Applying Machine Learning with Tree Ensemble Methods and SHAP Values based on Routine Circulating Biomarkers to Detect Left Atrial Morphological and Functional Remodeling in Hypertension

doi:10.21203/rs.3.rs-3399684/v1

Download PDF

Research Article

Applying Machine Learning with Tree Ensemble Methods and SHAP Values based on Routine Circulating Biomarkers to Detect Left Atrial Morphological and Functional Remodeling in Hypertension

https://doi.org/10.21203/rs.3.rs-3399684/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Introduction: Hypertension induces left atrial (LA) dysfunction and stiffness. Machine learning (ML) has been increasingly used in clinical diagnosis and prognosis prediction. To detect LA stiffness using ML with tree ensemble methods and SHAP values based on clinical biomarkers which were routinely measured in hypertension.

Methods: 351 hypertensive patients were enrolled and measured LA volume (LAV) using the biplane modified Simpson’s method and LA reservoir strain (LAS-S) using 2D speckle-tracking echocardiography. The LA stiffness index (LASI) was defined as the ratio of E/eʹ to LAS-S. Four tree-based ML algorithms, including XGBoost, GBDT, Random Forest (RF), and LightGBM were used to discriminate the increased LASI (≥0.29) and LAV index (LAVI) ( ≥ 28 mL/m2) based on the routine circulating biomarkers including 38 features. We also used the SHAP values to evaluate features importance and interactions.

Results: The top 20 selected variables were used as inputs for four ML models, GBDT presented the highest AUC/ROC (0.85, 95% CI 0.70-0.94) for predicting LASI, and RF model exhibited the best AUC/ROC (0.75, CI 0.57-0.92) for predicting LAVI. SHAP summary plot was applied on GBDT or RF model to identify feature contribution to LA stiffness and LA enlargement, and SHAP also revealed the interactions between variables.

Conclusions: tree-based ML models with the SHAP method combining routine circulating biomarkers predicted LA stiffness with high accuracy. ML models can be useful to screen hypertensive patients with preclinical cardiac TOD, in order to improve personalized medical care at low cost.

left atrial remodeling

machine learning

SHAP value

hypertension

The left atrium plays an important role in modulating left ventricular (LV) function and cardiovascular performance. Left atrial (LA) enlargement and dysfunction have emerged as the surrogate of LV diastolic dysfunction and the crucial predictors of cardiovascular adverse events^{[1, 2]}. Hypertension can lead to LA enlargement and phasic dysfunction. 2D speckle-tracking echocardiography (2DSTE) derived strain measurements can detect LA phasic function, including LA reservoir, conduit, and booster pump function. Among them, LA reservoir strain (LA_S−S), also known as LA global peak longitudinal strain, has been considered as a reliable and reproducible mechanical deformation parameter and reflects LA relaxation and compliance^[3]. LA stiffness index (LASI), defined as the ratio of E/eʹ to LA_S−S, reflects the LA function as well as LA-LV coupling, and has been used for the differential diagnosis and prognosis evaluation of heart failure^{[4, 5]}. LASI has also been recognized as the early marker of target organ damage (TOD) in hypertension^[6].

Machine learning (ML), which is an artificial intelligence (AI) based computational statistics, is used to select predicting variables more objectively and handle non-linear effects more accurately than traditional statistical methods. Recently, ML has been used in cardiovascular studies for risk stratification, auxiliary clinical diagnosis, and prognosis prediction^[7]. The interpretability is a barrier to the implementation of ML especially the black-box ML models. If the performance of ML model is judged purely by simple metrics, such as the classification accuracy which is an incomplete description of most real-world tasks, it might be a better choice to find out the reason of decision making before we trust the evaluation metrics. The Interpretable ML (IML) makes it easier to comprehend the certain decisions and predictions of the model. Model-agnostic interpretable methods such as SHapley Additive exPlanations (SHAP) are flexible and easy to work with because the they can be applied to any ML model.

To enable the data-driven analysis of LA structural and functional specific biomarkers, we quantified clinical routine circulating biomarkers in hypertension. LA morphology and function were evaluated using echocardiography with volume and 2DSTE based strain measurement. We analyzed clinical and echocardiographic data using tree-based ML algorithms to build and validate the predictive model for LA enlargement and stiffness, and we applied SHAP to explain the model.

Study population

A total of 351 essential hypertensive patients (18–75 years old) were recruited from the Hypertension and Heart Failure Ward at the Cardiac Department of the First Affiliated Hospital of Dalian Medical University from October 2019 to September 2020. The major exclusion criteria were as follows: heart failure with reduced ejection fraction (EF), coronary heart disease (history of angina pectoris or myocardial infarction, or coronary computed angiography showing more than 50% stenosis of epicardial coronary arteries), secondary hypertension, cardiac valvular stenosis and moderate or severe valvular regurgitation, atrial fibrillation or atrial flutter, severe hepatic and renal dysfunction, and malignant tumor.

