Algorithms for an Integrated Disease Database Management

Abstract

The project goal was to make a comparison between diferent types of algorithms to prove what will be the best in similar cases. Aditional we were looking at which one will be the best for frontend use. A simple classification of diseases according to symptoms. We use 2 files from the data set, first with diseases and symptoms and the other with the symptoms themselves and their weights. I use the first set fully. It contains 41 diferent diseases and 132 symptoms. In this setup, we will use 2 KNN algorithms, a soft classifier, and a decision tree. We compare their performance and execution time. The results were predictable, KNN has the best accuracy but was the slowest, and the decision tree was a little bit worst accuracy but very fast.

Keywords

• eol>Prediction model
• Classification algorithms
• Disease prediction
• Python

1. Introduction

2. Assumptions for the algorithms

Modern computer science is heading towards the creation Each of the algorithms should be prepared to meet the of intelligent systems that facilitate control, management following criteria: or optimization of certain user processes [ 1, 2, 3, 4, 5 ]. 1. Prepared according to the mathematical descripArtificial neural networks [ 6 ] play a key role here, which tion of the algorithm; form the basis of machine learning [7, 8] as well as pro- 2. Optimized for the performance on our data set; vide a number of useful tools for detecting certain features [ 9, 10, 11, 12, 13 ]. Optimization processes often re- 3. Returns the most likely disease; quire the use of very efective tools which, thanks to the 4. Should be used easy to implement or use in the eficiency of modern computers, allow to imitate the be- front-end site; havior of the animal community that most often aims to obtain food [14, 15, 16] and proper healthcare [ 17, 18, 19 ]. 3. Dataset and Data processing An interesting and extremely useful task at the moment with the use of heuristic algorithms is the reduction of en- The data was taken from the Kaggle platform from a ergy consumption [ 20 ]. Contemporary IT solutions com- database called "Disease Symptom Prediction"[ 26 ]. At bining IoT and artificial intelligence methods increase the very beginning, We started to "clean up" and wrote a the quality of life [ 21, 22 ] facilitate care for the elderly script that removed duplicates from almost 5,000 records [ 23 ] and are also used to detect road damage [ 24, 25 ]. in the database and managed to extract 442 unique ones!

The government and health insurance providers might Therefore, even the most stupid and bad solutions on the gain from disease prediction, among other stakeholders. Kaggle platform gave 100% or very close to this result. Patients who are at risk for certain illnesses or disorders That is why in the end We also tested manually/visually can be identified. The quality of treatment can be im- to be sure that the results are not just empty percentages. proved and possible hospital admissions can be avoided if clinicians take the necessary steps to avoid or minimize the risk. Also in the age of Covid and virtual contact with doctors, it can be a good alternative for the first quick diagnosis. • We replace the symptoms with the corresponding numbers which are the indexes of the 2nd file so there is no risk of mixing things up. • The last and most dificult step (after the first KNN clusterization which will be discussed below) is to change the data frames so that the columns become symptoms and the values are 0 when there is no symptom and 1 when there is. • The graph on fig. 4 shows how symptoms are distributed according to disease.

4. K Nearest Neighbors Algorithm

The K Nearest Neighbors (KNN) algorithm is the simplest and slowest classification algorithm. classification. This becomes a problem when dealing with large data sets. Find the k nearest elements (neighbors) to a new element and assign this element to the group to which most of its neighbors. To improve the performance of the KNN algorithm a common technique is to standardize or normalize the data. Its application causes all dimensions for which the distance is calculated to have equal significance. Otherwise, a situation could arise in which a single dimension would dominate the other dimensions. measures. In this case, we have data that is based on symptoms and there is no strength of their or time of occurrence, so we don’t have to worry about that. The KNN algorithm uses metrics to determine the nearest neighbors. In this case, I used the Minkowski distance. (, ) = ︃( ∑︁ | − | )︃ 1

5. Soft classifier

In order to build a soft classifier we first need to take care of a suitable data format. Then, we create a dictionary from the columns with 0-1 values. The dictionary building itself is done as follows: the algorithm calculates average values for the given column and then checks how many values in the column are below and above the average for the given disease. Based on this data, it completes (1) the dictionary with values. The static method which is responsible for this is group class.

