=Paper=
{{Paper
|id=Vol-3302/paper19
|storemode=property
|title=Predictability of Kidney Dialysis Survivability
|pdfUrl=https://ceur-ws.org/Vol-3302/paper11.pdf
|volume=Vol-3302
|authors=Arvinte Rareș,Diana Trandabăț
|dblpUrl=https://dblp.org/rec/conf/iddm/RaresT22
}}
==Predictability of Kidney Dialysis Survivability==
https://ceur-ws.org/Vol-3302/paper11.pdf
Predictability of Kidney Dialysis Survivability
Arvinte Rareș, Diana Trandabăț
a
Alexandru Ioan Cuza University, Faculty of Computer Science, Bulevardul Carol I, Nr.11, 700506, Iași,
Romania
Abstract
In medicine, vast amount of data is stored detailing the actions undertaken for each patient,
data very rarely used by hospitals, mostly having statistical and informative purposes for
doctors and hospital management. Such data could be successfully used to monitor a patient
more accurately, both outside and inside the hospital. Although there exist internal
monitoring processes, they can be expensive or limited to actions that are inherited from
simple, public statistics, when in fact increased efficiency could be seen if using more
complex data. In this paper, we use patient data to predict kidney failure, a work which may
inspire others to look for and find patterns in medical health records.
Keywords 1
Neural network; Kidney; dialysis; survivability
1. Introduction
The human body is a miracle of nature. It consists of the skeleton, muscles and organs. But what
happens when an organ doesn't work? Specifically, when the kidneys no longer function properly?
Nowadays, modern medicine found a solution to replace its function to help people stay alive because
kidneys have the critical mission of cleansing human blood. This method is dialysis which, as
specified above, has the role of taking over the mission of the kidneys.
Its role is to eliminate toxins from the blood and it works as follows: the patient's blood passes
through a filter that contains a fluid used in dialysis to stop toxins and at the exit the blood is cleansed.
The motivation behind choosing this topic is the fundamental idea of helping other people and this
came to light when a doctor from Iasi proposed to help the medical community through an application
or study in helping people on dialysis, a topic of increased interest in the last years [1]. This way,
doctors can help patients better and increase their rate to survive.
This application should help a doctor to make better decisions about the patient's health in the
context of dialysis. Through the patient's analyzes, it will be possible to generate graphs to keep the
doctor up to date with a patient’s health condition. This will be possible with the help of a neural
network that is trained with data about dialysis patients, method already studied in the literature [2].
In this paper we will present the application and explain its applicability in society and in future
projects. We will also follow the exposition of the technologies used in its creation and possible
development ideas in the future but also the description of the stages and difficulties encountered until
reaching a final form of the application. By describing the application, we will follow the detailing of
its functionalities, the technologies used in the open-source environment and the programming
language used, namely Python [3].
The application works on the premise of creating a neural network that is trained and classifies
dialysis patients. After if there is a data set of a single patient who presents his long term analyzes can
be made an accuracy of his evolution. In the future, the application can incorporate also graphic
support and various statistics for the doctor, as suggested in [4].
IDDM-2022: 5th International Conference on Informatics & Data-Driven Medicine, November 18–20, 2022, Lyon, France
EMAIL: raresarv1@gmail.com; diana.trandabat@gmail.com
ORCID: 0000-0002-8494-4441; 0000-0002-9874-0642;
© 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
2. State of the art
Dialysis removes residues and fluids from the body that the kidneys are unable to eliminate.
Dialysis also has the role of maintaining the body's balance, correcting the levels of various toxic
substances in the blood. Without dialysis, all patients with complete renal impairment would die from
the accumulation of toxins in the bloodstream [5].
After consulting [6] we concluded that a neural network would be the most appropriate solution.
