Heart diseases lead to disorders of the entire circulatory system. The risk of their occurrence depends on age. They are caused by poor diet, obesity, excessive stress, smoking, lack of physical activity or genetic predisposition. Heart diseases and circulatory system are the biggest threat to life. The development of technology, the use of appropriate algorithms along with relevant data allows us to increase the chances of detecting more quickly whether a patient has heart disease using given patient parameters. This work concerns the creation of a classification model for the risk of heart disease. It works based on sex, age, blood pressure, cholesterol etc. The model works with data from a data set Heart Failure Prediction Dataset [1] downloaded from kaggle. The best model was selected after a series of comparative tests of diferent classification systems.
Artificial intelligence plays an increasingly important
role in the operation of information systems [
Heart diseases lead to disorders of the entire circulatory system. The risk of their occurrence depends on age. They are caused by poor diet, obesity, excessive stress, smoking, lack of physical activity or genetic predisposition. Heart diseases and circulatory system are the biggest threat to life. The development of technology, the use of appropriate algorithms along with relevant data allows us to increase the chances of detecting more quickly whether a patient has heart disease using given patient parameters.
Each of the algorithms should be prepared to meet the following criteria: 1. Prepared according to the mathematical description of the algorithm; 2. Optimized for the performance on our data set; 3. Categorizes the case from the test set, determines whether the case has a heart disease or not; 4. Minimize the number of false negatives cases; 5. Returns an array containing information about: • Whether the data supplied to the function
was normalized; • The number of correct predictions for each
of the cases; • The number of incorrect predictions for
each of the cases; • Length of the computed set; • Confusion matrix in the following schema:
[[TP, FN], [FP, TN]] 6. Based on the selected algorithm, creation a client application for the assessment of the risk of the heart disease.
The set consists of 918 rows and 12 columns (11 features), 3 of them are categorical, 3 are factorial and 5 are numerical. A detailed description is provided below: 1. Age - Age of the examined person in years; 2. Sex - Sex of the examined person, originally the 3.2. Statistical data for the primary data 3. ChestPainType - Type of chest pain, originally the table contains markings: 1; 2; • TA (eng. Typical Angina) - changed to 0; • ATA (eng. Atypical Angina) - changed to • NAP (eng. Non-Anginal Pain) - changed to • ASY (eng. Asymptomatic) - changed to 3; pressed in mm/Hg; mmol/dl; pressed in mg/dl: 4. RestingBP - (eng.Resting Blood Pressure) - ex5. Cholesterol - Serum cholesterol expressed in 6. FastingBS - (eng. Fasting Blood Sugar) - ex• 1 - if the test level was above 120 mg/dl; • 0 - if it was below the mentioned level; 7. RestingECG - (eng. Resting Electrocardiogram Results) - original table contains markings: 8. MaxHR - (eng. Maximum Heart Rate) - highest heart rate recorded, a numerical value between 60 and 202 beats per minute; 9. Exercise Angina: • Y - changed to 1; • N - changed to 0; 10. Oldpeak - ST segment depression caused by physical action, numerical value; 11. ST_Slope - (eng. The slope of the ST section), originally the table contains the indications: • Up (eng. Upsloping - ST section elevation) - changed to 0;
set The data set we worked on was created by combining five other data sets containing data on heart disease. By merging and generalizing them, one of the largest publicly available data sets of its kind was created.
• Cleveland: 303 observations • Hungarian: 294 observations • Switzerland: 123 observations • Long Beach VA: 200 observations • Stalog (Heart) Data Set: 270 observations
There are some inaccuracies in the reported resting blood pressure value, which is 0 [mm/Hg]. This may suggest that some data have been misspelled. In the case of blood pressure, the value of 0 occurs in only one case for a 55-year-old man, according to an article published in the Toronto Metropolitan University Pressbooks the normal average value for his age group is between 110 and 145 [mm/Hg], so we omitted this result from our work. A value equal to 0 [mmol/dl] is also found for Cholesterol, but here it appears 172 times, which rather rules out confusion for us, it is more likely that its level was indeterminate. Another noticeable fact is the age of the respondents. The average is 53 years old, there are only 80 people under 40 years old (8.72% of the respondents), this shows that the risk of heart disease symptom increases dramatically with increasing age.
The pair plot on figure 1 shows that there are not many outliers (now clearly a 55-year-old man with an outlier for resting blood pressure). Green points indicate cases with confirmed hearing diseases. The pairs plot shows that there are not many outliers (now clearly a 55-yearold man with an outlier for resting blood pressure).
The correlation heat map on figure 2 shows a negative relationship between the presence of heart disease and maximum heart rate, and a similar relationship is found between heart rate and age. On the other hand, we can observe a positive correlation between resting blood pressure and age, between age and the occurrence of heart disease (we have already noticed this feature during the statistical analysis of the data), the last clear positive relationship is shown by the value of cholesterol with the value of maximum heart rate.
The pie chart on figure 3 shows the overall distribution of heart disease. It can be seen here that as many as one in five subjects was eventually diagnosed. When creating the test sets, we tried to get as close as possible to the percentage distribution of cases so that our test set would match the total as closely as possible, we tried to keep a diference of 1-2 percentage points from the total when generating the test set.