6. Decision tree

A decision tree is a supervised machine learning tool that may be used to classify or predict data based on how queries from the past have been answered. The model is supervised learning in nature, which means that it is trained and evaluated using data sets that include the required categorisation. The decision tree might not always ofer a simple solution or choice. Instead, it may provide the data scientist choices so they can choose wisely on their own. Decision trees mimic human thought processes, making it typically simple for data scientists to comprehend and evaluate the findings. A decision tree is drawn upside down with its root at the top. Each tree node can be split into branches. The end of the branch that doesn’t split anymore is the leaf.

7. Support Vector Machine

Algorithm 1: Data clustering algorithm.

Data: Input data set sample Data: Input data set data Data: Input k number of classifications Data: Input m Minkowski distance

Result: Class 1 Create classes from zeros dictionary ; 2 Create distances with an empty list ; 3 foreach x in range (0, len (data), 1) do 4 distances.append (minkowskiDistance (sample, data.iloc [x], m)) 5 end 6 data = data.assign (dist = distances) ; 7 data = data.sort_values(by=["dist"]) ; 8 data = data.drop (["dist"], axis = 1) ; 9 foreach i in range (0, k, 1) do 10 classes [data.iloc [i] ["Disease"]] + = 1 11 end 12 return max (classes, key = classes.get)

8.1. KNN algorithm pseudocodes

8. Algorithms

Support Vector Machine, or SVM, is a prominent Supervised Learning technique that is used for both classification and regression issues. However, it is mostly utilized Algorithm 2: Algorithm returning the accuracy in Machine Learning for Classification dificulties. The of the kNN classifier.

SVM algorithm’s purpose is to find the optimum line or Data: Test input decision boundary for categorizing n-dimensional space Data: Train input so that we may simply place fresh data points in the Data: Input k number of classifications proper category in the future. A hyperplane is the opti- Data: Input m Minkowski distance mal choice boundary. SVM selects the extreme vectors Result: Accuracy that aid in the creation of the hyperplane. These ex- 1 Create correct with the value 0 ; treme examples are referred to as support vectors, and 2 foreach i in range (0, len (test), 1) do the method is known as the Support Vector Machine. 3 if clustering (test.iloc [i], train, k, m) == test.iloc [i] .Disease then

correct + = 1 ; 4 5 end 6 return str (correct / len (test) * 100) + % ;

8.2. soft classifier pseudocodes

66–72 Algorithm 3: Algorithm for building a soft set.

Data: Input data set data

Data: Input symptoms dataset

Result: groupByClass soft dictionary 1 Create columnNames with column list from ; 2 Create uniqueClasses with a list of unique classes from ; 3 Create groupByClass with dictionary ; 4 Create a size with number of columns from ; 5 Create i ranging from 0 to the length of the unique classes ; 6 Create count with a value of 0 ; 7 Create a type with the dictionary ; 8 Create a cell from 0 to length ; 9 Create j ranging from 1 to the length of the columns of ; 10 foreach i in range (0, len (uniqueClasses), 1) do 11 count = 0 ; 12 type = symptoms [i]: 0 for i in range (len (symptoms)) foreach cell in range (0, len (data), 1) do 13 if data.at [cell, columnNames [0]] == uniqueClasses [i] then 14 count + = 1 ; 15 foreach j in range (1, size, 1) do 16 type [columnNames [j]] + = data.at [cell, columnNames [j]] ; 17

end 18 19 20 21 22 23 24 end 25 return groupByClass ; end if count == 0 then

continue ; type = k: v / count for k, v in type.items () ; ProcessingData.toOneOrZeroDict (type) ; groupByClass [i] = type ; 8.3.

Decision tree

For the decision tree, we use DecisionTreeClassifier from sklearn. The params were: nodes=40, criterion=’entropy’, random state=0, max depth=6, min samples leaf=1.

