Construction and validation of a nomogram to predict left ventricular hypertrophy in low‐risk patients with hypertension

Abstract

Electrocardiography (ECG) is an accessible diagnostic tool for screening patients with hypertensive left ventricular hypertrophy (LVH). However, its diagnostic sensitivity is low, with a high probability of false‐negatives. Thus, this study aimed to establish a clinically useful nomogram to supplement the assessment of LVH in patients with hypertension and without ECG‐LVH based on Cornell product criteria (low‐risk hypertensive population). A cross‐sectional dataset was used for model construction and divided into development (n = 2906) and verification (n = 1447) datasets. A multivariable logistic regression risk model and nomogram were developed after screening for risk factors. Of the 4353 low‐risk hypertensive patients, 673 (15.4%) had LVH diagnosed by echocardiography (Echo‐LVH). Eleven risk factors were identified: hypertension awareness, duration of hypertension, age, sex, high waist‐hip ratio, education level, tea consumption, hypochloremia, and other ECG‐LVH diagnostic criteria (including Sokolow–Lyon, Sokolow–Lyon products, and Peguero–Lo Presti). For the development and validation datasets, the areas under the curve were 0.724 (sensitivity = 0.606) and 0.700 (sensitivity = 0.663), respectively. After including blood pressure, the areas under the curve were 0.735 (sensitivity = 0.734) and 0.716 (sensitivity = 0.718), respectively. This novel nomogram had a good predictive ability and may be used to assess the Echo‐LVH risk in patients with hypertension and without ECG‐LVH based on Cornell product criteria.

1. INTRODUCTION

Left ventricular hypertrophy (LVH) is characterized by an increase in left ventricular mass caused by various cardiovascular risk factors, such as hypertension, aging, diabetes mellitus, and kidney disease. Considerable evidence suggests that LVH is an independent predictor of cardiovascular disease. ¹ In addition, the severity of LVH in patients with hypertension is associated with an increased risk of death. ² Therefore, detecting LVH in patients with hypertension is crucial for preventing adverse cardiovascular events, ³ and early detection and reversal can improve the overall prognosis of hypertension. However, the pathological and physiological changes caused by LVH are not accompanied by obvious symptoms or signs. A sudden deterioration due to the underlying cause or the presence of abnormal heart sounds and other symptoms usually prompts screening and diagnosis. Consequently, LVH is usually an unidentified condition.

Electrocardiography (ECG) is an inexpensive and easily accessible diagnostic tool for LVH screening. However, whether patients with hypertension should be routinely examined using imaging such as echocardiography or magnetic resonance imaging is controversial. ⁴ Echocardiography remains a relatively gold standard imaging method for evaluating LVH and its main variables. Nevertheless, the hypertension guidelines do not recommend routine echocardiography in patients with hypertension and instead suggest ECG as the first choice for screening patients with hypertension for cost‐benefit reasons. ⁵ However, all ECG diagnostic criteria have high specificity but poor sensitivity. The sensitivity for diagnosing mild LVH may be as low as 7%−35%, and the sensitivity for diagnosing moderate or severe LVH is only 10%−50%. ⁶ According to a study on the Italian population, approximately 8.18% of patients with hypertension and no ECG abnormalities had echocardiographic LVH (Echo‐LVH). ⁷ The prevalence of Echo‐LVH in American patients with hypertension and without ECG‐LVH was as high as 37.1%. ⁴ Furthermore, a cross‐sectional analysis of patients with hypertension who visited a Spanish clinic revealed that three‐quarters of the patients with Echo‐LVH were missed when using the Cornell product criteria for ECG. ⁸ Therefore, even if an ECG examination is performed in a clinic, false‐negative LVH diagnoses may occur. According to some studies, echocardiography is useful in guiding the prognosis of patients with hypertension and without ECG‐LVH. ⁹ The diagnosis of Echo‐LVH may represent the predictive value of LVH for prognosis. However, low‐risk patients with hypertension ⁴ and without ECG‐LVH are often overlooked when evaluating the prognosis of patients with hypertension or target organ damage, resulting in missed opportunities for diagnosis in clinical practice.

