Neural-network classification of cardiac disease from ³¹P cardiovascular magnetic resonance spectroscopy measures of creatine kinase energy metabolism

Since the mid-1980s numerous noninvasive localized phosphorus (³¹P) cardiovascular magnetic resonance spectroscopy (CMRS) studies have been performed on patients with heart disease to measure key energy metabolites–phosphocreatine (PCr) and adenosine triphosphate (ATP) [1]. ATP is the source of the cellular energy that fuels all cellular processes including myocardial contraction. PCr is the heart’s primary energy reserve, producing ATP from the phosphorylation of adenosine diphosphate (ADP) via the creatine kinase (CK) reaction: PCr + ADP < =>Cr + ATP, forming creatine (Cr) as a byproduct. The majority of human cardiac ³¹P CMRS studies use the PCr/ATP ratio as an index of this cardiac energy reserve [1]. Significant reductions in myocardial PCr/ATP are reported for a wide range of cardiac diseases including acute ischemia [2], dilated cardiomyopathy (DCM) [3‐6], hypertrophic cardiomyopathy (HCM) [7‐11], heart failure (HF) [3‐11], valve disease [7, 8], obesity [12], diabetes [13] and in patients with heart transplants [14]. Although there is a trend for greater reductions in worse illness [4, 5, 8‐10], the scatter and overlap amongst PCr/ATP values has nevertheless precluded its use as a metabolic biomarker for any specific condition or for grading disease severity, at least on an individual basis [1].

In the 1990s, methodologies that combine internal or external concentration references with magnetic resonance imaging (MRI) volumetry [10, 15‐18], have enabled the measurement of absolute concentrations of human cardiac creatine kinase (CK) metabolites, revealing deficits in patients with cardiomyopathy, HF and myocardial infarction (MI) [6, 10, 11, 17]. In addition, the development of fast ³¹P saturation-transfer techniques over the past decade have provided measurements of the pseudo first-order kinetic reaction rate, k_f (in s^-1) for the CK reaction [19, 20]. Reduced myocardial k_f was observed in patients with left ventricular hypertrophy (LVH) and HF, but not in those without HF [11]. This suggested that reductions in CK k_f might be specific to HF. By combining k_f and PCr concentration ([PCr]) measurements in the same cardiac ³¹P CMRS exam, the forward CK flux of ATP supplied to the heart can be obtained [6]. This is given by the product, k_f [PCr].

In nonischemic cardiomyopathy (NICM), CK flux declines with progression to HF [1, 21]. Indeed, the reduction in CK flux independently predicted cardiac events (HF hospitalization, transplantation, left-ventricular (LV) assist device requirement, cardiac death) in HF patients that occurred several years post-CMRS [22]. More recently it was shown that CK flux correlates with cardiac mechanical work [23], providing a mechanistic link between declining CK flux, mechanical work and cardiac events. While these findings support the ‘energy starvation hypothesis’ for HF [24, 25] and suggest that such measurements might index disease severity, the scatter among individual CK metabolite concentrations, k_f, and flux data, again preclude the discrimination of disease type or severity on an individual basis, using any one ³¹P CMRS metric alone.

Today, trained machine learning techniques employing multi-layer neural networks are emerging as powerful tools for identifying and classifying a vast array of multi-parametric input data in applications ranging from bioinformatics to sociology to driving cars. Once designed, a neural network can be retuned and retrained as additional data become available, without exhaustively searching for the best classifier. Here, we apply neural-network analysis to test whether a combination of cardiac [PCr], [ATP], k_f and the CK flux supplying ATP to the heart, as measured noninvasively with ³¹P CMRS, can distinguish several common myocardial diseases and their clinical severity. The study includes–to the best of our knowledge–the entire known data pool of subjects for which complete sets of [PCr], [ATP], k_f and CK flux measurements exist [6, 11, 21‐23, 26, 27]. The work tests the hypothesis that cardiac CK metabolism is altered in patterns that may be specific to different forms of heart disease and/or to the clinical severity of disease, specifically in HF. If true, cardiac CK metabolic data acquired with ³¹P CMRS could be used with a neural network classifier as a metabolic biomarker to gauge disease type and severity. Moreover, the existence of such relationships would link certain common clinical cardiac conditions to compromised CK energy supply in individual patients.

Table 2 presents the confusion-matrix documenting the accuracy of the first network for classifying healthy and heart disease patients. The results show an overall accuracy of 84% for discriminating heart disease solely on the basis of CK metabolism alone. The network correctly distinguished 90% of the HF data, and about 80% of the other classes except for MI, which was difficult to differentiate from HF (accuracy ~ 70%). It predicted the primary diagnosis of HF in the 12 HCM cases that were omitted from the training set (and Table 2), with 100% accuracy.

