Predictive Tools to Evaluate Cardiovascular Events in Chronic Heart

Failure Patients

Maria Carmela Groccia

1 a

, Rosita Guido

1 b

, Domenico Conforti

1 c

and Angela Sciacqua

2 d

Department of Mechanical, Energy and Management Engineering, University of Calabria,

Ponte Pietro Bucci 41C, 87036 Rende (Cosenza), Italy

Department of Medical and Surgical Sciences, University Magna Graecia of Catanzaro, Italy

Keywords:

Predictive Models, Knowledge Discovery, Machine Learning, Chronic Heart Failure, Decision Tree,

Respiratory Rate.

Abstract:

In this paper, a Knowledge Discovery task has been implemented with the aim of developing models for

predicting cardiovascular worsening events in Chronic Heart Failure (CHF) patients. A set of patients suffering

from CHF were enrolled and carefully evaluated through a ﬁve-year follow-up. Several predictive models were

developed on the collected data and then compared. Among these, the decision tree based predictive model has

been analysed by clinical experts. The decision tree is among all the trained and tested models the most simple

and interpretable mainly by clinicians because it discovers if-then rules. The extracted rules are compliant

with previous clinical studies. Nevertheless, the decision tree achieved lower performance compared to the

other predictive models, which conversely to the decision tree are not “clinician friendly” because they do not

provide an explanation of the classiﬁcation decisions.

1 INTRODUCTION

Chronic Heart Failure (CHF) is a complex syndrome

caused by the inability of the heart to pump a suf-

ﬁcient amount of blood around the body. Typical

symptoms of CHF patients include breathlessness, fa-

tigue, and ankle swelling (McDonagh et al., 2021;

Morrissey et al., 2011). Based on the symptoms, the

New York Heart Association (NYHA) classiﬁcation

distinguishes the CHF patients in four classes; from

NYHA I, without any limitation to physical activity,

to NYHA IV, where patients have inability to carry

out any type of activity without discomfort (Bredy

et al., 2018). CHF confers high risk for cardiovas-

cular (CV) worsening events that cause recurrent hos-

pitalizations and high mortality rate even in patients

with mild symptoms (Dunlay et al., 2009). An early

prediction of CV worsening events could offer ben-

eﬁts for a preventive treatment, limit serious conse-

quences and improve the quality of care (Ponikowski

et al., 2014). Therefore, it could have a relevant ad-

https://orcid.org/0000-0001-7570-8458

https://orcid.org/0000-0003-1744-2166

https://orcid.org/0000-0002-4816-4333

https://orcid.org/0000-0003-1674-5317

vantage on the reduction of hospitalizations and asso-

ciated costs.

Machine Learning (ML) and Knowledge Discov-

ery (KD) techniques can be applied to allow early pre-

diction of CV worsening events. These methodolo-

gies allow to learn knowledge from past experiences

through identifying patterns in the data (Fayyad et al.,

1996). ML algorithms are able to automatically learn

these patterns from past data and apply it to future

predictions.

Previous research applied ML and KD tech-

niques in CHF domain for predicting adverse events

(Tripoliti et al., 2017; Groccia et al., 2018). In

(Tripoliti et al., 2017) a review of several models

for predicting the presence of adverse events, such

as destabilizations, re-hospitalizations, and mortality

is presented. In (Groccia et al., 2018) a temporal

weighting approach was applied to risk prediction of

major cardiovascular worsening events in CHF pa-

tients taking into account the chronology of events.

In this paper, a KD task has been designed and im-

plemented to extract new predictive models that can

help clinicians to early detect CV worsening events in

patients with CHF. The KD task was conducted on a

real dataset collected over a ﬁve years follow-up. Sev-

eral ML algorithms were trained obtaining different

Groccia, M., Guido, R., Conforti, D. and Sciacqua, A.

Predictive Tools to Evaluate Cardiovascular Events in Chronic Heart Failure Patients.

