Article ID: CJ-21-0131
Background: The heterogeneity of B-type natriuretic peptide (BNP) levels among individuals with heart failure and preserved ejection fraction (HFpEF) makes predicting the development of cardiac events difficult. This study aimed at creating high-performance Naive Bayes (NB) classifiers, beyond BNP, to predict the development of cardiac events over a 3-year period in individual outpatients with HFpEF.
Methods and Results: We retrospectively enrolled 234 outpatients with HFpEF who were followed up for 3 years. Parameters with a coefficient of association ≥0.1 for cardiac events were applied as features of classifiers. We used the step forward method to find a high-performance model with the maximum area under the receiver operating characteristics curve (AUC). A 10-fold cross-validation method was used to validate the generalization performance of the classifiers. The mean kappa statistics, AUC, sensitivity, specificity, and accuracy were evaluated and compared between classifiers learning multiple factors and only the BNP. Kappa statistics, AUC, and sensitivity were significantly higher for NB classifiers learning 13 features than for those learning only BNP (0.69±0.14 vs. 0.54±0.12 P=0.024, 0.94±0.03 vs. 0.84±0.05 P<0.001, 85±8% vs. 64±20% P=0.006, respectively). The specificity and accuracy were similar.
Conclusions: We created high-performance NB classifiers for predicting the development of cardiac events in individual outpatients with HFpEF. Our NB classifiers may be useful for providing precision medicine for these patients.
The clinical outcomes of heart failure with preserved ejection fraction (HFpEF) remain poor,1,2 and the comorbidities of this syndrome increase the severity of the disease.3–10 To prevent the deterioration of HFpEF, physicians need to clarify the pathophysiology and exacerbating factors, predict aggravation, and individualize therapeutic strategies for each patient with HFpEF; therefore, precision medicine for HFpEF is recommended.11 If exacerbation of HFpEF can be predicted in the outpatient clinic, it may be prevented. The Meta-Analysis Global Group in Chronic Heart Failure (MAGGIC) risk model derived from a large international database of patients with HF is a representative model for predicting the 1- and 3-year mortality rates of patients with HF; however, most of the enrolled patients had HF with reduced EF (HFrEF) or mid-range EF (HFmrEF).12 Rich et al validated the MAGGIC score in patients with HFpEF and reported that its performance was the same as that of B-type natriuretic peptide (BNP).13 In patients with HFpEF, BNP is easy to use as a predictor, but in that study13 its performance was only fair (area under the receiver operating characteristics curve (AUC) was 0.74). The BNP level in patients with HFpEF can be lower than that in those with HFrEF under the condition of compensated HF, and it is also influenced by several other factors.14–16 The heterogeneity of BNP levels among individuals with HFpEF can make interpreting results difficult15 and may influence the predictability for cardiac events in HFpEF patients. From the point of view of prediction, much attention has been paid to machine learning (ML). Using ML, Angraal et al created classifiers predicting the development of death and HF hospitalization within 3 years in each patient with HFpEF, with an AUC of 0.72 and 0.76, respectively.17 Therefore, the predictive performance for the development of cardiac events within 3 years has been only fair in these previous reports,13,17 and more powerful indicators to predict cardiac events of HFpEF are needed for use in outpatient clinics. Recently, we clarified that cardiac function assessed by echocardiography and jugular venous pulse (JVP) is associated with cardiac events of HFpEF.7,8 The prediction of the development of cardiac events in HFpEF may be improved by learning features closely associated with the events. To create such classifiers, we need to select a specific algorithm from among the ML algorithms. Using conditional probabilities, the Naive Bayes (NB) method, a ML algorithm, provides insights into how the probabilities of cardiac events can be estimated from the observed data.18 Indeed, the NB algorithm is utilized in clinical settings.19 As the NB algorithm can consider simultaneous information from numerous attributes, this concept may be suitable for the prediction of cardiac events of HFpEF, which has many aspects. NB classifiers created using many risk factors and/or prognosticators may outperform BNP prediction and achieve more precise prediction for the development of cardiac events within 3 years. Therefore, we aimed at creating high-performance NB classifiers to predict the development of cardiac events within 3 years from the day when outpatients with HFpEF underwent a medical interview, blood sample test, echocardiography, and JVP examination.