This paper presents the behavior of a neural network, a decision tree, and a logistic regression on a
sample of 193 patients undergoing hemodialysis (HD) in Hasheminejad Kidney. Center (HKC) of
Tehran, which is one of Iran’s largest renal hospitals. This study is similar to mine difference being
that they only explain how these algorithms would work using a software framework developed with
Tanagra tool which is used for data mining related work. Instead, we built a neural network using
tensorflow. It was good because this article is on exactly the same topic as ours, the difference being
made by the approach and the data set [7].
The work in [8] addresses the problem from the perspective that vitamin D is the main cause and
that people with this deficiency should be identified, prioritizing their monitoring. The authors
propose three methods, namely multivariate logistic regression, neural networks and decision trees.
As in our dataset, they included issues that would influence decision making such as whether or not
the patient is a smoker. Their results on a data set of approximately 900 people indicate that the
multivariate logistic regression and neural network have similarly good results to the decision tree.
In the work [9], the authors approach a topic similar to ours, on a different data set, more evenly
distributed. Among the characteristics chosen for the neural network, many of them are similar with
the ones we use, which makes us lean towards the idea of some obligatory factors that must be
considered when it comes to dialysis, indicating that some criteria need to be used with a greater
weight than other factors.
3. Architecture
The available data set is an .xls file containing the data of 4114 dialysis patients, structured in 240
columns. A sample of the initial data can be seen in Figure 1.
Figure 1: Data snippet
In the first phase, we took these data and processed them in the form of a pandas type dataset.
After we tried to create a simple neural network to test the behavior of the data in practice. While
running we noticed that some columns contain incomplete data. As an initial solution the missing data
were populated with the value 0 but the next problem was the calendar data which indicates when
certain types of analyzes were done so we removed them from the data frame. The only problem
remains the fact that we did not know which is the output column so we tried to apply the K-means
algorithm which, after discussions with a specialist, we had come to the conclusion of placing people
in four risk groups.
We applied K-means and each person was assigned a risk group from zero to three. After this
grouping we saved it in the data set in a column called "target". Following the passage of this data set
with the "target" column as output through the neural network, an accuracy of 43% was obtained. The
results do not look so good so we resumed discussions with an expert in the field who suggested
cleaning the data, namely from the 240 columns many of them were apparently used to compensate
for the missing data that we had filled in with zero at a previous step. The overall architecture can be
seen in Figure 2.
Figure 2: Data conversion process
After removing duplicate columns, we used the per column average of the values to fill in the
missing data. After passing through the neural network, the results increased to 73%, but a closer look
revealed that these results varied between 14% and 73%, the reason for this being represented by the
vast number of different parameters. Following other discussions with the medical experts, we
decided to use the “all_cause_death” column as output, and removed all calendar data, except the start
and end date of dialysis, which we combined into a single column in which we keep evidence of the
number of days the patient undergo dialysis, as suggested in [10]. This column contains values of zero
(alive) or one (dead). The results were approximately the same because the number of surviving
patients is much higher than the number of the dead ones, the data being highly unbalanced.
Figure 3: Final architecture
We have chosen a method to group the numeric columns in 4 groups, more precisely the numeric
columns that do not have values only between zero and three, managing by this to reduce each
numeric class number of possible combinations and to obtain 84% accuracy. Finally, after changing
the loss function from categorical cross-entropy to binary cross-entropy, which is better suited for our
output, we obtained the final result of 94% accuracy in testing.
The final form of the application works as follows (Figure 3). At first it receives an .csv file at the
input which is opened with pandas and automatically converted into a pandas data frame. After that,
also with the help of the pandas, the columns with the missing data are filled in with the average per
column. Next is the division of the data frame into train data, test and evaluation after these new
branches of the initial data frame are converted into tensorflow data frame. The next step is to create
the feature layer that serves as input for the neural network. At this step we have three types of
features, namely embed, numeric and bucket sized features. Embed is used for columns consisting of
strings while numeric is for columns consisting of numbers and bucket sized for certain numeric
columns that have been optimized in terms of their input. The network model consists of the first
layer which is the input layer and contains all the features followed by two dense layers with relu
activation. Than an output layer with sigmoid activation because the output is binary. The model is
compiled with the loss function of binary cross-entropy and adam optimization with the function lr =
1e - 3 this comes automatically from the tensorflow package.