Figure 3(a) illustrates the results of the ROC analyses of the (first) network’s classification of each disease using the four CK metabolic parameters. The AUCs are all greater than 90%. Figure 3(b) is a three-dimensional (3D) plot of the independent variables [PCr], [ATP] and k_f depicting the network’s true and false classifications. A 3D video plot (see Additional file 1) shows that the fitted volumes associated with each output disease type are intertwined on the [PCr]–[ATP] axis (diagonal), indicating that the PCr/ATP ratio, by itself, is a poor discriminator of disease type.

The effect of successively eliminating in turn, each CK metabolic parameter and re-tuning the network in the second set of studies is shown in Fig. 4. All variables contributed to discrimination, but [ATP] contributed least to overall accuracy. The effect of omitting metabolic variables on the ROC–AUC analysis, for each disease type (Fig. 5) shows that: (i) successively removing variables reduces the AUC; (ii) patients with HCM but no HF were hardest to classify; and (iii) [ATP] provided the least added-value as a discriminator, as evidenced by the fact that including [ATP] instead of any other variable produced the greatest loss in AUC.

Table 3 shows the confusion matrix for the (third) network designed to distinguish HF patients documented with DCM from those with HCM, based just on their CK metabolism. Again an accuracy of about 80% was achieved. The network predicted the one DCM subject whose NYHA class was uncertain as class II HF (presently not confirmable), with a 50/50 chance of being either DCM or HCM.

Table 4 presents the confusion matrix results for classifying NYHA HF severity using the (fourth) CK metabolism-based neural network classifier. The overall accuracy for discriminating NYHA HF class on the basis of CK metabolism alone was 86%, with the caveat that the number of NYHA class IV patients is very limited.

When all the 3 T data are omitted from the first neural network that was used for Table 2 in order to assess whether the results are confounded by differences in the 1.5 T and 3 T methodologies, the accuracy of the resulting confusion matrix (Table 5a) improved slightly to 86% (vs 84% in Table 2). Omitting the 1.5 T data instead of the 3 T data yielded a similar 82% accuracy using this same un-retuned (first) network (Table 5b). Retraining and optimizing this network for just the 1.5 T data alone, improved the overall accuracy to 90% (Table 5c). This may be due to some systematic differences (including variability in scatter) between the 1.5 T and 3 T data and/or differences in the sub-groups sampled.

Rebalancing the network used for classifying disease with randomly interpolated data to equalize class size to that of HF, resulted in the same overall accuracy of 84% (Table 6a). This reflects an improvement by about 10% in the accuracy of detecting the formerly lesser-sampled LVH, MI and control groups, at the expense of HF patients whose accuracy declined by 14%. Similarly, rebalancing the network for distinguishing HF severity by augmenting the NYHA Classes I and III data to match the Class II data pool-size, resulted in essentially the same overall accuracy (85% vs. 86% in Table 4), with performance improving for Classes I and III, at the expense of Class II (Table 6b). Applying the balanced retrained network to the original data reduced overall performance a little to 78%, mainly to the detriment of the small HCM non-HF group (accuracy, 67%). Finally, when the network that was trained with the original data from Table 2, was plied with the entire set of augmented HF data (n = 168), its overall accuracy for detecting HF was 85%.

This is the first study to apply neural network techniques to test whether cardiac CK energy metabolism can differentiate the type of myocardial disease and the severity of disease in individual patients. The source data comprised all of the known complete sets of individual cardiac [PCr], [ATP], k_f and CK flux data from healthy subjects and heart disease patients extant in the world that we know of [1, 6, 11, 21‐23, 26, 27]. The results show that a three-layer neural-network can differentiate healthy, HF, non-HF MI and HCM cardiac diseases with an accuracy of 84% and an AUC of > 90% for each category, based on their CK metabolism alone (Table 2, Fig. 3). Among HF patients a neural network was able to differentiate DCM and HCM (Table 3), as well as the severity of HF as indexed by NYHA class (Table 4), both with about 80% or higher overall accuracy. This suggests that CK metabolic data acquired noninvasively with resting-state cardiac ³¹P CMRS, could be used with a neural network classifier as a metabolic biomarker to stratify disease type and severity with clinically-relevant accuracy. Moreover, from a mechanistic standpoint, the fact that a neural network can classify disease based on CK metabolism alone, means that there must exist heretofore unidentified differences in CK metabolism that are specific to disease type and severity which enable the network to do so.