To date, risk assessment methods based on ECG‐LVH, such as Sokolow–Lyon, Cornell product, and Peguero–Lo Presti criteria, have been developed sequentially, with varying assessment performance. ¹⁰ , ¹¹ Further, when different ethnic groups and age subgroups use the same ECG diagnostic criteria, the prediction efficiency differs significantly. ¹² A survey of 7415 Chinese adults found that the Cornell product criteria have relatively optimal diagnostic performance despite its low sensitivity. ¹³ However, according to Cornell product criteria, there is no ideal and feasible evaluation method for LVH in low‐risk patients with hypertension and without ECG‐LVH. ⁴ This study aimed to establish a diagnostic prediction model that can be used to indicate the LVH risk among low‐risk patients with hypertension, as a supplementary diagnostic assessment for those without an ECG‐LVH diagnosis. Moreover, we aimed to focus on the potential LVH population to improve the overall prognosis of patients with hypertension.

2. METHODS

2.1. Study population

The Northeast China Rural Cardiovascular Health Study (NCRCHS), a prospective dataset, has been conducted in rural areas since 2013, and we analyzed data from a cross‐sectional survey of this study. In the first stage of research, counties, including Dawa (the southern counties), Zhangwu (northern), and Liaoyang (eastern), were randomly selected from different regions of Liaoning Province. Three towns were randomly selected in the second stage (one from each county), and then 26 villages were randomly selected from each town (8−10 rural villages each). Patients who were pregnant, those with malignant tumors, and those with mental disorders were excluded. Of the included participants, 11 956 aged > 35 years were recruited randomly using a multistage, stratified, cluster‐sampling strategy. Participants without hypertension (n = 5815); without echocardiogram or ECG (without echocardiogram n = 197; without ECG n = 123); with missing major variables (n = 26); and with a pericardial disease, valvular heart disease, congenital heart disease, or severe cardiac arrhythmia (n = 773) were excluded. Patients with hypertension (n = 660) and ECG‐LVH (positive according to Cornell product criteria) and with left anterior branch block or left bundle branch block (n = 9) were also excluded, leaving only low‐risk patients with hypertension (n = 4353). The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of China Medical University (Shenyang, China; AF‐SDP‐7‐1, 0‐01). Written informed consent was obtained from the patients (or parent or legal guardian) before commencement of the study. The 4353 eligible participants were divided into development (n = 2906) and validation datasets (n = 1447) in a 2:1 ratio (Figure 1).

FIGURE 1.

Study population screening and nomogram analysis process. DBP, diastolic blood pressure; ECG, electrocardiogram; LASSO, least absolute shrinkage and selection operator; LVH, left ventricular hypertrophy; ROC, receiver operating characteristic; SBP, systolic blood pressure.

2.2. Acquisition of variables

The TRIPOD statement ¹⁴ served as the reference for the construction, validation, and update of our model. In a face‐to‐face visit, standardized questionnaires were used to collect detailed information, such as demographic characteristics, family per capita income, living habits, disease history, and medication history. In a sitting position, the systolic (SBP) and diastolic blood pressure (DBP) were measured (sphygmomanometer: HEM‐907; Omron, Tokyo, Japan). Prior to BP measurements, all smoking and drinking stimulants, such as tea, coffee, and alcohol, and excessive exercise were prohibited. The final BP value was determined by taking the average of three measurements. Fasting blood samples (fasting for more than 12 h) were collected, and serum lipids, serum creatinine, electrolyte levels, and other biochemical indicators were analyzed and measured using automated equipment (Olympus AU640; Olympus, Tokyo, Japan). Detailed methods of the physical examination and data collection process are discussed elsewhere. ¹⁵ At least two echocardiography experts (from the Cardiovascular Ultrasound Department of the First Affiliated Hospital of China Medical University) conducted the echocardiogram analysis; a third expert was consulted for any cases of disagreement. The echocardiography examination was not attributed to specific clinical indications, and all participants were examined. The following indicators were measured using the parasternal long‐axis view: left ventricular (LV) end‐diastolic internal dimension (LVEDd), posterior wall thickness dimension (PWTd), and interventricular septal thickness dimension (IVSTd). The LV mass (LVM) was calculated using the American Society of Echocardiography formula as follows: LVM (g) = 0.8 × [1.04 × {(LVEDd + PWTd + IVSTd)3 − (LVEDd)3}] + 0.6 g. ¹⁶ The LVM index (LVMI) was calculated by dividing the LVM by the body surface area. The MUSE Cardiology Information System (Version 7; GE Healthcare, Chicago, IL, USA) was used to automatically analyze the ECG in advance for preliminary screening and filtering. Two cardiovascular physicians determined and confirmed the final clinical diagnosis of the ECG.

