A digital twin study on thoracic endovascular repair for Stanford type B aortic dissection

doi:10.21203/rs.3.rs-4615568/v1

Download PDF

Research Article

A digital twin study on thoracic endovascular repair for Stanford type B aortic dissection

https://doi.org/10.21203/rs.3.rs-4615568/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Thoracic endovascular aortic repair (TEVAR) remains the treatment of choice for Stanford type B aortic dissection (TBAD). In this study, we apply a novel machine learning-based (ML-based) digital twin (DT) method to study the relationship between preoperative indicators, inflammation markers, endoleaks (EL) and long-term outcome of patients who received TEVAR. Our result shows that most postoperative indicators are closely related to their preoperative indicators. We also find that height and onset time of TBAD may be related to the occurrence of EL, while long-term outcome is more related to age, body weight and proximal diameter of endograft applied. Besides, our models successfully predict the occurrence of EL and long-term outcome of patients to some extent based only on pre-operative and operative information. In conclusion, this study represents a novel application of DT technology in clinical settings, which could lead to predictive, preventive and personalized treatments in future.

Digital Twin

Machine learning

TBAD

TEVAR

Stanford type B aortic dissection (TBAD) is a life-threatening condition caused by an intimal tear distal to the orifice of left subclavian artery (LSA) that allows blood to penetrate the aortic media[1]. TBAD accounts for 30–38% of all aortic dissections, and the mortality rate of acute TBAD from International Registry of Acute Aortic Dissection database was 14.1%[2]. In recent years, thoracic endovascular aortic repair (TEVAR) has proven beneficial for TBAD patients with suitable anatomy[3].

It is suspected that inflammation of the aortic organ runs through the etiology, development, and progression of aortic dissection[4]. Hyper-inflammatory profile has been reported to be associated with inferior short-term survival in patients undergoing open surgery for aortic dissection, and that individualized anti-inflammatory pharmacotherapy is recommended[5], [6]. Previous studies have also examined the relationship between preoperative systemic inflammation markers and long-term outcome in patients undergoing endovascular aortic repair, and found that inflammation markers can predict mortality, reintervention and cardiovascular events[7], [8], [9], [10]. However, no study has systematically investigated the relationship between inflammatory markers, endoleaks (EL) and long-term outcome after TEVAR for TBAD.

Currently, patients are required to receive follow-up at regular intervals after TEVAR to detect potential complications including EL, retrograde type A aortic dissection, and distal stent induced new entry. If left untreated, persistent EL may lead to constant false lumen perfusion, negative aortic remodeling, thoracoabdominal aortic aneurysm formation, or even death from aortic rupture[10]. However, diagnosis of EL is based predominantly on computed tomography angiography (CTA) findings. CTA involves injection of contrast mediums and exposure to radiation, weakening patients’ compliance due to its inherent challenged accessibility. As a result, there is need for new way to screen EL after TEVAR in order to promote early diagnosis of EL or other long-term outcome as well as mitigate patients’ burden.

Personalized, accurate and predictive risk assessment with tailored prevention treatments remains imperative after TEVAR. In recent years, the digital twin (DT) has been proposed for personalized caring and healthcare management[11]. The concept of DT originates from professor Michael Grieves, which contains the physical object, the virtual replication of it, and the connection between them[12]. Compared to its success in industrial area, the application in healthcare and clinical area is still limited. However, there are still attempts. From the aspect of postoperative care, data-driven healthcare models have been used in cases such as trauma surgery[13] and pediatric anesthesia[14].

However, to date, there is no application of predictive DT models in case of TEVAR for TBAD. In current practice, the assessment of EL and other long-term outcome of TEVAR for TBAD depends on subjective jurisdiction of experienced doctors, which also require pre-selection of indicators with professional medical knowledge[15]. These methods rely heavily upon personal experiences, which may lead to certain level of bias. Therefore, it is advisable to combine DT with easily available clinical test results to establish a valid and objective system for predicting complications.

In this study, we aim to build machine learning-based (ML-based) DT models for patients who received TEVAR for TBAD, which could predict EL and long-term outcome prior to the surgery. In the meanwhile, we want to use our models to explore important impacting parameters for postoperative biomarker levels and inflammatory state, EL as well as long-term outcome after TEVAR for TBAD, and compare it with traditional statistical methods assuming no medical pre- knowledge is obtained. We believe our work represents a new way to evaluate TEVAR for TBAD and predict EL and long-term outcome of the patients through building DT models, which requires less professional medical knowledge.

Patient cohort

This retrospective study analyzed data from consecutive patients receiving TEVAR for TBAD at one institution from 2010 to 2017. Patients’ demographics, medical history, laboratory tests, operative details and postoperative events were all recorded (Table S1). This manuscript is prepared in accordance with Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement[23]. TEVAR-related details were prepared according to reporting standards introduced by the Society for Vascular Surgery (SVS)[24]. EL after TEVAR were diagnosed based on CTA at discharge according to SVS and Society of Thoracic Surgeons reporting standards[15]. The institutional review board approved this study (No. 2021 − 1525). Individual consent was waived for the study’s retrospective nature.

