Development of a Machine-learning Based Diagnosis Procedure to Distinguish Aortic Dissection from Non-ST- Elevation Myocardial Infarction

doi:10.21203/rs.3.rs-4153593/v1

Background: Similar symptoms have been observed in Non-ST elevation myocardial infarction (NSTEMI) and aortic dissection (AD), making diagnosis challenging. Recognizing the distinction between them is essential for prompt treatment. This study was to establish a model based on machine learning (ML) to improve diagnosis accuracy; (2) Methods: 353 individuals' clinical characteristics and laboratory results (193 AD, 160 NSTEMI) were analyzed. The Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was used to identify significant indicators. Four ML models were constructed, and the Voting algorithm was used to conduct an ensemble analysis. Decision Curve Analysis (DCA) assessed the clinical value. And collected a new validation set of 36 AD and 48 NSTEMI patients to assess the generalizability of the optimal model. Shapley Additive explanations (SHAP) was used to evaluate feature contribution; (3) Results: With an accuracy of 92%, recall of 94%, F1-score of 91.43%, and an AUC of 0.95 (95CI%: 0.91-0.99) on the test set, the ensemble Voting model was recognized as the optimal model. DCA provided evidence of the model's clinical value in AD prediction. The SHAP indicated that Troponin T and D-dimer were crucial predictors; (4) Conclusions: We successfully established a machine-learning based diagnosis approach for timely distinguish of AD and NSTEMI. Based on our results, the Voting model performed the best in terms of predicting efficacy. In addition, we used SHAP to provide a personalized risk assessment for the development of the prediction results. This diagnosis model may assist the emergency department to quickly avoiding misdiagnosis of AD with NSTEMI.

This research involving biomedical studies on human subjects was conducted in accordance with the principles outlined in the Declaration of Helsinki. The study protocol and informed consent procedures were reviewed and approved by the Institutional Ethics Committee. The approval reference number for this study is B-2023-060.

Aortic Dissection

Myocardial Infarction

Machine Learning

The primary artery that supplies blood from the heart to the rest of the body, the aorta, has an inner layer that can break. This disease is known as aortic dissection (AD), and it is extremely deadly. Blood can pass between the blood vessel's layers attributable to the tear. leading to the separation of the layers and potentially resulting in aortic rupture, a life-threatening possibility[1]. Failure to diagnose and treat AD promptly can lead to a significantly higher mortality rate compared to timely surgical intervention, making early diagnosis crucial for reducing mortality from AD[2]. A prevalent cardiovascular disease with significant risks of complications and death is acute coronary syndrome (ACS). and the incidence of Non-ST-Elevation Myocardial Infarction (NSTEMI) within this group is on the rise[3]. The high incidence and mortality associated with NSTEMI emphasize the importance of accurate early diagnosis[4]. Patients with AD and NSTEMI often present with similar clinical symptoms, predominantly chest pain, unlike those with ST-Elevation Myocardial Infarction (STEMI), who can be distinguished by definitive features on an electrocardiogram (ECG). The similarities in ECG findings for AD and NSTEMI can make rapid differentiation between the two conditions challenging[5]. Patients with AD require effective immobilization, blood pressure and ventricular rate control, and urgent preparation for surgery to avoid the risks associated with unnecessary transport and movement that could increase the risk of aortic rupture[6]. For NSTEMI patients, effective coronary vasodilation, anticoagulation, and antiplatelet therapy are needed, along with urgent vascular reconstruction[7]. Misdiagnosing a patient with AD as having NSTEMI and treating them accordingly can lead to very serious consequences[8]. Therefore, timely, accurate, and convenient differential diagnosis is crucial to improving patient survival rates.

While the diagnosis of AD is made utilizing non-invasive imaging methods like magnetic resonance angiography (MRA) and computed tomography angiography (CTA)[9], These techniques cost time, and they're not consistently available at the patient's bedside, especially in primary care settings. Necessary patient transfers and positional changes may increase the risk of aortic dissection rupture[10]. Furthermore, a number of laboratory indicators have been established as biomarkers for AD, including D-dimer, Matrix Metalloproteinases (MMPs), and genetic markers; however, the diagnostic utility of these markers has yet to be limited[11, 12]. In constructing disease risk assessments and predictive models, we aim to minimize the number of required indicators while maintaining accuracy in risk assessment to maximize the benefits and efficiency of the process.