8.4. Support Vector Machine

For SVM we choose the SVC implementation for C-Support vector classification also from sklearn. The input params were: kernel=’linear’, C=1

9. Select the best algorithm

Algorithm Accuracy Time KNN v1 82.2% 10.4s KNN v2 100% 31.2s Soft classifier 97.74% 4.4s Decision tree 84,96% 0.7s SVM 87.22% 0.6s KNN v1 was KNN with dataset with orginal shape. KNN v2 was KNN with rashaped dataset.

Each classifier was tested 30 times to ensure that the

References

results we obtained were as near to their true accuracy neural networks applied for prediction of plant leaf as possible. KNN on a well-shaped data set was, in our diseases , Sensors 21 ( 2021 ) 4749 .

opinion, the best of the tested algorithms . It has the [7] A. T. Özdemir , B. Barshan , Detecting falls with best accuracy but takes a lot of time. The compromise wearable sensors using machine learning techbetween speed and accuracy was a decision tree. It has a niques , Sensors 14 ( 2014 ) 10691 - 10708 .

near 100% score and also good performance , we can also [8] K. G. Liakos , P. Busato , D. Moshou , S. Pearson, easily save weights for future use in the front-end. The D. Bochtis, Machine learning in agriculture: A quickest was SVM and the accuracy was not the worst review , Sensors 18 ( 2018 ) 2674 .

so if we need really quick classification we need to take [9 ] O. Dehzangi , M. Taherisadr , R. ChangalVala, ImuSVM into consideration. based gait recognition using convolutional neural networks and multi-sensor fusion , Sensors 17 ( 2017 ) 2735 .

10. Conclusion and future work [10] C. Napoli , G. De Magistris, C. Ciancarelli , F. Corallo , F. Russo , D. Nardi , Exploiting wavelet recurTo sum it all up, we have a wide range of algorithms to rent neural networks for satellite telemetry data choose and the proven algorithms performed well on the modeling, prediction and control, Expert Sysdataset . Further testing with more algorithms , or con- tems with Applications 206 ( 2022 ). doi: 10 .1016/ necting fast algorithms with some simple decision model, should be undertaken to increase their performance . Ex- [11] jH.. eGs.wHao.n2g0,2M2 ..1B1.7L8e3e , 1K .. R. Park, Convolutional panding the collection with new features and situations neural network-based finger-vein recognition using would be another way to improve the utility of our ef- nir image sensors , Sensors 17 ( 2017 ) 1297 .

fort. For mobile phones, performance is very important . [12] C. Ciancarelli , G. De Magistris, S. Cognetta , For example, we are not able to use KNN in some health D . Appetito , C. Napoli , D. Nardi , A gan aprealeted applications but we can use weights from soft proach for anomaly detection in spacecraft telemeclassifier or SVM . tries, Lecture Notes in Networks and Systems 531 LNNS ( 2023 ) 393 - 402 . doi: 10 .1007/ References 978-3- 031 -18050-7_ 38 . [13] G. Capizzi , G. Lo Sciuto , C. Napoli , E. Tramontana,

[1] Q.- b . Zhang, P. Wang , Z.-h. Chen , An improved M. Woźniak, A novel neural networks-based texparticle filter for mobile robot localization based on ture image processing algorithm for orange defects particle swarm optimization, Expert Systems with classification , International Journal of Computer Applications 135 ( 2019 ) 181 - 193 . Science and Applications 13 ( 2016 ) 45 - 60 .

[2] S. Illari , S. Russo , R. Avanzato , C. Napoli , A cloud- [14] T. Qiu , B. Li , X. Zhou , H. Song , I. Lee , J. Lloret, oriented architecture for the remote assessment A novel shortcut addition algorithm with particle and follow-up of hospitalized patients, in: CEUR swarm for multisink internet of things , IEEE TransWorkshop Proceedings , volume 2694 , 2020 , pp. 29 - actions on Industrial Informatics 16 ( 2019 ) 3566 - 35 . 3577 .

