancer

More Related Content

What's hot (17)

Similar to EUSFLAT 2019: explainable neuro fuzzy recurrent neural network to predict colorectal cancer (20)

More from Servio Fernando Lima Reina (17)

Recently uploaded (20)

EUSFLAT 2019: explainable neuro fuzzy recurrent neural network to predict colorectal cancer