Machine learning (ML), a branch of artificial intelligence (AI) research, has seen widespread application in key aspects of human life in recent years. Studies have shown that ML algorithms perform well in predicting cardiovascular disease risk[13], analyzing images[14], and in diagnostics[15]. The efficiency of ML models in data processing makes them a powerful tool for assisting in the diagnosis of AD and NSTEMI. Appropriate ML algorithms are expected to increase the accuracy of diagnoses and the efficiency of clinical practice, providing strong informational support for doctors making treatment decisions[16].

In this study, we constructed four ML models based on clinical data from patients with AD and NSTEMI at the Chest Pain Center of the First Affiliated Hospital of Shantou University Medical College and comprehensively evaluated the diagnostic performance of each model. Considering that each base classifier may have different biases and variances, which can lead to overfitting[17], we used the Voting ensemble method to improve the overall predictive performance, thereby obtaining more reliable prediction outcomes[18]. To assess the model's performance on future data, we incorporated a newly collected dataset of AD and NSTEMI patients into the Voting validation[19].

Our research offers fresh perspectives on how ML diagnostic models might be used to distinguish between AD and NSTEMI. However, the interpretability of these risk prediction models and their application in actual clinical practice remain limited[20]; Consequently, in order to intuitively explain the elements influencing the patients' estimated risk, we also employed the Shapley Additive explanations (SHAP) methodology[21]. Considering the model's clinical applicability, we further conducted a clinical decision curve analysis (DCA)[22] to evaluate its decision-making impact in a real-world clinical setting.

This comprehensive approach not only highlights the potential of ML in improving diagnostic accuracy but also emphasizes the importance of interpretability and clinical relevance when deploying such models.

2.1.1 Participants

The study utilized clinical data from the Chest Pain Center at the First Affiliated Hospital of Shantou University Medical College. It included records of 193 patients with AD and 160 patients with NSTEMI between January 2017 and December 2020. The initial dataset comprised 29 clinical and laboratory characteristics. Additionally, a time-matched validation dataset was created using data from newly registered AD and NSTEMI patients. The study was conducted with the approval of the Ethics Committee of Shantou University Medical College.

2.1.2 Inclusion Criteria

The following criteria were established for this study's inclusion: patients had to meet the diagnostic standards for AD and NSTEMI, and possess a complete set of clinical data. This encompassed results from Computed Tomography (CT) imaging, Coronary Angiography (CAG), routine electrocardiograms (ECG), and echocardiograms among other examinations. Diagnoses for all patients were conclusively determined based on clinical guidelines, imaging outcomes, clinical symptoms, and a comprehensive assessment of physical signs for either NSTEMI or AD.

2.2 Research Methodology

2.2.1 Diagnostic Approach for Diseases

Patients in this study were categorized into two groups upon discharge from the medical facility: AD and NSTEMI. AD diagnosis relied on the sudden onset of acute back or chest pain, often described as a ripping sensation. Imaging techniques like Computed Tomography Angiography (CTA) or Magnetic Resonance Imaging (MRI) were crucial for identifying aortic wall dissection. Elevated D-dimer levels in laboratory tests also provided diagnostic clues. For NSTEMI diagnosis, patients typically exhibited chest pain, while biochemical markers, especially elevated cardiac troponin levels, indicated myocardial injury. ECG could reveal T-wave inversion, ST-segment depression, and non-specific changes. Imaging studies like echocardiography or nuclear imaging might detect regional wall motion abnormalities or perfusion defects. A comprehensive evaluation of the patient's medical history was conducted before reaching a final diagnosis.

2.2.2 Study Variables

A total of 29 variables were considered in the study: (1). General data: Age and gender. (2). Laboratory tests, including: White Blood Cell Count (WBC), Neutrophil Percentage (NE%), Lymphocyte Percentage (L%), Monocyte Percentage (MO%),Eosinophil Percentage (E%), Basophil Percentage (B%),Red Blood Cell Count (RBC), Hemoglobin (HGB), Platelet Count (PLT), Mean Platelet Volume (MPV), Platelet Distribution Width (PDW),Triglycerides (TG), High-Density Lipoprotein (HDL), Low-Density Lipoprotein (LDL), Glucose (GLU), D-Dimer (D-dimer), Creatine Kinase-MB (CK-MB), Myoglobin (Myo), Troponin T (cTnT), TyG index (TyG = ln(TG(mg/dl)×GLU(mg/dl)/2)), H-TyG index (TG×LDL×GLU/HDL), NE/L ratio, PLT/L ratio, MO/L ratio, HDL/(TG×LDL) ratio, LDL/HDL ratio, Systemic Immune-Inflammation Index (SII = PLT×NE/L), Triglyceride-glucose index (TyG = ln(TG(mg/dl)×GLU(mg/dl)/2)). Among these, the TyG index has been validated in some studies as an independent risk factor for predicting the severity of arteriosclerosis[23], and the Systemic Immune-Inflammation Index (SII) has been associated with the prognosis of patients with coronary heart disease and heart failure[24].