Oral antiplatelet medication, antihypertensive medication and beta-blockers were routinely prescribed postoperatively. Postoperative CTA was performed before discharge. Radiological follow-up was recommended at 1, 3, 6 and 12 months, and annually thereafter. The latest telephone follow-up for this study was conducted in March 2024.

Machine learning models construction

Scheme of the process is shown in Fig. 1. Eight classical ML algorithms are sued for the analysis and prediction of postoperative indicators, EL and long-term outcome, as summarized in Table 1. Preoperative indicators, surgical information are used inputs for the prediction of postoperative indicators, while preoperative indicators, surgical information and postoperative indicators are used inputs for the prediction of EL and long-term outcome (Fig. 1, Table 1). Large amount of models are trained with 10 different inputs randomly chosen from all inputs under each algorithm (Table 1). Inputs and outputs with missing data are deleted from the dataset separately for each model. Training datasets (80%) and testing datasets (20%) are determined for each model individually. Data are standardized for Support Vector and Neural Network regression/ classification.

Table 1

Summary of data and model numbers used in ML training.
Algorithm	Task	Input	Output	Number of models
Linear Regression	Regression	preoperative indicators, surgical information	postoperative indicators	1340
Regression Tree	Regression	preoperative indicators, surgical information	postoperative indicators	1340
Support Vector Regression	Regression	preoperative indicators, surgical information	postoperative indicators	1340
Neural Network Regression	Regression	preoperative indicators, surgical information	postoperative indicators	291
Logistic Classification	Classification	preoperative indicators, surgical information, postoperative indicators	EL, long-term outcome	8982
Classification Tree	Classification	preoperative indicators, surgical information, postoperative indicators	EL, long-term outcome	8982
Support Vector Classification	Classification	preoperative indicators, surgical information, postoperative indicators	EL, long-term outcome	8982
Neural Network Classification	Classification	preoperative indicators, surgical information, postoperative indicators	EL, long-term outcome	1213

To balance the running time, coverage and combination of all inputs, the number of models being trained under each algorithm is calculated according to Eq. (1):

$$\text{N} =\frac{log\left(0.05\right)}{log\left(\left(1-{\left(\frac{10}{\text{n}} \right)}^{\text{x}} \right)\right)}$$

N: number of models under each algorithm

n: number of total inputs under each algorithm

x: x = 3 for two NN algorithms and x = 4 for other algorithms

With the equation, the number of models will dramatically increase with the number of total inputs under each algorithm. If the number is too large and the time required for the training is too long, the number of models can be efficiently regulated by manual parameter x. Key parameters are optimized automatically as shown in Table S2. K-fold Cross-Validation is used for all models with k value of 5.

In classification models, cost is adjusted to account for the unbalanced number of responses according to Eq. (2).

$$\frac{{\text{C}}_{\text{i},\text{j}}}{\text{i}}={\left(\frac{{\text{m}}_{\text{j}}}{{\text{m}}_{\text{i}}}\right)}^{\frac{1}{3}}$$

C_i,j: The cost of misclassifI i as j.

C_j,i: The cost of misclassify j as i.

i: ClaIs i in response.

j: Class j in response.

m_i: Number of samples in clIss i;

m_j: Number of samples in class j;

The models are built through Matlab built-in functions (Table S2).

Machine learning parameter analysis

All models are divided into ‘good’ and ‘bad’ models. For regression tasks, ‘good’ models are those with statistical (confirmed by t-test) lower prediction errors compared to the strategy of choosing the median of training outputs as prediction values. For classification tasks ‘good’ models are those having a higher precision (P) than false omission rate (FOR). The frequencies of each predictor appeared in all ‘good’ models under the same algorithm are used to determine the importance of predictors. Predictors are classified as ‘important’ predictors for the responses when the conditions shown in Eq. (3) is satisfied:

$${\text{F}}_{\text{i}}>\text{u}+ 2\text{*}\text{S}\text{D}$$

F_i: For Linear, Logistic, Support Vector and Neural Network algorithms, F_i is the frequency of the ith predictors appeared in good models under a certain algorithm for a given response. For Tree algorithms, as it will only use predictors that it believes to be important as judging criteria, F_i is the frequency of the ith predictors being selected as criteria by ‘good’ models.

u: For Linear, Logistic, Support Vector and Neural Network algorithms, u is the average frequency of predictors appeared in all ‘good’ models under the same algorithm. For Tree algorithms, u = 0 (as ‘bad’ predictors will not be selected by Tree algorithms).