Figure 4: Neural Network architecture
Finally, the model (Figure 4) is fitted and then the test data are executed. In the end, graphs are
generated for the analysis of the results and also a method available at the end to test with a patient his
medical evolution based on many analyzes created manually.
4. Data description
The initial data set was an XLS file with 4,114 patients and their analyzes along with the dialysis
period and whether the person died or not. There are 240 columns of which for the columns “hb”,
“ektv”, “calcium”, “phos”, “tgo”, “albumin”, “crp” have rows whose data are missing or are
incomplete so for these columns initially there are another 20 identical columns in which an average
of the values is applied to fill in these missing values but we considered it necessary to eliminate these
duplicated columns and to complete the incomplete or missing data with the average per column, thus
reaching the remaining 100 columns. Then there were columns that indicated when the respective
medical analysis was done so we eliminated those and finally we joined the start and end date of
dialysis in a single column that keeps track of the number of days on dialysis. Thus, the final data set
will be 4114 patients with 83 features in an .csv file. From the beginning the data were anonymous,
patients being represented by an id column that was finally removed because it does not help
anything.
The data respects the principles of GDPR (general data protection regulation), there are also
columns that represent the id of the hospital which in the end were deleted as well. Figure 5 shows
several data distribution reports to present an overview of the data used. After consultation with a
healthcare professional, all remaining features are essential in classifying patients. In the initial data
set the features can be divided into five categories, namely demographic, clinical, laboratory data,
laboratory data to compensate for missing values and situational. And in the final data set only three
categories remain, namely demographic, clinical, laboratory data. Demographic traits are traits that
represent aspects of one's appearance. These include traits such as the patient's weight or age.
Following are clinical features that are in principle those related to the patient's condition such as
whether he suffers from certain comorbidities or has certain addictions such as smoking. Laboratory
data and laboratory data to compensate for missing values are the values of patient analyzes. Here we
have features from the spectrum of several types of tests from blood to dialysis, some examples would
be the level of fat-mass, the number of white blood cells, calcium or albumin levels. This is where
most of the features are included, with the largest share in the decision-making factor. Finally,
situational traits are related to identification data such as the patient's ID or date of birth, which in the
final data are no longer taken into account and do not present a factor for making a decision.
Figure 5: Statistic data
5. Modules
The application is divided into several modules to clean the data and prepare the neural network to
be executed successfully. It does not have a graphical interface being only taken to the test stage.
Below we describe the modules and how they work to understand why we chose this division.
Data frame reader module: The first module is to read the data frames according to the required
situation such as the file format. For this we use the pandas library which provides functions for
reading xls files or csv files. This module, in addition to reading the file and converting to a panda
dataframe, also deals with filling in the missing data with the media on the column. This is also done
with pandas.
Conversions module: This module deals with the conversion of data, columns or data frames into
new data types. There are three functions in this module. The first function is to convert an array of
strings of a pandas data frame into one of numbers of a pandas data frame. This function receives as
input the data frame, the string list, the column name and the last index. This function can be used for
any column that contains strings but for our data it had to be used only for the "renal_diagnosis_icd"
column. Also, this function is used only to be able to do k-means clustering because in this method we
use only numpy data (only accept numbers). At output we have a copy of the data frame with changes
done. The next function is a simple one that receives a pandas data frame and converts it into a numpy
object. And this function is used only for k-means that act as input for it. The final function is the
most important in the module, it receives a pandas data frame and converts it to tensorflow data set.
This function also removes the output column from the data frame because it is not relevant for
tensorflow data frame to contain this column. The output of this function return a tensorflow data
frame that will represent the input for the neural network.