2.2.3 Construction and Evaluation of the Predictive Model

Following the identification of important feature elements among the possible independent variables, a training set (consisting of 70% of the original dataset of patients experiencing chest discomfort) and a test set (30%) were created. A variety of ML classification models were rigorously analyzed and compared across these datasets to identify the optimal predictive model. The time-matched validation dataset was then applied to the selected model to assess its performance. Furthermore, the contribution of each characteristic to the model's predictions was measured using SHAP values. Lastly, the clinical usefulness of the best model was established utilizing DCA. The particular actions comprised (Fig. 1):

Selection of feature factors: Initially, the R software program (version 4.1.6) was used for variable selection and model complexity modification when conducting the LASSO regression analysis.

ⅰ.New Dataset Establishment: Based on the feature factors selected from the LASSO regression analysis, a new dataset was established.

ⅱ.Data Division: Using the random number method in Python (version 3.9.13), the patients with chest pain were randomly separated into a group for training (247 individuals) and a test group (106 individuals) at a ratio of 7:3.

ⅲ.Comprehensive Analysis of Multi-classification Models: Four models were constructed: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVC), and K-Nearest Neighbors (KNN), and trained and tested using Python (sklearn version 1.0.2). The optimal model was selected by analyzing performance metrics on both the training and test sets, including the confusion matrix and the area under the Receiver Operating Characteristic (ROC) curve.

ⅳ.Application of Ensemble Algorithm Voting: To reduce potential biases of single algorithms, the ensemble algorithm Voting was employed to obtain more fair and balanced diagnostic outcomes, and was evaluated on the time validation dataset.

ⅴ.Further Performance Evaluation: An additional 84 cases were collected for a new time validation dataset to assess the performance of the optimal model.

ⅵ.Decision Curve Analysis (DCA): DCA was performed on the optimal model to determine its potential for clinical practice application. By integrating these steps, we were able to determine which model performs best in terms of clinical applicability, thus providing significant support for medical decision-making.

ⅶ.SHAP Value Analysis: SHAP values were plotted using Python (shap version 0.43.0) to explain the predictive importance of the model and the contribution of each feature.

All factors were included in the statistical analysis of this study in order to assess the differences between patients who had NSTEMI and those who had AD. Continuous variables were reported as medians with interquartile ranges (IQR), and the Mann-Whitney U test was used to evaluate group differences. The chi-squared test was used to compare categorical variables. P-values < 0.05 were deemed statistically significant for the purposes of this study. All statistical analyses were performed using SPSS (version 25.0), R (version 4.1.6), and Python (version 3.9.13) software. The selection of these analytical tools provided us with a robust and flexible platform for analysis, ensuring accurate assessment and interpretation of the data to yield reliable conclusions.

4.1 Selection of Suspected Risk Characteristic Factors for AD and NSTEMI

We performed LASSO regression analysis on 29 potential variables [see Additional file 1] to identify the most impactful risk characteristic factors for the diagnosis of AD and NSTEMI (Fig. 2). The LASSO regression method is particularly advantageous for dealing with variables that exhibit high collinearity and effectively prevents model overfitting by shrinking regression coefficients[25]. With the optimal regularization parameter (λ = 0.721) chosen, at the point of minimum mean squared error, the original 29 independent variables were effectively reduced to 9 variables with the highest diagnostic relevance. These variables include age, White Blood Cell count (WBC), Lymphocyte Percentage (L%), Eosinophil Percentage (E%), Platelet Count (PLT), Platelet Distribution Width (PDW), Triglycerides (TG), D-Dimer, and High sensitivity Cardiac Troponin T (HS-cTnT), each contributing significantly to the predictive power of the model.

4.2 Baseline Characteristics of the Training and Test Cohorts

We divided the data into the training and test groups in a 7:3 ratio using Python's random split function (version 3.9.13). All baseline characteristics did not exhibit statistically significant differences between the two groups, as demonstrated in Table 1 (P > 0.05). This finding indicates the sample division's balanced nature and provides a solid basis for further analyses. Subsequently, Based on the 9 key characteristic factors identified by the LASSO regression, we constructed a new dataset and supplemented it with a new time-validation dataset, as presented in Table 2.