The box plot on figure 4 shows us the distribution of heart disease incidence by gender and age, here we see that many more early positive cases have been recorded among men, in ladies the threshold at which the chance increases is much higher.
The first action taken to process the data was to check the completeness of the collection. It contained 918 rows, each of which contained complete data, which was checked with the appropriate function.
The next step was to convert the factored values Y/N, M/F to 1 and 0, respectively. Then categorical values were converted to numeric using our own function category_to_numeric(). In order to conduct a comparison of how the algorithms perform on the normalized and non-normalized data set, a copy of the data set was written (non-standardised - heart_nnorm.csv). The columns that can be normalized are Age, RestingBP, Cholesterol, MaxHR and Oldpeak. The operation was performed using a custom function - normalize() using the min-max formula. After the data was normalized, it was written to a file heart_norm.csv for further use.
It was decided to create three algorithms from scratch: • KNN - an algorithm that checks the nearest neighbors of a case and determines what class its neighbors are. On this basis, the algorithm deduces what class is the searched case. The experimental method was used to determine the hyperparameter k, its value was set to the 5 nearest neighbors. • Classification using soft sets - the data set was adapted for use, then using our own functions we checked which category best fits the processed case. • Naive Bayes classifier - it is a simple probabilistic model based on the assumption of mutual independence of predictors. Vectors of mean values and standard deviations were calculated. On this basis a function was built to predict the class of cases.
2. Sensitivity recall (SEN): 3. Specificity (SPE): 4. Precision (PRE): = = =
( + )
( + )
( + ) 1 = (2 * * )
( + ) 5. F-score (F1), a measure of the accuracy of the test; 184
Actual condition
To make the results obtained on the normalized and unnormalized set using a single algorithm as good as possible for comparison we presented them in pairs on one page.
Algorithm execution time: 18 sec, 10.10 elem/s
Algorithm execution time: 19 sec, 9.37 elem/s 184 Predicted condition Compared to the version running on normalized data, coAncdtiutaioln 899FTPP 7106 TFNN we can see much larger diferences in quality measures.
The overall performance is much worse. Such a diference is caused by the lack of normalization, to which In each of the measured measures, we see quite high per- this algorithm is very sensitive. This is because features centages, none of the determinants difers significantly with larger values will completely cover those with small from the others. A high PRE value means that the classifier has done well to identify true positive cases. Ad- vraanlugees.28F.o0r-e7x7a.m0)pwlei,llthheavfeeaatumreucAhghei(gohpeerrawtienigghotnthtahne ditionally, which was extremely important for us, the algorithm classified only 16 cases as false negative. In FastingBS, which ranges from 0.0 to 1.0. the case of KNN, we used the Minkowski metric with m = 2.
(1) (2) (3) (4) (5)
Algorithm execution time: 23 sec, 7.76 elem/s
184
Actual condition The time is much longer than the KNN algorithm, the results are worse than KNN. The SPE and PRE indices stand out the most. The PRE index is at the highest level among all the algorithms, which tells us that this version of the algorithm did very well with true positive cases, as it classified as many as 96 of them. However, the number of 41 false negative cases is unacceptable.
Algorithm execution time: 18 sec, 9.72 elem/s 184
Actual condition
Predicted condition 84 TP 20 FN 23 FP 57 TN Compared to the normalized version, the results again show large diferences between each other. The overall trend is worse in all indicators. Of note is the rather low execution time of the algorithm.
Both Bayes classifiers do not have a given time because their execution is almost instantaneous. This is due to the 1–8
184
Actual condition lack of need to calculate the mean value and standard deviation each time, which we can calculate only once and very quickly using the built-in functions of the pandas package. For the normalized data we see very similar results in each of the indicators but they are slightly lower than those in KNN. None of them deviate from the rest.
For non-normalized data, the scores obtained by this algorithm are very low. This is similarly to the KNN case due to incorrectly set weights on the learning set. Due to such low scores, this version of the naive algorithm will not be considered further.
Each of the classifiers was used 50 times (the same normalized and non-normalized test set was always used during a single trial), so that the results we obtained were as close as possible to their actual accuracy. In our opinion, the best among the tested algorithms was KNN on normalized data. It features very high quality scores and very high closeness. It found slightly fewer TP cases than the algorithm operating on soft sets, but the number of FN cases is significantly lower which is very important for disease classification. I would say it is better to make a false positive diagnosis and refer for further testing than to make a false negative diagnosis. Ultimately this argument convinced us to choose KNN algorithm as the best algorithm to predict the occurrence of heart disease in our database. It is also very important that the data on which the model would work in the future should be normalized, without this the efectiveness of the algorithm will fall dramatically.
The application was written based on the flask library, due to its easy and light-to-use structure. The application is created on the basis of the website using the Bootstrap 10. Conclusion and future work In summary, the design assumptions were successfully met, the proven algorithms were successful on the dataset. To improve their performance, further tests should be performed with more algorithms, or with two or more algorithms simultaneously. An additional factor to increase the usefulness of our work would be to expand the set with more features and cases. With more variety and quantity, the algorithm could learn better to identify classes and calculate the risk of heart disease.