Miscellaneous module: This module contains three functions the first is to extract all the elements
of a column from the data frame. It receives as input the data frame and the column name and at the
output we have a list with all the elements of the column. This function is used only for the
"renal_diagnosis_icd" column. The next function is to create a list that contains a number of equal
intervals used to optimize the neural network. This function receives the elements of a numeric
column and the number of intervals, calculates the length of an interval and saves in a list the
beginning of each interval. The last function receives the date frame as input and returns a list of all
the column names that will later be used to divide the columns into feature types.
K-means module: There is a dedicated module for k-means that contains the whole process of
performing a clustering with this method. It contains two functions the first of which calculates the
silhouette score which is used to decide the number of clusters, it uses the sklearn library to calculate
this score. It receives as input the data frame and the number of iterations (clusters) and executes k-
means for each number of clusters and for each score it is checked if it is higher than the maximum
found. If it is higher it becomes the new maximum and the iteration number becomes the number of
clusters. The other function is actually the k-means clustering used with the help of sklearn. In the end
its also plot a graph with all its points and centroid.
K-means executor module: This module handles the whole process of calling all the functions
needed to read, convert and execute the cluster.
Neural network module: The neural network module contains the whole process of creating the
features (from tensorflow data frame) and the model that will be executed in the execution module.
With the help of the tensorflow library, a division is made into features that will be used as the first
layer of the neural network. Three types of features are used, the first being embed feature with the
only column of this type being "re-nal_diagnosis_icd". Then numeric features that are divided into
simple and bucketsized features. Here we use the interval division method described above to create
these buckets. Simple numeric features are those that contain integer data in the range of zero - three.
This condition was chosen ad hoc after manual data analysis. The rest of the data is of the bucket
features type. After these they are used as the first layer of the neural network followed by two dense
layers and a dense layer used as output. The compilation is done with the Adam optimizer using the
binary cross-entropy loss function. It also contains a function for creating the neural network model
that will receive these features created above. And at the end of the module is a test function used to
test the system with a new patient with multiple analyzes.
Execution module: The execution module contains the whole process of executing the neural
network and graph creation. Here the train, test and evaluated data are converted from pandas data
frame to tensorflow data frame using the conversion module. Also here is the training of the neural
network, its evaluation and testing. Here the statistical functions from the statistics module are called,
but also here the execution of the testing of a new patient is called.
Statistics module: This module deals with any means of creating graphs or displaying statistics. It
contains functions for example the creation of statistics on the test data following the evaluation of the
system or the evolution of the patient over time or the confusion matrix.
File manipulation module: This module deals with manipulating the data to bring it into an ad-hoc
final form of our choice. It contains functions for converting a file from xls to csv or creating copies
of a file with the addition of a new column or even converting columns that contain time periods into
numbers or merging two time columns into a numeric column. This functions were used in the early
stages of the project when we used centroid cluster clustering as output. At the end of the project it is
no longer used because the "all_cause_death" column is used as output.
6. Results
The factors considered for evaluating the results consisted in a number of metrics: loss,
performance classification, accuracy, precision, recall, AUC and F-score. After training the neural
network and validating it with the validation data, the results obtained can be interpreted from Figure
6.
Figure 6: MAE graph and MSE graph
The first image shows the evolution of mean absolute error (MAE) which helps us to get an
overview of how the error rate evolves with each epoch. If the line becomes constant it means that the
neural network is no longer learning. The MAE score after 20 epochs is 0.074.
The other image shows the evolution of mean squared error (MSE) which evaluates the quality of
predictions. The closer it is to zero, the better the quality. It is observed that the value after 20 epochs
is close to zero. The result value is 0.067. After training we tested the system with test data (20% of
the total set date). Accuracy is a metric for measuring how often the neural network correctly
classifies an input from the given set date.
Table 1
Metrics
Metric Values
Accuracy 0.929
Precision 0.870
Recall 0.641
AUC 0.843
MAE 0.068
MSE 0.077
F-score 0.738
As can be seen in Table 1, the accuracy obtained is 92.9%, which indicates a fairly good score.