2.3. Definitions of the variables

Hypertension was defined as SBP ≥140 mm Hg, DBP ≥90 mm Hg, use of antihypertensive drugs within 14 days, or a clear diagnosis from a previous physician. According to the definition standards reported in previous studies, low‐risk hypertension was defined as hypertension without an ECG‐LVH diagnosis (according to Cornell product criteria). ⁴ Diabetes mellitus was defined as a fasting blood glucose level > 7.0 mmol/L or a previous diagnosis by a physician. ¹⁷ Physical labor intensity was classified as mild, moderate, or severe based on the intensity of physical activity during agricultural occupations. ¹⁸ Current smoking was defined as smoking more than one cigarette per day, and current drinking was defined as drinking at least once a week. ¹⁹ Dyslipidemia was diagnosed as follows: hypercholesterolemia (≥6.2 mmol/L), hypertriglyceridemia (≥2.3 mmol/L), increased low‐density lipoprotein cholesterol (≥4.1 mmol/L), and reduced high‐density lipoprotein cholesterol (<1 mmol/L). ²⁰ The estimated glomerular filtration rate for chronic kidney disease was calculated using the epidemiological collaborative equation. ²¹ The aspartate transaminase and alanine transaminase levels of >35 U/L and >40 U/L, respectively, were considered abnormal. High waist‐to‐hip ratio (WHR) ²² was defined as WHR ≥0.90 for men and ≥0.80 for women. Body mass index was calculated by dividing the weight (kg) by the square of the height (m); it was then divided into the following categories: low (< 18.5 kg/m²), normal (18.5−23.9 kg/m²), overweight (24.0−27.9 kg/m²), and obesity (≥28 kg/m²). ²³ Hyperuricemia was defined as serum uric acid levels > 420 μmol/L. ²⁴ Anemia was defined as hemoglobin levels < 110 g/L. The serum ion cut‐off values were defined as below the following thresholds: 2.25 mmol/L (calcium), 0.97 mmol/L (phosphorus), 0.87 mmol/L (magnesium), 5.0 mmol/L (potassium), 145 mmol/L (sodium), and 105 mmol/L (chlorine). High leukocyte and platelet counts were defined as >10 × 10⁹ and >350 × 10⁹, respectively. Echo‐LVH was defined as LVMI > 115 g/m² for men and >95 g/m² for women. ¹⁶ The LVH ECG criteria were as follows:

1. Sokolow–Lyon ¹⁰ (SL LVH): SL voltage SV1 + RV5/V6 ≥3.5 mV.

2. SL product ¹¹ (SL product LVH): (SV1 + RV5/V6) × QRS duration ≥371 000 uV·ms.

3. Cornell ¹⁰ (Cornell LVH): Cornell voltage SV3 + RaVL > 2.8 mV for men and > 2.0 mV for women.

4. Cornell product ¹¹ (Cornell product LVH): (SV3 + RaVL) × QRS duration (men) and (SV3 + RaVL + 8 mm) × QRS duration (women) > 244 000 uVm·ms.