DOI: 10.5220/0010829900003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 5: HEALTHINF, pages 475-481

ISBN: 978-989-758-552-4; ISSN: 2184-4305

475

predictive models. Models performance have been

evaluated and compared. Decision tree based model

has been deeply analysed by clinicians exploiting the

possibility to extract simple rules compliant with the

previous clinical studies.

2 METHODS

A KD task has been designed and implemented to

analyse the collected data and to develop models for

predicting CV worsening events in CHF patients. The

KD analysis was deﬁned as a predictive task stated as

supervised binary classiﬁcation problem.

Supervised ML approaches learn from a given

dataset a function f that predicts an output variable

(or class label) y from a feature vector x containing

N input variables, such that y = f (x) (Mitchell, 1997;

Jo, 2021). In the considered classiﬁcation problem,

the class label is a categorical variable with only two

values.

2.1 Collected Data

The collected real dataset contains clinical informa-

tion of 50 patients with an established diagnosis of

CHF and NYHA classes I, II and III. The data was

collected at the CHF ambulatory of the Geriatrics Di-

vision at the “Mater Domini” University Hospital in

Catanzaro, Italy. Patients were followed up every 3

months on an outpatient basis for an average of ﬁve

years. All patients gave their availability and written

consent for participation at the pilot study.

At ﬁrst outpatient visit, personal data and medical

history including date of birth, gender, NYHA class,

etiology, cardiovascular history, use of medications,

other diseases were collected. At each outpatient

visit, vital signs such as Heart Rate (HR), Body Tem-

perature (BT), Systolic Blood Pressure (SBP), Dias-

tolic Blood Pressure (DBP), Respiratory Rate (RR)

and weight were recorded. Dates of speciﬁc events

(i.e., date of the visits, date at which a CV wors-

ening event occurred) were reported too. During

the follow-up, 19 patients presented a CV worsening

event. Among these, 8 patients had more than one

event. Demographic and clinical characteristics of the

patients are summarized in Table 1.

2.2 Preprocessing

In the original format, the dataset is organized in a

wide format and consists of 50 rows. Each row con-

tains, in its ﬁrst columns, the patients data that don’t

change across time: personal data and medical his-

tory. The remaining columns contain the clinical pa-

rameters recorder at each visit and the dates of spe-

ciﬁc events. To perform the classiﬁcation task, the

dataset was converted in a long format. In this for-

mat, each patient has data in multiple rows. Each row

represents a patient’s visit.

The class label was deﬁned based on the occur-

rence or not of CV worsening events between two

consecutive outpatient visits. An instance is a row of

the dataset. 31 instances were designated as positive

Table 1: Demographic and clinical characteristics of the pa-

tients. NYHA: New York Heart Association; CHF: chronic

heart failure; PTCA: percutaneous transluminal coronary

angioplasty; ICD: implantable cardioverter deﬁbrillator;

TIA: transient ischaemic attack; COPD: chronic obstructive

pulmonary disease; CV:cardiovascular.

Characteristic

All patients

n=50

Age (years ± SD) 72.5 ± 14.2

Gender

Male 36 (72%)

Female 14 (28%)

NYHA Class

I 3 (6%)

II 38 (76%)

III 9 (18%)

CHF etiology

Ischemic heart disease 23 (46%)

Idiopathic dilatation 9 (18%)

Hypertension 4 (8%)

Valvular diseases 8 (16%)

Valvular diseases + Hypertension 4 (8%)

Alcoholic habit 2 (4%)

Cardiovascular history

Instable angina 1 (2%)

PTCA 1 (2%)

By-pass 7 (14%)

Atrial ﬂutter 13 (26%)

Pacemaker 3 (6%)

Cardiac resynchronization 1 (2%)

ICD 2 (4%)

Mitral insufﬁciency 21 (42%)

Aortic insufﬁciency 4 (8%)

Hypertension 28 (56%)

TIA 2 (4%)

Other diseases

Diabetes 11 (22%)

Hypothyroidism 1 (2%)

Renal failure 4 (8%)

COPD 5 (10%)

Asthma 1 (2%)

Sleep apnea 4 (8%)

Pulmonary ﬁbrosis 1 (2%)

Gastrointestinal diseases 4 (8%)

Hepatic diseases 3 (6%)

CV worsening events 19 (38%)

HEALTHINF 2022 - 15th International Conference on Health Informatics

476

instances (patients with CV worsening events) and the

remaining 762 instances were designated as negative

instances (patients without events).