The precision and recall are calculated as follows: precision is the total of true positives (TP) divided
by the sum of true positives and false positives (FP) and the recall is the total of true positives divided
to the sum of the total of true positives and false negatives (FN). Precision shows us the probability
that a patient is classified correctly with one of the two values while the recall indicates the
probability that a patient is cataloged with a certain tag. After testing the neural network, the precision
obtained was 87% and a recall of 64%.
F-score its weighted harmonic mean helps us to have an overview of the accuracy of the test data.
Its value is between zero and one. The closer it is to one we can say that the system does a better job if
new data comes at the input. The F-score obtained after the test was 0.738 which indicates that the
system is doing quite well if would receive new data as input.
ROC or receiver operating characteristic curve is a graph that shows the performance of the
classifier in the ratio true positive rate (TPR) versus the false positive rate (FPR) [11].
Above we talked about ROC to lay a foundation in explaining the AUC. AUC (area under the
ROC Curve) as the name implies refers to the area under the graph which tells us how well the neural
network distinguishes between the classes of a classifier, in our case it shows how well it
distinguishes between living or dead patients. The AUC value is between zero and one and the closer
it is to one the better the system makes the differences. The AUC after testing was 0.843.
Confusion matrix helps to understand how classified data is distributed so that it is understood how
well the system handles. Out of the total number of 4114 patients, 823 are test patients. Following the
evaluation, 682 were classified as true negatives, 82 true positives, 13 false positives, 46 false
negatives.
As can be seen, the data are quite unbalanced because many patients are classified as alive or true
negatives in our case. In the first instance this was a problem but in the end after we applied methods
to stop this by balancing the values of the features we managed to ignore this and therefore the
program makes few mistakes. After testing the system, we decided to check his behavior for a new
patient with multiple tests and to perform an evolution of his disease. Figure 7 shows the evolution of
the patient's disease depending on the number of days since he is on dialysis, thus demonstrating how
the patient's risk of dying over time can be interpreted.
Figure 7: Patient evolution in time
Figure 8 shows the distribution of patients according to the time spent on dialysis in relation to
their near death. This figure indicates how the neural network classifies patients according to time so
one can see the results of this approach.
Figure 8: Results in comparison to evidence days
Comparing our results to the ones discussed in the state of the art sections, we can infer, for
example, that in [6] the aim of the paper is to compare several data mining techniques to predict
whether a person would survive kidney dialysis. Because we only approached the neural network
method, we cand only compare with their neural network method. First of all, it should be noted that
their data set is small, balanced and has only eight parameters taken into account. They also used the
Tanagra tool which provides a suite of tools for the mining date. Like us, they used the accuracy to
determine the performance, their score being 93.852% compared to the one we obtained, namely
92.9%, but compared to them we did not average all the runs (the accuracy varies between 91% and
96%) this variation is due to the fact that our data is more balanced than their data.
The paper [9] addresses neural networks and logistic regression, the comparison metric chosen for
comparison is AUC. We will also refer only to the results obtained on the neural network. They used
Neuroshell 2 version 3.0 to make the network, on a data set of 3629 dialysis patients and got an AUC
of 0.7602 compared to our neural network which got an AUC of 0.8437. Our opinion is that this result
reflects the optimizations made for bucket sized features.
7. Conclusions and further work
Above we discussed the problem addressed and the solution we chose, namely a way to predict the
evolution of a patient on dialysis based on his probability of dying. We presented the architecture of
the neural network built using python and tensorflow and the results obtained by it. Our conclusion is
that this issue has been addressed in the past and we agree with the opinion of other authors who have
concluded that neural networks have the best results [12].
The difference brought by this paper is the data set tested and the results obtained. We could say
that we have improved the results obtained by them. we believe that computer science should have a
greater impact on such areas that work with human life, not to replace decision-making but to be
introduced as an assistant for it thus increasing the speed of response and consideration of opinions
which at the moment may have escaped for example a doctor.
