http://orcid.org/0000-0003-2297-3308
Figures
Abstract
Early prediction of the potential for neurological recovery after resuscitation from cardiac arrest is difficult but important. Currently, no clinical finding or combination of findings are sufficient to accurately predict or preclude favorable recovery of comatose patients in the first 24 to 48 hours after resuscitation. Thus, life-sustaining therapy is often continued for several days in patients whose irrecoverable injury is not yet recognized. Conversely, early withdrawal of life-sustaining therapy increases mortality among patients who otherwise might have gone on to recover. In this work, we present Canonical Autocorrelation Analysis (CAA) and Canonical Autocorrelation Embeddings (CAE), novel methods suitable for identifying complex patterns in high-resolution multivariate data often collected in highly monitored clinical environments such as intensive care units. CAE embeds sets of datapoints onto a space that characterizes their latent correlation structures and allows direct comparison of these structures through the use of a distance metric. The methodology may be particularly suitable when the unit of analysis is not just an individual datapoint but a dataset, as for instance in patients for whom physiological measures are recorded over time, and where changes of correlation patterns in these datasets are informative for the task at hand.
We present a proof of concept to illustrate the potential utility of CAE by applying it to characterize electroencephalographic recordings from 80 comatose survivors of cardiac arrest, aiming to identify patients who will survive to hospital discharge with favorable functional recovery. Our results show that with very low probability of making a Type 1 error, we are able to identify 32.5% of patients who are likely to have a good neurological outcome, some of whom have otherwise unfavorable clinical characteristics. Importantly, some of these had 5% predicted chance of favorable recovery based on initial illness severity measures alone. Providing this information to support clinical decision-making could motivate the continuation of life-sustaining therapies for these patients.
Citation: De-Arteaga M, Chen J, Huggins P, Elmer J, Clermont G, Dubrawski A (2019) Predicting neurological recovery with Canonical Autocorrelation Embeddings. PLoS ONE 14(1): e0210966. https://doi.org/10.1371/journal.pone.0210966
Editor: Xiao Hu, University of California San Francisco, UNITED STATES
Received: May 31, 2018; Accepted: January 3, 2019; Published: January 28, 2019
Copyright: © 2019 De-Arteaga et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The code required to reproduce the results can be found in the following git repository: https://github.com/mariadea/CAE.git. Data cannot be shared publicly due to legal restrictions. The data are owned by Dr. Elmer, a study author. He is an employee of the University of Pittsburgh, which contractually has control of research data generated and used by its employees. The data for this manuscript were shared across academic institutions for the purpose of published analysis under the auspices of a Data Use Agreement, a fully executed contract under Pennsylvania State Law, which did not permit dissemination of the shared data. The University of Pittsburgh Office of Research did not grant permission to post these data because of ongoing work to protect its intellectual property derived from these data. Study data are available upon request. To request data access, contact the study authors at elmerjp@upmc.edu, or the University of Pittsburgh Office of Research at offres@pitt.edu or 412-624-7400. We confirm that we did not have special access privileges that others would not have. Those interested in gaining access to the data have the possibility of entering a data use agreement with the University of Pittsburgh, which would grant them access to the same data we use in our study.
Funding: This work has been partially supported by Defense Advanced Research Projects Agency (DARPA) under awards FA8750-12-2-0324 and FA8750-17-2-0130. Dr. Elmer is supported by National Institutes of Health (NIH) through grants 5K12HL109068 and 1K23NS097629.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Cardiac arrest is the most common cause of death in high-income nations [1]. In the United States alone, over 350,000 people suffer out-of-hospital cardiac arrest each year [2]. Despite advances in care, only a minority of those that are resuscitated and survive to hospital admission are discharged alive, and even fewer enjoy a favorable neurological recovery [2, 3]. Among non-survivors, the most common proximate cause of death is withdrawal of life-sustaining therapy based on perceived poor neurological prognosis [3, 4]. This decision may be motivated by the rarity of favorable recovery, the emotional and financial hardship placed on families faced with the prospect of even a few days of intensive care, or fear of survival with severe disability.
Unfortunately, accurate neurological prognostication after cardiac arrest is challenging, particularly in the first 3 to 5 days after resuscitation [5]. Life-sustaining therapy is still often withdrawn before prognosis is certain, unnecessarily reducing rates of favorable recovery [3, 6, 7, 8]. At the same time, patients with brain injury that is ultimately deemed irrecoverable are often supported for days while providers gather sufficient data to make such an assessment.