5. RaVL ²⁵ (RaVL LVH): RaVL ≥1.1 mV.

6. Peguero–Lo Presti ²⁶ (PLP LVH): PLP voltage SD + SV4 ≥2.8 mV for men and ≥2.3 mV for women.

2.4. Statistical analysis

Continuous variables are described as means ± standard deviations, whereas categorical variables are presented as n (%). Student's t‐test or the chi‐square test was used to test for differences between groups. The Shapiro–Wilk and Kolmogorov–Smirnov tests were used to determine the normality of the datasets. We used a two‐step method to filter the variables used in model construction. First, the risk factors for LVH were screened based on the results of single‐factor binary logistic regression analysis (p < .20). Following the preliminary screening, the least absolute shrinkage and selection operator (LASSO) was used to screen candidate variables further to avoid model overfitting and severe collinearity among variables. The repeated sampling method was used for internal bootstrapping verification of the nomogram model. Next, the effectiveness of the nomogram model (discrimination, calibration, and clinical usefulness) was evaluated in the training dataset and subsequently in the validation dataset. The area under the receiver operating characteristic curve (AUC), the corresponding 95% confidence interval (CI), and the sensitivity and specificity under the optimal cut‐off probability were used to assess the discriminability of the prediction model. The calibration ability of the model was tested using the Hosmer–Lemeshow χ ² test, and a calibration curve that reflected the consistent relationship between the observed and predicted probability was plotted. The net benefits of using the nomogram model represent clinical usefulness, as determined by decision curve analysis (DCA).

R software (version 4.1.1; R Foundation, Vienna, Austria) was used for statistical analysis. The following packages were used: “glmnet,” “rms,” “pROC,” “nricens,” “PredictABEL,” “rmda,” and “rms.” The performance of the original and updated nomogram models was measured using category‐free net reclassification improvement (NRI) and integrated discrimination improvement (IDI). In addition, DCA curves were used to intuitively determine the advantages and disadvantages of implementing the two nomogram models. All tests in the analysis were two‐tailed, and p < .05 was considered statistically significant.

3. RESULTS

3.1. Demographic characteristics

Table 1 shows the baseline demographic characteristics of the development and validation datasets. The participants in the two datasets had an average age of 56.49 ± 10.01 and 56.61 ± 10.20 years, respectively (p = .71). The development dataset included 1356 men (46.66%). The mean SBP of the development and validation datasets were 156.88 ± 18.27 mm Hg and 156.82 ± 18.04 mm Hg (p = .916), and the mean DBP were 88.18 ± 10.44 mm Hg and 88.12 ± 10.51 mm Hg (p = .839), respectively. The prevalence of Echo‐LVH (15.45% vs. 15.48%; p = .98) and other variables did not differ significantly between the two datasets.

TABLE 1.

Baseline characteristics of the development and validation datasets.