Editorial p ????
In this study, after receiving approval from the institutional human subject review committee, all data from the hospital’s echocardiographic and JVP database and medical records were retrospectively obtained. Between January 2014 and December 2017, 3,301 consecutive outpatients underwent echocardiography (Vivid 7, General Electric Healthcare, Wauwatosa, WI, USA) for examination of cardiac disease. For 1,842 patients, echocardiographic examination and electrocardiography were simultaneously recorded, and then electrocardiogram, phonocardiogram, and JVP measurement were also simultaneously performed in the echocardiographic examination. All data were stored using a hard-disk memory system (echoPAC PC, General Electric Healthcare) for later analyses. A flowchart of this study is shown in Figure 1. In the present study, we defined patients with HFpEF as those with LVEF ≥50%, ≥2 positive variables of LV diastolic dysfunction, having symptoms and/or signs of HF, and a BNP level >35 pg/mL.20,21 Patients were excluded if they lacked data on echocardiographic variables (LVEF, mitral e’, left atrial volume index [LAVI], right ventricular [RV] systolic pressure, and tricuspid annular plane systolic excursion [TAPSE]), JVP waveform, or blood sample tests (BNP, creatine, or hemoglobin) (Figure 1). Patients were also excluded if they had no symptoms or signs of HF, normal LV diastolic function, LVEF <50%, HF due to a specific cause (constrictive pericarditis, cardiac amyloidosis, hypertrophic cardiomyopathy, moderate or severe valvular heart disease, congenital heart disease, pulmonary arterial hypertension, acute coronary syndrome within 6 months, uncontrolled angina pectoris, idiopathic pulmonary arterial hypertension, or acute decompensated heart failure), or were unable to be followed up for 3 years. In total, we retrospectively enrolled 289 outpatients in the present study. The data from 234 patients obtained during the period from January 2014 to December 2016 were used as original data to create the NB classifiers and data from the 55 other patients obtained during the period from January 2017 to December 2017 were used to validate the classifiers’ performance. All patients received chronic medication for 4 months and were followed up for 3 years. Informed consent was given by all patients and the study complied with the Declaration of Helsinki.
Study flowchart showing patient enrollment in the present study. BNP, B-type natriuretic peptide; HFpEF, heart failure with preserved ejection fraction; JVP, jugular venous pulse; LAVI, left atrial volume index; LVEF, left ventricular ejection fraction; TR, tricuspid regurgitation.
Cardiac function was evaluated as in our previous report.7,8,22 The JVP waveform was used to evaluate RV distensibility. It was recorded by well-trained cardiac sonographers with the patient lying supine. A pulse-wave transducer (TY-306, Fukuda Denshi, Tokyo, Japan) was placed over the neck, above and to the right of the junction of the right clavicle and the manubrium sterni, and held in place manually. The JVP waveform was recorded for ≥30 s and digitized at a sampling interval of 600 Hz. Using an off-line moving average technique (Matlab version 14, Mathworks, Natick, MA, USA), respiratory baseline fluctuations (0.1–0.5 Hz) were excluded from the jugular waveform to measure the relative depth of the nadirs of the ‘X’ and ‘Y’ descent.22 According to the established significance of the JVP waveform,23–25 2 cardiologists who were blinded to the clinical data judged whether the JVP had a dominant ‘Y’ descent, where the nadir of the ‘Y’ descent was deeper than that of the ‘X’ descent, reflecting a less-distensible RV. LV end-diastolic and end-systolic volumes were measured using a modification of Simpson’s method. The LVEF was calculated as the stroke volume divided by the end-diastolic volume. To evaluate the diastolic properties of the LV, we measured the early diastolic velocities (e’) using pulsed-wave tissue Doppler from the apical view. We measured the septal and lateral E/e’, and averaged the values for more reliable assessment of LV relaxation and filling pressure.21 If patients had atrial fibrillation (AF), we estimated velocity measurements from 10 consecutive cardiac cycles.21 The LAVI was obtained using the biplane method from both the apical 4- and 2-chamber views.26 In addition, the tricuspid regurgitant jet velocity was assessed using the continuous Doppler technique to measure the RV systolic pressure. The peak pressure gradient from the RV to the right atrium (RA) was calculated from the peak tricuspid regurgitant velocity (V) using a modified Bernoulli equation (pressure gradient=4 V2). The peak RV pressure was then calculated by adding the peak pressure gradient to the RA pressure, which was estimated from the echocardiographic characteristics of the inferior vena cava.27
Definition of Cardiovascular Events for HFpEFCardiac events were defined as sudden death, death from HF, or hospitalization for deterioration of HFpEF within 3 years from the day of echocardiographic examination. These cardiac events were adjudicated by cardiologists at the hospital.