Multiple modalities which might inform early prognostication have been explored [9, 10, 11, 12]. Of particular interest is the rich electroecephalographic (EEG) data that may be obtained. Research indicates that EEG signals can improve prediction accuracy [10, 13, 14]. Qualitatively, some EEG patterns such as seizures suggest severe brain injury [15]. Quantitatively, patterns with strong correlations between channels or over time are suggestive of diffuse cortical damage seen after non-survivable brain injury [14, 16]. Within EEG, as in many biological systems, entropy is a marker of information content [17]. By contrast, strong spatial or temporal correlations are an ominous predictor of severe brain injury [10, 14, 16]. Because these correlations may be subtle and/or complex, they may be inapparent to providers qualitatively interpreting the EEG, leading to growing interest in quantitative EEG analysis. Fig 1 shows an example of an EEG of a post-arrest patient with mild brain injury who goes on to enjoy a favorable recovery and an example of an EEG of a patient with severe brain injury, for which correlations across channels are very strong. Motivated by this, our goal is to characterize patients in terms of their multivariate, non-linear structures of correlation and use the resulting featurization to identify patients who likely have the potential for favorable neurological recovery.
(Left) EEG of a post-arrest patient who goes on to recover. (Right) EEG of a patient with poor neurological prognosis.
We propose Canonical Autocorrelation Analysis (CAA) as a method for automated discovery of multiple-to-multiple correlation structures within a set of features. Through the introduction of a distance metric between CAA correlation structures, we are able to define Canonical Autocorrelation Embeddings (CAE), a feature space embedding in which each individual/object is represented by the set of its multivariate correlation structures. In this feature space embedding, traditional machine learning algorithms that rely on distance metrics, such as nonparametric clustering and k-nearest neighbors (k-nn), can be applied to compare correlation structures.
This methodology is particularly fitting to tasks where multiple potentially correlated data points are recorded over space or time for each individual or unit of study. For example, in clinical medicine several vital signs or other physiological measures may be repeatedly sampled for each patient being monitored. Because physiological processes are interdependent and interact, analyzing the correlation structure between several such processes may reveal otherwise unrecognized patterns that may characterize the current state of the patient [16, 18]. In this work, we demonstrate the utility of CAE by presenting a specific clinical example: predicting future neurological recovery in a cohort of comatose survivors of cardiac arrest using sets of quantitative EEG autocorrelations.
Because of the difficulty identifying patients with potential for recovery and desire to limit futile care described above, patients may be at risk for withdrawal of life-sustaining therapy when their potential to recover goes unrecognized [3]. To reduce this risk, we propose a decision-support system that provides recommendations to the clinician whenever it is confident that a patient is likely to have a positive neurological recovery and defers in all other instances. Therefore, rather than always providing a recommendation, the algorithm only does so when it is confident life-sustaining therapies should be continued, a prediction it reaches based on patterns in multivariate correlation structures of the EEG that the clinician might not have observed. In all other cases the algorithm will defer to the clinician’s judgment.
In the remainder of this paper, Section 1 presents a brief review of related work. Section 2.1 discusses the task and data in more detail. In Section 2.1, CAA and CAE are introduced, as well as the use of a k-nn algorithm in the resulting embedded space. Section 3 contains the experimental results, Section 4 discusses our findings and Section 5 summarizes the conclusions and future work.
1 Related work
Canonical Correlation Analysis (CCA) is a statistical method first introduced by [19], useful for exploring relationships between two sets of variables. It is used in machine learning, with applications to medicine, biology and finance, e.g., [20, 21, 22, 23]. Sparse CCA, an ℓ1 variant of CCA, was proposed by [23, 24]. This method adds constraints to guarantee sparse solutions, which limits the number of features being correlated. Given two matrices and
, CCA aims to find linear combinations of their columns that maximize the correlation between them. Usually, X and Y are two disjoint matrix representations for one set of objects, so that each matrix is using a strictly different set of variables to describe them. Assuming X and Y have been standardized, the constrained optimization problem is shown in Eq 1. When c1 and c2 are small, solutions will be sparse and thus only a few features are correlated.
(1)
The extension of Sparse CCA for discovery of multivariate correlations within a single set of features to study brain imaging has been previously explored in [20, 21]. Using the notion of autocorrelation, the authors attempt to find underlying components of functional magnetic resonance imaging (fMRI) and EEG, respectively, that have maximum autocorrelation. The types of data used in these works are ordered, both temporally and spatially.
Canonical Autocorrelation Analysis (CAA), the methodology we propose, is a generalized approach to discovering multiple-to-multiple correlations within a set of features. Fig 2 illustrates the different use cases of Sparse CCA and CAA. The proposed formulation also allows for the user to select sets within which correlations are forbidden, which is useful when trivial correlations should be avoided. Moreover, we introduce a distance metric between canonical autocorrelation structures, which gives substantially more power to CAA-based methodology, making it useful for various learning tasks, such as clustering and classification on datasets and distributions.
(Left) Sparse CCA finds sparse multiple-to-multiple linear correlations between subsets of the features in matrix X and subsets of features in matrix Y. (Right) CAA extends this to cases where it is not known a priori how to group the features into two sets.