Table 1

Differences in Variables Between the Training and Test Groups
Variables	Training Set (n = 247)	Testing Set (n = 106)	Z	P-value
Age(years)	50[25,75]	65[54,73]	-0.704	0.481
Triglyceride(mmol/L)	1.22[0.87,1.81]	1.23[0.93,1.67]	-0.705	0.941
Ddimer(ug/ml)	1.08[0.42,3.16]	0.81[0.27,3.31]	-0.722	0.47
WBC(10^9/L)	10.72[8.43,13.69]	10.63[7.81,14.01]	-0.839	0.402
Lymphocyte(%)	13.6[8.2,23.5]	14[9.98,26.15]	-1.501	0.133
Eosinophils(%)	0.5[0.1,1.6]	0.6[0.1,1.98]	-1.574	0.116
Platelet(10^9/L)	208[171,249]	215[175,255.5]	0.580	0.58
PDW	11.5[10.4,13.3]	11.8[10.45,14.2]	-0.925	0.355
HS-cTnT(pg/ml)	27.44[8.59,496.9]	31.51[6.53,524.13]	-0.428	0.668
Note: The Z-value was calculated using the Mann-Whitney U test, a non-parametric test for comparing two independent samples. A negative Z-value indicates that the median in the testing set is lower than the median in the training set, while a positive Z-value indicates that the median in the testing set is higher than the median in the training set. The P-value was calculated using the Mann-Whitney U test, which assesses the significance of the difference between the medians of two groups. In this dataset, a P-value > 0.05 suggests that the observed difference is not statistically significant.

Table 2

2022 Temporal Validation Dataset for AD and NSTEMI
Variables	NSTEMI(N = 48)	AD (N = 36)	Z	P-value
Age(years)	68.9 ± 10.8	63.47 ± 10.71	2.285	0.025
Triglyceride(mmol/L)	1.28[0.92,1.97]	1.14[0.74,1.68]	-1.098	0.272
Ddimer(ug/ml)	0.33[0.21,0.67]	1.67[0.79, 3.68]	-5.641	< 0.001
WBC(10^9/L)	7.95[6.70,9.39]	11.19[7.70,16.33]	-3.457	0.001
Lymphocyte(%)	22.8[18.28.88]	13.36[7.33,24.98]	-3.231	0.001
Eosinophils(%)	1.4[0.8,2.35]	0.31[0.1,1.88]	-2.276	0.023
Platelet(10^9/L)	228[194.75,264.75]	204[167.5,248.25]	-1.808	0.071
PDW	12.6[11.43,16.33]	15.35[11.73,16.68]	-1.302	0.193
HS-cTnT(pg/ml)	71.87[14.59,452.78]	12.38[8.04,22.31]	-4.709	< 0.001
Note: Age (years): Z-test for independent samples. Triglyceride (mmol/L), D-dimer (ug/ml), WBC (10^9/L), Lymphocyte (%), Eosinophils (%), Platelet (10^9/L), PDW, HS-cTnT (pg/ml) for Mann-Whitney U test. P-value < 0.05 suggests that the observed difference is statistically significant.

4.3 Multi-Model Analysis in Machine Learning

In this study, we input the split training dataset into four different machine learning algorithms: LR, RF, SVC, and KNN for model training[26]. To comprehensively evaluate the performance of these models, we employed metrics from the confusion matrix (Accuracy, Precision, Recall, F1-Score) (Table 3) and the Area Under the Curve (AUC). Guided by these metrics, the RF model demonstrated its superiority throughout the test and training sets, achieving the highest AUC value of 0.97 (95CI%: 0.93-1.00). In the test set, the RF achieved an accuracy of 0.9057 and an F1 score of 0.9074, outperforming the other models (Table 4). Based on these evaluation metrics, we can assert that the RF exhibits the most outstanding performance among this suite of models[27](Figs. 3a and 3b).

Table 3

Confusion Matrix Metrics
Metrics	Description
		Positive	Negative
Confusion Metrics	Positive	True Positive	False Positive
	Negative	False Negative	True Negative

Note: True Positives (TP): The quantity of samples that the model accurately predicts as positive. False Positives (FP), often referred to as Type I errors or false alarms, are the number of samples that the model mistakenly predicts as positive. True Negatives (TN): The quantity of samples that the model accurately predicted to be negative. False Negatives (FN): The quantity of samples that the model mispredicted to be negative; sometimes referred to as a Type II mistake or miss.