NB MethodTo create NB classifiers for predicting the development of cardiac events over the next 3 years in individual outpatients with HFpEF, we used R and downloaded several packages.28–32 In the classification using NB, the relationships between dependent events can be described using a probabilistic theorem, as in the following formula:
P(A/B) = P(B/A) * P(A) / P(B)
where P(A/B) is the posterior probability, P(B/A) is the likelihood, P(A) is the prior probability, and P(B) is the marginal likelihood. Prior probabilities of cardiac events in outpatients with HFpEF were calculated using data from the patients in this study. The NB method must use frequency to learn from the data. To optimize discretization of numeric features, we referred to previous reports.7,16,21,33–38 In the case of numeric features being unable to be discretized referring to previous reports, the cutoff value of the numeric feature for the cardiac events was generated from receiver operating characteristics curves and the optimal threshold was automatically identified as the value that minimized the expression [(1-sensitivity)2 + (1-specificity)2]. Parameters having less to do with cardiac events (coefficient of association <0.1) were excluded from the features. We used the step forward method to search for the best model. According to a previous report,17 we defined the best model in this study as that with the maximum AUC among classifiers. A 10-fold cross-validation method was used to validate the generalization performance of the classifiers; therefore, 10 classifiers were created.18 The mean kappa statistics, AUC, sensitivity, specificity, and accuracy were also evaluated and compared between the best model and only the BNP learning model.
Validation StudyWe performed an independent validation analysis on 55 of 289 patients (data obtained during the period of January 2017 to December 2017). The development of cardiac events in these patients was predicted using the NB classifiers, and we then evaluated the kappa statistics, sensitivity, specificity, and accuracy.
Statistical AnalysisNormally distributed data are expressed as the mean±standard deviation and non-normally distributed data are expressed as the median (interquartile range). The unpaired t-test or Mann-Whitney U test was used to compare numerical data between groups, and the chi-square test or Fisher’s exact test was used to compare nonparametric data between groups. One-way analysis of variance or the Kruskal-Wallis test was used to compare numerical data among groups. Significance was established at P<0.05. The statistical analyses were carried out using EZR (Saitama Medical Center, Jichi Medical University, Saitama, Japan).39 The coefficient of association, mean kappa statistics, AUC, sensitivity, specificity, and accuracy were estimated using R.28,29,32
The characteristics of the original and validation cohorts are shown in Table 1. The number of prior hospitalizations and usage of loop diuretics, average mitral e’, and RV systolic pressure were significantly lower in the validation cohort than in the original cohort. Hemoglobin was significantly higher and RV outflow tract diameter was significantly larger in the validation cohort than in the original cohort. Other features were similar between groups.