Other methods for finding sparse representations of data comprised in a single matrix include the well-known Sparse Principal Component Analysis (Sparse PCA). While CAA resembles Sparse PCA in the sense that it finds sparse representations of data contained in one matrix, Sparse PCA maximizes retained variance of data in one-dimensional projections, while CAA finds two-dimensional projections where correlation across two subsets of features is maximized. CAA specifically seeks projections composed by pairs of strongly correlated linear combinations of features, enabling discovery of hidden characteristic correlations in data, which cannot be easily found with other methods such as Sparse PCA. S2 File explores the difference between the two methods in more detail and from a theoretical perspective.
Extraction of informative projections has been tackled in the past [25, 26]. Our work differs from the existing methodology in two primary ways. First, each of the CAA projection axes is defined by a linear combination of features, rather than a single feature, which helps discover complex structures if they exist. Secondly, rather than finding projections where classes are well-separated, the proposed methodology is unsupervised and it is aimed at characterizing objects or individuals that have a batch of data points associated to them, yielding an embedding where standard machine learning methodologies can be used with minor modifications. In that sense, the extracted projections are different both in their form and in their purpose.
The comparison of correlation structures and principal components has been explored in the literature for decades. Most prominently, [27] discusses comparison of principal components between groups. To do so, they propose a metric inspired by the concept of congruence coefficient [28], which corresponds to the cosine of the angle between the two p-dimensional vectors. Also related to our task is [29], where a metric between covariance matrices is proposed. The notion of a distance metric between canonical autocorrelation structures differs from these because CAA finds a factorization of the correlation matrix where each portion of the correlation matrix is expressed as the outer product of a pair of orthonormal vectors, which define a bi-dimensional space in which the projected data follows a linear correlation. Section 2.3 discusses the proposed metric.
Learning to defer has been studied in the literature as a means to effectively combine algorithmic and human decision-making [30, 31]. When decision-makers are knowledgeable domain experts, as is the case of clinicians providing care to comatose survivors of cardiac arrest, it is desirable to provide a framework in which the algorithm only provides suggestions when confident, and defers to the human in all other cases. Our work incorporates a deferral notion, and the proposed system only provides recommendations for a subset of cases where it is confident of its predictions.
2 Methods
2.1 Data sources
This study was approved by the University of Pittsburgh Institutional Review Board with a waiver of informed consent. The data used in this case study are derived from 451 comatose survivors of cardiac arrest treated at a single academic medical center between 2010 and 2015 [10, 32]. For each patient, raw EEG data (recorded at 256Hz across 20 electrode channels distributed in space across the scalp) were signal-processed using commercially available FDA-approved software (Persyst(R) Version 12, Persyst Development Corp, Prescott AZ), using standard clinical signal processing engines. The resulting quantitative EEG (qEEG) measures were summarized at a resolution of 1Hz and are available for continuous EEG recordings averaging about 36 hours per patient. The total number of qEEG features is 66 and include seizure probability, amplitude-integrated EEG for the left and right hemispheres of the brain, epileptiform spike detections, suppression ratio, summary frequency measures, and other metrics physicians find informative. The raw EEG data were not available. The full list of features can be found in S1 Table.
Also available for each patient are time-invariant clinical characteristics and outcomes, including survival to hospital discharge. For those who lived, the quality of their functional recovery at discharge was measured using two standard outcome scales: Cerebral Performance Category and modified Rankin Scale. We considered “favorable recovery” to be either a Cerebral Performance Category of 1 or 2 or a modified Rankin Scale score of 0-2 at hospital discharge. For those who died, the proximate cause of death is known. Fig 3 shows this information in detail. The data used in our experiments is limited to patients who survived hospital discharge and who were monitored for at least 36 hours, which corresponds to a total of 80 patients, half of whom had a favorable recovery. Table 1 includes the demographic and clinical characteristics of this cohort. The reasons for limiting our analysis to this subset are explained more in detail in Section 3.
(Left) Survival and outcome. (Right) Cause of death.
2.2 Canonical Autocorrelation Analysis
The goal of CAA is to find multivariate sparse correlations within a single set of variables. In the Sparse CCA framework, this could be understood as having identical matrices X and Y. Applying Sparse CCA when X = Y results in solutions u = v, corresponding to Sparse PCA solutions for X [24]. We overcome this issue by introducing a penalty for overlapping feature support. The resulting optimization problem for CAA is shown in Eq 2.
(2)
This can be understood as a new generalization of the Penalized Matrix Decomposition [24]. Note that the equality constraint in Eq 2 can be seen as a weighted L1 penalty when either u or v are fixed. Replacing the equality constraint by an inequality constraint gives a biconvex problem, while resulting in the same solution. Therefore, we can solve it through alternate convex search [33], as shown in Algorithm 1.
Algorithm 1: CAA via alternate convex search
1 Initialize v s.t. ||v||2 = 1;
2 repeat
3
4 s.t. , ||u||1 ≤ c1,
5
6 s.t. , ||v||1 ≤ c1,
7 until u, v converge;
8 d ← uTXTXv;
At each iteration, the resulting convex problem can be solved through the Karush-Kuhn-Tucker (KKT) conditions. The pseudo-code for solving the convex problems at each iteration of the alternate convex search is provided in Algorithm 2, where we solve for u without loss of generality. For a detailed derivation see S1 File.
