Deep Patient Predicts Patients Future from Health Records

This section presents the deep patient method and describes the pipeline implemented to evaluate the benefits of this representation in the task of predicting future diseases.

Deep Patient Representation

Figure 1 shows the high-level conceptual framework to derive the deep patient representation. EHRs are first extracted from the clinical data warehouse, pre-processed to identify and normalize clinically relevant phenotypes, and grouped in patient vectors (i.e., raw representation, Fig. 1A). Each patient can be described by just a single vector or by a sequence of vectors computed in, e.g., predefined temporal windows. The collection of vectors obtained from all the patients is used as input of the feature learning algorithm to discover a set of high level general descriptors (Fig. 1B). Every patient in the data warehouse is then represented using these features and such deep representation can be applied to different clinical tasks (Fig. 1C).

(A) Pre-processing stage to obtain raw patient representations from the EHRs. (B) The raw representations are modeled by the unsupervised deep architecture leading to a set of general and robust features. (C) The deep features are applied to the entire hospital database to derive patient representations that can be applied to a number of clinical tasks.

We derived the patient representation using a multi-layer neural network in a deep learning architecture (i.e., deep patient). Each layer of the network is trained to produce a higher-level representation of the observed patterns, based on the data it receives as input from the layer below, by optimizing a local unsupervised criterion (Fig. 2). Every level produces a representation of the input pattern that is more abstract than the previous level because it is obtained by composing more non-linear operations. This process is loosely analogous to neuroscience models of cognition that hierarchically combine lower-level features to a unified and compact representation. The last network of the chain outputs the final patient representation.

Each layer of the neural network is trained to produce a higher-level representation from the result of the previous layer.

Denoising Autoencoders

We implemented our framework using a stack of denoising autoencoders (SDA), which are independently trained layer by layer; all the autoencoders in the architecture share the same structure and functionalities¹⁸. Briefly, an autoencoder takes an input and first transforms it (with an encoder) to a hidden representation through a deterministic mapping:

parameterized by , where is a non-linear transformation (e.g., sigmoid, tangent) named “activation function”, W is a weight coefficient matrix, and b is a bias vector. The latent representation y is then mapped back (with a decoder) to a reconstructed vector , such as:

with and (i.e., tied weights). The hope is that the code y is a distributed representation that captures the coordinates along the main factors of variation in the data. When training the model, the algorithm searches the parameters that minimize the difference between x and z (i.e., the reconstruction error ).

Autoencoders are often trained to reconstruct the input from a noisy version of the initial data (i.e., denoising) in order to prevent overfitting. This is done by first corrupting the initial input x to get a partially destroyed version through a stochastic mapping . The corrupted input is then mapped, as with the basic autoencoder, to a hidden code and then to the decoded representation z (see the Supplementary Appendix A online for a graphical representation). We implemented input corruption using the masking noise algorithm¹⁸, in which a fraction y of the elements of x chosen at random is turned to zero. This can be viewed as simulating the presence of missed components in the EHRs (e.g., medications or diagnoses not recorded in the patient records), thus assuming that the input clinical data is a degraded or “noisy” version of the actual clinical situation. All information about those masked components is then removed from that input pattern, and denoising autoencoders can be seen as trained to fill-in these artificially introduced blanks.

The parameters of the model θ and θ′ are optimized over the training dataset to minimize the average reconstruction error,

where is a loss function and N is the number of patients in the training set. We used the reconstruction cross-entropy function as loss function, i.e.,

Optimization is carried out by mini-batch stochastic gradient descent, which iterates through small subsets of the training patients and modifies the parameters in the opposite direction of the gradient of the loss function to minimize the reconstruction error. The learned encoding function is then applied to the clean input x and the resulting code y is the distributed representation (i.e., the input of the following autoencoder in the SDA architecture or the final deep patient representation).