| Cohort to create NB classifiers (n=234) |
Cohort to validate NB classifiers (n=55) |
P value | |
|---|---|---|---|
| Medical interview | |||
| Age, years | 77 (69–85) | 77 (69–82) | 0.271 |
| Male sex | 109 (47) | 28 (51) | 0.668 |
| No. of prior hospitalizations for HFpEF | |||
| 0/1/≥2 | 185/28/21 | 39/14/2 | 0.033 |
| Underlying disorders | |||
| Hypertension | 206 (88) | 53 (96) | 0.085 |
| Diabetes mellitus | 55 (24) | 15 (27) | 0.68 |
| Prior coronary revascularization | 69 (29) | 11 (20) | 0.212 |
| NYHA II/III | 171/63 | 40/15 | 1 |
| Symptoms and signs of HF | 234 (100) | 55 (100) | 1 |
| Dyspnea on exertion | 226 (97) | 55 (100) | |
| Leg edema | 102 (44) | 18 (33) | |
| Neck vein dilatation | 68 (29) | 12 (22) | |
| Pleural effusion | 52 (22) | 7 (13) | |
| Medications | |||
| β-blockers | 123 (53) | 37 (67) | 0.068 |
| Calcium-channel blockers | 128 (55) | 30 (55) | 1 |
| ACE inhibitors/ARBs | 165 (71) | 43 (78) | 0.331 |
| Loop diuretics | 114 (49) | 15 (27) | 0.006 |
| Blood tests | |||
| Hemoglobin, g/dL | 12.4 (10.8–13.6) | 13.2 (11.8–14.1) | 0.018 |
| eGFR, mL/min/1.73 m2 | 59±23 | 66±19 | 0.069 |
| BNP, pg/mL | 163 (74–318) | 123 (68–208) | 0.228 |
| ECG, echocardiography, and JVP | |||
| AF | 56 (24) | 14 (25) | 0.95 |
| LVEF, % | 65 (60–72) | 69 (62–73) | 0.162 |
| LVEDD, mm | 48 (44–52) | 46 (43–50) | 0.115 |
| LAVI, mL/m2 | 37 (34–43) | 38 (35–42) | 0.953 |
| Average mitral e’, cm/s | 8.9 (7.0–9.4) | 7.7 (6.1–8.8) | 0.024 |
| Average mitral E/e’ ratio | 10 (8–13) | 11 (8–14) | 0.094 |
| RVOT, mm | 26 (23–29) | 28 (25–31) | 0.019 |
| RVSP, mmHg | 32 (26–39) | 27 (23–30) | <0.001 |
| TAPSE, mm | 20±3 | 20±4 | 0.623 |
| Less-distensible RV | 91 (39) | 20 (36) | 0.847 |
| Cardiac events | 74 (32) | 10 (18) | 0.070 |
| Hospitalization for worsening HFpEF | 71 | 9 | |
| Death due to HF after hospitalization | 23 | 4 | |
| Death due HF | 1 | 0 | |
| Sudden death | 2 | 1 | |
Data are the number of patients (%), mean±SD, or median (interquartile). ACE, angiotensin-converting enzyme; AF, atrial fibrillation; ARB, angiotensin-receptor blocker; BNP, B-type natriuretic peptide; ECG, electrocardiogram; eGFR, estimated glomerular filtration rate; HFpEF, heart failure with preserved ejection fraction; JVP, jugular venous pulse; LAVI, left atrial volume index; LVEDD, left ventricular end-diastolic dimension; LVEF, left ventricular ejection fraction; NHYA, New York Heart Association; RVOT, right ventricular outflow tract; RVSP, right ventricular systolic pressure; TAPSE, tricuspid annular plane systolic excursion.