Algorithm 2: CAA alternate convex search iteration via KKT conditions
1 ;
2 if then
3 return
4 else
5 Binary search to find λ2 s.t. ;
6 return
7 end
To find multiple pairs of CAA canonical vectors, Algorithm 1 can be repeated iteratively, replacing XTX with a matrix from which the already found correlations are removed, as shown in Eq 3, where d = uTXTXv.
(3)
In order to enable the discovery of non-linear correlations by extending the feature space with subsequent powers of the original features [34], we modify the optimization problem to extend the concept of disjoint support to sets of features. This also prevents the discovered correlations to be dominated by relationships between features that are already known to be correlated by design. Assuming each feature xi has a subset Si of associated indices of other features that should not be included as correlates of xi, the resulting optimization problem follows Eq 4.
(4)
The new constraint for disjoint support can still be understood as a weighted-L1 penalty at each iteration of the biconvex optimization algorithm. Hence, the problem can still be solved as discussed above, with the only difference that the parameters of the soft-thresholding operator will change.
2.3 Canonical Autocorrelation Embeddings
CAA allows us to find bi-dimensional projections where the data closely follows a linear distribution. Each axis of these projections corresponds to a linear combination of the original features, and their respective coefficients are represented in a pair of vectors . We call each pair u, v a CAA canonical space, and each CAA model may consist of one or more of such canonical spaces.
Since the correlations discovered by CAA are defined by pairs of vectors in , we can measure the distance between two CAA canonical spaces in terms of Euler angles defining the rotation from one pair of axes to the other. Measuring the angle between two vectors is equivalent to measuring the arc between them, and ||u||2 = ||v||2 = 1 ∀i, therefore, the distance between two CAA canonical spaces C1 and C2 can be defined as shown in Eq 5. This yields an embedding that we refer to as Canonical Autocorrelation Embedding (CAE).
(5)
It is easy to show that this metric satisfies the necessary conditions for a well-defined distance, see S3 File for the proof.
Even though we believe that Eq 5 provides a good distance metric that captures what we desire to measure, we do not claim this is the only nor necessarily the best such metric, and it is appropriate to continue exploring alternatives. S4 File contains a short discussion of why “principal angles”, a metric that is commonly used to measure distance between subspaces and which naturally comes to mind in this setting, is actually not well-suited in this case.
2.4 K-Nearest correlations
Having formulated a distance metric between pairs of CAA canonical spaces enables us to employ a range of distance-based machine learning algorithms, such as k-means, hierarchical clustering, or k-nn, to leverage similarities among correlation structures present in data. One additional complexity in our case is that each subset of data being compared may be represented by more than one CAA canonical space, and therefore more than one point in the embedding.
This setting can be incorporated into the k-nn framework by calculating the class probability for each correlation structure through the votes of their k nearest neighbors, and then aggregating over all correlations associated to an object using log-odds, as shown in Eq 6, where np,i,j denotes the class label of the jth neighbor of the ith correlation of patient p.
(6)
However, it is likely that some type of correlation structures will be common to both classes, while others are discriminative. To reduce noise and allow for those discriminative correlations to lead the decision, we incorporate a threshold t, so that log-odds are only calculated over those correlation structures with a class probability that is discriminative enough, as shown in Eq 7. Incorporating this threshold also enhances interpretability of the comparisons, as it reduces the number of structures that are used for making a prediction, making it easier for practitioners to understand which correlations appear relevant for the task at hand. The parameters k, indicating the number of neighbors, and t can be tuned through cross-validation.
(7)
3 Results
Our principal goal is to help improve care given to comatose survivors of cardiac arrest through a decision support system that can boost the accuracy and timeliness of prognostication. To do so, we propose a new way to characterize patients using their latent multivariate correlation structures, and use the resulting featurization of data to build predictive models. The results presented in this section leverage data collected over a five year period at an academic medical center to provide a proof of concept of the proposed methodology.
As seen in Fig 3, the main cause of death for this patient population is withdrawal of life-sustaining therapy due to perceived poor neurological prognosis. However, as mentioned in Section 1, it is possible that in some cases treatment might be withdrawn too early, a decision which nearly invariably leads to death and precludes favorable recovery. Including those patients in our training set could result in the model replicating mistakes clinicians may be making, leading to a self-fulfilling prophecy. Considering this and the fact that our goal is to predict positive neurological outcome rather than survival alone, we train our model using only those patients who lived, making our target label whether they had a good or a poor neurological outcome. This also reduces the risk for unaccounted treatment effects, since presumably all patients who are kept on live support receive a minimum standard of appropriate care, while therapeutic nihilism may influence outcomes for patients for whom life-supporting therapies are interrupted.