Routine circulating biomarkers

After participating patients fasted for at least 8 h, venous blood was collected and analyzed to determine the levels of fasting plasma glucose (FPG), total cholesterol (TC), triglyceride (TG), low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), Lipoprotein a (LPa), serum creatinine (Scr), cystatin c (CC) and uric acid (UA) (Fully Automatic Biochemistry Analyzer 7600, Hitachi, Ltd., Japan). Routine blood, including red blood cell distribution width SD (RDW_SD) and CV (RDW_CV), and 24h urinary sodium and potassium were also tested. Biomarkers of renin-angiotensin-aldosterone system, including plasma concentrations of renin, angiotensin II and aldosterone, were measured using chemiluminescent immunoassay, and the aldosterone to renin ratio (ARR) was calculated. Glycated hemoglobin A1c (HbA1c) was analyzed with high-performance liquid chromatography. Inflammation parameters included hypersensitivity c-reactive protein (hs-CRP), neutrophil/lymphocyte ratio (NLR) and platelets/lymphocytes ratio (PLR). Myocardial stress and damage biomarkers included B-type natriuretic peptide (BNP) and cardiac troponin I (cTnI). The plasma concentrations of thyroid stimulating hormone (TSH), free thyroxin (FT4), and free triiodothyronine (FT3) were also measured. Renal function was evaluated by the estimated glomerular filtration rate (eGFR) using the following formula: eGFR (ml/min/1.73 m²) = 186 × (Scr/88.402)^−1.154 × age^− 0.203(×0.742 for females).

Echocardiography

Transthoracic echocardiography was performed for all participating patient with a Vivid E9 ultrasound system (GE Vingmed Ultrasound, Horten, Norway) equipped with an M5S phased array transducer (2.5–5.0 MHz), and all measurements followed the guideline of the American Society of Echocardiography (21). (Supplementary material provides measurement methods for more relevant indicators).

Variables analysis

The variables analysis procedures were completed using the site-packages in Python 3.7.3 (Python Software Foundation, Delaware, USA). (Supplementary material provides description and explanation of more variable index processing).

Characteristics of the participating population

The overall participants included 351 hypertensive patients (55.3% male patients), and the mean age was 51.79 ± 12.84 ys. Supplementary data online, Tables S1 and S2 show the clinical and echocardiographic features of the study participants by LASI and LAVI groups. Regarding LA stiffness, all patients were divided into two groups based on the median of LASI. Regarding LA enlargement, the cut-off value of LAVI was 28 mL/m², and LA enlargement was present in 154 (43.9%) patients.

Feature selection

After the missing data were filled, all data were split into training set and test set. 90% patients (315 individuals) were included in training set, and 10% patients (36 individuals) were in test set. RF, GBDT, LightGBM and XGBoost models were established for all clinical and echocardiographic features, aiming to find the best model according to the performance on test set, and we could get the candidate features based on the feature importance of such models.

Figure 1 shows the ML classifiers workflow. Table 1 (all) shows AUC/ROC and AUC/PR metric of ML models for predicting LA stiffness and LA enlargement based on all features. For predicting LA stiffness, both default and optimized RF model presented the highest AUC/ROC (0.82 and 0.87), and the highest AUC/PR (0.83 and 0.87) (shown in Fig. 2A and Supplemental Fig. S1). For LAVI, both default and optimized RF model also presented the highest AUC/ROC (0.67 and 0.76), and the highest AUC/PR (0.54 and 0.63) (shown in Fig. 2B and Supplemental Fig. S2), but lower than that for LASI.