Input errors were corrected and a new variable that

contains the age of patients at admission was created.

The dataset is imbalanced because the number of

instances of patients with CV worsening events is

much lower than the instances of the patients without

events.

With the aim to create predictive models that use

few and simple clinical parameters, only the vital

signs measured at each outpatient visit were included

in the training set. Clinical parameters, i.e., HR, RR,

DPB, and SBP are used both to monitoring CHF and

as a primary tool regarding patient status.

The dataset was randomly divided into training

(70%) and test (30%) set. The training set was used to

build the predictive models. The test set instead, was

used to evaluate the performance of each model on

unseen data. A resampling approach has been adopted

to balance the classes in the training set.

2.3 Models Building

Several ML algorithms such as Support Vector Ma-

chine (SVM), Artiﬁcial Neural Network (ANN),

ıve Bayes, Decision Tree and Random Forest were

implemented to develop the predictive models.

SVM is a classiﬁer based on statistical learning theory

(Cortes and Vapnik, 1995; Burges, 1998). It searches

for an optimal hyperplane, in an N-dimensional space,

that separates patterns of classes by maximizing the

margin. In non-linearly separable dataset, the SVM

maps inputs into high-dimensional feature spaces us-

ing a kernel function in order to transform it in a lin-

ear separable dataset. The most popular kernels used

in SVM classiﬁcation tasks are polynomial kernels

and Radial Basis Function (RBF), also called Gaus-

sian kernels.

ANN is a computational model, consisting of a num-

ber of artiﬁcial neural units called perceptron. They

emulate biological neural networks (Krenker et al.,

2011). In this work, we used a type of a fully con-

nected, feed-forward artiﬁcial neural network named

Multilayer Perceptron (MLP). MLP consists of neu-

rons arranged in layers: one input layer, one output

layer, and one or more hidden layers.

ıve Bayes is a probabilistic classiﬁcation algorithm

based on the Bayes Theorem with strong (na

ıve) in-

dependence assumptions between the features (Rish,

2001). This algorithm is based on the assumption that

a particular feature in a class is unrelated to the pres-

ence of any other feature.

Decision Tree is a non-parametric supervised learn-

ing method (Quinlan, 1986). The goal is to create

a model that predicts the value of a target variable

by learning simple decision rules inferred from the

data features. A Decision Tree consists of nodes and

branches. Each node represents an input attribute and

a split point on that attribute. The leaf nodes contain

an output attribute which is used to make a prediction.

Given a new input, the tree is traversed by evaluating

the speciﬁc input started at the root node of the tree.

One of their main advantage is that they are simple to

understand and interpret, and they can be visualised.

Random forest (Breiman, 2001) consists of individ-

ual decision trees that operate as an ensemble. Each

tree is built by applying bagging, which is the general

technique of bootstrap aggregating. A simple major-

ity vote of all trees gives the ﬁnal result. The inter-

pretability of a single decision tree is lost in random

forest because many decision trees are aggregated.

The tuning model hyper-parameters has been opti-

mized using a 5-fold cross validation. In k fold cross

validation, the dataset is split randomly into k equal

sized folds. K iterations are performed and, at each

iteration one of the k folds is used as the validation

set while all remaining folds are used as the training

set. In this process each instance is used for testing

exactly once. The resampling was performed only on

the folds used as training data as discussed in (Santos

et al., 2018).

