Background
Type A acute aortic dissection (AAD) requires emergency replacement of the ascending aorta and/or the aortic arch with or without aortic valve replacement and in-hospital mortality ranges from 7 to 30% in recent series [
1,
2]. Among 526 patients enrolled from 1996 to 2001 by the International Registry of AAD investigators, 30-day mortality was 25.1% on average [
1]. A large list of pre-, intra- and immediate post-operative factors may independently contribute to increase the mortality risk at short-term (see [
2] for extensive review). These include: history of aortic valve replacement, migrating chest pain, hypotension and/or shock, cardiac tamponade, limb ischemia, the length of extracorporeal circulation and chronic renal failure. There has also been an effort to investigate whether surgical techniques may contribute to modify the risk; however inconsistent results were obtained as to the role of retrograde, anterograde or selective cerebral perfusion after circulatory arrest [
1,
2]. More recently, anatomo-surgical parameters [
3] and biological indexes, such as D-dimer values above a given threshold [
4], were assessed as diagnostic tools, but no study was performed to clarify their potential predictive role. On the other hand, it is largely unknown whether the assessed short-term risk factors may also predict long-term (say 1- to 2-year) mortality in Type A AAD patients.
Aim of the study was therefore to see whether selected risk factors assessed previously for prediction of 30-day mortality risk among Type A AAD patients [
1,
2], may also contribute to index long-term prediction using neural networks known to have a larger global accuracy as compared to standard models such as logistic regression [
2,
5]. In addition, to improve discrimination between cases and non cases [
6], which is essential once new risk equations are tested in general and in cardiac surgical outcome studies [
7‐
10] in particular, support vector machines (SVM) were also used [
11,
12] for the first time on this material.
Discussion
This is the first investigation to adopt neural networks and support vector machines to assess the relatively long-term predictive role of a quite large series of potential risk factors including pre-operative, operative and immediately post-operative variables in AAD Type A patients. The presence of Marfan habitus, the length of circulatory arrest, an intervention on the ascending aorta plus hemi-arch and immediate post-operative chronic renal failure were the risk factors selected in common by these methods with a very high global accuracy (ROC AUC > 0.82). Although the factors selected were not new, their combination might be used in practice to enable the construction of risk charts whereby levels of risk might be defined. However, it is clear that the corresponding cells of these charts need to contain a sufficient number of cases and non cases, which is presumably possible only after large multi-centre and/or multinational cooperative efforts will be undertaken. The evidence presented here might contribute to stimulate cooperation to reach this aim.
The presented rules provided very good predictive and discrimination properties, however only Marfan habitus was a parameter that could be used pre-operatively. Determination a priori about which patients are not candidates for surgery is therefore not possible using the evidence of this investigation. Nevertheless, as there were 2 operative parameters contributing to increase long-term mortality risk, it is important that attention is paid to keep the length of circulatory arrest at the minimal level and to consider that an intervention on the aorta plus hemi-arch conveys an independent risk of lower survival. On the other hand, all efforts should be done to reduce the incidence of post-operative chronic renal failure.
The incidence of AAD Type A has been estimated at from 5 to 30 per million people per year in the United States, which is 880 to 147 times less than the incidence of acute myocardial infarction, but still provides an important clinical problem and sometimes a dilemma for the differentiating difficulties between these presentations [
1‐
3]. Although biological thresholds of plasma molecules such as D-dimer are actively looked for in order to improve diagnosis [
4], this may not have an impact on prediction before the results of larger studies are obtained. Therefore, risk profiling remains crucial. Based on results obtained by the IRAD investigators, short-term mortality could be reduced from as high as 58% in medically treated patients to the current average figure of 25.1% (and sometimes less) when surgery is performed [
1]. Risk factors may contribute to better management and a more defined risk assessment [
1,
2]: in-hospital mortality was as high as 31.4% in unstable patients presenting with cardiac tamponade, shock, congestive heart failure, cerebro-vascular accident, stroke, coma, acute myocardial and/or mesenteric ischemia and acute renal failure at the time of operation, whereas stable patients may present with a mortality as low as 16.7%.