Table 1

AUC of ROC and PR results of LASI and LAVI detection combined over repeats and folds on test set
	AUC/ROC (95% CI)		AUC/PR (95% CI)
	Optimized	Default	Optimized	Default
LASI (all)
RF	0.87 (0.74, 0.96)	0.82 (0.70, 0.92)	0.87 (0.66, 0.96)	0.83 (0.64, 0.92)
GBDT	0.74 (0.58, 0.87)	0.72 (0.63, 0.88)	0.76 (0.57, 0.88)	0.81 (0.64, 0.91)
XGBoost	0.81 (0.63, 0.91)	0.77 (0.58, 0.89)	0.84 (0.68, 0.92)	0.80 (0.62, 0.90)
LightGBM	0.83 (0.64, 0.93)	0.82 (0.65, 0.93)	0.85 (0.69, 0.94)	0.84 (0.67, 0.94)
LAVI (all)
RF	0.76 (0.56, 0.92)	0.67 (0.45, 0.91)	0.63 (0.21, 0.84)	0.54 (0.17, 0.84)
GBDT	0.66 (0.42, 0.87)	0.63 (0.37, 0.85)	0.53 (0.16, 0.78)	0.37 (0.16, 0.66)
XGBoost	0.58 (0.35, 0.78)	0.54 (0.28, 0.72)	0.31 (0.15, 0.50)	0.28 (0.14, 0.44)
LightGBM	0.65 (0.38, 0.83)	0.63 (0.42, 0.80)	0.47 (0.16, 0.75)	0.40 (0.17, 0.70)
LASI (part)
RF	0.80 (0.64, 0.93)	0.77 (0.61, 0.87)	0.80 (0.62, 0.92)	0.79 (0.62, 0.90)
GBDT	0.85 (0.70 0.94)	0.82 (0.70, 0.92)	0.86 (0.73, 0.94)	0.81 (0.64, 0.93)
XGBoost	0.84 (0.67, 0.95)	0.80 (0.62, 0.91)	0.86 (0.72, 0.95)	0.84 (0.68, 0.93)
LightGBM	0.83 (0.68, 0.93)	0.80 (0.65, 0.91)	0.85 (0.71, 0.94)	0.83 (0.69, 0.93)
LAVI (part)
RF	0.75 (0.57, 0.92)	0.74 (0.55, 0.91)	0.60 (0.22, 0.88)	0.54 (0.20, 0.80)
GBDT	0.69 (0.40, 0.88)	0.58 (0.35, 0.83)	0.54 (0.17, 0.82)	0.33 (0.15, 0.68)
XGBoost	0.60 (0.30, 0.80)	0.54 (0.33, 0.77)	0.33 (0.16, 0.52)	0.31 (0.14, 0.63)
LightGBM	0.66 (0.45, 0.84)	0.60 (0.37, 0.83)	0.49 (0.19, 0.73)	0.38 (0.15, 0.69)

Table 2 shows the other evaluation metrics for predicting LA stiffness and enlargement based on all features, including F1, Sensitivity, Precision, Specificity and NPV. For LASI, we selected RF as the ML method for feature selection. As shown in Fig. 3A, the top 20 features included CC, age, BNP, NLR, HbAIc, FPG, RDW_SD, HsCRP, eGFR, apoB, HT-duration, ARRL, TG, BMI, RenL, apoA, AldL, LPa, UA, and LDL-C. For LAVI, we also selected RF for feature selection. As shown in Fig. 3B, the top 20 features included BNP, ARRL, RenL, FT4, age, eGFR, HT-duration, CC, AldL, HsCRP, TSH, PLR, FT3, UA, NLR, 24h urinary potassium, AngIIL, LPa, RDW_SD, FPG.

Table 2

F1, sensitivity, precision, specificity, and negative predictive value of LASI and LAVI on test set prediction from optimized models based on all features, combined over repeats, and folds
	F1	Sensitivity (TPR)	Precision (PPV)	Specificity (TNR)	Negative predictive value (NPV)
LASI
RF	0.71 (0.52, 0.88)	0.68 (0.40, 0.94)	0.78 (0.47, 0.92)	0.81 (0.44, 0.96)	0.74 (0.54, 0.93)
GBDT	0.65 (0.38, 0.82)	0.66 (0.24, 0.88)	0.69 (0.45, 0.93)	0.73 (0.48, 0.95)	0.71 (0.44, 0.90)
XGBoost	0.66 (0.30, 0.84)	0.57 (0.20, 0.77)	0.80 (0.54, 0.93)	0.86 (0.56, 0.96)	0.70 (0.57, 0.85)
LightGBM	0.71 (0.52, 0.86)	0.69 (0.47, 0.89)	0.74 (0.53, 0.93)	0.78 (0.57, 0.95)	0.75 (0.56, 0.89)
LAVI
RF	0.47(0.22, 0.68)	0.55 (0.18, 0.85)	0.51 (0.16, 0.86)	0.71 (0.39, 0.96)	0.80 (0.56, 0.90)
GBDT	0.42(0.18, 0.67)	0.49 (0.13, 0.78)	0.42 (0.17, 0.80)	0.70 (0.37, 0.96)	0.77 (0.59, 0.92)
XGBoost	0.44(0.16, 0.67)	0.60 (0.18, 0.88)	0.37 (0.17, 0.63)	0.58 (0.12, 0.84)	0.77 (0.38, 0.94)
LightGBM	0.47 (0.21, 0.70)	0.58 (0.25, 0.86)	0.41 (0.15, 0.67)	0.66 (0.26, 0.85)	0.80 (0.60, 0.95)

ML performance and Model selection

The top 20 selected variables were used as inputs for four tree-based models (RF, GBDT, XGBoost and LightGBM). After the hyperparameters tuning based on 10-fold cross validation on the training set, we could get the performance of all models on test set. As shown in Table 1 (part), for predicting LA stiffness, GBDT exhibited the best AUC/ROC (0.85, 95% CI 0.70–0.94) (Fig. 4A), and AUC/PR (0.86, 95% CI 0.73–0.94) (shown in Supplemental Fig. S3). For predicting LA enlargement, similar as the model which included all features, RF showed the best AUC/ROC (0.75, CI 0.57–0.92) (shown in Fig. 4B), and AUC/PR (0.60, CI 0.22–0.88) (shown in Supplemental Fig. S4). Table 3 shows the other major evaluation metrics of the test set for all ML models including the top 20 selected features for LASI and LAVI.

