Machine learning–based predictive models have been found to be highly accurate in many clinical decision support tasks. Examples include predictions of unforeseen complications [1], patient severity assessment through mortality predictors [2] and automated coding for billing [3], and prediction of patient outcomes [4], to name a few. The key ingredients of these models are the features used to describe patients for whom predictions are required. The traditional approach for building these features is to handcraft them typically in collaboration with a domain expert. However, with the growing amount, complexity, and diversity of clinical information sources, such manual feature engineering is practically infeasible. For instance, in electronic medical records (EMRs), patient information may be distributed among laboratory tests, nursing notes, radiology images and reports, genomic data, and other data sources.

Representation learning aims to learn features or representations from the given input sources automatically, often in an unsupervised manner. This obviates the need for manual feature engineering and is particularly useful with unstructured data sources such as clinical notes. These real-valued vectorial representations can be used as features directly in machine learning models for various downstream tasks such as prediction or cluster detection. Such representation learning has been found to be effective in several predictive models, for example, disease category prediction [5] and mortality prediction [6].

Previous studies have primarily used clinical notes to learn patient representations. Clinical notes are a rich source of information containing detailed subjective and objective evaluations of patient conditions during the hospital stay. Some previous studies have also combined other structured tables from EMR with features extracted from notes to obtain patient representations [1,5] or to mine clinical information such as drug mentions [7]. Many of these studies have used variants of deep neural architecture based on autoencoders to obtain unsupervised patient representations.

When information from multiple heterogeneous sources is available, predictive models benefit from latent representations that systematically model correlated shared structures. The aim of multi-view learning is to effectively build such latent representations, where views refer to measurements for the same subjects that differ in source, datatype, or modality; heterogeneous data sources within EMR provide such multiple views of patients. A general technique for multi-view representation learning from arbitrary collections of heterogeneous data sources is collective matrix factorization (CMF) [8]. CMF can be used to obtain patient representations from multi-view EMR data and can also be used to seamlessly integrate auxiliary information from external sources.

One such auxiliary source of information is a clinical knowledge graph (KG) that has been found to be valuable for improving both the accuracy and interpretability of predictive models. These KGs have clinical entities (eg, diseases, drugs, and biomolecules) as nodes and different kinds of relations (eg, treats, predisposes, and causes) as edges. They can be automatically created from various sources such as biomedical literature and web-based health portals. Representation learning methods have also been developed for graph inputs that can automatically learn vectorial representations of nodes to incorporate the global structural and semantic properties of the graph. These node representations can then be used in machine learning models for graph analytics such as community detection or node classification. Owing to its wide applicability, a large number of graph representation learning techniques have been developed for various classes of graphs, including KGs.

In this paper, we analyze patient representation learning in light of 2 recent advances in CMF and KG representation learning. A deep autoencoder-based architecture, called deep CMF (DCMF), was developed for CMF, which was found to outperform classical nonneural variants of CMF in several tasks [9]. Using DCMF, which provides a seamless way of integrating heterogeneous data, we evaluate the effectiveness of patient representations when the input data are augmented with additional information from literature-derived KGs. The generality of DCMF allows many different ways of using KG as inputs; however, not all of them are equally effective. Recently, it has been shown that many graph representation learning methods can be reformulated as a matrix factorization problem. Leveraging this formulation within the context of CMF and DCMF, we infuse auxiliary information during patient representation learning. To our knowledge, this is the first study to use this technique to obtain clinical KG representations and use it within the DCMF framework to obtain patient representations.

Furthermore, the DCMF architecture can easily be extended to create a task-specific end-to-end model that learns to simultaneously find effective patient representations and predictions. We also compare the efficacy of such a model to that of a 2-stage process of first learning unsupervised representations and then independently learning a predictive model.

We rigorously evaluate patient representation learning using DCMF-based methods and autoencoders for 2 clinical decision support tasks on EMR data comprising 28,563 patient episodes. The first task is that of primary diagnosis category prediction, which is performed during coding from discharge summaries when a patient is discharged from the hospital for billing and reimbursement purposes. The second task is that of mortality (risk of death) prediction, which can be used to identify high-risk patients and prioritize their care.