For each patient, their entire qEEG record is available, with lengths varying from less than an hour to more than a week. We aim to predict recovery as early as possible, but the earlier we attempt prediction, the more challenging it is. For the purposes of this experiment, we target prediction after 36 hours of monitoring. We use CAA to characterize a two hour epoch between hours 34 and 36. We choose 36 hours because we are interested in a period where the EEG is relatively static so that we do not need to account for temporal trends within the analyzed epoch. Clinically, patients are cooled down for 24 hours then allowed to rewarm at about 0.25-0.5C/hr. Both temperature and medications used to suppress shivering can alter the EEG [10]. At 36 hours, patients are back to a normal body temperature. The specific question the proposed model answers is: can the correlations present during this epoch predict whether the patient will go on to enjoy a favorable recovery? We consider only two hours because it can be expected that each patient’s state fluctuates over time, and the resulting variance could obfuscate important patterns of correlation. Identifying temporal trends, or inferring meta-correlation structures that describe these trends, is an important subject of future work beyond the scope of current analysis. Fig 4 illustrates the process of characterization of multiple patients’ EEG data with CAA.
In order to avoid spurious results, we only consider CAA canonical projections that yield correlations with R2 > 0.25. Moreover, to ensure that only reasonably close neighbors are used for matching, we prune connections by only considering distances smaller than , a threshold that corresponds to a 90° rotation over one axis. Empirical results of k-Nearest Correlations with CAE obtained through 10-fold cross-validation, with tuning parameters k and t in an internal 10-fold cross-validation loop within each training fold, are presented in Figs 5 and 6.
X-axis in log-scale to emphasize low FPR region.
(A) CAE, TPR vs. FPR. (B) CAE, TNR vs. FNR. (C) Logit on sets, TPR vs. FPR. (D) Logit on sets, TNR vs. FNR.
For baseline comparison, we use a popular approach: extract features based on metrics calculated over windows of time and apply standard classifiers to the resulting featurization [35, 36, 37]. We calculate quartiles for each input feature over two hours preceding the 36-hour mark, and provide them as features to logistic regression with lasso regularization [38] and k-nn with Euclidean distance. We refer to these as logistic regression on sets and k-nn on sets, respectively. To emphasize the importance of considering a window of time rather than a snapshot, we also compare against the same two algorithms taking as input the last data point after 36 hours of monitoring, that is, the recording at one time step. We refer to this approach as logistic regression on points and k-nn on points, respectively. The parameters are chosen through 10-fold cross-validation. The results are included in Figs 5 and 6.
Finally, we apply the resulting system to those patients who were withdrawn from life-sustaining therapies. Amongst 31 patients who received life-sustaining therapies for at least 36 hours before withdrawal of life-sustaining therapies, five patients would have been marked by our classifier as very likely to recover at a threshold of FPR equal to 0.025.
4 Discussion
Recall that the proposed decision-support system is one that only makes recommendations when it has strong indications that the patient is likely to have a positive neurological recovery, and defers in all other cases. We evaluate the performance of the algorithm at the thresholds at which it would make recommendations, which we characterize through low false positive rates (FPR). The true positive rate (TPR) at a given FPR indicates what portion of positives–TPR—would be retrieved while assuring that no more than a given rate of negatives–FPR—will be incorrectly labeled as positive.
Due to the gravity of errors in this scenario, the tolerance for false positives should be extremely low. Receiver Operator Curves (ROC) shown in Figs 5 and 6 display true positive rates at different false positive rates, with the x-axis in log-scale to emphasize the low FPR region. While Area Under the Curve (AUC) is reported, it is important to note that this performance metric is not particularly relevant in our case (nor in any other case in which there is a fixed threshold at which decisions are made). AUC is a measure that allows us to aggregate performance over all possible FPR thresholds, but what we really care about is the performance at the thresholds that are chosen for deployment.
The results presented in Fig 5 show that the proposed methodology has predictive power, and the comparison to k-nn using Euclidean distance highlights the role of CAE. While the performance of all other methods at low FPR is no better than random, the performance of CAE at low FPR is promising, with a TPR of 0.325 and corresponding 95% confidence interval [0.125, 0.46] at a FPR of 0.025 (Fig 6A). This means that with very low probability of making a Type I error, we are able to confidently identify at least 12.5% of the patients who will go on to have a positive neurological recovery. If the deployment setup changed, an ensemble model including CAE and logistic regression could be used to draw benefits from both of its components: high recall at low FPR of CAE, and overall good separability between outcome classes of logistic regression.
Even though consensus guidelines advocate maintaining life-sustaining therapies for at least 72 hours after cardiac arrest [7, 8], the burden associated to continuing life-support for patients who will not have a positive neurological recovery still often leads clinicians to withdraw treatment earlier [3]. Thus, the ability of CAE to confidently identify patients that will likely recover with a good outcome has the potential to save lives.