Table 3

F1, sensitivity, precision, specificity, and negative predictive value of LASI and LAVI on test set prediction from optimized models, based on top selected features, averaged over repeats, and folds
	F1	Sensitivity (TPR)	Precision (PPV)	Specificity (TNR)	Negative predictive value (NPV)
LASI
RF	0.65 (0.38, 0.81)	0.60 (0.31, 0.87)	0.74 (0.50, 0.90)	0.81 (0.65, 0.96)	0.70 (0.48, 0.88)
GBDT	0.73 (0.57, 0.85)	0.70 (0.50, 0.88)	0.78 (0.50, 0.93)	0.81 (0.57, 0.96)	0.75 (0.58, 0.91)
XGBoost	0.71 (0.52, 0.88)	0.66 (0.44, 0.89)	0.80 (0.60, 0.93)	0.85 (0.59, 0.96)	0.74 (0.59, 0.91)
LightGBM	0.70 (0.50, 0.84)	0.67 (0.40, 0.90)	0.77 (0.47, 0.93)	0.80 (0.50, 0.96)	0.74 (0.47, 0.92)
LAVI
RF	0.44 (0.15, 0.73)	0.51 (0.13, 0.91)	0.50 (0.19, 0.83)	0.73 (0.32, 0.96)	0.79 (0.60, 0.94)
GBDT	0.44 (0.20, 0.64)	0.50 (0.20, 0.73)	0.42 (0.17, 0.67)	0.70 (0.33, 0.93)	0.77 (0.50, 0.92)
XGBoost	0.42 (0.11, 0.64)	0.49 (0.13, 0.83)	0.40 (0.10, 0.67)	0.70 (0.04, 0.88)	0.76 (0.50, 0.92)
LightGBM	0.44 (0.18, 0.67)	0.47 (0.17, 0.82)	0.45 (0.17, 0.75)	0.77 (0.44, 0.93)	0.79 (0.62, 0.90)

In summary, for LASI, GBDT model exhibited the best AUC/ROC, AUC/PR, F1, Sensitivity, and NPV, but lower Precision and Specificity than XGBoost. Therefore, the GBDT model was taken to predict LA stiffness. The RF model was taken for predicting LA enlargement. The RF model showed the best AUC/ROC, AUC/PR and Sensitivity, Precision and NPV, but slightly lower F1and Specificity than LightGBM. Overall, the power of ML models for predicting LA enlargement was not as good as predicting LA stiffness.

Supplementary Table S3 shows the optimized hyperparameters GBDT for LASI and RF for LAVI.

Model explanation based on SHAP

SHAP summary plot was applied on GBDT and RF model to identify feature contribution to LA stiffness and LA enlargement, and the top 20 features in the ensemble model were displayed in Fig. 5. In both figures, each feature was analyzed independently, and each point represented one patient. The position in the x-axis demonstrated risk factors (> 0) or protective factors (< 0). The point color corresponded to the value of each variable, from blue to red representing low to high value. As shown in Fig. 5A, age, CC, BNP, FPG, TG, ARR, BMI and HsCRP have more significant impacts on LA stiffness. As shown in Fig. 5B, BNP, ARR, FT4, the plasma concentration of Rennin, CC, age, eGFR, and TSH have more significant impacts on LA enlargement.

SHAP values also revealed the interactions between variables (shown in Supplementary Fig. S5 and Fig. S6). Age played the most important role in LASI with the cutoff value around 52 years old, and the older the patient, the stiffer the left atrium. Furthermore, age and NLR interacted, and the increased NLR seemed to partly offset the beneficial effect of the younger age on LA stiffness. There were also the important interaction effects between CC and UA, BNP and CC, as well as FPG and age. Furthermore, as shown in supplementary S5 B and C, the interaction effect of CC and UA between 0.8-1.0 of CC, and the interaction effect of BNP and CC between 20–50 of BNP, were minimal (close to 0) on LASI. As shown in Supplementary Figure S6, BNP played the most important role in LAVI with the cutoff value around 36.3. Furthermore, there were significant interaction effects between the BNP and the concentration of aldosterone, ARR and FT4, FT4 and age, as well as the concentration of renin and TSH.