Despite this, we also recognize that each base classifier may possess different biases and variances, which could lead to overfitting issues. To improve the overall predictive performance and obtain more stable prediction results, we employed the Voting ensemble method. The model, after Voting ensemble, experienced a slight decrease in AUC value to 0.95 (95CI%: 0.91–0.99) in the test set. However, it showed improvements in accuracy, recall, precision, and F1 score, reaching 0.9151, 0.9412, 0.8889, and 0.9143 respectively, overall surpassing the performance of individual bas e classifiers[28].

Table 4

Confusion Matrix Analysis Results for the Four Algorithms
		Recall	Precision	F1 Score
K-Nearest Neighbors	0.7453	0.8431	0.6935	0.7611
Logistic Regression	0.8868	0.9020	0.8679	0.8846
Support Vector Machine	0.6792	0.9412	0.6076	0.7385
Random Forest	0.9057	0.9608	0.8596	0.9074

Furthermore, the AUC metric focuses on the predictive accuracy of the model rather than determining its clinical utility[29]. To assess the model's actual value in clinical applications, we conducted DCA on the Voting ensemble model[30]. The analysis indicated that at a threshold of 0.5, the model maintains a good balance in predicting positive and negative samples[22]. The DCA curve remains above the net benefit line across most of the threshold range, demonstrating the model's potential clinical value in predicting AD. While the net benefit of the model decreases as the threshold approaches 1, considering other metrics, the model's predictions remain meaningful and valuable (Fig. 4)[31].

In conclusion, the Voting ensemble model is identified as the optimal model, incorporating the strengths of the individual base classifiers while reducing the risk of overfitting and providing more robust prediction results[32].

In order to verify the model's generalization ability on new data, we collected data from 2022 as an external validation set and assessed the performance of the model. On this validation set, our model continued to demonstrate high performance[29], with an AUC of 0.947 (95CI%: 0.899–0.986), accuracy of 0.847, precision of 0.780, F1 score of 0.831, and recall of 0.889 (Figs. 5). Compared to the test set, there was a slight decrease in performance on the validation set, which may be related to differences in data distribution between the new time period and the earlier datasets.

4.5 Model Interpretation with SHAP

To explain the impact of selected variables on the prediction of AD versus NSTEMI, we used the SHAP method[33]. Since the Voting model cannot directly calculate SHAP values, we employed a RF as the base model for analysis[34]. Figure 6 (a) shows the nine key variables ranked by importance, while Fig. 6 (b) presents their impact on the model's predictions using mean absolute SHAP values[35]. Notably, D-dimer and Troponin T have a significant positive effect on predictions, followed by white blood cell count. These insights help medical professionals understand how the model distinguishes between AD and NSTEMI based on biomarker expressions.

In this study, we analyzed data from 353 patients at the Chest Pain Center of the First Affiliated Hospital of Shantou University Medical College. We categorized the patients into two groups: AD and NSTEMI. We used Lasso regression and a Voting ensemble model to analyze the data, and validated the model's performance with new cases. The SHAP method was also employed to improve interpretability and enhance the distinction between AD and NSTEMI diagnoses. We discuss the methods, results, and their application in clinical decision-making.

Lasso regression played a crucial role in selecting predictive variables for AD and NSTEMI. By reducing unnecessary features, the model focused on significant predictors such as White Blood Cell count (WBC) and Neutrophil percentage (NE%). From 29 potential variables, we selected 9 key predictive variables, providing a solid foundation for subsequent models[36].

We utilized various machine learning algorithms, including Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVC), and K-Nearest Neighbors (KNN), to capture complex patterns. The Voting ensemble model outperformed other algorithms by synthesizing multiple models and reflecting the intrinsic structure of the data[37]. Model performance was evaluated using metrics such as the confusion matrix, area under the ROC curve (AUC), and Decision Curve Analysis (DCA). These metrics confirmed the accuracy of the Voting model and its value in clinical practice[38]. The SHAP approach quantified each feature's contribution to the model's predictions, enhancing transparency and interpretability. This enabled clinicians to understand and trust the model's decision-making process, crucial for medical decision-making[39].