Coefficients of association between cardiac events and features are shown in Table 2. A total of 18 parameters had a coefficient of association ≥0.1 and were applied as learning features of the classifiers.
| Coefficient of association |
|
|---|---|
| BNP | 0.585 |
| NYHA classification | 0.561 |
| No. of prior hospitalizations for HFpEF | 0.513 |
| Loop diuretics | 0.477 |
| Less-distensible RV | 0.419 |
| RVSP | 0.397 |
| Anemia | 0.375 |
| AF | 0.329 |
| TAPSE | 0.329 |
| eGFR | 0.308 |
| Age | 0.262 |
| LAVI | 0.262 |
| Average mitral E/e’ ratio | 0.242 |
| Average mitral e’ | 0.194 |
| RVOT | 0.182 |
| LVEDD | 0.127 |
| Prior coronary revascularization | 0.117 |
| Male/female | 0.101 |
| Hypertension | 0.080 |
| β-blockers | 0.053 |
| ACE inhibitors/ARBs | 0.037 |
| LVEF | 0.037 |
| Diabetes mellitus | 0.009 |
| Calcium-channel blockers | 0.009 |
Data are the coefficients of association. Abbreviations are the same as in Table 1.
Results of the step forward method are shown in Figure 2. Classifiers learning 1 feature, or BNP, had an average AUC of 0.84. Classifiers learning 13 features (age, New York Heart Association classification, number of prior hospitalizations for HFpEF, usage of loop diuretics, estimated glomerular filtration rate (eGFR), anemia, BNP, AF, LAVI, average mitral E/e’ ratio, RV systolic pressure, TAPSE, and RV distensibility) had the highest AUC among the classifiers.
Step forward method to find a high-performance model. Naive Bayes classifiers learning 13 features had the highest AUC (Red bar). AUC, area under the receiver operating characteristics curve.
The frequencys of cardiac events with each feature are shown in Table 3. To use the cutoff points identified in previous reports, numerical data were well discretized.
| P value | ||||
|---|---|---|---|---|
| Age (years) | ||||
| <70 | 70–79 | ≥80 | ||
| Cardiac events (%) | 10±1 | 17±1 | 57±2 | <0.001 |
| NYHA | ||||
| II | III | |||
| Cardiac events (%) | 16 (16–16) | 74 (74–75) | <0.001 | |
| No. of prior hospitalizations for HFpEF | ||||
| 0 | 1 | ≥2 | ||
| Cardiac events (%) | 19 (19–20) | 70 (70–74) | 85 (84–89) | <0.001 |
| Usage of loop diuretics | ||||
| No | Yes | |||
| Cardiac events (%) | 10±1 | 54±1 | <0.001 | |
| eGFR (mL/min/1.73 m2) | ||||
| <30 | 30–89 | ≥90 | ||
| Cardiac events (%) | 68 (67–69) | 29 (28–29) | 11 (10–12) | <0.001 |
| Anemia | ||||
| Non | Mild | Moderate and severe | ||
| Cardiac events (%) | 17 (16–17) | 32 (31–33) | 57 (56–58) | <0.001 |
| BNP (pg/mL) | ||||
| <100 | 100–299 | ≥300 | ||
| Cardiac events (%) | 4 (3–4) | 28 (27–28) | 72 (72–74) | <0.001 |
| AF | ||||
| No | Yes | |||
| Cardiac events (%) | 23±1 | 59±2 | <0.001 | |
| LAVI (mL/m2) | ||||
| <40 | ≥40 | |||
| Cardiac events (%) | 22±1 | 47±1 | <0.001 | |
| Average mitral E/e’ ratio | ||||
| <14 | ≥14 | |||
| Cardiac events (%) | 26±1 | 53±2 | <0.001 | |
| RVSP (mmHg) | ||||
| <35 | ≥35 | |||
| Cardiac events (%) | 18±1 | 56±2 | <0.001 | |
| TAPSE (mm) | ||||
| <17 | ≥17 | |||
| Cardiac events (%) | 64±2 | 24±1 | <0.001 | |
| RV distensibility | ||||
| Distensible | Less-distensible | |||
| Cardiac events (%) | 16±1 | 56±1 | <0.001 | |
Data are the mean±standard deviation or median (interquartile). Abbreviations are the same as in Table 1.