The main finding of this study is that ML with tree ensemble methods and SHAP values based on routine clinical and laboratory data has shown high accuracy of the prediction of LA remodeling, particularly for LA stiffness, in hypertensive population. These ML classifiers might be useful to screen hypertensive patients with preclinical target organ damage (TOD), leading to closer clinical monitoring and preventive strategies in order to delay the progression of TOD in hypertension.

Routine clinical biomarkers correlated with LA remodeling

As we all known, the typical cardiac TOD of hypertension is LV hypertrophy^[8]. In addition, LA remodeling in hypertension has also received extensive attention in clinical and research. LA size and function have been proved to be the robust predictors of cardiovascular outcome^[1]. LA remodeling, including structural and functional remodeling, is a complex process involving multiple mechanisms. In addition to hemodynamic disorders, neurohormonal factors also play an important role. Renin angiotensin system activation, particularly the increase of angiotensin II and aldosterone levels, contributed to atrial remodeling both in animal experiment^[9] and in clinical trial^[10]. In a community cohort with 4547 participants, Jenifer et al.^[11] observed renin suppression was correlated with increased LAVI, and they also found the association of an increased risk for occurrence of atrial fibrillation with higher aldosterone levels. Chen et al’s study^[12] found LV global longitudinal strain was significantly correlated with ARR and the concentration of aldosterone both in primary aldosteronism and essential hypertension.

In the present study, using ML with tree ensemble methods, we found ARR and concentrations of renin and aldosterone also played the important roles in predicting LA stiffness and enlargement. Moreover, we revealed metabolic elements, including TG, body mass index (BMI), FPG, HbA1c, and UA contributed to predict LA remodeling.

BNP is secreted by cardiomyocytes and has been considered as a biomarker related to cardiac structure and function. Not only LV but also LA volume and pressure load regulate the secretion of BNP. BNP secretion is more susceptible to volume load, which leading to myocardial cell stretch. BNP has been accepted as a diagnostic tool to detect LV systolic and diastolic dysfunction. Besides that, It has been reported that BNP is also produced in the atrial wall^[13]. In this study, we found BNP also contributed to the prediction of LA remodeling, particularly for LA enlargement.

In addition, this study also confirmed other clinically relevant biomarkers in LA remodeling. Cystatin C, which is a cysteine protease inhibitor, has been recognized as a more accurate biomarker of eGFR than serum creatinine. Furthermore, cystatin C has been reported to have the association with the cardiovascular disease independent of renal function^{[14, 15]}, the subclinical TOD in white-coat hypertension^[16], and has also been considered as an indicator of cardiac remodeling, including LV structural and functional remodeling and LA enlargement^{[17, 18]}, because it is involved in extracellular matrix remodeling. NLR and PLR are inflammatory markers and are easily accessible in clinical routine blood test. For the past few years, the associations of NLR or PLR with cardiovascular diseases have been studied extensively, and found they correlated with adverse outcomes in patients with cardiovascular diseases^{[19, 20]} and contributed to LA thrombosis and dysfunction in atrial fibrillation^{[21, 22]}. Moreover, Alterations in thyroid hormones has also been reported to affect both LV and LA function^[23–25]. Increased thyroid hormone was associated with more cardiac fibrosis^[26]. Pervious clinical studies have found FT4, but not TSH or FT3, was associated significantly with LV diastolic dysfunction^[27], LA enlargement^[28] and independently predicted atrial fibrillation recurrence^[29], even FT4 is in the normal range^{[30, 31]}. In the current study, using ML with tree ensemble methods, we found CC, NLR/PLR, and FT4 played relatively more important roles in predicting LA remodeling.

ML tree model and SHAP: the advantages

The data type of our study is tabular data, which is complicated and heterogeneous including many types of features like numerical and sparse categorical features^[32]. In contrast to image or language data, tabular data is an important and useful data type in medical research, many pitfalls, including noise, value ranges, non-availability of values, etc. make it a hard task for many Deep Learning (DL) methods such as neural network (NN) although DL models have multiple advantages than traditional ML models^[33]. Some researchers believed that the NN models were struggling to handle the numerous uninformative features present in tabular data^[34]. In recent years, the brilliant performances of tree-based models have presented many achievements cin tabular data competition such as the Kaggle competition (https://www.kaggle.com/kaggle-survey-2021). Besides, as for the good self-interpretability and human understandability of tree-based model, medical workers found it a more reasonable method to apply tree-based model rather than a black-box one. In this study, we therefore applied some standard or state-of-the-art tree-based models like RF, GBDT, XGBoost and LightGBM to handle this task.

With the vigorous development of AI technology, the interpretability of ML or DL models have gradually come into people's eyes. Model-agnostic methods like SHAP and LIME are highly flexible as for their strong applicability to various models^[35]. Compared with other model-agnostic interpretable methods, SHAP has a fast implementation for tree-based models, which makes it possible to conduct a global model interpretation. In addition to the advantage of interpretability, SHAP is good at establishing the interactions among features and making clear visualization of the interpretation. Therefore, we believe that SHAP plays a vital role in our study.