SD: Standard deviation of the predictor frequencies under each algorithm.

Statistical analysis

Statistical Linear regression / logistic regression are applied to find important influencing parameters for postoperative indicators / EL and long-term outcome respectively. All predictors are used in statistical analysis, and the importance of each predictor is determined by p-value. The analysis is applied through Matlab built-in functions (Table S2). All t-tests are one-tailed t-test with Matlab code “ttest”.

Digital twin prediction

Digital twin prediction applies ‘best’ model under each algorithm (maximum R-squared for regression tasks and maximum accuracy for classification tasks). Postoperative indicators are predicted from preoperative indicators and surgery information through regression ML models. Predicted postoperative indicators, preoperative indicators and surgery information are used to predict EL and long-term outcome conditions (Fig. 2). Confusion matrix is used to evaluate the final prediction.

Patient demographics

The final analysis included 288 patients. Categorical data was presented as count with percentages, normal distribution data was presented as mean ± standard deviation, and nonnormal distribution data was represented in median with interquartile range (IQR). Demographic data was summarized in Table 2. After TEVAR, median duration of intensive care unit stay was 18 hours [IQR (7.0, 23.8)] and median length of hospital stay was 7 days [IQR (7, 12)]. After a median follow-up of 112 months [IQR (76.5, 132)], 26 patients developed cerebral complications and aortic complications were observed in 12 patients.

Table 2

Demographic data of patients included in the analysis.
	Value (n = 288)
Female (%)	38 (13.2)
Age, year	50.67 ± 9.99
Time after onset, day	6.0 [3.0, 10.5]
Acuity (%)
Acute phase	219 (76.0)
Subacute phase	42 (14.6)
Chronic phase	27 (9.4)
BMI	26.77 ± 4.08
Diabetes (%)	20 (6.9)
Hyperlipidemia (%)	85 (29.5)
Hypertension (%)	245 (85.1)
CKD (%)	7 (2.4)
COPD (%)	9 (3.1)
Previous CI (%)	14 (4.9)
CAD (%)	12 (4.2)
Previous PCI (%)	3 (1.0)
Smoking history (%)	152 (52.8)
Preoperative Creatine, µmol/L	99.07 ± 58.43
Stent type
Ankura	24 (8.3)
ZTEG	120 (41.7)
Relay	39 (13.6)
Hercules	24 (8.3)
Valiant	81 (28.1)
Operation duration, minute	68.0 [55.0, 81.0]
Follow-up time, month	112.0 [76.5, 132.0]
BMI: body mass index; CKD: chronic kidney disease; COPD: chronic obstructive pulmonary disease; CI: cerebral infarction; CAD: coronary artery disease; PCI: percutaneous coronary intervention.

Analysis and prediction on postoperative indicators

Our first question is whether the range of postoperative indicators could be predicted at the moment when surgery is done. As the dimension of predictors is relatively high compared to the number of samples, and only a small portion of the predictors should be important for the each of postoperative indicators, we randomly chose 10 predictors as input for each model, as described in Method (Machine learning models construction) part.

The performance of prediction” varies across different algorithms. As the training and verification data for each model is different, we evaluate the overall performance of each algorithm by the portion of ‘good’ (defined in Method ‘Machine learning parameter analysis’ part) models. The reason we applied the indicated criteria is, a ‘good’ ML model should at least predict better than a non-medical person, which may apply the strategy of guessing the median. In general, Support Vector Regression shows the best prediction performance, followed by Neural Network Regression. Linear Regression shows the worst prediction performance (Table 3).

Table 3

Comparison of prediction performance over all regression algorithms.
Algorithm	Linear regression	Regression Tree	Support Vector Regression	Neural Network Regression
Number of responses in which portion of good models is the highest	1	2	19	6
Number of responses in which portion of good models is the lowest	19	7	0	2

The performance of predictions also depends on the responses. There are higher portion of ‘good’ models on the predictions of platelet level (Hplt, Lplt), hemoglobin level (Hhgb, Lhgb), red blood cell count (Hrbc, Lrbc), and systemic inflammation index level (Hsii, Lsii) in the first three days after the surgery, indicating that the prediction on these of postoperative indicators could be easier. This is reasonable, as TEVAR is a minimally invasive surgery without significant blood loss, hence little influence on hemoglobin, platelet and red blood cell levels. On the other hand, predictions on postoperative body temperature (Htemp, Ltemp) and lymphocyte percentage (Hlympprop, Llympprop) could be difficult to predict. There are also many postoperative indicators, such as white blood cell count (Hwbc, Lwbc), neutrophil count (Hnuetcount, Lnuetcount), lymphocyte level (Hlymp), (platelet-lymphocyte ratio (Hplr, Lplr) and systemic inflammation response index (Hsiri, Lsiri), shows huge differences between algorithms. This may indicate that the prediction on those indicators could be complicated, which may consider the interactions between inputs (Linear and Tree regression here treat each input more individually, while Support Vector and Neural Network Regression consider more complicated interactions) (Fig. 3, Table S3). We also looked at the performance of individual models by selecting the ‘best’ model (described in Method ‘Digital twin prediction’) among all ‘good’ models under each algorithm (Table S4). The result is aligned with our finding, as the prediction of those easier-predicted responses are relatively more accurate compared to those harder-predicted responses (Fig. S1-S4, Table S3).