The utility of DCMF-based patient representations, obtained from only EMR data and a combination of KGs and EMR data in these 2 tasks, is empirically analyzed and discussed.

We extended the unsupervised DCMF model to incorporate task-specific supervision. This allowed us to learn entity representations that are influenced by the target variables provided for the predictive task. Furthermore, this creates a predictive model that can seamlessly learn from arbitrary collections of matrices. We assumed that the predictive task, for example, regression or classification, is with respect to one entity only. In the case of clinical tasks, this entity is most often patients. All other data, such as EMRs and KGs, can be used as inputs from which a predictive model for patients can be built. Examples include predicting the length of stay (regression) or the risk of an unforeseen complication (classification).

The DCMF architecture is extended by adding an additional task-specific layer that takes as input the latent representation of the entity for which labels are provided. This layer is provided with labels during training and is trained along with the rest of the network. Let e_p be the specific entity (eg, patients) for which task-specific labels y_T are provided for a task T. Let be the bottleneck of the autoencoder corresponding to the entity e_p. The network is constructed as described above with the addition of a single network layer that takes as input and has an activation layer depending on the task and loss function (eg, sigmoid for classification and linear for regression). There is a task-specific loss l_T(y_T,y’) associated with this layer that is also task dependent (eg, cross-entropy for classification and mean-squared error for regression), where y’ denotes the network’s predictions. The supervised latent representations are now learned by solving the following equation:

Collective training of all autoencoders is performed in exactly the same way as in DCMF but with the new loss function as given above. During prediction, new inputs for entity e_p may be given along with all other auxiliary data, and the additional layer’s outputs can be used as predictions.

Figure 1 shows a schematic of the model. There are 5 input matrices containing pairwise relations across 6 entities. The graph at the top shows the associations between entities and matrices. One of the entities (shaded) is associated with the labels for a classification task. The network comprises 6 autoencoders, as shown at the bottom, 1 for each entity. The input to the autoencoders is from the concatenated matrix corresponding to each entity (shown in the input transformation part). The bottleneck layer from the first autoencoder is used as input to a network layer that uses the provided labels during training. Note that this illustration shows a specific example of 5 matrices; however, the DCMF model can be used with any collection of input matrices.

Any graph may be represented by its adjacency matrix. However, factorization of this adjacency matrix may not yield effective representations. We also observed this empirically in our experiments. Another way of using KGs is to first obtain graph embeddings and then use the embeddings within the CMF. We experimented with TransE [40] and found that this did not yield effective representations. To obtain good representations, we used the technique used previously by Liu et al [29]. The key idea was to compute the similarities between the nodes in the graphs and obtain representations by factorizing the similarity matrices.

The global resource allocation (GRA) similarity, between 2 nodes in a graph, was proposed by Liu et al [29] with the aim of having similar embeddings for similar nodes and generalizing previous metrics. We found similarities between diseases, medications, and procedures (separately) from the SemMedDB KG using the GRA similarity. These similarity matrices are provided as input to CMF-based methods that internally factorize all the matrices collectively, as shown in Figure 2.

We now provide an intuitive explanation of GRA similarity and explain why it is a good measure for clinical KGs; a more technical description can be found in the study by Liu et al [29]. The similarity between 2 nodes i and j is computed based on the paths that exist between them. Such a global measure can be applied to any 2 nodes in the graph, irrespective of their distance within the graph. In contrast, local measures, such as the number of common neighbors, often yield ineffective embeddings as many node pairs may have the same scores. This is particularly true for dense clinical KGs.