Compare the different ML results of LASI and LAVI

To our knowledge, this study was the first to perform ML with tree models and SHAP values to select significant biomarkers based on routine clinical laboratory test, and the results of prediction models made sense in clinical. In the current study, routine clinical biomarkers included renin-angiotensin-aldosterone system, inflammation parameters, myocardial stress and damage biomarkers, metabolic elements, and parameters correlated renal function as well as thyroid hormones. Our findings also showed that the most important predictors for LASI were age and CC, as opposed to that, BNP and ARR were the most significant predictors for LAVI. LASI was not only a parameter of LA function but also an indicator of LA-LV coupling, and it can be considered as an early marker of TOD in hypertension^[6]. Besides being affected by the elevated LV filling pressure and LV diastolic dysfunction, the increase LASI also indicated the myocardial fibrosis leading to the deterioration of the intrinsic myocardial function. Advancing age-related deposition of collagen causes cardiac interstitial fibrosis, and then leads to the reduction of myocardial compliance. The increase of cystatin C has also been proved to be associated with the alterations in myocardial collagen metabolism and diastolic dysfunction. By contrast, LA enlargement is more susceptible to volume load. BNP is produced in response to myocardial stretch in situations of volume or pressure overload. The current study also revealed ARR, the relevant indicator of potassium, sodium and water balance, was the important biomarker of LA volume. Furthermore, in the current study, the predictive effect of ML with tree ensemble models was better for predicting LA stiffness than LA enlargement. Clinically, the methods of assessing LA stiffness are more difficult and complex than those for assessing LA enlargement. Therefore, these ML classifiers might be useful to pre-select the potential patients with LA stiffening for further evaluation.

Limitations

Firstly, the number of participating patients was limited, and this is a single-center study during a specific time period. Secondly, this study only included inpatients data, further study including outpatients’ data is needed to verify the conclusions. Thirdly, the ML methods applied in this study are tree-based model, and therefore the multivariate redundant variables are not removed. Fourthly, many ML packages we applied in our research does not accept missing data and the missing values must be imputed, this might partly disrupt the distribution of real-world data. Finally, echocardiographic evaluation is not the gold standard for the LA stiffness.

ML with tree ensemble methods and SHAP values combined with routine circulating biomarkers has shown high accuracy of the prediction of LA remodeling, particularly for LA stiffness, in hypertensive population.

These ML classifiers might be useful to pre-select patients who require further echocardiographic and 2DSTE-based strain examination, and to screen hypertensive patients with preclinical cardiac TOD, in order to improve personalized medical care at low cost.

Ethics approval and consent to participate

This study complied with the principles of the Declaration of Helsinki, and was approved by the Ethics Committee of the First Affiliated Hospital of Dalian Medical University. Written informed consent was provided before enrolment.

Consent for publication

All the authors consent to publish this paper in BMC Medical Informatics and Decision Making and agree to pay the page charges.

Availability of data and materials

If other readers need these data and materials, they can apply to corresponding writer by a valid reason.

Competing Interests

no conflict of interest here.

Funding

NO founds

Author Contributions

Shaobo Wang and Youjun Liu used computers to complete the data processing of machine learning and explained the model. Yu Pan completed the collection of clinical blood circulation markers. Shaobo Wang and Panyu jointly completed the writing of the paper. Tingting and Qiaobing Sun completed the echocardiography For the measurement of relevant indicators, Zengtao Jiao provided electronic medical record information services, and Yinong Jiang and Yan Liu completed the conception and writing guidance of the paper.

Data Availability Statement

All data generated or analyzed during this study are included in this article. Further enquiries can be directed to the corresponding author.