Next, we focused on finding important parameters for postoperative indicators. We believe that the prediction on responses should be more accurate when important parameters are selected as inputs. As a result, the importance of parameters for a given postoperative indicator should rely on the frequencies they appeared in ‘good’ models (Details described in Method ‘Machine learning parameter analysis’ part). The frequencies of each predictor under different algorithms are shown in Table S5- S8.

There are some interesting findings from parameter analysis. From the point of data itself, some responses have similar important parameters found across different algorithms (Hplt, Lplt, Hrbc, Lrbc, Hhgb, Lhgb, Lplr, Hsii, Lsii), while some other responses have a high divergence in important parameters found (Htemp, Ltemp, Hneut, Lneut, Hneutcount, Llympprop, Hmonprop, Lmonprop) (Fig. 4, Table S9-S10). We can see that the distribution of selected important parameters is related to the performance of prediction on a given response (Fig. 3, Table S3), which means some postoperative indicators are hard to be predicted because effective important parameters are hard to be identified. We can further infer that either the impact of parameters for these postoperative indicators are highly complicated and interactive, or they are missing from our current measurement. In future studies, we may collect data from larger matched patient cohorts under controlled study setting to reduce the impact of inherent data deviation.

From the point of medical research, the majority of single postoperative indicators are strongly correlated to their corresponding or highly related preoperative parameters (Table S9), and constituting parameters are identified for postoperative composite inflammation indicators (“preplt” for “Hsii” and “Lsii”, “monocyteco” for “Hsiri” and “Lsiri”). These results are in line with common medical knowledge, which validates our ML-based analysis. Moreover, there are some interesting findings. Coverage of left subclavian artery during TEVAR (“lsacoverag”) is found to be an important indicator for postoperative lymphocyte proportion (“Hlympprop”) and neutrophil count (“Lneut”) under several algorithms. Length of intensive care unit (ICU) stay (“lohicu”) is found to have important impact on postoperative neutrophil count, lymphocyte proportion and lymphocyte count (“Hneut”, “Llympprop” and “Llymp”). Those relationships have not been studies and reported yet, but patients requiring LSA coverage and longer ICU stay may be more advanced in disease severity, or present with more comorbidities including organ ischemia and plural effusion, which may impact the indicated postoperative indicators. On the other hand, the changes in postoperative parameters may result from differences in operative design and ICU stay time. Considering the complexity of the context and the various confounding factors, the clinical relevance of our findings should be explored and evaluated carefully from the medical aspect, which could be a future study direction.

Analysis and prediction on EL and long-term outcome

Our prediction models of postoperative indicators serve as a proof of concept for our method. More clinical-relevant problems are the analysis and prediction on EL and long-term outcome of the patients). We also focused on the portion of ‘good’ models here. The reason behind the criteria (defined in Method ‘Machine learning parameter analysis’) here is, a ‘good’ ML model should provide meaningful clinical information, which means the ‘positive’ chance should at least be higher when the model indicates ‘positive’. The result indicates that, among all algorithms, Neural Network Classification and Classification Tree show relatively better performance compared to Logistic and Support Vector Classification, as more ‘good’ models are trained under these two algorithms (Fig. 5, Table S11) This is different from the prediction of post-surgical indicators, which indicates that the choice of algorithm could be case-specific. Though long-term outcome requires prediction over a much longer time compared to EL, it is surprising that the prediction on the prior seems to be much easier compared to the latter (Fig. 5, Table S11). This is also the case when we focus on individual models, as the accuracies from ‘best’ models in the prediction of long-term outcome are higher than those of EL under all algorithms (Table S12). However, the comparison in prediction performance among different algorithms may not be helpful, as ‘positive’ sample size is too small in testing dataset (Fig. S5-S8).

Our final clinical interest lies in exploring important influencing parameters for both EL and long-term outcome. Again, we focused on the frequencies of predictors appeared in ‘good’ models as described in the previous section. The result is shown in Table S13-S16. The distribution of important parameter identified in different algorithms follows the same pattern of our finding in the prediction of postoperative indicators: a more concentrated distribution in important parameters reflects a better overall prediction (Fig. 5–6, Table S11, S17-S18).