Evaluation Design

Feature learning algorithms are usually evaluated in supervised applications to take advantage of the available manually annotated labels. Here we used the Mount Sinai data warehouse to learn the deep features and we evaluated them in predicting patient future diseases. The Mount Sinai Health System generates a high volume of structured, semi-structured and unstructured data as part of its healthcare and clinical operations, which include inpatient, outpatient and emergency room visits. Patients in the system can have as long as 12 years of follow up unless they moved or changed insurance. Electronic records were completely implemented by our health system starting in 2003. The data related to patients who visited the hospital prior to 2003 was migrated to the electronic format as well but we may lack certain details of hospital visits (i.e., some diagnoses or medications may not have been recorded or transferred). The entire EHR dataset contains approximately 4.2 million de-identified patients as of March 2015, and it was made available for use under IRB approval following HIPAA guidelines. We retained all patients with at least one diagnosed disease expressed as numerical ICD-9 between 1980 and 2014, inclusive. This led to a dataset of about 1.2 million patients, with every patient having an average of 88.9 records. Then, we considered all records up to December 31, 2013 (i.e., “split-point”) as training data (i.e., 34 years of training information) and all the diagnoses in 2014 as testing data.

EHR Processing

For each patient in the dataset, we retained some general demographic details (i.e., age, gender and race), and common clinical descriptors available in a structured format such as diagnoses (ICD-9 codes), medications, procedures, and lab tests, as well as free-text clinical notes recorded before the split-point. All the clinical records were pre-processed using the Open Biomedical Annotator to obtain harmonized codes for procedures and lab tests, normalized medications based on brand name and dosages, and to extract clinical concepts from the free-text notes¹⁹. In particular, the Open Biomedical Annotator and its RESTful API leverages the National Center for Biomedical Ontology (NCBO) BioPortal²⁰, which provides a large set of ontologies, including SNOMED-CT, UMLS and RxNorm, to extract biomedical concepts from text and to provide their normalized and standard versions²¹.

The handling of the normalized records differed by data type. For diagnoses, medications, procedures and lab tests, we simply counted the presence of each normalized code in the patient EHRs, aiming to facilitate the modeling of related clinical events. Free-text clinical notes required more sophisticated processing. We applied the tool described in LePendu et al.²², which allowed identifying the negated tags and those related to family history. A tag that appeared as negated in the note was considered not relevant and discarded⁵. Negated tags were identified using NegEx, a regular expression algorithm that implements several phrases indicating negation, filters out sentences containing phrases that falsely appear to be negation phrases, and limits the scope of the negation phrases²³. A tag that was related to family history was just flagged as such and differentiated from the directly patient-related tags. We then analyzed similarities in the representation of temporally consecutive notes to remove duplicated information (e.g., notes recorded twice by mistake)²⁴.

The parsed notes were further processed to reduce the sparseness of the representation (about 2 million normalized tags were extracted) and to obtain a semantic abstraction of the embedded clinical information. To this aim we modeled the parsed notes using topic modeling²⁵, an unsupervised inference process that captures patterns of word co-occurrences within documents to define topics and represent a document as a multinomial over these topics. Topic modeling has been applied to generalize clinical notes and improve automatic processing of patients data in several studies (e.g., see^5,26,27,28). We used latent Dirichlet allocation as our implementation of topic modeling²⁹ and we estimated the number of topics through perplexity analysis over one million random notes. We found that 300 topics obtained the best mathematical generalization; therefore, each note was eventually summarized as a multinomial of 300 topic probabilities. For each patient, we eventually retained one single topic-based representation averaged over all the notes available before the split-point.

Dataset

All patients with at least one recorded ICD-9 code were split in three independent datasets for evaluation purposes (i.e., every patient appeared in only one dataset). First, we held back 81,214 patients having at least one new ICD-9 diagnosis assigned in 2014 and at least ten records before that. These patients composed validation (i.e., 5,000 patients) and test (i.e., 76,214 patients) sets for the supervised evaluation (i.e., future disease prediction). In particular, all the diagnoses in 2014 were used to evaluate the predictions computed using the patient data recorded before the split-point (i.e., prediction from the patient clinical status). The requirement of having at least ten records per patient was set to ensure that each test case had some minimum of clinical history that could lead to reasonable predictions. We then randomly sampled a subset of 200,000 different patients with at least five records before the split-point to use as training set for the disease prediction experiment.