In a previous report we investigated 30-day mortality among 208 patients from 2 Italian Centres [
2] using a series of demographic, pre-operative, operative and post-operative characteristics, selected from 37 such variables considered in the Literature as potential predictors of short-term mortality after AAD Type A. When logistic or neural network models were produced in one Centre and applied to the data from the second Centre, for external validation [
13‐
15], there were predictors which were selected in common: the presence of pre-operative shock, intubation and neurological symptoms, immediate post-operative presence of dialysis in continuous and the quantity of bleeding in the first 24 hours post-operation. By neural network model only, the length of extracorporeal circulation and post-operative chronic renal failure were detected as independent predictors of 30-day mortality. Different from the IRAD Registry investigators [
1] we showed [
2] that operative and immediate post-operative factors should be considered to predict short-term mortality. They contributed significantly to obtain a large overall accuracy, which might be explained in part by these factors being continuous [
19]. On the other hand, similar to studies investigating predictive performance of short-term mortality after coronary artery bypass surgery [
9,
10], neural networks had a better performance when compared to standard methods such as logistic regression [
2,
5].
When the performance and/or reliability of predictive models is limited, or of low sensitivity and specificity, their capability may be hampered to identify high risk subjects who deserve individualized treatment [
13]. The neural network method stems [
14,
15] from its potential for improved predictive performance by exploring, hidden layers to find nonlinearities, interactions and nonlinear interactions among predictors. The attraction of neural networks is quite evident from the impressive growth of results published [
15]. However, there are relatively few comparative reports on the performance and accuracy of neural networks, which was assessed only versus multiple logistic function, to predict events in clinical [
9] or epidemiological [
5,
18] cardiovascular studies.
There has been some controversy as to whether new risk predictors, or series of old and newer ones, can add to the prediction of events, including mortality, in terms of clinical utility, impact or discrimination [
6]. Although in clinical and epidemiological experiences discrimination metrics (such as ROC AUC) are quite well established methods [
2,
5,
18,
20,
21], it has been pointed out that ROC AUC are insensitive in comparing models [
6], which may be circumvented however by making comparisons with fixed number of covariates [
5]. To evaluate and compare predictive risk models there have been therefore new methods to be proposed, based primarily on stratification into clinical categories on the basis of risk and attempts to assess the ability of new models to more accurately reclassify individuals into higher or lower risk strata [
22,
23]. Risk reclassification for single factors can be then examined by using models with and without each risk factor in turn or measuring the net reclassification improvement, that is the difference in proportions moving up and down risk strata among case patients versus control participants [
6,
23]. Whatever reclassification method is selected it is important to understand that when length of follow-up differs (as in the present series) among individuals and/or the cohort is relatively small it may be impossible to apply them [
6]. Moreover and more importantly, reclassification methods depend on the particular categories used [
6]: in our case it is far from established if a 5%, 10%, 20%, 30% or more are adequate categories of long-term risk of AAD Type A. To compare with established experiences in preventive cardiology [
20,
24] or coronary by-pass surgery [
25], the sensitivity and specificity of the abovementioned thresholds should be accurately assessed, which again calls for large amount of data being collected and therefore improved multi-centre collaboration.
Conclusions
The classification provided by neural network models and related SVM may represent a compromise to cope with the necessity to assess the clinical relevance of variables used for predictive purposes in AAD Type A patients, but also in different areas of research. These methods may also go beyond the classical contention of standard predictive models, namely that only predictors that are statistically significant are typically used [
6]. Indeed, with SVM a high discrimination is obtained by using a large number of variables, most with little informative content if used alone. As we have shown, however, it is extremely important not only to train but to validate these methods, which demands further study and the accumulation of very large data sets. Our results may well stimulate these efforts.
An important take-home message for clinicians should be that with neural networks and SVM, by concentrating on a few risk factors such as those described here, it is possible to predict long-term mortality in AAD Type A patients with a global good accuracy. We produced an html tool (see Additional File
3) based on the neural network solution reported here, whereby it is easy to appreciate that increasing from 60 to 80 min the circulatory arrest time, the patient long-term risk category evolves from false (survival) to true (dead) at an assessment strength (roughly the degree of certitude) of 1/3. By further increasing circulatory arrest times to 120 and 180 min, the assessment strengths become 2/3 and almost 1, respectively. Although Surgeons know well and from decades that this is a hardly steerable variable in the clinical practice, a dimensional outcome predictive assessment might be obtained using our tool immediately after the operation is finished, which may have an impact for further clinical decision making. The other variables described in the present study might also be used for predictive assessments so that a very large combination of clinical presentations could be easily modeled.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
FM and PEP participated equally in the design of the study. PEP performed the statistical analysis, drafted the manuscript and coordinated the implementation of it. AS collected the data, obtained the follow-up information and participated in the draft of the manuscript. MT1 participated in the data collection and the draft of the manuscript. FT collected the data, obtained the follow-up information and participated in the statistical analysis and the draft of the manuscript. MC and MT2 conceived of the study, and participated in its design and coordination and helped to draft the manuscript. All authors read and approved the final manuscript.