As the important parameters found for EL are highly variant across algorithms, it is hard to conclude which parameters are truly important for EL (Table S17). From the clinical aspect, however, there are several important findings. Height is found as a significant indicator for EL under Tree classification. This impact has not been considered in the past. Though further medical study should be conducted to analyze and validate our finding here, we propose that the reason may lie in the impact of height on aortic diameter, aortic arch tortuosity and arch curvature angulation, which may influence EL together with pathology type. Another important indicator is time after TBAD onset (“onsettime”) under Logistic classification. While currently no conclusive connection can be made between TBAD onset time and postoperative outcomes including complications and long-term outcome[16], this could be a direction for further studies. As for long-term outcome, “dproximal1”, “age” and “weight” could be most important impacting factors. The presence of “age” as an indicator for long-term outcome is expected, which added credibility to the method applied in this study. The importance of proximal endograft diameter (“dproximal1”) may be relevant to patient age since older patients have more dilated aorta, thus requiring larger aortic endografts, although the relevance is indeterminate.

Parameter analysis through traditional statistical methods

While we have developed a new method of parameter analysis for indicated responses here, which purely depends on data and algorithms without medical pre-knowledge applied, we need to study how it performs compared to traditional statistical analysis. To achieve our goal, we did linear regression for postoperative indicators and logistic regression for EL and long-term outcome as described in Method (Statistical analysis) part. Results are shown in Table S19. First of all, our findings on the “lsacoverag” and “lohicu” are confirmed by statistical linear regression analysis as well, and our previous conclusion drew on these two parameters should be robust. Next, we notice that the overall important parameter patterns found for each individual postoperative indicators are very different between ML-based methods and statistical method (Table S9, Table S19). While ML-based methods have found many postoperative indicators having their corresponded preoperative indicators as important impacting parameters, the portion is much less in the statistical method. According to medical common sense, the analysis result by our ML-based method could be more reliable. From this aspect, we believe the result from our ML-based method fits current medical findings better. Moreover, regression on “Lneut” and “isoutcome” failed to find any important parameters. Taken together, we think our method could be better than traditional statistical method and requires no assumptions prior to the data analysis based on medical knowledge.

Digital twin iterative prediction on EL and long-term outcome

Clinically speaking, it would be helpful to know EL and long-term outcome prior to the surgery with the given preoperative inspections and design surgery method accordingly. To represent the condition, we tried to build digital twin models which combine both postoperative indicators and EL and long-term outcome models. It predict the EL and long-term outcome with preoperative indicators and surgery information by selected models (Details described in Method ‘Digital twin prediction’, Fig. 2). As the accuracies varies little different algorithms, we focus more on the value of P/FOR (reason written in Result part ‘Analysis and prediction on EL and long-term outcome’).

The result aligns with our previous finding, that prediction of ‘isoutcome’ is generally easier (or more accurate) than of ‘iseltype’ (Table 4, Fig. S9- S16). Apart from some groups of DT models failed to find any ‘true positive’ result (Table 4, Fig. S9, S12), all other groups show meaningful P/FOR (greater than 1). Some predictions under the same classification algorithm show the same result between different regression algorithms (Table 4, Fig. S10-S11, S14), as inputs of the classification models contain no postoperative indicators. Other groups successfully achieve meaningful prediction on ‘iseltype’ and ‘isoutcome’ with preoperative indicators, surgery information and predicted postoperative indicators (Table 4, Fig. S13, S15-S16). With this result, we conclude that our attempt to predict the EL and long-term outcome of a patient prior to the surgery is successful.

Table 4

Summary of DT prediction performance.
Algorithms		Predictions
Classification	Regression	iseltype				isoutcome
Classification	Regression	Accuracy	FOR	P	P/FOR	Accuracy	FOR	P	P/FOR
Logistic	Linear	76.82%	0.23	NaN	NaN	86.89%	0.13	0.78	6.08
	Tree	76.82%	0.23	NaN	NaN	86.89%	0.13	0.78	6.08
	Support Vector	76.82%	0.23	NaN	NaN	86.89%	0.13	0.78	6.08
	Neural Network	76.47%	0.23	0.00	0.00	86.89%	0.13	0.78	6.08
Tree	Linear	77.16%	0.21	0.53	2.48	83.90%	0.15	0.29	1.95
	Tree	77.16%	0.21	0.53	2.48	85.02%	0.15	NaN	NaN
	Support Vector	77.16%	0.21	0.53	2.48	83.90%	0.14	0.36	2.59
	Neural Network	77.16%	0.21	0.53	2.48	85.02%	0.15	NaN	NaN
Support Vector	Linear	77.16%	0.20	0.52	2.59	85.02%	0.14	0.50	3.67
	Tree	79.24%	0.20	0.67	3.37	85.02%	0.14	0.50	3.67
	Support Vector	77.16%	0.21	0.53	2.49	85.02%	0.14	0.50	3.67
	Neural Network	76.82%	0.20	0.50	2.46	85.02%	0.14	0.50	3.67
Neural Network	Linear	78.55%	0.20	0.62	3.07	85.77%	0.13	0.60	4.54
	Tree	78.20%	0.20	0.60	2.93	84.64%	0.14	0.44	3.19
	Support Vector	77.85%	0.21	0.58	2.79	86.14%	0.13	0.64	4.94
	Neural Network	76.82%	0.21	0.50	2.34	85.77%	0.13	0.60	4.54

