Page 8 - Read Online

Page 4 of 12 Sathyan et al. Complex Eng Syst 2022;2:18 I http://dx.doi.org/10.20517/ces.2022.41

[2]
recommendations . This in turn helps the patients to trust the model results and system recommendations.
This can also increase the probability that the patient will accept and follow the recommended medical treat-
ment. Moreover, XAI can decrease the probability of error in the diagnostic process since it helps clinicians to
focus on the relevant data and help them to better understand the model recommendations.

XAI is an evolving field. As mentioned before, at this current stage, even state-of-the-art XAI algorithms
have disadvantages. In literature, we can find approaches that aim to improve some aspects. One of the main
challenges of using XAI in healthcare environments is the need to remain neutral regarding preferences. We
canfindabonafideapproachcalledscientificexplanationinAI(sXAI)thatcanbeusedinthefieldofmedicine
and healthcare [19] . An additional approach based on integrated Electronic Medical Records (EMR) medical
systemsisdescribedin [20] . Theapproachfocusesonexplainabilityandinteroperabilityfromthehumanaspect.
Ensemble of machine learning (ML) can also increase the level of interpretability, as can be seen in [21] . In [21] ,
theauthoruseensembleofMLforlogicdrivingofanthropometricmeasurementsinfluencingbodymassindex
(BMI). Additional evidence for the implementations of several XAI models is mentioned in [22] . The paper
showshowintegratingXAImodelshelpstoincreasethepersuasiveandcoherencelevelsinthedecisionmaking
of clinicians and medical professionals teams. The usage of XAI has shown an improvement in transparency
and reliability in the field of neuroscience field [23] .

In this paper, we apply some XAI concepts to a use case applicable to the medical field. Our work focus on XAI
implementation for breast cancer diagnostics. Our research uses the commonly researched UCI breast cancer
dataset. We focus on breast cancer since it is the most common type of cancer amongst women [24] . The usage
of XAI for diagnostics and prediction of breast cancer can impact and help a large number of patients. The
UCI breast cancer dataset includes 569 data points [25] . Each data point consists of 32 attributes that include
the ID number, the diagnosis, and 30 features used as predictors in this work. The 30 predictors include the
mean, standard deviation and the mean of 3 largest values of 10 features: (1) radius (mean of distances from
center to points on the perimeter); (2) texture (standard deviation of gray-scale values); (3) perimeter; (4) area;
(5) smoothness; (6) compactness; (7) concavity; (8) concave points; (9) symmetry; and (10) fractal dimension.

3. METHODOLOGY
3.1. LIME
LIME is one of the methodologies that is used to explain the predictions made by machine learning classifier
models [26] . It can explain individual predictions made by text classifiers as well as classifiers that are modeled
on tabular data.

In this work, we are focusing on using LIME to explain decisions made by a neural network classifier that
works on tabular dataset. The process of LIME to explain individual predictions are as follows:
1. For each instance that needs to be explained, LIME perturbs the observation times.
2. For tabular data, the statistics for each variable in the data are evaluated.
3. The permutations are then sampled from the variable distributions within the neighborhood of the original
data point for which an explanation is being sought.
4. In our case, the original model is a neural network. The trained neural network model is used to predict
the outcome of all permuted observations.
5. Calculate the distance from the perturbed points to the original observation and then convert it to a simi-
larity score.
6. Select features best describing the original model outcome for the perturbed data.
7. Fit a simple model (linear model) on the perturbed data, explaining the original model outcome with the
features from the permuted data weighted by its similarity to the original observation.
8. Extract the feature weights from the simple model and use these as explanations.