The similarity score depends on (1) the number of paths, (2) the length of the paths, and (3) the node degrees of the intermediate nodes in each path. For each path between i and j, its contribution is equal to the reciprocal of the product of the degrees of the intermediate nodes of the path. Let p^l(i,j) be a path of length l between nodes i and j, and let the intermediate nodes be i₁,i₂,...i_{l–2}. Let k(i) denote the degree of node i, that is, the number of edges incoming to or outgoing from i. The contribution of a path c(p^l) is defined as follows:

In this manner, paths that contain high-degree nodes have higher denominators, and their contributions are decreased. This is justified as high-degree nodes connect many different nodes and thus affect many paths. Therefore, paths that do not contain such high-degree nodes should contribute to the higher similarity between the nodes. The final GRA similarity is the sum of the contributions over all paths weighted by a factor that decays exponentially with path length:

By exponentially decaying the weights, shorter paths are assigned higher weights. Thus, both the number and length of the paths are accounted for in the similarity measure.

Liu et al [29] showed that this technique generalizes and outperforms many previous graph embedding methods. To our knowledge, ours is the first study to use this technique to obtain clinical KG representations and use it within a collective matrix factorization setting to obtain patient representations.

Figure 3 shows a schematic of the experimental settings. We considered 3 views: 1, 2, and 3. View 1 comprises data extracted from clinical notes that have been used for patient representation learning in several previous studies. In view 2, data from SemMedDB KGs were extracted as described above and added to the data from view 1. In view 3, structured data from the EMR were also added to obtain patient representations. In the following section, we evaluate the performance of representations learned from these 3 views in 2 clinical decision support tasks.

We first present the results of the diagnosis category prediction and then mortality prediction. For each task, we visually present the results in 2 ways: one organized by view and another organized by method. The former allowed us to compare methods within each view, and the latter allowed us to compare views within each method.

Table 3 shows the results of the diagnosis category prediction. In view 1, predictions using supervised DCMF yielded >30% improvement in macro-F1 scores compared with classifiers with SDAE-based representations. In views 2 and 3, considerable improvement, ranging from 82% to 1955% in macro-F1 scores, was observed over other methods that separately learned representations and classifiers. In view 1, the accuracy and weighted F1-score of supervised DCMF were comparable with those obtained from classifiers trained on SDAE-based representations. However, with the addition of knowledge matrices in view 3, which can be performed seamlessly, supervised DCMF surpassed their performance.

Figure 4 shows the results of the diagnosis category prediction across the 3 views. In view 1, we observed that neural representations from SDAE and DCMF outperformed nonneural representations from CMF. The supervised DCMF outperformed all other methods. The addition of information from KGs in view 2 improved the performance of DCMF, both unsupervised and supervised, in all 3 metrics. The addition of structured EMR data in view 3 further improved the performance.

Figure 5 shows the same results of diagnosis category prediction as seen in Figure 4 but is organized based on the method. SDAE representations cannot be used in augmented multi-view settings but outperform CMF-based representations even when the CMF uses more data in views 2 and 3. This is likely because of the better representation learning capability of the neural networks. We also see that the DCMF learned better representations from all 3 views. However, although the addition of KG matrices in view 2 improved performance over view 1, further addition of data in view 3 deteriorated performance. However, with the addition of supervision from the labels, supervised DCMF was able to learn better with increasing performance across the 3 views.

Table 4 shows the results of mortality prediction. We observed that supervised DCMF outperformed SDAE-based models by >16% in AUC and >13% in macro-F1 in view 1, where data were obtained from clinical notes. In views 2 and 3, where data from KGs and EMRs were cumulatively added to clinical notes, supervised DCMF outperformed all the baselines by similar margins. These results demonstrate the advantage of end-to-end learning using supervised DCMF over other methods that separately learn representations and classifiers.

Figure 6 shows the AUC and F1 scores obtained by the methods across the 3 views. In view 1, the SDAE representations outperform those from CMF. Results with the logistic regression classifier were marginally better than those from the random forest, with SDAE, CMF, and DCMF representations. In view 1, DCMF representations have performance comparable with that of SDAE. Supervised DCMF outperformed all other methods by a large margin. The addition of KG matrices in view 2 improved the performance of the unsupervised DCMF-based classifier. The addition of structured EMR data in view 3 improved the performance of the supervised DCMF.