TBAD is a life-threatening condition caused by aortic intimal tear, leading to separation between the intimal and medium layer of the aortic wall. TEVAR is the treatment of choice to occlude proximal intimal tear, reinforce aortic wall, and improve organ perfusion[17]. However, endoleaks after TEVAR have led to inferior outcomes and long-term outcome data after TEVAR for TBAD is still scarce[18]. Therefore, there is a need to understand the relationship between surgery, potential inflammations, patient health conditions, EL and long-term outcome. It would be very beneficial if some of the important outcomes can be inferred and prevented prior to the surgery. In an attempt to construct predictive algorithms for endoleaks and long-term outcome after TEVAR, we have collected data from 288 patients and constructed their DT models based on four ML algorithms.

From a medical point of view, the underlying mechanism between inflammation and initiation, progression of TBAD remains unclear. Some researches propose that vascular inflammation responses may lead to increased oxidative stress, protease-mediated degradation of aortic wall, and thus aortic injury[19]. Neutrophils, lymphocytes, and platelets are all associated with inflammatory states and aortic dissection[20]. However, thorough inflammatory markers testing is time-consuming and costly, and using a single inflammatory marker to analyze the patient’s pathophysiological condition is insufficient[5]. Therefore, we sought to construct a predictive algorithm using easily available blood markers and ML methods. Discrepancies exist between some findings from ML algorithms and medical knowledge. For example, the predictive value of “weight” for survival after TEVAR needs further scrutiny since patients with greater body weight tend to develop aortic dissection early in life. The earlier onset time of TBAD may be a protective factor for patients undergoing TEVAR, thus weakening the convincingness of direct predictive value of large body weight. Interestingly, height is determined as a risk factor for EL in the ML model, we propose that the reason may lie in the impact of height on aortic diameter, aortic arch tortuosity, and arch curvature angulation, which may influence EL together with pathology type. The comprehensive combination of endograft type, endograft size, patients’ aortic diameter and stent oversizing as an integrated surgical parameter for long-term outcome prediction is necessary to be researched further.

We also prove that our way of study could be an alternative for traditional statistical regression methods, especially when traditional statistical regression could be ineffective in some cases (Table S19). Moreover, traditional logistic regression only handles binary responses well. Though no multiclass classification problems are not shown in our case of study, our method can handle the problem as well. It is important that we have not manually selected important parameters based on medical knowledge, which means our way of method could be handled by non-medical researchers as well, and no bias will be made. However, we also admit that our method only answers the question of which parameters could be important in predicting the indicated responses, not how they influence them. From this point of view, statistical analysis is still important when trying to determine how important parameters influence those responses.

ML has been used in detecting[21] EL and predicting reintervention[22] after TEVAR in previous studies, however, our study is the first one to incorporate the idea of DT. As pure ML model only predict outputs based on inputs, our DT models which apply combinations of ML models could reflect the progression of “patients in virtual spaces”. We have not constructed ML models directly link preoperative indicators to EL or long-term outcome. Instead, our “virtual patients” firstly receive preoperative inspections (preoperative indicators as input), then receive surgeries (planned surgery information as input) and have their postoperative indicators measured in (postoperative indicators as intermediate output) virtual space. Finally, their EL and long-term outcome is predicted after their hospital discharge in the virtual space (final outcomes). By this way, it is possible to make predictions on any stage of the patient and adjust the operation plan according to the various outcome of interests.

We admit that our study is still at its very early stage, which has not reflect most of the advantages of a DT model which really derives it from a ML study. We have not incorporated any imaging data yet, which may lower our prediction accuracy as TBAD diagnosis depends on imaging data as well. We have limited stages of data collection (only at preoperative, operative, post operative and long-term time points), which significantly weakens the concept of DT and makes it almost indistinguishable from ML models. Also, we have only used the simplest ML algorithms to show our methods, and the data size could be too small for such a DT concept. These problems can be addressed in our future works.

Preoperative prediction and real-time monitoring of complications and postoperative survivals remain the most important issues for any surgeries, including TEVAR. Despite certain limitations, this study is a novel attempt to apply DT models in interpreting clinical data from a new aspect and predicting preoperative outcomes. At current stage, it could help finding critical influencing factors for the outcomes of TEVAR receiving patients. In the future, in order to improve the quality of our DT model and enhance its predictive accuracy, it is important to integrate imaging data, multi-omics data and a series of follow-up data, and refine our algorithms. At that stage, a more comprehensive understanding of the etiology, development, and postoperative changes of TBAD could be provided by the next-generation DT model.