Table 2: Hyper-parameters tuning. LR: Learning Rate;

Mom: Momentum; Ep: No. epochs; CF: Conﬁdence Fac-

tor; Iter: No. iterations.

Model

Hyper-

Param

space

Step

Best

value

SVM

linear kernel

C [1,5] 1 1

SVM

poly kernel

d [1,5] 1 3

C [1,10] 0.5 10

SVM

RBF kernel

γ [0.01,1.00] 0.01 0.03

C [1,10] 0.5 10

ANN

LR [0.1,1.0] 0.1 0.3

Mom [0.1,1.0] 0.1 0.2

Ep [400,600] 100 500

Decision

tree

CF [0.10,0.50] 0.05 0.25

Random

Forest

Iter [100,400] 100 100

The hyper-parameters were optimized by search-

ing the best value in a deﬁned range for each ML

model. Table 2 reports the hyper-parameters tuning of

the tested algorithms. In the third column of the table

Predictive Tools to Evaluate Cardiovascular Events in Chronic Heart Failure Patients

477

there is the range of values deﬁned as search space.

The incremented value is denoted in the fourth col-

umn as Step. The last column shows the best value.

We tested SVM with three kernel functions, i.e., lin-

ear kernel, polynomial kernel, and RBF kernel.

Waikato Environment for Knowledge Analysis

(WEKA) software, version 3.8.2, was used to build

the predictive models by using classiﬁcation algo-

rithms (Eibe et al., 2016). We used SMO (Sequential

Minimal Optimization) algorithm for SVM and J48

for decision tree.

2.4 Models Evaluation

The Area under the ROC curve (AUC), sensitivity,

speciﬁcity, and Geometric Mean (G-mean) are used

to evaluate and compare the predictive performance

of the build ML models on the test set.

The confusion matrix was used to deﬁne the met-

rics discussed in this section. Table 3 shows the struc-

ture of confusion matrix. Let P and N, be the num-

ber of positive and negative instances, respectively.

T P and T N, are the number of instances correctly

predicted as positive and negative, respectively; FP

and FN are the number of instances predicted as pos-

itive and negative whereas they belong to the opposite

class, respectively.

Table 3: Confusion Matrix for binary classiﬁer.

predicted

positive

predicted

negative

actual positive TP FN

actual negative FP TN

AUC: measures the classiﬁer’s ability to avoid false

classiﬁcation. It is the area under the curve of the true

positive ratio vs. the false positive ratio and indicates

the probability that the model will rank a positive case

more highly than a negative case.

Sensitivity: measures the proportion of positive in-

stances that are correctly identiﬁed, i.e, it is the ability

to predict a CV worsening event. It is deﬁned as

Sens =

T P

T P + FN

Speciﬁcity: measures the proportion of negatives that

are correctly identiﬁed, i.e, it is the ability to predict

patients without CV worsening events. It is deﬁned as

Spec =

T N

T N + FP

G-mean: takes into account the balance of the classi-

ﬁer’s performance on the two classes. It is deﬁned as

the geometric mean of sensitivity and speciﬁcity as

G-Mean = sqrt(Sens ∗ Spec)

3 RESULTS AND DISCUSSION

Table 4 shows the performance of the predictive mod-

els on the test set.

Table 4: Results of the predictive models on the test set.

Model AUC Spec Sens G-mean

SVM

linear kernel

0.77 0.79 0.75 0.77

SVM

poly kernel

0.70 0.90 0.50 0.67

SVM

radial kernel

0.77 0.79 0.75 0.77

Naive Bayes 0.73 0.83 0.50 0.64

ANN 0.81 0.86 0.50 0.66

Random Forest 0.73 0.93 0.12 0.34

Decision Tree 0.57 0.89 0.25 0.47

Following the speciﬁc indications of the clinical

domain experts’ involved in this study, a deep clinical

assessment of the decision tree model has been devel-

oped since it is easier to understand by clinicians.