The average likelihood ratios of our NB classifiers with each feature are shown in Figure 3. The average likelihood ratios of a higher category of RV systolic pressure, lower category of TAPSE, and less-distensible RV were the same as or greater than those of the larger category of LAVI and higher category of the average E/e’ ratio. Moreover, the average likelihood ratios of the moderate or severe category of anemia and of the lowest category of eGFR were the same as or greater than that of cardiac dysfunction.
Average likelihood ratio of 13 features represented on a logarithmic scale. AF, atrial fibrillation; BNP, B-type natriuretic peptide; eGFR, estimate glomerular filtration rate; LAVI, left atrial volume index; NYHA, New York Heart Association; RV, right ventricular; RVSP, right ventricular systolic pressure; TAPSE, tricuspid annular plane systolic excursion.
The performance of the NB classifiers is shown in Table 4. Kappa statistics, AUC, and sensitivity were significantly higher for NB classifiers learning 13 features than for those learning only BNP (0.69±0.14 vs. 0.54±0.12 P=0.024, 0.94±0.03 vs. 0.84±0.05 P<0.001, 85±8% vs. 64±20% P=0.006, respectively). The specificity and accuracy were similar between the 2 groups (85±9% vs. 89±8% P=0.355, 85±6% vs. 81±8% P=0.099, respectively).
| Learning 13 features |
Learning only BNP |
P value | |
|---|---|---|---|
| Kappa statistic | 0.69±0.14 | 0.54±0.12 | 0.024 |
| AUC | 0.94±0.03 | 0.84±0.05 | <0.001 |
| Sensitivity (%) | 85±8 | 64±20 | 0.006 |
| Specificity (%) | 85±9 | 89±8 | 0.355 |
| Accuracy (%) | 85±6 | 81±8 | 0.099 |
Data are the mean±standard deviation or median (interquartile). AUC, area under the receiver operating characteristics curve; BNP, B-type natriuretic peptide.
The results of the validation study are shown in Table 5. Kappa statistics, sensitivity, specificity, and accuracy were 0.73 (95% confidence interval (CI), 0.50–0.96), 90% (95% CI, 56–100), 91% (95% CI, 79–98%), and 91% (95% CI, 80–97%), respectively.
| Predicted by NB classifiers | Total | ||
|---|---|---|---|
| Event (−) | Event (+) | ||
| Actual | |||
| Event (−) | 41 | 4 | 45 |
| Event (+) | 1 | 9 | 10 |
| Total | 42 | 13 | 55 |
Data are numbers. NB, Naive Bayes.
Based on the MAGGIC risk model, we also created supplemental NB classifiers to predict 1-year cardiac events (Supplementary Tables 1–3).
Using the NB algorithm, we created classifiers to predict the development of cardiac events for HFpEF within 3 years of medical evaluation. To predict events development, the classifiers learned 13 features available from medical interviews, blood sample tests, and examination of electrocardiography, echocardiography, and JVP measurements in the outpatient clinic, and had higher Kappa statistics, AUC, and sensitivity than those that learned only the BNP level. To our knowledge, this is the first report of the development of cardiac events in individual HFpEF patients being predicted by high-performance NB classifiers, and which was superior to prediction by BNP.
The sensitivity of the NB classifiers learning 13 features was significantly higher than that of NB classifiers learning only BNP, which was the main factor for the difference in performance. There is a strong correlation between the level of BNP and LV end-diastolic wall stress.15 According to Laplace’s law, the wall stress increases according to the LV end-diastolic dimension and decreases according to the LV wall thickness. As some HFpEF is characterized by a small LV volume and LV hypertrophy, LV wall stress is not high compared with HFrEF exhibiting LV remodeling. The level of BNP is usually lower in patients with HFpEF than in those with HFrEF under the condition of compensated HF.15 Indeed, we recently identified a subgroup of HFpEF patients in whom the complications of RV afterload mismatch and renal dysfunction were associated with the poorest outcomes, and their BNP level was the highest among the groups,8 although not as high as expected considering the disease severity. Although the performance of BNP is the same as that of the MAGGIC score,13 its level may not reflect the severity of the disease in some patients with HFpEF. Thus, it can become difficult to predict cardiac events for HFpEF using only the BNP level, and ML of other features is required for precise prediction. As shown in Figure 2, learning more features increased the AUC and, as shown in Figure 3, several comorbidities, such as AF, moderate or severe anemia, severe chronic kidney disease, and RV dysfunction, also increased the posterior probabilities of the development of cardiac events. Thus, learning both BNP and these features improved the sensitivity of the NB classifiers.