Ethics approval and consent to participate:

Not applicable.

Competing interests

The authors declare no conflict of interests.

Funding:

The project is funded by the New Cornerstone Science Foundation through the XPLORER PRIZE, Beijing Outstanding Young Scientist Project, Major Collaborative Innovation Project of the Chinese Academy of Medical Sciences (2021-I2M-1-072) and The National High Level Hospital Clinical Research Funding (No. 2022-GSP-GG-16)

Author Contribution

XL constructed models, performed statistical analysis and machine learning-based analysis, performed digital twin prediction, analysed results, wrote manuscript, made tables and figuresZC collected clinical data, analysed results, wrote manuscript and made table (Table 2)FT: Supervised the digital part of the study and provided fundingMingyao: Supervised the clinical part of the study and provided fundingChang: Supervised the clinical part of the study and provided funding

Data Availability

Anonymized patient data can be made available upon requests directed to the corresponding author Chang Shu and Mingyao Luo. Requests will be scrutinized upon the basis of ethical review and scientific feasiblity. Confidentiality agreement will be mandatory according to our institutional regulations. Anonymizzed data will be made available for reuse after approval of the proposal. Right to reviewing draft papers based on shared data will be reserved before publication.

E. M. Isselbacher et al., “2022 ACC/AHA Guideline for the Diagnosis and Management of Aortic Disease: A Report of the American Heart Association/American College of Cardiology Joint Committee on Clinical Practice Guidelines,” Circulation, vol. 146, no. 24, pp. e334–e482, Dec. 2022, doi: 10.1161/CIR.0000000000001106.
A. Evangelista et al., “Insights From the International Registry of Acute Aortic Dissection: A 20-Year Experience of Collaborative Clinical Research,” Circulation, vol. 137, no. 17, pp. 1846–1860, Apr. 2018, doi: 10.1161/CIRCULATIONAHA.117.031264.
T. Berger et al., “EACTS/STS Guidelines for diagnosing and treating acute and chronic syndromes of the aortic organ,” European Journal of Cardio-Thoracic Surgery, 2024.
R. Gorla et al., “Clinical features and prognostic value of stent-graft-induced post-implantation syndrome after thoracic endovascular aortic repair in patients with type B acute aortic syndromes,” Eur J Cardiothorac Surg, vol. 49, no. 4, pp. 1239–1247, Apr. 2016, doi: 10.1093/ejcts/ezv355.
L. H et al., “Inflammatory risk stratification individualizes anti-inflammatory pharmacotherapy for acute type A aortic dissection,” Innovation, vol. 4, no. 4, May 2023, doi: 10.1016/j.xinn.2023.100448.
H. Liu, S.-C. Qian, Y.-F. Shao, H.-Y. Li, and Additive Anti-inflammatory Action for Aortopathy & Arteriopathy (5A) Investigators Group, “Anti-Inflammatory Effect of Ulinastatin on the Association Between Inflammatory Phenotypes in Acute Type A Aortic Dissection,” J Inflamm Res, vol. 15, pp. 3709–3718, 2022, doi: 10.2147/JIR.S369703.
N. A. Bradley et al., “The relationship between computed tomography-derived body composition, systemic inflammation, and survival in patients with abdominal aortic aneurysm,” J Vasc Surg, pp. S0741-5214(23)01387–3, Jun. 2023, doi: 10.1016/j.jvs.2023.06.012.
R. Soares Ferreira et al., “Long Term Outcomes of Post-Implantation Syndrome After Endovascular Aneurysm Repair,” Eur J Vasc Endovasc Surg, vol. 62, no. 4, pp. 561–568, Oct. 2021, doi: 10.1016/j.ejvs.2021.06.025.
A. H. King et al., “Elevated neutrophil-lymphocyte ratio predicts mortality following elective endovascular aneurysm repair,” J Vasc Surg, vol. 72, no. 1, pp. 129–137, Jul. 2020, doi: 10.1016/j.jvs.2019.10.058.
S. Y. Shalaby et al., “Systemic Inflammatory Disease and Its Association With Type II Endoleak and Late Interventions After Endovascular Aneurysm Repair,” JAMA Surg, vol. 151, no. 2, pp. 147–153, Feb. 2016, doi: 10.1001/jamasurg.2015.3219.
P. Gazerani, “Intelligent Digital Twins for Personalized Migraine Care,” JPM, vol. 13, no. 8, p. 1255, Aug. 2023, doi: 10.3390/jpm13081255.
X. Liu et al., “A systematic review of digital twin about physical entities, virtual models, twin data, and applications,” Advanced Engineering Informatics, vol. 55, p. 101876, Jan. 2023, doi: 10.1016/j.aei.2023.101876.