The decision tree extracted from the KD task has

undergone a post-processing process by the support of

clinical domain experts in order to extract few simple

rules that could be directly used by clinicians in their

daily practice. The tree is reported in Figure 1. The

label 0 identiﬁes the absence of risks, while the label

1 identiﬁes risk of CV worsening events.

The three rules constructed by the decision tree

to predict CV worsening events in CHF patients are

described below.

Rule 1. The ﬁrst rule suggests a test on RR (root

node). The tree identiﬁes the RR as the most rele-

vant parameter for predicting an event. In particular,

a CHF patient with RR greater than 20 apm may have

an event.

Rule 2. The second rule suggests a new event for

patients with a RR ≤ 20 apm, HR > 88 bpm and

DBP ≤ 72 mmHg.

Rule 3. If RR ≤ 20 apm and HR ≤ 88 bpm, the

third rule suggests reconsidering again these values.

If 56 < HR ≤ 75 bpm with a RR ≤ 18 apm and

DBP ≤ 52 mmHg, a new event may occur with any

SBP value.

Regarding the third classiﬁcation rule, the SHIFT

study (Borer et al., 2012) already showed how HR de-

crease induced by ivabradine led to a decrease in hos-

pitalization rate in patients suffering from CHF with

reduced systolic function. Therefore, the model is

consistent with currently published data. While the

HR cut-off was not foreseen in the study, the classi-

ﬁcation rule of our model identiﬁes what is the HR

HEALTHINF 2022 - 15th International Conference on Health Informatics

478

Figure 1: Decision tree.

value associated with an acute destabilization event.

The clinical importance that emerged through this

model for the RR is undoubtedly interesting. With

regard to the algorithms published in the past, a re-

cent meta-analysis suggests that weight was the most

used parameter to monitor CHF patients (Klersy et al.,

2009). According to our analysis, weight does not

give meaningful contributions to the model and was

eliminated from the classiﬁcation tree. Generally, in

this type of model RR is not taken into consideration.

Yet RR is an excellent clinical indicator, not only for

respiratory system, but also for hemodynamic equilib-

rium. Indeed, worsening of respiratory diseases neg-

atively affects cardiac function. On the other hand,

during the early stages of decompensation lung inter-

stitial congestion can occur, thereby triggering the ac-

tivation of J receptors which in turn stimulates pul-

monary ventilation. It is clinically relevant to take

this variable into consideration for the prediction of

a worsening event. In this model, RR is considered

in more than one classiﬁcation rule. In particular, the

ﬁrst rule considers exclusively if RR is higher than 20

acts per minute. From a clinical point of view, RR

is really useful since tachypnea is an indicator of CV

distress, which arises as an attempt of compensatory

mechanism that in the long term further promotes de-

compensation.

Moreover, the second classiﬁcation rule is based

on a group of parameters, including RR, HR and DBP.

Of course, all these considerations should not come

as a surprise, because clinical parameters can always

be affected by compensatory mechanisms. The latter

are initially crucial to maintain an adequate cardiac

output; however, compensatory mechanisms can con-

tribute to further worsen decompensation. We take

into consideration the activation of both sympathetic

nervous system and renin-angiotensin system. In ad-

dition, we recall the importance of HR as an early

clinical indicator of decompensation, closely associ-

ated with sympathetic activation and with the conse-

quent positive chronotropic effects expressing an ini-

tial compensation mechanism, which however favors

a further worsening of the overall clinical status. In

fact, HR is the main inducer of myocardial oxygen

consumption, which can concur to promote decom-

pensation. These concepts explain why the use of

beta-blockers is a cornerstone of CHF treatment.

Finally, it is useful to discuss the role of DBP. In-

deed, the second and the third classiﬁcation rules im-

ply that an event is predicted when the DBP is ≤ 72

mmHg and ≤ 52 mmHg, respectively. Also in this

case the system is in agreement with clinical litera-

ture. A previous Japanese study (Tsujimoto and Ka-

jio, 2018) had shown that a low DBP value was as-

sociated with the onset of CV events and with an in-

crease in the number of hospitalizations due to HF.