Indeed there was no one unique disease-specific CK-metabolic defect, as evidenced by the incremental effect of successively removing each parameter and retuning the network (Fig. 4). Instead, nuanced variations in multiple parameters permitted accurate neural network discrimination. The elimination analysis showed that [PCr] and k_f were the first and second most important input parameters for classifying heart disease in the neural network (Figs. 4 and 5). The pair, [PCr] and k_f, was better than CK flux alone, but using all three, namely CK flux, k_f, and [PCr], was better still. This is unexpected because CK flux is not independent, being the product of k_f and [PCr] as noted earlier. In order for CK flux to add such value as a classifier compared to its component factors, k_f and [PCr] must vary in a co-dependent way that is exacerbated by disease. This is consistent with the observation that CK flux independently predicted cardiac events whereas individual k_f, [PCr] (and PCr/ATP) measures alone, did not [22]. [ATP] was the least important parameter for discriminating disease, an observation likely reflecting the tight control of [ATP] that is required to preserve cellular viability. With [ATP] as the ‘last to go’, the CK reaction, oxidative phosphorylation, glycolysis and other metabolic pathways apparently make whatever adjustments that are needed to preserve [ATP] and prevent demise.

The main limitation of this work is the number of studies, particularly the NYHA class IV HF, and non-HF HCM and MI sub-categories. Although all extant complete data sets were included, they derive from research studies that were not designed for the specific investigation reported herein. In that respect the study might be considered as blinded compared to the prospective acquisition of data to test a preconceived hypothesis. Unfortunately, the availability of new data to test the potential of this cardiac ³¹P CMRS quartet to serve as a metabolic biomarker for clinical disease, is presently limited to the extent that cardiac ³¹P MRS is not yet a clinical procedure.

The synthetic augmentation of metabolic data by 2.3-fold for disease type and by 1.7-fold for HF severity using the SMOTE method, provided results that were consistent with the smaller original data sets. This suggests that the overall level of performance realized here should be reasonably representative of what is achievable with the addition of ³¹P CMRS data of comparable quality from future studies. Nevertheless, the primary purpose of data augmentation was to balance class-size to avoid over-fitting and/or bias of the neural network towards a majority class. Indeed, its overall effect was to smooth out performance differences between the majority and lesser-sampled classes (Table 6 vs. Tables 2 and 4), which was the intended purpose [29]. There is no substitute for original data, especially when the underlying functional inter-relationship between metabolic parameters, disease type and severity is not known. That the successive dropping of metabolic parameters degraded predictive accuracy (Fig. 4), means that [PCr], [ATP], k_f and CK flux have to be treated as a parametric quartet. Thus in general, network performance might not be expected to survive simulations that involve independent randomization of the component parameters, at least until the nature of their mechanistic link is better understood. Disease heterogeneity is a potentially confounding factor as hypertrophy, DCM and MI are common co-morbidities that underlie HF. They would therefore be expected to confound the neural network’s ability to distinguish them. In that sense, the accuracy with which the network was able to discriminate disease type as demonstrated here, is quite surprising. Also surprising was its ability to differentiate NYHA class, which involves a degree of subjectivity in clinical assessment. Indeed all of the network’s misclassification errors involved sliding into an immediately adjacent NYHA class (Table 4).

Other potential confounding variables were age and gender differences, and the use of different 1.5 and 3 T methodologies among sub-groups of the study population. As can be seen in Table 1, the former are fairly minor with the ages of control and HF patients being around 40–50 years. Although the MI and non-HF HCM patients were older, these patients had normal k_f and [ATP] when compared with their original study controls [11, 21]. Nevertheless, a trend for reduced metabolite concentrations (and probably PCr/ATP) with age over 40 [1] has been reported elsewhere and should be noted. Regarding methodological differences between the 1.5 and 3 T CMRS protocols, the re-analysis of separated 1.5 T and 3 T data should eliminate any such confounding effects (Table 5). Indeed, the ability of the neural network to classify disease based on CK metabolism alone was unimpaired for those diseases for which CK metabolic data were available from both 1.5 T and 3 T methodologies. These results can only further support the primary conclusion that a neural network can differentiate individual disease type and severity based on CK metabolism alone.

Neural network analysis employing only cardiac CK metabolic parameters can accurately classify cardiac disease and severity, suggesting a role for noninvasive quantitative ³¹P CMRS as a metabolic biomarker. In HF, the degree of compromise in CK energy supply as identified by the neural network, predicted the clinical severity of HF symptoms in individual patients. The ability to accurately classify disease based on CK metabolism suggests that altered CK metabolism is an intrinsic and nuanced factor in many common heart diseases. If the relationship between compromised CK metabolism and clinical severity is causal, then therapies that increase CK energy supply could well reduce symptoms and support a role for the CK reaction as a potentially important new therapy target [22, 26].