	Development dataset	Validation dataset
Variables	n = 2906	n = 1447	p‐value
Age (year)	56.49 ± 10.01	56.61 ± 10.20	.710
Sex, n (%)			.890
Female	1550 (53.34%)	775 (53.56%)
Male	1356 (46.66%)	672 (46.44%)
BMI (kg/m2)	25.50 ± 3.62	25.52 ± 3.59	.875
WHR (ratio)	0.87 ± 0.08	0.87 ± 0.07	.712
SBP (mm Hg)	156.88 ± 18.27	156.82 ± 18.04	.916
DBP (mm Hg)	88.18 ± 10.44	88.12 ± 10.51	.839
Mean pulse (times/min)	78.83 ± 13.78	79.44 ± 14.35	.180
FPG (mmol/L)	6.10 ± 1.80	6.16 ± 1.89	.367
Hb (g/L)	140.39 ± 17.61	140.08 ± 19.75	.597
PLT count (109/L)	215.25 ± 61.81	216.15 ± 60.60	.651
ALT (U/L)	23.53 ± 17.01	23.70 ± 18.20	.757
AST (U/L)	23.21 ± 12.31	23.18 ± 12.59	.923
eGFR (mL/min/1.73 m2)	91.51 ± 15.96	90.87 ± 15.65	.215
UA (mmol/L)	296.43 ± 85.28	297.63 ± 86.90	.664
TC (mmol/L)	5.41 ± 1.08	5.42 ± 1.14	.719
TG (mmol/L)	1.81 ± 1.60	1.79 ± 1.69	.700
HDL‐C (mmol/L)	1.43 ± 0.40	1.43 ± 0.42	.925
LDL‐C (mmol/L)	3.07 ± 0.83	3.07 ± 0.88	.938
LVMI (BSA)	86.80 ± 18.49	86.14 ± 19.64	.273
Current smoking, n (%)			.258
No	1902 (65.45%)	972 (67.17%)
Yes	1004 (34.55%)	475 (32.83%)
Current drinking, n (%)			.936
No	2222 (76.46%)	1108 (76.57%)
Yes	684 (23.54%)	339 (23.43%)
Marital status, n (%)			.133
Unmarried	27 (0.93%)	18 (1.24%)
Married or remarried	2632 (90.57%)	1283 (88.67%)
Divorced	14 (0.48%)	13 (0.90%)
Widowed	233 (8.02%)	133 (9.19%)
Education (years), n (%)			.456
≤6	1580 (54.37%)	804 (55.56%)
>6	1326 (45.63%)	643 (44.44%)
Diabetes mellitus, n (%)			.699
No	2481 (85.38%)	1229 (84.93%)
Yes	425 (14.62%)	218 (15.07%)
Sleep duration (h per day), n (%)			.214
<6	482 (16.59%)	270 (18.66%)
6−8	1767 (60.81%)	850 (58.74%)
>8	657 (22.61%)	327 (22.60%)
Physical labor intensity, n (%)			.095
Light	1108 (38.13%)	601 (41.53%)
Moderate	570 (19.61%)	270 (18.66%)
Heavy	1228 (42.26%)	576 (39.81%)
Snoring, n (%)			.671
No	1611 (55.44%)	812 (56.12%)
Yes	1295 (44.56%)	635 (43.88%)
Sokolow–Lyon ECG‐LVH, n (%)	418 (14.38%)	197 (13.61%)	.492
Sokolow–Lyon product ECG‐LVH, n (%)	193 (6.64%)	88 (6.08%)	.409
Cornell ECG‐LVH, n (%)	36 (1.24%)	19 (1.31%)	.836
RaVL ECG‐LVH, n (%)	30 (1.03%)	14 (0.97%)	.840
Peguero–Lo Presti ECG‐LVH, n (%)	707 (24.33%)	376 (25.98%)	.234
Antihypertensive drugs, n (%)			.074
No	2176 (74.88%)	1047 (72.36%)
Yes	730 (25.12%)	400 (27.64%)
Echo‐LVH, n (%)			.980
No	2457 (84.55%)	1223 (84.52%)
Yes	449 (15.45%)	224 (15.48%)

Abbreviations: ALT, alanine aminotransferase; AST, aspartate aminotransferase; BMI, body mass index; BSA, body surface area; DBP, diastolic blood pressure; Echo‐LVH, left ventricular hypertrophy under echocardiography; eGFR, estimated glomerular filtration rate; FPG, fasting plasma glucose; Hb, hemoglobin; HDL‐C, high‐density lipoprotein cholesterol; LDL‐C, low‐density lipoprotein cholesterol; LVMI, left ventricular mass index; PLT, platelet; SBP, systolic blood pressure; TC, total cholesterol; TG, triglyceride; UA, uric acid; WHR, waist‐to‐hip ratio.

3.2. Selection of model variables

Univariate logistic analysis of the development dataset was performed to screen the variables with p < .25, including ECG‐related (ECG‐LVH by other criteria) and other non‐ECG‐related variables, and a LASSO‐penalized logistic analysis was conducted to reduce model overfitting and the number of variables included in the model. The SBP and DBP were alternative variables for the updated model; consequently, they were not included in the screening process. Table S1 shows the univariate logistic analysis of BP. Cross‐validation was used to find an ideal λ value; when the mean square error was taken as the minimum, 25 variables corresponding to lambda.min were entered into the model. Further, when the standard error was used, a model with excellent performance and relatively simplified independent variables corresponding to lambda.1se was produced, and 11 variables were entered into the model. Figure 2 shows the variable screening process (LASSO was used to screen 38 independent variables) and cross‐validation using LASSO‐logistic regression. The model was screened for hypertension awareness and duration, age, sex, WHR, education, tea consumption, hypochloremia, and ECG‐LVH diagnostic criteria (including SL LVH, SL product LVH, and PLP LVH). In the development dataset, the 11 variables listed above were the primary risk factors for Echo‐LVH [Figure S1, Forest plot of Model A].

FIGURE 2.