Hoit BD. Left atrial size and function: role in prognosis. J Am Coll Cardiol. 2014;63:493–505. 10.1016/j.jacc.2013.10.055.
Thomas L, Marwick TH, Popescu BA, Donal E, Badano LP. Left Atrial Structure and Function, and Left Ventricular Diastolic Dysfunction: JACC State-of-the-Art Review. J Am Coll Cardiol. 2019;73:1961–77. 10.1016/j.jacc.2019.01.059.
Weber J, Bond K, Flanagan J, Passick M, Petillo F, Pollack S, et al. The Prognostic Value of Left Atrial Global Longitudinal Strain and Left Atrial Phasic Volumes in Patients Undergoing Transcatheter Valve Implantation for Severe Aortic Stenosis. Cardiology. 2021;146:489–500. 10.1159/000514665.
Reddy YNV, Obokata M, Egbe A, Yang JH, Pislaru S, Lin G, et al. Left atrial strain and compliance in the diagnostic evaluation of heart failure with preserved ejection fraction. Eur J Heart Fail. 2019;21:891–900. 10.1002/ejhf.1464.
Bytyci I, Bajraktari G, Fabiani I, Lindqvist P, Poniku A, Pugliese NR, et al. Left atrial compliance index predicts exercise capacity in patients with heart failure and preserved ejection fraction irrespective of right ventricular dysfunction. Echocardiography. 2019;36:1045–53. 10.1111/echo.14377.
Zhao Y, Sun Q, Han J, Lu Y, Zhang Y, Song W, et al. Left atrial stiffness index as a marker of early target organ damage in hypertension. Hypertens Res. 2021;44:299–309. 10.1038/s41440-020-00551-8.
Quer G, Arnaout R, Henne M, Arnaout R. Machine Learning and the Future of Cardiovascular Care: JACC State-of-the-Art Review. J Am Coll Cardiol. 2021;77:300–13. 10.1016/j.jacc.2020.11.030.
Williams B, Mancia G, Spiering W, Agabiti Rosei E, Azizi M, Burnier M, et al. 2018 ESC/ESH Guidelines for the management of arterial hypertension. Eur Heart J. 2018;39:3021–104. 10.1093/eurheartj/ehy339.
Reil JC, Tauchnitz M, Tian Q, Hohl M, Linz D, Oberhofer M, et al. Hyperaldosteronism induces left atrial systolic and diastolic dysfunction. Am J Physiol Heart Circ Physiol. 2016;311:H1014–H23. 10.1152/ajpheart.00261.2016.
Wang D, Xu JZ, Chen X, Xu TY, Zhang W, Li Y, et al. Left atrial myocardial dysfunction in patients with primary aldosteronism as assessed by speckle-tracking echocardiography. J Hypertens. 2019;37:2032–40. 10.1097/HJH.0000000000002146.
Brown JM, Wijkman MO, Claggett BL, Shah AM, Ballantyne CM, Coresh J, et al. Cardiac Structure and Function Across the Spectrum of Aldosteronism: the Atherosclerosis Risk in Communities Study. Hypertension. 2022;101161HYPERTENSIONAHA12219134. 10.1161/HYPERTENSIONAHA.122.19134.
Chen ZW, Huang KC, Lee JK, Lin LC, Chen CW, Chang YY, et al. Aldosterone induces left ventricular subclinical systolic dysfunction: a strain imaging study. J Hypertens. 2018;36:353–60. 10.1097/HJH.0000000000001534.
Inoue S, Murakami Y, Sano K, Katoh H, Shimada T. Atrium as a source of brain natriuretic polypeptide in patients with atrial fibrillation. J Card Fail. 2000;6:92–6. 10.1016/s1071-9164(00)90010-1.
Keller T, Messow CM, Lubos E, Nicaud V, Wild PS, Rupprecht HJ, et al. Cystatin C and cardiovascular mortality in patients with coronary artery disease and normal or mildly reduced kidney function: results from the AtheroGene study. Eur Heart J. 2009;30:314–20. 10.1093/eurheartj/ehn598.
Levin A, Lan JH, Cystatin C, Disease C. Causality, Association, and Clinical Implications of Knowing the Difference. J Am Coll Cardiol. 2016;68:946–8. 10.1016/j.jacc.2016.06.037.
Androulakis E, Papageorgiou N, Lioudaki E, Chatzistamatiou E, Zacharia E, Kallikazaros I, et al. Subclinical Organ Damage in White-Coat Hypertension: The Possible Role of Cystatin C. J Clin Hypertens (Greenwich). 2017;19:190–7. 10.1111/jch.12882.
Zivlas C, Triposkiadis F, Psarras S, Giamouzis G, Skoularigis I, Chryssanthopoulos S, et al. Left atrial volume index in patients with heart failure and severely impaired left ventricular systolic function: the role of established echocardiographic parameters, circulating cystatin C and galectin-3. Ther Adv Cardiovasc Dis. 2017;11:283–95. 10.1177/1753944717727498.
Sakuragi S, Ichikawa K, Yamada K, Tanimoto M, Miki T, Otsuka H, et al. Serum cystatin C level is associated with left atrial enlargement, left ventricular hypertrophy and impaired left ventricular relaxation in patients with stage 2 or 3 chronic kidney disease. Int J Cardiol. 2015;190:287–92. 10.1016/j.ijcard.2015.04.189.