Some limitations of our study should be noted. The restricted nature of the dataset, highlighting the need for validation on data from additional research centers[40]. Furthermore, exploring other advanced machine learning or deep learning methods could enhance the approach. The applicability of the developed model to other diseases also requires further research and investigation[41].The combination of Lasso regression, the Voting ensemble model, and the SHAP method provides a new perspective for the differential diagnosis of AD and NSTEMI. It highlights the importance of integrating methods and enhancing model interpretability in clinical prediction. Future research should address limitations and expand the application of machine learning in clinical decision-making to optimize medical decisions.

In this study, we established a machine learning-based diagnostic approach employing 29 clinical variables and standardized laboratory biomarkers to timely differentiate between AD and NSTEMI. According to our findings, the Voting ensemble model demonstrated superior predictive efficacy. Additionally, we leveraged SHAP to provide personalized risk assessments for the predictions made. This diagnostic model can assist emergency department clinicians in rapidly avoiding the misdiagnosis of AD with concurrent NSTEMI.

AD	Aortic dissection
NSTEMI	Non-ST elevation myocardial infarction
ACS	Acute coronary syndrome
LASSO	Least Absolute Shrinkage and Selection Operator
ML	Machine Learning
AI	Artificial Intelligence
LR	Logistic Regression
RF	Random Forest
SVC	Support Vector Classification
KNN	K-Nearest Neighbors
Shap	Shapley Additive Explanations
DCA	Decision Curve Analysis
CT	Computed Tomography
CAG	Coronary Angiography
ECG	Electrocardiograms
CTA	Computed Tomography Angiography
MRI	Magnetic Resonance Imaging
MRA	Magnetic Resonance Angiography

Ethics approval and consent to participate:

The study protocol and informed consent procedures were reviewed and approved by the Institutional Ethics Committee. The approval reference number for this study is B-2023-060.

Consent for publication:

Not applicable.

Competing interests:

The authors declare no competing interests.

Funding:

This research was funded by the China Zhongguancun Precision Medicine science and technology foundation.

Author Contribution

M.H. model building, R language writing and wrote the main manuscript text; L.L. database maintenance; X.F. Proofreading of articles and pictures, Y.W. Project design, financial support, and revision the article. All authors reviewed the manuscript.

Availability of data and materials:

Data is contained within the Supplementary Material.