The screening process for predictive factors of LVH. (A) Screening process; (B) Screening and cross‐validation process using LASSO‐penalized logistic analysis. The red line segment represents lambda.1se; the black line segment represents lambda.min. LASSO, least absolute shrinkage and selection operator; LVH, left ventricular hypertrophy.

3.3. Construction and validation of predictive nomograms (Models A and B)

Using the 11 independent risk factors, a nomogram (Model A) was constructed to predict the probability of LVH in low‐risk patients with hypertension (hypertension without an LVH diagnosis according to Cornell product criteria) (Figure 3A and Figure S1). SBP and DBP were risk factors for LVH in the univariate logistic regression, indicating that different levels of hypertension severity may correspond to different LVH risk (Table S1). Therefore, Model B was updated with parameters that include the SBP and DBP levels. Table S2 and Figure S2 show the updated binary multivariate logistic regression model (Model B). Figure 3(B) depicts the updated nomogram (Model B).

FIGURE 3.

Nomograms of two models for predicting the probability of LVH. (A) Model A. (B) Model B. DBP, diastolic blood pressure; HTN, hypertension; low Cl, hypochloremia; LVH, left ventricular hypertrophy; PLP, Peguero–Lo Presti; SBP, systolic blood pressure; SL, Sokolow–Lyon; SL product, Sokolow–Lyon product; WHR, waist‐to‐hip ratio.

In the development and validation datasets, the AUCs of the LVH diagnostic prediction Model A were 0.724 (95% CI, 0.699−0.75) and 0.70 (95% CI, 0.663−0.737), respectively (Figure 4A, B). In contrast, the AUCs of Model B were 0.735 (95% CI, 0.709−0.76) and 0.716 (95% CI, 0.680−0.752) in the development and validation datasets, respectively (Figure 4A, B). In the development datasets, the optimal cut‐off probability of Model A was 13.4% (sensitivity = 0.606, specificity = 0.731, positive predictive probability = 0.292, negative predictive probability = 0.910), and that of Model B was 17.3% (sensitivity = 0.734, specificity = 0.626, positive predictive probability = 0.264, negative predictive probability = 0.928). A score below this cut‐off is considered negative for LVH, whereas a score above is positive. In the validation datasets, the optimal cut‐off probabilities of Models A and B were 15.9% and 17.1%, respectively.

FIGURE 4.

Evaluation and validation of models for predicting LVH. ROC curves show the discrimination for LVH of the nomograms in the (A) development and (B) validation datasets. Model A and B calibration plots show the LVH risk prediction in the development (C) and validation (D) datasets. The consistency between the predicted risk and actual LVH diagnosis was relatively ideal. The 45° diagonal grey line indicates a predicted state where the two completely coincide. Decision curve analysis of the nomograms of the two models for the development (E) and validation (F) dataset. The x‐axis represents the threshold probability, and the y‐axis represents the net benefit. The horizontal black line represents the hypothesis that all patients with hypertension did not have LVH, whereas the solid grey line represents the assumption that all patients with hypertension had LVH. AUC, area under ROC curve; CI, confidence interval; LVH, left ventricular hypertrophy; ROC, receiver operating characteristic; TN, true negative; TP, true positive.

3.4. Calibration of the different nomograms (Models A and B)

The predicted and actual probabilities of LVH were relatively consistent in both datasets (Figure 4C, D). The Hosmer–Lemeshow test further confirmed the consistency of predicted and actual probabilities (all p > .05).

3.5. Clinical usefulness

Figure 4(E, F) presents the DCA results. Predicting LVH based on the nomogram is superior to the two methods of not predicting or completely predicting LVH within the prediction threshold range of 10%−50%, with a cost‐benefit ratio of 1:10.

3.6. Comparison of models: NRI and IDI

NRI represents the degree of improvement in the proportion of correctly re‐discriminated outcome events. There were differences in the discrimination of non‐LVH between the two models, with the updated Model B having better discrimination (NRI−, p < .05; Table 2). To some extent, updated Model B predicted Echo‐LVH more accurately. IDI considers the conditions of various cut‐off points, which can be used to reflect the improvement of the new model compared to the old one, and to some extent, complements the shortcomings of NRI. IDI > 0 in the validation dataset indicated that the predictive discrimination ability of the updated Model B was better in the validation dataset (p < .05; Table 2).