Bhat T, Teli S, Rijal J, Bhat H, Raza M, Khoueiry G, et al. Neutrophil to lymphocyte ratio and cardiovascular diseases: a review. Expert Rev Cardiovasc Ther. 2013;11:55–9. 10.1586/erc.12.159.
Afari ME, Bhat T. Neutrophil to lymphocyte ratio (NLR) and cardiovascular diseases: an update. Expert Rev Cardiovasc Ther. 2016;14:573–7. 10.1586/14779072.2016.1154788.
Fukuda Y, Okamoto M, Tomomori S, Matsumura H, Tokuyama T, Nakano Y, et al. In Paroxysmal Atrial Fibrillation Patients, the Neutrophil-to-lymphocyte Ratio Is Related to Thrombogenesis and More Closely Associated with Left Atrial Appendage Contraction than with the Left Atrial Body Function. Intern Med. 2018;57:633–40. 10.2169/internalmedicine.9243-17.
Yalcin M, Aparci M, Uz O, Isilak Z, Balta S, Dogan M, et al. Neutrophil-lymphocyte ratio may predict left atrial thrombus in patients with nonvalvular atrial fibrillation. Clin Appl Thromb Hemost. 2015;21:166–71. 10.1177/1076029613503398.
Shenoy R, Klein I, Ojamaa K. Differential regulation of SR calcium transporters by thyroid hormone in rat atria and ventricles. Am J Physiol Heart Circ Physiol. 2001;281:H1690–6. 10.1152/ajpheart.2001.281.4.H1690.
Ozturk S, Dikbas O, Ozyasar M, Ayhan S, Ozlu F, Baltaci D, et al. Evaluation of left atrial mechanical functions and atrial conduction abnormalities in patients with clinical hypothyroid. Cardiol J. 2012;19:287–94. 10.5603/cj.2012.0051.
Ayhan S, Ozturk S, Dikbas O, Erdem A, Ozlu MF, Baltaci D, et al. Detection of subclinical atrial dysfunction by two-dimensional echocardiography in patients with overt hyperthyroidism. Arch Cardiovasc Dis. 2012;105:631–8. 10.1016/j.acvd.2012.07.003.
Davis PJ, Leonard JL, Davis FB. Mechanisms of nongenomic actions of thyroid hormone. Front Neuroendocrinol. 2008;29:211–8. 10.1016/j.yfrne.2007.09.003.
Malhotra Y, Kaushik RM, Kaushik R. Echocardiographic evaluation of left ventricular diastolic dysfunction in subclinical hypothyroidism: A case-control study. Endocr Res. 2017;42:198–208. 10.1080/07435800.2017.1292524.
Tang RB, Liu DL, Dong JZ, Liu XP, Long DY, Yu RH, et al. High-normal thyroid function and risk of recurrence of atrial fibrillation after catheter ablation. Circ J. 2010;74:1316–21. 10.1253/circj.cj-09-0708.
Cappola AR, Arnold AM, Wulczyn K, Carlson M, Robbins J, Psaty BM. Thyroid function in the euthyroid range and adverse outcomes in older adults. J Clin Endocrinol Metab. 2015;100:1088–96. 10.1210/jc.2014-3586.
Sousa PA, Providencia R, Albenque JP, Khoueiry Z, Combes N, Combes S, et al. Impact of Free Thyroxine on the Outcomes of Left Atrial Ablation Procedures. Am J Cardiol. 2015;116:1863–8. 10.1016/j.amjcard.2015.09.028.
Pei Y, Xu S, Yang H, Ren Z, Meng W, Zheng Y, et al. Higher FT4 level within the normal range predicts the outcome of cryoballoon ablation in paroxysmal atrial fibrillation patients without structural heart disease. Ann Noninvasive Electrocardiol. 2021;26:e12874. 10.1111/anec.12874.
Borisov V, Leemann T, Seßler K, Haug J, Pawelczyk M, Kasneci G. Deep neural networks and tabular data: A survey. arXiv preprint arXiv:211001889. 2021.
Sahoo D, Pham Q, Lu J, Hoi SC. Online deep learning: Learning deep neural networks on the fly. arXiv preprint arXiv:171103705. 2017.
Grinsztajn L, Oyallon E, Varoquaux G. Why do tree-based models still outperform deep learning on tabular data? arXiv preprint arXiv:220708815. 2022.
Lundberg SM, Erion G, Chen H, DeGrave A, Prutkin JM, Nair B, et al. From Local Explanations to Global Understanding with Explainable AI for Trees. Nat Mach Intell. 2020;2:56–67. 10.1038/s42256-019-0138-9.

No competing interests reported.

Supplements..docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Applying Machine Learning with Tree Ensemble Methods and SHAP Values based on Routine Circulating Biomarkers to Detect Left Atrial Morphological and Functional Remodeling in Hypertension

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

Study population

Routine circulating biomarkers

Echocardiography

Variables analysis

Results

Characteristics of the participating population

Feature selection

ML performance and Model selection

Model explanation based on SHAP

Discussion

Routine clinical biomarkers correlated with LA remodeling

ML tree model and SHAP: the advantages

Compare the different ML results of LASI and LAVI

Limitations

Conclusions

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1