Tchana-Sato V, Sakalihasan N, Defraigne JO. [Aortic dissection]. Rev Med Liege. 2018;73(5–6):290–5.
Zhu Y, Lingala B, Baiocchi M, Tao JJ, Toro Arana V, Khoo JW, Williams KM, Traboulsi AA, Hammond HC, Lee AM, et al. Type A Aortic Dissection-Experience Over 5 Decades: JACC Historical Breakthroughs in Perspective. J Am Coll Cardiol. 2020;76(14):1703–13.
Murugiah K, Wang Y, Nuti SV, Li X, Li J, Zheng X, Downing NS, Desai NR, Masoudi FA, Spertus JA, et al. Are non-ST-segment elevation myocardial infarctions missing in China? Eur Heart J Qual Care Clin Outcomes. 2017;3(4):319–27.
Dondo TB, Hall M, Timmis AD, Gilthorpe MS, Alabas OA, Batin PD, Deanfield JE, Hemingway H, Gale CP. Excess mortality and guideline-indicated care following non-ST-elevation myocardial infarction. Eur Heart J Acute Cardiovasc Care. 2017;6(5):412–20.
Wang ZY, Wang YJ, Liu XL, Zhao X, Mao CY. Aortic dissection with acute non-st-segment elevation myocardial infarction as the first manifestation. J Biol Regul Homeost Agents. 2017;31(3):711–5.
Morello F, Santoro M, Fargion AT, Grifoni S, Nazerian P. Diagnosis and management of acute aortic syndromes in the emergency department. Intern Emerg Med. 2021;16(1):171–81.
Case BC, Weintraub WS. Non-ST-Segment-Elevation Myocardial Infarction: When Is Rapid Revascularization Critical? J Am Heart Assoc. 2021;10(19):e023645.
Xue Y, Chong H, Zhu X, Fan F, Wang D, Zhou Q. Aortic dissection patients mimic acute coronary syndrome with preoperative antiplatelet therapy. J Thorac Dis. 2019;11(8):3385–90.
Baliga RR, Nienaber CA, Bossone E, Oh JK, Isselbacher EM, Sechtem U, Fattori R, Raman SV, Eagle KA. The role of imaging in aortic dissection and related syndromes. JACC Cardiovasc Imaging. 2014;7(4):406–24.
Hagan PG, Nienaber CA, Isselbacher EM, Bruckman D, Karavite DJ, Russman PL, Evangelista A, Fattori R, Suzuki T, Oh JK, et al. The International Registry of Acute Aortic Dissection (IRAD): new insights into an old disease. JAMA. 2000;283(7):897–903.
Balmforth D, Harky A, Adams B, Yap J, Shipolini A, Roberts N, Uppal R, Bashir M. Is there a role for biomarkers in thoracic aortic aneurysm disease? Gen Thorac Cardiovasc Surg. 2019;67(1):12–9.
Hahne K, Lebiedz P, Breuckmann F. Impact of d-Dimers on the Differential Diagnosis of Acute Chest Pain: Current Aspects Besides the Widely Known. Clin Med Insights Cardiol. 2014;8(Suppl 2):1–4.
Cho SY, Kim SH, Kang SH, Lee KJ, Choi D, Kang S, Park SJ, Kim T, Yoon CH, Youn TJ, et al. Pre-existing and machine learning-based models for cardiovascular risk prediction. Sci Rep. 2021;11(1):8886.
Dey D, Slomka PJ, Leeson P, Comaniciu D, Shrestha S, Sengupta PP, Marwick TH. Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art Review. J Am Coll Cardiol. 2019;73(11):1317–35.
Liao B, Jia X, Zhang T, Sun R. DHDIP: An interpretable model for hypertension and hyperlipidemia prediction based on EMR data. Comput Methods Programs Biomed. 2022;226:107088.
D'Ascenzo F, De Filippo O, Gallone G, Mittone G, Deriu MA, Iannaccone M, Ariza-Solé A, Liebetrau C, Manzano-Fernández S, Quadri G, et al. Machine learning-based prediction of adverse events following an acute coronary syndrome (PRAISE): a modelling study of pooled datasets. Lancet. 2021;397(10270):199–207.
Charilaou P, Battat R. Machine learning models and over-fitting considerations. World J Gastroenterol. 2022;28(5):605–7.
Burgman MA, Regan HM, Maguire LA, Colyvan M, Justus J, Martin TG, Rothley K. Voting systems for environmental decisions. Conserv Biol. 2014;28(2):322–32.
Wang MH, Li NF, Luo Q, Wang GL, Heizhati M, Wang L, Wang L, Zhang WW. Development and validation of a novel diagnostic nomogram model to predict primary aldosteronism in patients with hypertension. Endocrine. 2021;73(3):682–92.
Petch J, Di S, Nelson W. Opening the Black Box: The Promise and Limitations of Explainable Machine Learning in Cardiology. Can J Cardiol. 2022;38(2):204–13.
Shah RV, Grennan G, Zafar-Khan M, Alim F, Dey S, Ramanathan D, Mishra J. Personalized machine learning of depressed mood using wearables. Transl Psychiatry. 2021;11(1):338.
Allyn J, Ferdynus C, Bohrer M, Dalban C, Valance D, Allou N. Simplified Acute Physiology Score II as Predictor of Mortality in Intensive Care Units: A Decision Curve Analysis. PLoS ONE. 2016;11(10):e0164828.
Zhao J, Fan H, Wang T, Yu B, Mao S, Wang X, Zhang W, Wang L, Zhang Y, Ren Z, et al. TyG index is positively associated with risk of CHD and coronary atherosclerosis severity among NAFLD patients. Cardiovasc Diabetol. 2022;21(1):123.