[3] M. A. Sanchez , O. Castillo , J. R. Castro , Generalized [15] V. Marcotrigiano , G. Stingi, S. Fregnan , P. Magatype-2 fuzzy systems for controlling a mobile robot relli , P. Pasquale, S. Russo , G. Orsi, M. Montagna , and a performance comparison with interval type- C. Napoli , C. Napoli , An integrated control plan 2 and type-1 fuzzy systems, Expert Systems with in primary schools: Results of a field investigaApplications 42 ( 2015 ) 5904 - 5914 . tion on nutritional and hygienic features in the

[4] G. Lo Sciuto , G. Susi, G. Cammarata, G. Capizzi , A apulia region (southern italy) , Nutrients 13 ( 2021 ). spiking neural network-based model for anaerobic doi : 10 .3390/nu13093006. digestion process, in: 2016 International Sympo- [16] M. Ren , Y. Song , W. Chu , An improved locally sium on Power Electronics, Electrical Drives, Au- weighted pls based on particle swarm optimization tomation and Motion (SPEEDAM) , IEEE, 2016 , pp. for industrial soft sensor modeling, Sensors 19 996 - 1003 . ( 2019 ) 4099 .

[5] C. Napoli , F. Bonanno , G. Capizzi, Exploiting solar [17] S. Russo , S. Illari , R. Avanzato , C. Napoli , Reducwind time series correlation with magnetospheric ing the psychological burden of isolated oncological response by using an hybrid neuro-wavelet ap- patients by means of decision trees , in: CEUR Workproach, Proceedings of the International astronom- shop Proceedings , volume 2768 , 2020 , pp. 46 - 53 . ical union 6 ( 2010 ) 156 - 158 . [18] G. Lo Sciuto , S. Russo , C. Napoli , A cloud-based

[6] V. S. Dhaka , S. V. Meena , G. Rani , D. Sinwar , M. F. lfexible solution for psychometric tests validation , Ijaz, M. Woźniak , A survey of deep convolutional administration and evaluation , in: CEUR Workshop Proceedings , volume 2468 , 2019 , pp. 16 - 21 .

[19] S. Russo , C. Napoli , A comprehensive solution for psychological treatment and therapeutic path planning based on knowledge base and expertise sharing , in: CEUR Workshop Proceedings , volume 2472 , 2019 , pp. 41 - 47 .

[20] M. Woźniak , A. Sikora , A. Zielonka , K. Kaur , M. S. Hossain , M. Shorfuzzaman , Heuristic optimization of multipulse rectifier for reduced energy consumption , IEEE Transactions on Industrial Informatics 18 ( 2021 ) 5515 - 5526 .

[21] M. Woźniak , A. Zielonka , A. Sikora , M. J. Piran , A. Alamri , 6g-enabled iot home environment control using fuzzy rules , IEEE Internet of Things Journal 8 ( 2020 ) 5442 - 5452 .

[22] G. C. Cardarilli , L. Di Nunzio , R. Fazzolari , D. Giardino , M. Re , A. Ricci , S. Spanò , An fpga-based multi-agent reinforcement learning timing synchronizer , Computers and Electrical Engineering 99 ( 2022 ) 107749 .

[23] M. Woźniak , M. Wieczorek , J. Siłka , D. Połap , Body pose prediction based on motion sensor data and recurrent neural network , IEEE Transactions on Industrial Informatics 17 ( 2020 ) 2101 - 2111 .

[24] M. Woźniak , A. Zielonka , A. Sikora , Driving support by type-2 fuzzy logic control model , Expert Systems with Applications 207 ( 2022 ) 117798 .

[25] L. Canese , G. C. Cardarilli , L. Di Nunzio , R. Fazzolari , H. Famil Ghadakchi , M. Re, S. Spanò, Sensing and detection of trafic signs using cnns: an assessment on their performance , Sensors 22 ( 2022 ) 8830 .

[26] P. Patil , Disease symptom prediction, 2020 . URL: https://www.kaggle.com/datasets/itachi9604/ disease -symptom-description-dataset?select= dataset .csv.