Although we cannot directly compare performance, our classifiers had several differences as compared with previous reports.13,17 The MAGGIC risk score is a representative model to predict cardiac death of HF; however, most of the entry patients (≈86%) had HFrEF or HFmrEF.12 Because the MAGGIC risk score was built using forward stepwise regression, the features having strong effects for the prediction of death in patients with HFrEF and HFmrEF might be selected. For example, usage of angiotensin-receptor blockers or β-blockers can reduce the probabilities of the development of cardiac death in the MAGGIC score, but the effectiveness for HFpEF has not been fully established yet.40–42 Our NB classifiers learned the features of patients with HFpEF, and therefore, specialize in predicting the development of cardiac events in patients with HFpEF. The performance of the classifiers may have been improved due to the differences in learning features and algorithm. As all ML algorithms are only as good as the input data,18 the important point is what features are taught. Our NB classifiers learned cardiac function, especially RV distensibility and afterload, which was unique compared with previous classifiers.17 These features are more associated with the development of cardiac events than other cardiac morphological features and functions, as shown in Table 2. Learning RV function may play an important role in outperforming classifiers learning only BNP. Moreover, most of our included 13 features were associated with HFpEF as a risk factor or prognosticator in prior studies.7,16,21,33–37
Selecting the ML algorithm is important for creating high-performance classifiers. Many ML algorithms select features in order of importance and ignore features that have weak effects;18 however, features judged as meaningless in the prediction of cardiac events by artificial intelligence may be important in clinical practice. Indeed, classifiers have been created using random forest, with a maximum depth of only 6 for each tree,17 which may be insufficient for more precise prediction of the development of cardiac events for HFpEF, which has many aspects. From a medical point of view, beneficial classifiers to predict the development of cardiac events in individual outpatients with HFpEF may include the following: features strongly related to HFpEF, those that explain the pathophysiology of HFpEF, are therapeutic targets for HFpEF, and are measurable and available parameters from noninvasive examinations in the outpatient clinic. NB classifiers are best applied to problems in which information from numerous attributes has to be considered simultaneously in order to estimate the overall probability of an outcome.18 The NB algorithm can alleviate problems due to dimensionality and the need for data sets that increase exponentially with the number of features, which reduces the accuracy of prediction. It also requires relatively few examples for training. Thus, the NB algorithm may be suitable for small data sets and predicting the development of cardiac events in this heterogeneous syndrome.
It is well known that accurate prior probabilities and appropriate discretization of numeric features are required to create high-performance NB classifiers for precise prediction.18 Previous studies reported that the prevalence of the development of cardiac events for the next 3 years in patients with HFpEF was approximately 30%.40,42,43 Our prior probability of cardiac events was almost the same as that in previous studies.40,42,43 It was therefore reasonable to use our prior probabilities to create the NB classifiers. Discretization of numerical features is a key point in creating high-performance classifiers by the NB method.18 It was possible to discretize our numerical features using discretization algorithms to find optimal cutoff points in order to maximize the performance classifiers; however, such cutoff points may be meaningless from a medical point of view. Based on the cutoff points supported by previous reports,7,16,21,33–38 we were able to discretize most of the numerical features. Arbitrary discretization of numerical features was also prevented. Moreover, our 8 numerical features were discretized well, as shown in Table 3, and significantly aided in the creation of high-performance classifiers. We were therefore able to create high-performance NB classifiers that included important features. Moreover, using a cohort of 234 outpatients with documented HFpEF and a validation cohort of 55 independent outpatients with HFpEF, we demonstrated the feasibility and validity of the NB classifiers. Our NB classifiers may be applicable in clinical practice.