We used ICD-9 codes to state the diagnosis of a disease to a patient. However, since different codes can refer to the same disease, we mapped the codes to a disease categorization structure used at Mount Sinai, which groups ICD-9s into a vocabulary of 231 general disease definitions³⁰. This list was filtered to retain only diseases that had at least 10 training patients and manually polished by a practicing physician to remove all the diseases that could not be predicted from the considered EHR labels alone because related to social behaviors (e.g., HIV) and external life events (e.g., injuries, poisoning), or that were too general (e.g., “other form of cancers”). The final vocabulary included 78 diseases, which are reported in the Supplementary Appendix B online.

Finally, we created the training set for the feature learning algorithms using the remaining patients having at least five records by December 2013. The choice of having at least five records per patient was done to remove some uninformative cases and to decrease the training set size and, consequently, the time of computation. This lead to a dataset composed of 704,587 patients and 60,238 clinical descriptors. Descriptors appearing in more than 80% of patients or present in fewer than five patients were removed from the dataset to avoid biases and noise in the learning process leading to a final vocabulary of 41,072 descriptors. Overall, the raw patient dataset used for feature learning was composed by 200 million non-zero entries (i.e., about 1% of all the entries in the patient-descriptor matrix).

Patient Representation Learning

SDAs were applied to the dataset of 704,857 patients to derive the deep patient representation. All the feature values in the dataset were first normalized to lie between zero and one to reduce the variance of the data while preserving zero entries. We used the same parameters in all the autoencoders of the deep architecture (regardless the layer) since this configuration usually leads to similar performances as having different parameters for each layer and is easier to evaluate^18,31. In particular, we found that using 500 hidden units per layer and a noise corruption factor lead to a good generalization error and consistent predictions when tuning the model using the validation data set. We used a deep architecture composed by three layers of autoencoders and sigmoid activation functions (i.e., “DeepPatient”). Preliminary results on disease prediction using a different number of layers are reported in the Supplementary Appendix C online. The deep feature model was then applied to train and test sets for supervised evaluation; hence each patient in these datasets was represented by a dense vector of 500 features.

We compared the deep patient representation with other well-known feature learning algorithms having demonstrated utility in various domains including medicine¹². All of these algorithms were applied to the scaled dataset as well and performed only one transformation to the original data (i.e., shallow feature learning). In particular, we considered principal component analysis (i.e., “PCA” with 100 principal components), k-means clustering (i.e., “K-Means” with 500 clusters), Gaussian mixture model (i.e., “GMM” with 200 mixtures and full covariance matrix), and independent component analysis (i.e., “ICA” with 100 principal components). In particular, PCA uses an orthogonal transformation to convert a set of observations of possibly correlated variables into a set of linearly uncorrelated variables called principal components, which are less than or equal to the number of original variables. The first principal component accounts for the greatest possible variability in the data, and each succeeding component in turn has the highest variance possible under the constraint that it is orthogonal to the preceding components. K-means groups unlabeled data into k clusters, in such a way that each data point belongs to the cluster with the closest mean. In feature learning, the centroids of the cluster are used to produce features, i.e., each feature value is the distance of the data point from each cluster centroid. GMM is a probabilistic model that assumes all the data points are generated from a mixture of a finite number of Gaussian distributions with unknown parameters. ICA represents data using a weighted sum of independent non-Gaussian components, which are learned from the data using signal separation algorithms. As done for DeepPatient, the number of latent variables of each model was identified through preliminary experiments by optimizing learning errors or expectations as well as prediction results obtained in the validation set. We also included in the comparison the patient representation based on the original descriptors after removal of the frequent and rare variables (i.e., “RawFeat” with 41,072 entries).

Future Disease Prediction

To predict the probability that patients might develop a certain disease given their current clinical status, we implemented random forest classifiers trained over each disease using a dataset of 200,000 patients (one-vs.-all learning). We used random forests because they often demonstrate better performances than other standard classifiers, are easy to tune, and are robust to overfitting^32,33. By preliminary experiments on the validation dataset we tuned every disease classifier to have 100 trees. For each patient in the test set (and for all the different representations), we computed the probability to develop every disease in the vocabulary (i.e., each patient was represented by a vector of disease probabilities).