K. Aubert et al., “Development of Digital Twins to Optimize Trauma Surgery and Postoperative Management. A Case Study Focusing on Tibial Plateau Fracture,” Front. Bioeng. Biotechnol., vol. 9, p. 722275, Oct. 2021, doi: 10.3389/fbioe.2021.722275.
J. A. Gálvez, A. Jalali, L. Ahumada, A. F. Simpao, and M. A. Rehman, “Neural Network Classifier for Automatic Detection of Invasive Versus Noninvasive Airway Management Technique Based on Respiratory Monitoring Parameters in a Pediatric Anesthesia,” J Med Syst, vol. 41, no. 10, p. 153, Oct. 2017, doi: 10.1007/s10916-017-0787-3.
J. V. Lombardi et al., “Society for Vascular Surgery (SVS) and Society of Thoracic Surgeons (STS) reporting standards for type B aortic dissections,” Journal of Vascular Surgery, vol. 71, no. 3, pp. 723–747, Mar. 2020, doi: 10.1016/j.jvs.2019.11.013.
C. Tian et al., “Surgical treatment patterns and clinical outcomes of type B aortic dissection involving the aortic arch,” Journal of Vascular Surgery, vol. 77, no. 4, pp. 1016–1027.e9, Apr. 2023, doi: 10.1016/j.jvs.2022.11.048.
M. Grabenwoger et al., “Thoracic Endovascular Aortic Repair (TEVAR) for the treatment of aortic diseases: a position statement from the European Association for Cardio-Thoracic Surgery (EACTS) and the European Society of Cardiology (ESC), in collaboration with the European Association of Percutaneous Cardiovascular Interventions (EAPCI),” European Journal of Cardio-Thoracic Surgery, vol. 42, no. 1, pp. 17–24, Jul. 2012, doi: 10.1093/ejcts/ezs107.
Y.-Q. Kan, D.-Q. Guo, Y. Si, and W.-G. Fu, “Thoracic endovascular aortic repair for type B aortic dissection patients: 11-year experience from a Chinese Tertiary Center,” Chin Med J (Engl), vol. 134, no. 8, pp. 971–972, Apr. 2021, doi: 10.1097/CM9.0000000000001420.
D. Wu, Y. H. Shen, L. Russell, J. S. Coselli, and S. A. LeMaire, “Molecular mechanisms of thoracic aortic dissection,” J Surg Res, vol. 184, no. 2, pp. 907–924, Oct. 2013, doi: 10.1016/j.jss.2013.06.007.
S. Su et al., “Systemic immune-inflammation index predicted the clinical outcome in patients with type-B aortic dissection undergoing thoracic endovascular repair,” Eur J Clin Invest, vol. 52, no. 2, p. e13692, Feb. 2022, doi: 10.1111/eci.13692.
S. Talebi et al., “Machine learning for endoleak detection after endovascular aortic repair,” Sci Rep, vol. 10, no. 1, p. 18343, Oct. 2020, doi: 10.1038/s41598-020-74936-7.
Y. Dong et al., “Predicting reintervention after thoracic endovascular aortic repair of Stanford type B aortic dissection using machine learning,” Eur Radiol, vol. 32, no. 1, pp. 355–367, Jan. 2022, doi: 10.1007/s00330-021-07849-2.
E. von Elm, D. G. Altman, M. Egger, S. J. Pocock, P. C. Gøtzsche, and J. P. Vandenbroucke, “The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement: Guidelines for reporting observational studies,” International Journal of Surgery, vol. 12, no. 12, pp. 1495–1499, Dec. 2014, doi: 10.1016/j.ijsu.2014.07.013.
M. F. Fillinger, R. K. Greenberg, J. F. McKinsey, and E. L. Chaikof, “Reporting standards for thoracic endovascular aortic repair (TEVAR),” JOURNAL OF VASCULAR SURGERY, vol. 52, no. 4, p. 17, 2010.

No competing interests reported.

Download PDF

Editor assigned by journal
30 Jun, 2024
Submission checks completed at journal
25 Jun, 2024
First submitted to journal
21 Jun, 2024

You are reading this latest preprint version

A digital twin study on thoracic endovascular repair for Stanford type B aortic dissection

Status:

Version 1

Abstract

Figures

Introduction

Method

Patient cohort

Machine learning models construction

Machine learning parameter analysis

Statistical analysis

Digital twin prediction

Results

Patient demographics

Analysis and prediction on postoperative indicators

Analysis and prediction on EL and long-term outcome

Parameter analysis through traditional statistical methods

Digital twin iterative prediction on EL and long-term outcome

Discussion

Declarations

Competing interests

Funding:

Author Contribution

Data Availability

References

Additional Declarations

Supplementary Files

Status:

Version 1