Also, what we have done can be seen as a starting point for the modernization of medical
infrastructures in Romania, it can help to develop applications for the benefit of doctors, thus helping
the community and saving more lives. It should be added in conclusion that this application was made
in python with tensorflow and not with an online application that provides tools to make this possible
The application is purely demonstrative, for the future we will try to create a graphical interface
and a database in which to store future patients so that it can be used in hospitals. Not to replace
doctors but to help them make a decision taking into account the choices made by the program. Thus,
the program can show a doctor the evolution of the patient and he can act as such. The program can
also interpret for the doctor the data provided so as to help him even more, highlighting the features
that have changed in the patient and where to work with him, the benefits provided as a
recommendation because in the end something subjective must to make the decision to solve the
problem is not an objective machine [13]. Also, if this application is to be further developed in the
future, it can motivate people to try new methods and improve the system, thus increasing the chance
of success and to be able to be introduced in hospitals to help save human lives.
8. References
[1] Tomašev, N., Glorot, X., Rae, J. W., Zielinski, M., Askham, H., Saraiva, A., ... & Connell, A.
(2019). A clinically applicable approach to continuous prediction of future acute kidney injury.
Nature, 572(7767), 116-119.
[2] Chan, L., Vaid, A., & Nadkarni, G. N. (2020). Applications of machine learning methods in
kidney disease: hope or hype?. Current opinion in nephrology and hypertension, 29(3), 319.
[3] Van Rossum, G. (2007, June). Python Programming Language. In USENIX annual technical
conference (Vol. 41, p. 36).
[4] Kim, H. W., Heo, S. J., Kim, J. Y., Kim, A., Nam, C. M., & Kim, B. S. (2021). Dialysis
adequacy predictions using a machine learning method. Scientific reports, 11(1), 1-7.
[5] Thongprayoon, C., Hansrivijit, P., Bathini, T., Vallabhajosyula, S., Mekraksakit, P., Kaewput,
W., & Cheungpasitporn, W. (2020). Predicting acute kidney injury after cardiac surgery by
machine learning approaches. Journal of Clinical Medicine, 9(6), 1767.
[6] Lakshmi, K. R., Nagesh, Y., & Krishna, M. V. (2014). Performance comparison of three data
mining techniques for predicting kidney dialysis survivability. International Journal of Advances
in Engineering & Technology, 7(1), 242
[7] Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., ... & Kudlur, M. (2016). Ten-
sorflow: A system for large-scale machine learning. In 12th {USENIX} Symposium on
Operating Systems Design and Implementation ({OSDI} 16) (pp. 265-283).
[8] Bhan, I., Burnett-Bowie, S. A. M., Ye, J., Tonelli, M., & Thadhani, R. (2010). Clinical measures
identify vitamin D deficiency in dialysis. Clinical Journal of the American Society of
Nephrology, 5(3), 460-467.
[9] Tangri, N., Ansell, D., & Naimark, D. (2008). Predicting technique survival in peritoneal dial-
ysis patients: comparing artificial neural networks and logistic regression. Nephrology Dialysis
Transplantation, 23(9), 2972-2981.
[10] Verplancke, T., Van Looy, S., Steurbaut, K., Benoit, D., De Turck, F., De Moor, G., &
Decruyenaere, J. (2010). A novel time series analysis approach for prediction of dialysis in criti-
cally ill patients using echo-state networks. BMC medical informatics and decision mak-ing,
10(1), 4.
[11] Flach, P. A. (2003). The geometry of ROC space: understanding machine learning metrics
through ROC isometrics. In Proceedings of the 20th international conference on machine learn-
ing (ICML-03) (pp. 194-201).
[12] Cassol, C., & Sharma, S. (2021). Nephrology Lagging Behind in Machine Learning Utilization.
Kidney Medicine, 3(5), 693-695.
[13] Zhou, X. J., Zhong, X. H., & Duan, L. X. (2022). Integration of Artificial Intelligence And
Multi-omics in Kidney Diseases. Fundamental Research.