Xia Y, Xia C, Wu L, Li Z, Li H, Zhang J. Systemic Immune Inflammation Index (SII), System Inflammation Response Index (SIRI) and Risk of All-Cause Mortality and Cardiovascular Mortality: A 20-Year Follow-Up Cohort Study of 42,875 US Adults. J Clin Med 2023, 12(3).
McNeish DM. Using Lasso for Predictor Selection and to Assuage Overfitting: A Method Long Overlooked in Behavioral Sciences. Multivar Behav Res. 2015;50(5):471–84.
Jing X, Wang X, Zhuang H, Fang X, Xu H. Multiple Machine Learning Approaches Based on Postoperative Prediction of Pulmonary Complications in Patients With Emergency Cerebral Hemorrhage Surgery. Front Surg. 2021;8:797872.
Zhou Y, Han W, Yao X, Xue J, Li Z, Li Y. Developing a machine learning model for detecting depression, anxiety, and apathy in older adults with mild cognitive impairment using speech and facial expressions: A cross-sectional observational study. Int J Nurs Stud. 2023;146:104562.
Sun R, Hou X, Li X, Xie Y, Nie S. Transfer Learning Strategy Based on Unsupervised Learning and Ensemble Learning for Breast Cancer Molecular Subtype Prediction Using Dynamic Contrast-Enhanced MRI. J Magn Reson Imaging. 2022;55(5):1518–34.
Mustafić S, Brkić S, Prnjavorac B, Sinanović A, Porobić Jahić H, Salkić S. Diagnostic and prognostic value of procalcitonin in patients with sepsis. Med Glas (Zenica). 2018;15(2):93–100.
Van Calster B, Wynants L, Verbeek JFM, Verbakel JY, Christodoulou E, Vickers AJ, Roobol MJ, Steyerberg EW. Reporting and Interpreting Decision Curve Analysis: A Guide for Investigators. Eur Urol. 2018;74(6):796–804.
Liu Y, Wang M, Qiu X, Ma G, Ji M, Yang Y, Sun M. A novel clinical-imaging nomogram for predicting primary aldosteronism in patients with hypertension. Hypertens Res. 2023;46(12):2603–12.
Reshan MSA, Amin S, Zeb MA, Sulaiman A, Alshahrani H, Azar AT, Shaikh A. Enhancing Breast Cancer Detection and Classification Using Advanced Multi-Model Features and Ensemble Machine Learning Techniques. Life (Basel) 2023, 13(10).
Chowdhury SU, Sayeed S, Rashid I, Alam MGR, Masum AKM, Dewan MAA. Shapley-Additive-Explanations-Based Factor Analysis for Dengue Severity Prediction using Machine Learning. J Imaging 2022, 8(9).
Nordin N, Zainol Z, Mohd Noor MH, Chan LF. An explainable predictive model for suicide attempt risk using an ensemble learning and Shapley Additive Explanations (SHAP) approach. Asian J Psychiatr. 2023;79:103316.
Bloch L, Friedrich CM. Data analysis with Shapley values for automatic subject selection in Alzheimer's disease data sets using interpretable machine learning. Alzheimers Res Ther. 2021;13(1):155.
Wang Q, Qiao W, Zhang H, Liu B, Li J, Zang C, Mei T, Zheng J, Zhang Y. Nomogram established on account of Lasso-Cox regression for predicting recurrence in patients with early-stage hepatocellular carcinoma. Front Immunol. 2022;13:1019638.
Mastour H, Dehghani T, Moradi E, Eslami S. Early prediction of medical students' performance in high-stakes examinations using machine learning approaches. Heliyon. 2023;9(7):e18248.
Li Y, Lv X, Wang B, Xu Z, Wang Y, Sun M, Hou D. Predicting EGFR T790M Mutation in Brain Metastases Using Multisequence MRI-Based Radiomics Signature. Acad Radiol. 2023;30(9):1887–95.
Xue B, Li D, Lu C, King CR, Wildes T, Avidan MS, Kannampallil T, Abraham J. Use of Machine Learning to Develop and Evaluate Models Using Preoperative and Intraoperative Data to Identify Risks of Postoperative Complications. JAMA Netw Open. 2021;4(3):e212240.
Dian R, Guo A, Li S. Zero-Shot Hyperspectral Sharpening. IEEE Trans Pattern Anal Mach Intell. 2023;45(10):12650–66.
Prihoda D, Maamary J, Waight A, Juan V, Fayadat-Dilman L, Svozil D, Bitton DA. BioPhi: A platform for antibody design, humanization, and humanness evaluation based on natural antibody repertoires and deep learning. MAbs. 2022;14(1):2020203.

No competing interests reported.

Additionalfile1.pdf

Development of a Machine-learning Based Diagnosis Procedure to Distinguish Aortic Dissection from Non-ST- Elevation Myocardial Infarction

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

2.1.1 Participants

2.1.2 Inclusion Criteria

2.2 Research Methodology

2.2.1 Diagnostic Approach for Diseases

2.2.2 Study Variables

2.2.3 Construction and Evaluation of the Predictive Model

3. Statistical Analysis

4. Results

4.1 Selection of Suspected Risk Characteristic Factors for AD and NSTEMI

4.2 Baseline Characteristics of the Training and Test Cohorts

4.3 Multi-Model Analysis in Machine Learning

4.5 Model Interpretation with SHAP

5. Discussion

6. Conclusion

Abbreviations

Declarations

Competing interests:

Funding:

Author Contribution

Availability of data and materials:

References

Additional Declarations

Supplementary Files

Status:

Version 1