Notably, in this study DBP values < 70 mmHg were

taken into consideration, thereby indicating them as

more unfavorable than the 80 − 89 mmHg range.

Despite the knowledge extracted with the deci-

sion tree is in a form easily to be interpreted and the

extracted rules are compliant with previous clinical

Predictive Tools to Evaluate Cardiovascular Events in Chronic Heart Failure Patients

479

studies, this predictive model has lower performance

than other models. It has low sensitivity although the

reasons behind the low sensitivity could be linked to

the low sample size.

The SVM with linear and radial kernel had

the best performance on the test set for predict-

ing CV worsening events in CHF patients: AUC =

0.77, speci f icity = 0.79, sensitivity = 0.75 and G −

mean = 0.77. This indicate that SVM has a high abil-

ity to avoid false classiﬁcation.

4 CONCLUSIONS

This work presents and compares predictive models

based on ML algorithms for the early prediction of

CV worsening events in CHF patients using few clin-

ical parameters.

Among the predictive models, SVM with linear

and radial kernel had the best performance on the test

set. As we showed, the decision tree is among all

the trained and tested models the most simple and in-

terpretable mainly by clinicians because it discovers

if-then rules as clinicians do. Conversely, although

SVM has the best performances, it is not “clinician

friendly”. The inability of SVMs in providing a sim-

ple and understandable interpretation of the classiﬁca-

tion decisions is one of the main obstacles impeding

their application in the clinical practice.

As future work, we will reproduce these experi-

mental models on a larger size study and a shorter in-

terval occurring between two consecutive visits. A

follow-up based on shorter intervals could increase

both sensitivity and speciﬁcity of the models. Another

possible application could consider remote monitor-

ing, with the active help of either the patients them-

selves or their caregivers, to intervene as early as pos-

sible to avoid events and subsequent hospitalization.

In addition, techniques for rule extraction from SVM

could be adopted to ameliorate the aforementioned is-

sue.

ACKNOWLEDGEMENTS

This work has been partially supported by the indus-

trial research and development project “HEARTNET-

ICS - Advanced Analytics for Heart Diseases Man-

agement” (European Regional Development Fund,

Calabria Region Grant J58C17000150006, Italy,

2017-2020)

REFERENCES

Borer, J. S., B

ohm, M., Ford, I., Komajda, M., Tavazzi,

L., Sendon, J. L., Alings, M., Lopez-de Sa, E., Swed-

berg, K., and Investigators, S. (2012). Effect of ivabra-

dine on recurrent hospitalization for worsening heart

failure in patients with chronic systolic heart failure:

the shift study. European heart journal, 33(22):2813–

2820.

Bredy, C., Ministeri, M., Kempny, A., Alonso-Gonzalez,

R., Swan, L., Uebing, A., Diller, G.-P., Gatzoulis,

M. A., and Dimopoulos, K. (2018). New york heart

association (nyha) classiﬁcation in adults with con-

genital heart disease: relation to objective measures

of exercise and outcome. European Heart Journal-

Quality of Care and Clinical Outcomes, 4(1):51–58.

Breiman, L. (2001). Random forests. Mach Learn, 45:5–32.

Burges, C. J. (1998). A tutorial on support vector machines

for pattern recognition. Data mining and knowledge

discovery, 2(2):121–167.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine learning, 20(3):273–297.

Dunlay, S. M., Redﬁeld, M. M., Weston, S. A., Therneau,

T. M., Hall Long, K., Shah, N. D., and Roger, V. L.

(2009). Hospitalizations after heart failure diagnosis:

a community perspective. Journal of the American

College of Cardiology, 54(18):1695–1702.

Eibe, F., Hall, M. A., and Witten, I. H. (2016). The weka

workbench. online appendix for data mining: practi-

cal machine learning tools and techniques. In Morgan

Kaufmann.

Fayyad, U., Piatetsky-Shapiro, G., and Smyth, P. (1996).