TABLE 2.

Comparison between the original Model A and the updated Model B.

Risk of left ventricular hypertrophy
	Cut‐off probability = 13.4%
Index Category NRI	Development dataset	p‐value	Validation dataset	p‐value
NRI (95% CI)	0.0003 (−0.0307−0.0302)	.986	0.0384 (−0.0047−0.0818)	.091
NRI+ (95% CI)	−0.0156 (−0.0426−0.0095)	.302	0.0089 (−0.0353−0.0503)	.676
NRI− (95% CI)	0.0159 (0.0032−0.0287)	.015	0.0294 (0.0113−0.0451)	.001
IDI	0.02	.191	0.021	.028

Abbreviations: CI, confidence interval; IDI, integrated discrimination improvement; NRI, net reclassification improvement.

4. DISCUSSION

In this study, we screened for the risk factors of LVH and established two diagnostic models in the form of nomograms to predict the risk of LVH in the low‐risk hypertensive population. We used the LASSO strategy (which was necessary to avoid excessive model fitting and solve the problem of serious collinearity among variables) to compress the variable coefficients in the regression model for multiple risk factors. The resulting nomograms effectively supplement the diagnostic shortcomings of screening for hypertensive LVH based on ECG Cornell product criteria, providing more personalized evaluation and diagnostic methods for so‐called low‐risk patients with hypertension and without ECG‐LVH, thereby avoiding false‐negative LVH diagnoses. The model variable factors were simple, easy to obtain, and useful in clinical diagnosis. The model was particularly suitable for LVH evaluation of low‐risk (<50% predicted risk) groups, and the calibration of the two models was higher for this population. Moreover, we updated Model 1 with BP levels (Model 2) to customize LVH warnings for low‐risk patients with hypertension and varying BP levels.

ECG‐LVH diagnosis has long been a concern in the field of hypertensive cardiac target organ damage. ECG is currently the first requirement in terms of standard recommendations for patients with hypertension owing to its widespread popularity, low cost, and high LVH diagnostic specificity. Despite the progress of innovative ECG diagnostic criteria in recent years, such as the Seamens’ Sign ²⁷ and Peguero–Lo Presti criteria, which improved the sensitivity and accuracy of LVH diagnosis in Chinese populations, ²⁸ the overall sensitivity of ECG diagnosis remains inadequate. In particular, many missed LVH diagnoses remain after completion of ECG examinations. Artificial intelligence technology has been used to conduct preliminary research on this issue, although it is still in the early stages. For example, the LVH of 1120 inpatients in China was predicted using a deep learning model of a convolution neural network. However, the AUC was only 0.62 (with a sensitivity of 68% and specificity of 57%), which is still insufficient. ²⁹ Moreover, as most models lack demographic data and other cardiovascular risk factors, diagnostic efficiency has not improved. ³⁰ This is because, in addition to ECG indicators, other indicators, such as the system immune‐inflammation index (calculated as platelet × neutrophil/lymphocyte count), have been shown to be useful in predicting LVH. ³¹ Therefore, rather than using suboptimal ECG diagnostic methods, developing supplementary diagnostic methods for groups that may have false‐negative LVH diagnoses is crucial.

The nomogram diagnostic model presented herein included demographic factors, which was different from the traditional ECG diagnostic criteria. The updated Model B nomograph included the following factors: hypertension awareness, hypertension duration, SBP, DBP, age, sex, high WHR, education, tea consumption, hypochloremia, and ECG‐LVH criteria (including Sokolow–Lyon, Sokolow–Lyon product, and Peguero–Lo Presti). Many previous studies have revealed a link between BP and LVH. Prospective studies have found that every 19‐mm Hg increase in SBP increases the risk of ECG‐LVH by 49%, ³² although the impact of changes in DBP on LVH needs to be investigated further. Consequently, we included BP in our model. In addition, in our model, an age ≥65 years was given a value of 100 points. Aging cardiac tissue causes changes in cardiac structure and function, including progressive myocardial remodeling and LVH. ³³ Currently, the influence of sex on LVH in different populations is controversial. Some studies have found a higher prevalence of LVH in men. ³⁴ In contrast, another study suggested that the female sex is a risk factor for LVH, and most studies acknowledge that female hormones exacerbate the adverse effects of the renin‐angiotensin‐aldosterone system. ³⁵ , ³⁶ In our prediction model, women were assigned a score of 70 points (and men were assigned a score of 0 points), which may be a reflection of the older age of the study population, where most women have already experienced menopause. As traditional cardiovascular disease risk factors, high WHR and years of education were also included in the model. Furthermore, tea consumption was found to be associated with a low risk of LVH and was incorporated into the model. Previous studies have revealed that tea and its second‐generation metabolites reduce antioxidant stress, inhibit renin activity, enhance endothelial nitric oxide synthase activity, reduce vascular inflammation, and decrease smooth muscle contractility. ³⁷