Figure 7 shows the same results from Figure 6, but is organized based on each method. The performances of the unsupervised neural methods SDAE and DCMF are comparable. DCMF can use information from KG matrices to boost its performance. However, the addition of structured EMR data did not increase its performance. However, supervised DCMF is able to use additional data well and achieves the best performance overall with view 3.

Our experiments strongly suggest that end-to-end models that are trained in a supervised manner outperform models comprising 2 stages of unsupervised representation learning and an independently learned classifier. An end-to-end neural model also learns patient representations internally; however, these representations are influenced by task-specific labels used for supervision. How these supervised representations perform on tasks other than what they are trained for, that is, whether they are beneficial in transfer learning, remains to be examined. Thus, for a given clinical decision support task, if labels are available, our experiments indicate that an end-to-end model should be preferred.

DCMF provides a seamless way of integrating multiple sources of data for obtaining patient representations in both unsupervised and supervised settings. As a versatile learning method, it can be used with inputs from a single source (eg, clinical notes) as well as when inputs are from multiple sources (eg, clinical notes and structured EMR tables). Its performance in these settings is comparable with that of previous autoencoder-based representation learning methods. DCMF can also be used to obtain representations in augmented multi-view settings containing arbitrary collections of matrices, where most previous representation learning methods cannot be used directly. In such settings, its performance is considerably superior to that of the previous nonneural methods for CMF. Thus, it provides a framework for infusing valuable information from auxiliary information sources, such as KG, into patient representations.

Graph embeddings allow us to obtain vectorial representations of nodes in a graph in a way that incorporates the global structural and semantic properties of the graph. Such embeddings can be obtained for KGs as well. The technique for obtaining the embedding can be formulated as a factorization of a similarity matrix where the similarities between nodes are defined based on the number and structural characteristics of the paths between them. With this formulation, the factorization can become part of CMF, which enables us to learn patient representations from multiple clinical data sources as well as KGs. Such patient representations were found to improve downstream predictive performance, especially in supervised settings. Other ways of using KGs within DCMF were not found to be as effective; the 2 alternatives tested were directly using the adjacency matrices of the graphs and first obtaining graph embeddings and then using the embedding matrices within CMF.

Our experimental evaluation was conducted on 2 clinical decision support tasks: a binary classification task (mortality prediction) and a multilabel classification task (primary diagnosis category prediction). Furthermore, the evaluation was performed on a subset of data sources (clinical notes, laboratory investigations, and medications) from a single hospital. The trends in performance are expected to remain the same for other tasks (eg, regression tasks) and the addition of other data sources (eg, radiology images) but must be empirically verified.

The KG used is derived automatically from biomedical literature using NLP techniques. Inaccuracies because of NLP algorithms may lead to false positives (erroneous nodes and edges) and false negatives (incompleteness) in KG. Further investigation into the effects of these inaccuracies in the representations is required. Evaluation of KGs derived from other sources can also be performed. It is possible that the results may improve with decreasing inaccuracies in the KG.

Very little hyperparameter tuning was performed for the neural models. The results of all neural models are expected to improve with more tuning. The autoencoders used within the DCMF are simple feedforward networks. Other types of autoencoders, such as SDAE or variational autoencoders, may also be used, which may improve the performance of the DCMF.

In this study, we investigated the use of DCMF to obtain patient representations for 2 clinical decision support tasks. The key advantage of DCMF is its versatility: it can be used to obtain representations from a single view (eg, clinical notes), from multiple views (eg, notes and structured tables in EMR data), and in augmented multi-view settings where it can seamlessly integrate information from diverse sources such as EMR data and KGs. Most previous representation learning methods cannot be used with such augmented multi-view data. Furthermore, DCMF can be easily used to learn representations in both supervised and unsupervised settings. In our experiments, we found that DCMF-based representations lead to predictive accuracy that is comparable with or better than previous techniques. Thus, DCMF offers an effective way of integrating heterogeneous data sources and infusing auxiliary knowledge into patient representations.