The kdd process for extracting useful knowledge from

volumes of data. Communications of the ACM,

39(11):27–34.

Groccia, M. C., Lofaro, D., Guido, R., Conforti, D., and

Sciacqua, A. (2018). Predictive models for risk as-

sessment of worsening events in chronic heart failure

patients. In 2018 Computing in Cardiology Confer-

ence (CinC), volume 45, pages 1–4. IEEE.

Jo, T. (2021). Machine Learning Foundations: Supervised,

Unsupervised, and Advanced Learning. Springer Na-

ture.

Klersy, C., De Silvestri, A., Gabutti, G., Regoli, F., and Au-

ricchio, A. (2009). A meta-analysis of remote moni-

toring of heart failure patients. Journal of the Ameri-

can College of Cardiology, 54(18):1683–1694.

Krenker, A., Be

ster, J., and Kos, A. (2011). Introduction

to the artiﬁcial neural networks. Artiﬁcial Neural Net-

works: Methodological Advances and Biomedical Ap-

plications. InTech, pages 1–18.

McDonagh, T. A., Metra, M., Adamo, M., Gardner, R. S.,

Baumbach, A., B

ohm, M., Burri, H., Butler, J.,

Celutkien

e, J., Chioncel, O., Cleland, J. G. F., Coats,

A. J. S., Crespo-Leiro, M. G., Farmakis, D., Gi-

lard, M., Heymans, S., Hoes, A. W., Jaarsma, T.,

Jankowska, E. A., Lainscak, M., Lam, C. S. P., Lyon,

A. R., McMurray, J. J. V., Mebazaa, A., Mindham,

R., Muneretto, C., Francesco Piepoli, M., Price, S.,

Rosano, G. M. C., Ruschitzka, F., Kathrine Skibelund,

HEALTHINF 2022 - 15th International Conference on Health Informatics

480

A., and Group, E. S. D. (2021). 2021 ESC Guidelines

for the diagnosis and treatment of acute and chronic

heart failure: Developed by the Task Force for the di-

agnosis and treatment of acute and chronic heart fail-

ure of the European Society of Cardiology (ESC) With

the special contribution of the Heart Failure Associ-

ation (HFA) of the ESC. European Heart Journal.

ehab368.

Mitchell, T. (1997). Machine Learning. McGraw Hill.

Morrissey, R. P., Czer, L., and Shah, P. K. (2011). Chronic

heart failure. American Journal of Cardiovascular

Drugs, 11(3):153–171.

Ponikowski, P., Anker, S. D., AlHabib, K. F., Cowie, M. R.,

Force, T. L., Hu, S., Jaarsma, T., Krum, H., Rastogi,

V., Rohde, L. E., et al. (2014). Heart failure: prevent-

ing disease and death worldwide. ESC heart failure,

1(1):4–25.

Quinlan, J. (1986). Induction of decision trees. Mach Learn,

1:81–106.

Rish, I. (2001). An empirical study of the naive bayes clas-

siﬁer. Technical report.

Santos, M. S., Soares, J. P., Abreu, P. H., Araujo, H., and

Santos, J. (2018). Cross-validation for imbalanced

datasets: avoiding overoptimistic and overﬁtting ap-

proaches [research frontier]. ieee ComputatioNal iN-

telligeNCe magaziNe, 13(4):59–76.

Tripoliti, E. E., Papadopoulos, T. G., Karanasiou, G. S.,

Naka, K. K., and Fotiadis, D. I. (2017). Heart fail-

ure: diagnosis, severity estimation and prediction of

adverse events through machine learning techniques.

Computational and structural biotechnology journal,

15:26–47.

Tsujimoto, T. and Kajio, H. (2018). Low diastolic blood

pressure and adverse outcomes in heart failure with

preserved ejection fraction. International journal of

cardiology, 263:69–74.

Predictive Tools to Evaluate Cardiovascular Events in Chronic Heart Failure Patients

481