Serum chlorine has a distinct role in humoral homeostasis, affecting the heart, kidney, and neurohormonal system. ³⁸ Hyperchloremia promotes the development of hypertension and proteinuria in chronic kidney disease. ³⁹ In addition, hospital‐acquired hyperchloremia (but not hypochloremia) has been linked to an increased in‐hospital mortality. ⁴⁰ According to some studies, the myocardial anion transporters can transport chloride, and the negative chloride current or transporter protein is involved in the adaptive remodeling process of cardiac hypertrophy. ³⁷ We found that in patients with hypertension and without an ECG‐LVH diagnosis (based on Cornell product criteria), relatively low serum chloride levels may be associated with a low risk of LVH. However, further research is needed to confirm the underlying mechanisms.

Our research has the following advantages. First, participants were selected from the general population within rural communities and underwent comprehensive screening. The patients with hypertension were not recruited from clinics or wards, which helped in avoiding selection bias. Second, the population was randomly divided into development and validation datasets for model construction and internal and external validation, and the LASSO strategy was used to simplify the model and avoid overfitting. The modeling process was thus scientifically standardized. Third, the collection of epidemiological data were relatively comprehensive, and the indicators used in the nomogram are easily available in the primary healthcare sector, resulting in a practical and readily applicable model. In addition, as multiple ECH‐LVH diagnostic criteria may have complementary diagnostic value, they were also included in the model. Lastly, the AUC of the updated model B was optimal. Currently, there are no convenient and effective methods for evaluating LVH in the so‐called low‐risk hypertensive population. Thus, developing a diagnostic model for these patients is clinically valuable as LVH may have been missed in these patients.

Nonetheless, this study has some limitations. First, although our statistical analysis showed that the model had a high cost‐benefit ratio (1:10), compared with using ECG alone for LVH prediction, the model requires the collection of additional information, including disease history and an electrolyte test, for comprehensive evaluation. Thus, a cost‐benefits analysis should be performed in the future to fully compare the use of our model with the use of ECG alone. Second, various types of antihypertensive medication may influence the risk of LVH; however, as the general population of patients with hypertension in rural communities has poor treatment compliance, verifying the specific drug regimens is difficult. Consequently, we only investigated whether patients had been treated with antihypertensive medications. Thus, future research should explore the effects of different antihypertensive therapies on the risk of LVH in this population.

5. CONCLUSIONS

In conclusion, we developed a nomogram‐based diagnostic prediction model to supplement the shortcomings of existing criteria and assess the risk of LVH by combining traditional cardiovascular risk factors and other ECG diagnostic criteria. This model can be used to detect LVH in patients with hypertension and without ECG‐LVH (based on Cornell product criteria) and provide more personalized prevention and treatment strategies for cardiac target organ damage in patients with hypertension.

CONFLICT OF INTEREST STATEMENT

The authors declare that there are no conflicts of interest.

Supporting information

Supporting Information

Zhang X, He C, Lu S, et al. Construction and validation of a nomogram to predict left ventricular hypertrophy in low‐risk patients with hypertension. J Clin Hypertens. 2024;26:274–285. 10.1111/jch.14780

DATA AVAILABILITY STATEMENT

The deidentified participant data are available for validation analysis and will be shared by the corresponding author upon reasonable request. Only the data underlying this article are available; no additional related documents will be shared.