To appropriately estimate the potential impact of such a decision support system in terms of lives saved, it is useful to compare against physicians’ assessments to validate if the predictions made with the proposed approach are non-redundant to what doctors already know. Each patient in our dataset is classified by Pittsburgh Cardiac Arrest Category, a 4-level, validated prognostic indicator assigned in the first six hours of their stay [39]. This classification indicates whether the patient is awake with little brain injury (category i), in a mild to moderate brain injury with good heart and lung function (category ii), in a mild to moderate brain injury but poor heart and/or lung activity (category iii), or severe brain injury with loss of some brainstem reflexes (category iv). While patients in category i have an associated probability of survival of 80%, and 60% probability of having a positive neurological recovery, patients in category iv have probabilities of 10%, and 5%, respectively. At a FPR lower than 0.025, the proposed methodology correctly identified a category iv patient who later went on to have a positive recovery. This constitutes a preliminary indication that the patterns of correlations in neurological activity measured with EEG constitute novel findings and have the potential to improve reliability of prognostication.
As discussed in Section 3, amongst those patients whose cause of death is withdrawal of life-sustaining therapy for perceived neurological prognosis, 5 out of 31 patients who received life-sustaining therapies for at least 36 hours would have been marked by our system as likely to have a positive recovery. Two of these patients had received a Pittsburgh Cardiac Arrest Category of iv. The remaining three received an initial Pittsburgh Cardiac Arrest Category of ii. While we do not have ground truth regarding counterfactuals of what would have happened if life-sustaining therapies had been continued for these patients, these results provide further indication that CAE is not simply leveraging patterns that are already being used by physicians.
While it would also be desirable to identify patients who have a very small probability of neurological recovery, we note that neither of the models would be able to provide confident recommendations to withdraw life-sustaining therapies while guaranteeing low false negative rates (FNR). Fig 6B and 6D show the results for CAE and logit on sets at low FNR. These negative results may in part be explained by the fact that the available labeled data encodes positive/negative outcomes, but these are not limited to just neurological activity. A patient could have a positive neurological recovery but have other medical complications that limit function and thus result in a bad outcome label. Meanwhile, the positive recovery label is sure to indicate positive neurological recovery (as well as positive recovery in other areas).
A research direction that could further improve the performance of CAE is correlation trajectory modeling. While our model captures correlations observed within an interval of time, and in that sense it goes beyond a purely stationary approach, leveraging the sequential structures in data and using all data collected during a patients’ stay is desirable. Methodologically, this calls for the development of models for trajectory modeling of multivariate correlation structures. This could also encompass further exploration of additional distance metrics that could incorporate other types of information. By leveraging more information, such an approach would have the potential of providing earlier and more specific predictions.
An additional direction for performance enhancement comes from the fact that our characterization of brain activity with CAA is motivated by the importance clinicians place on correlations. However, the correlations they know to be informative are across raw EEG channel measurements, and it is likely that at the current level of data aggregation, a big portion of the information may be to some extent obfuscated. This does not constitute a risk in terms of the validity of the results presented in this paper, but it means that further promising results may be expected from characterizing correlations in raw EEG signals. Such models could also lead to biological insights that may not be easily derived with the current approach.
An important challenge that arises in this setting is the selective labels problem [40]. Selective labels is a common yet understudied problem that occurs whenever historical decision-making blinds us to the true outcome for certain instances. In the case of predicting neurological recovery, we may only observe the true outcome when the clinicians decide to extend life sustaining therapy, while we are blind to the conterfactual of what would have happened in those cases for which life sustaining therapy is withdrawn early. If patients for whom treatment was stopped early are significantly different from those for whom it was not, which is possibly the case, machine learning models trained only on the observed outcomes might have a lower-than-desired performance for that group.
5 Conclusions and future work
Cardiac arrest is a leading cause of death around the world, coma after cardiac arrest is common, and good neurological recovery is rare. Everyday, clinicians are tasked with making a prediction that determines whether they will continue life-sustaining therapies for their patients in coma or not. Motivated by the emphasis the clinicians place on potential informativeness of the correlation structures in EEG data, we have proposed a way to characterize and compare patients based on the latent structures of multivariate correlations, and use such information to predict positive neurological recovery. To do so, we have proposed a new formulation of Canonical Autocorrelation Analysis (CAA), a method that automatically finds subsets of features of data that form strong multiple-to-multiple correlations. We have also introduced Canonical Autocorrelation Embeddings (CAE) to enable the comparison of discovered correlation structures. CAE makes powerful and well established machine learning methodologies that rely on the use of distance metrics applicable to the task at hand.
The results presented in this paper constitute a proof of concept. Future work involves collecting more data to train and validate the model. It is reasonable to believe that there may be a substantial heterogeneity across patients, hence experiments using more data of more subjects are a necessary next step. Applying CAE to the raw EEG channels rather than to the aggregated featurizations of data would also be interesting to explore once that data becomes available.
In addition, we are developing machine learning methodology to tackle the selective labeling problem by incorporating clinicians’ domain knowledge while not reproducing their mistakes. Developing proper evaluation metrics to assess performance under selective labels, and finding ways to tackle the blindness that may result from this problem, is an important ingredient needed to successfully use machine learning to save lives in clinical settings.