Clinical ImplicationClassifiers having 85% accuracy are sufficient for predicting cardiac events for HFpEF and for screening severe HFpEF in the clinical setting. First, the development of cardiac events is predicted by the NB classifiers. When the classifiers judge that cardiac events associated with HFpEF will not develop, the physician should not only continue the current treatment for the patient, but also prevent the onset of features associated with increases in posterior probabilities by careful follow-up of the patient. On the other hand, when the classifiers judge that cardiac events associated with HFpEF may develop, the physician should reconsider or consult with a cardiologist about therapeutic strategies for HFpEF. Second, our NB classifiers also suggested that therapeutic strategies focusing on 1 feature may fail to improve the clinical outcomes of HFpEF because another feature associated with cardiac events of HFpEF will continue to increase the posterior probabilities. Thus, the physician has to examine the causes of the features associated with an increase in the posterior probabilities and perform appropriate therapies. Lastly, the posterior probabilities are reassessed after optimization of therapeutic strategies.
Study LimitationsSome limitations are the same as in our previous report.8 This was a retrospective study conducted at a single center. As we required satisfactory imaging of echocardiography and JVP, some patients, such as markedly obese patients with limited windows or fatty neck, may have been underrepresented.8 Patients with tachycardia may also have been excluded because of the difficulty in separating the E and A waves in the mitral inflow or the ‘X’ and ‘Y’ descent of the JVP.44 Moreover, the criterion of HFpEF is updated once every few years and patients meeting the latest criterion for HFpEF were enrolled retrospectively in this study; therefore, some patients were excluded from the original sample, which may have caused selection bias. It is well known that wild-type transthyretin amyloidosis is an underdiagnosed cause of HFpEF.45 Although we excluded patients with overt amyloid heart, those with subclinical amyloidosis may have been included in the present study. Because all patients enrolled in this study had HFpEF, our NB classifiers are not applicable for prediction of the development of cardiac events in patients with HFrEF and HFmrEF.
Methodological limitations of the NB algorithm are also present. The NB method assumes that all features in the dataset are equally independent. These assumptions are rarely true in most real-world situations. However, in most cases when these assumptions are violated, NB still performs well even if strong dependencies are found among the features. One explanation as to why NB works well despite its faulty assumptions is that it is not important to obtain a precise estimate of probability as long as the predictions are accurate.18,46 Moreover, the NB method uses simple principles, and conditional probabilities, so it is an easy-to-understand prediction algorithm for physicians who are unfamiliar with ML. Lastly, other complex algorithms, such as time-to-event analysis and/or neural networks, may be more suitable for the prediction of cardiac events. A more high-performance model may be able to be created using another feature selection method such as exhaustive feature search. Although the future performance of the created NB classifiers was demonstrated to be high by the AUC and was validated in another cohort, further prospective clinical studies are warranted to confirm our results and to improve their performance.
We created high-performance NB classifiers, compared with those that learned only the BNP level, to predict the development of cardiac events for 3 years in outpatients with HFpEF. Our NB classifiers may be useful for providing precision medicine for these patients.
H.A. received research grants from Sun Medical Technology Research Corp., Sumitomo Riko Company Limited, Century Medical, Inc., Teijin Pharma Limited., Nipro Corporation, and Medtronic Japan Co., Ltd. All other authors declare that they have no conflicts of interest.
Imizu Municipal Hospital Ethics/Clinical Trial Review Committee (reference No. is 020005).
Please find supplementary file(s);
http://dx.doi.org/10.1253/circj.CJ-21-0131