Supporting information
S1 Table. qEEG features.
Complete list of qEEG features available and used in this study.
https://doi.org/10.1371/journal.pone.0210966.s001
(PDF)
S1 File. CAA optimization.
Solution of CAA optimization problem via KKT conditions.
https://doi.org/10.1371/journal.pone.0210966.s002
(PDF)
S2 File. CAA and Sparse PCA.
Discussion of the relationship and differences between CAA and Sparse PCA.
https://doi.org/10.1371/journal.pone.0210966.s003
(PDF)
S3 File. CAA distance metric.
Proof of CAA well-defined distance metric.
https://doi.org/10.1371/journal.pone.0210966.s004
(PDF)
S4 File. Principal angles and CAA.
Discussion of why principal angles are not a well suited distance for CAA canonical spaces.
https://doi.org/10.1371/journal.pone.0210966.s005
(PDF)
References
- 1. Lozano R, Naghavi M, Foreman K, Lim S, Shibuya K, Aboyans V, et al. Global and regional mortality from 235 causes of death for 20 age groups in 1990 and 2010: a systematic analysis for the Global Burden of Disease Study 2010. The lancet. 2012;380(9859):2095–2128.
- 2. Benjamin EJ, Blaha MJ, Chiuve SE, Cushman M, Das SR, Deo R, et al. Heart disease and stroke statistics—2017 update: a report from the American Heart Association. Circulation. 2017;135(10):e146–e603. pmid:28122885
- 3. Elmer J, Torres C, Aufderheide TP, Austin MA, Callaway CW, Golan E, et al. Association of early withdrawal of life-sustaining therapy for perceived neurological prognosis with mortality after cardiac arrest. Resuscitation. 2016;102:127–135. pmid:26836944
- 4. Laver S, Farrow C, Turner D, Nolan J. Mode of death after admission to an intensive care unit following cardiac arrest. Intensive care medicine. 2004;30(11):2126–2128. pmid:15365608
- 5. Callaway CW, Donnino MW, Fink EL, Geocadin RG, Golan E, Kern KB, et al. Part 8: Post–Cardiac Arrest Care. Circulation. 2015;132(18 suppl 2):S465–S482.
- 6. Gold B, Puertas L, Davis SP, Metzger A, Yannopoulos D, Oakes DA, et al. Awakening after cardiac arrest and post resuscitation hypothermia: are we pulling the plug too early? Resuscitation. 2014;85(2):211–214. pmid:24231569
- 7.
Elmer J, Callaway CW. The brain after cardiac arrest. In: Seminars in neurology. vol. 37. Thieme Medical Publishers; 2017. p. 019–024.
- 8. Mulder M, Gibbs HG, Smith SW, Dhaliwal R, Scott NL, Sprenkle MD, et al. Awakening and withdrawal of life-sustaining treatment in cardiac arrest survivors treated with therapeutic hypothermia. Critical care medicine. 2014;42(12):2493. pmid:25121961
- 9. Bassetti C, Bomio F, Mathis J, Hess CW. Early prognosis in coma after cardiac arrest: a prospective clinical, electrophysiological, and biochemical study of 60 patients. Journal of Neurology, Neurosurgery & Psychiatry. 1996;61(6):610–615.
- 10. Elmer J, Gianakas JJ, Rittenberger JC, Baldwin ME, Faro J, Plummer C, et al. Group-based trajectory modeling of suppression ratio after cardiac arrest. Neurocritical care. 2016;25(3):415–423. pmid:27033709
- 11. Booth CM, Boone RH, Tomlinson G, Detsky AS. Is this patient dead, vegetative, or severely neurologically impaired?: assessing outcome for comatose survivors of cardiac arrest. Jama. 2004;291(7):870–879. pmid:14970067
- 12. Sandroni C, Cariou A, Cavallaro F, Cronberg T, Friberg H, Hoedemaekers C, et al. Prognostication in comatose survivors of cardiac arrest: an advisory statement from the European Resuscitation Council and the European Society of Intensive Care Medicine. Intensive care medicine. 2014;40(12):1816–1831. pmid:25398304
- 13. Hofmeijer J, Beernink TM, Bosch FH, Beishuizen A, Tjepkema-Cloostermans MC, van Putten MJ. Early EEG contributes to multimodal outcome prediction of postanoxic coma. Neurology. 2015;85(2):137–143. pmid:26070341
- 14. Elmer J, Rittenberger JC, Faro J, Molyneaux BJ, Popescu A, Callaway CW, et al. Clinically distinct electroencephalographic phenotypes of early myoclonus after cardiac arrest. Annals of neurology. 2016;80(2):175–184. pmid:27351833
- 15. Cloostermans MC, van Meulen FB, Eertman CJ, Hom HW, van Putten MJ. Continuous electroencephalography monitoring for early prediction of neurological outcome in postanoxic patients after cardiac arrest: a prospective cohort study. Critical care medicine. 2012;40(10):2867–2875. pmid:22824933
- 16. Hofmeijer J, Tjepkema-Cloostermans MC, van Putten MJ. Burst-suppression with identical bursts: a distinct EEG pattern with poor outcome in postanoxic coma. Clinical neurophysiology. 2014;125(5):947–954. pmid:24286857
- 17. Ignaccolo M, Latka M, Jernajczyk W, Grigolini P, West BJ. The dynamics of EEG entropy. Journal of biological physics. 2010;36(2):185–196. pmid:19669909
- 18. Menegazzi JJ, Callaway CW, Sherman LD, Hostler DP, Wang HE, Fertig KC, et al. Ventricular fibrillation scaling exponent can guide timing of defibrillation and other therapies. Circulation. 2004;109(7):926–931. pmid:14757695
- 19.
Hotelling H. Relations between two sets of variates. Biometrika. 1936; p. 321–377.
- 20. Friman O, Borga M, Lundberg P, Knutsson H. Exploratory fMRI analysis by autocorrelation maximization. NeuroImage. 2002;16(2):454–464. pmid:12030831
- 21. De Clercq W, Vergult A, Vanrumste B, Van Paesschen W, Van Huffel S. Canonical correlation analysis applied to remove muscle artifacts from the electroencephalogram. Biomedical Engineering, IEEE Transactions on. 2006;53(12):2583–2587.
- 22.
Todros K, Hero A. Measure transformed canonical correlation analysis with application to financial data. In: Sensor Array and Multichannel Signal Processing Workshop (SAM), 2012 IEEE 7th. IEEE; 2012. p. 361–364.
- 23. Witten DM, Tibshirani RJ. Extensions of sparse canonical correlation analysis with applications to genomic data. Statistical applications in genetics and molecular biology. 2009;8(1):1–27.
- 24. Witten DM, Tibshirani R, Hastie T. A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis. Biostatistics. 2009; p. kxp008. pmid:19377034
- 25.
El-Arini K, Moore AW, Liu T. Autonomous visualization. In: Knowledge Discovery in Databases: PKDD 2006. Springer; 2006. p. 495–502.
- 26.
Fiterau M, Dubrawski A. Projection retrieval for classification. In: Advances in Neural Information Processing Systems; 2012. p. 3023–3031.
- 27. Krzanowski W. Between-groups comparison of principal components. Journal of the American Statistical Association. 1979;74(367):703–707.
- 28. Korth B, Tucker LR. Procrustes matching by congruence coefficients. Psychometrika. 1976;41(4):531–535.
- 29.
Förstner W, Moonen B. A metric for covariance matrices. In: Geodesy-The Challenge of the 3rd Millennium. Springer; 2003. p. 299–309.
- 30.
Cortes C, DeSalvo G, Mohri M. Learning with rejection. In: International Conference on Algorithmic Learning Theory. Springer; 2016. p. 67–82.
- 31.
Madras D, Pitassi T, Zemel R. Predict Responsibly: Increasing Fairness by Learning To Defer. arXiv preprint arXiv:171106664. 2017.
- 32. Elmer J, Rittenberger JC, Coppler PJ, Guyette FX, Doshi AA, Callaway CW, et al. Long-term survival benefit from treatment at a specialty center after cardiac arrest. Resuscitation. 2016;108:48–53.
- 33. Gorski J, Pfeuffer F, Klamroth K. Biconvex sets and optimization with biconvex functions: a survey and extensions. Mathematical Methods of Operations Research. 2007;66(3):373–407.
- 34. De Branges L. The Stone-Weierstrass theorem. Proceedings of the American Mathematical Society. 1959;10(5):822–824.
- 35.
Chen L, Dubrawski A, Clermont G, Hravnak M, Pinsky MR. Modelling risk of cardio-respiratory Instability as a heterogeneous process. In: AMIA Annual Symposium Proceedings. vol. 2015. American Medical Informatics Association; 2015. p. 1841.
- 36.
Wiens J, Guttag J, Horvitz E. Learning evolving patient risk processes for c. diff colonization. In: ICML Workshop on Machine Learning from Clinical Data; 2012.
- 37. Güiza F, Depreitere B, Piper I, Van den Berghe G, Meyfroidt G. Novel methods to predict increased intracranial pressure during intensive care and long-term neurologic outcome after traumatic brain injury: development and validation in a multicenter dataset. Critical care medicine. 2013;41(2):554–564. pmid:23263587
- 38. Tibshirani R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B (Methodological). 1996; p. 267–288.
- 39. Coppler PJ, Elmer J, Calderon L, Sabedra A, Doshi AA, Callaway CW, et al. Validation of the Pittsburgh Cardiac Arrest Category illness severity score. Resuscitation. 2015;89:86–92. pmid:25636896
- 40.
Lakkaraju H, Kleinberg J, Leskovec J, Ludwig J, Mullainathan S. The Selective Labels Problem: Evaluating Algorithmic Predictions in the Presence of Unobservables. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM; 2017. p. 275–284.