In this paper we focus on prediction of health status of patients from the historical Electronic Health Records (EHR). We propose a multi-task framework that can monitor the multiple status of diagnoses. Patients' historical records are fed into a Recurrent Neural Network (RNN) which memorizes all the past visit information, and then a task-speci c layer is trained to predict multiple diagnoses. Experimental results show that prediction accuracy is reliable if compared to widely used approaches 1
Disease monitoring is often limited by physician experience, test time, economic barriers and so on. The Electronic Health Record (EHR) is a valuable source for exploratory analysis to monitor diseases and assist clinical decision making. However, due to the complexity of EHR data, the e cient mining of EHRs is not trivial.
Recent work has made rapid progress in utilizing EHRs for predictive
modeling tasks in healthcare, including predicting unplanned readmission [
In order to model the dependencies of diagnoses, deep leaning techniques,
such as recurrent neural networks, can be employed. Recent work [
In this paper, our goal is to predict the status of multiple diagnoses (or observations), with each diagnosis having multiple severity levels. We form our problem as multi-task learning, which rst learns a shared representation from all the features, and then performs task-speci c predictions. We propose an attentionbased RNN model to monitor patient's longitudinal health information. First, we use an RNN to memorize all the information from historical visits, and then attention mechanisms to measure visit importance. Based on the latent representation, we train multiple classi ers and each focuses on the prediction of a speci c task. We perform our model on two applications: predicting chronic states for bone health, and monitoring BloodTest values for cardiovascular disease. 2
The basic component of our framework is gated recurrent unit, which is a stateof-the-art deep learning architecture for modelling long range sequences. To further improve its performance, we apply attention mechanisms to measure the importance of historical sequences. To predict the status of multiple diagnoses, we add a multi-task classi cation layer on top of the learned representations.
We implement our RNN with Gated Recurrent Units (GRU) [
RNN can remember the past information for future prediction. However, it is limited to only a few latest steps, with more impact from later ones, and may not be able to discover major in uences from earlier timestamps. Therefore, we apply attention mechanisms to memorize the e ect from long-time dependencies, which have gained success in many tasks.
Our task is to predict the status of multiple measurement results at the time (t + 1) given the historical records from x1 to xt. Figure 1 shows a high-level overview of the proposed model. Given the information from time 1 to t, the i-th visit's health record xi is fed into an RNN network, which outputs a hidden state hi as the representation of the i-th visit. Along with the set of hidden states fhigit=11, we compute their relative importance t, and then obtain a context state ct. From the context state ct and the current hidden state ht, we can obtain an attentional hidden state h~t, which is used to predict diagnoses in the (t + 1)th visit. For the prediction, we use M softmax classi ers, which correspond to the M di erent diagnoses, to predict the severity level for each diagnosis. The representation ht contains the visit information of all the input features, and the task-speci c classi er focuses on the prediction of each diagnosis.
Task 1
Task 2
Task − 1 −
Task … … RNN
RNN
RNN −
Softmax Attention
RNN We conduct experiments on two real-word datasets, and evaluate the performance of the proposed attention-based RNN models compared to other prediction methods.
Study of Osteoporotic Fractures Dataset. The study of osteoporotic fracture
(SOF) [?] is the largest and most comprehensive study focused on bone diseases.
It includes 20 years longitudinal data about osteoporosis of 9,704 Caucasian
women aged 65 years and older. Potential risk factors and confounders belong
to several groups such as demographics, family history, and lifestyle. We process
people's bone health diagnoses of di erent areas using the bone mineral density
(BMD) values by comparison with young healthy references [
BloodTest Dataset. This dataset [
As a common issue of EHR, these datasets are irregularly sampled and sparse, so that data preprocessing is needed. For each person, we remove those visits without any monitored variables recorded, and remove patients with less than three visits. We use simple imputation to ll missing variables. For the SOF data, we ll the missing variables with the values in the previous visit. For the BloodTest data, we impute missing sequences (where a single variable is missing entirely) with a clinical normal value. The used datasets with statistics is shown in Table 1.
For each patient, we want to predict the diagnosis results of each visit based on his/her previous records. To validate the performance of the proposed models
Dataset SOF BloodTests Number of patients 5,318 2,055 Number of visits 22,313 18,758 Average number of visits per patient 4.19 9.13 Number of normal claims 25,145 221,642 Number of low abnormal claims 55,399 17,407 Number of high abnormal claims 31,021 79,837 Total number of features 42 17
Number of monitored diagnoses 5 17 in this diagnosis prediction task, we conduct experiments on two categories of methods: baselines and RNN-based models.
We set up two kinds of baselines. The rst baseline is to use the median value of each monitored variable from V1 to Vt to predict Vt+1 for continuous variables. This is based on a heuristic assumption that the most frequent state is more likely to occur. For each patient, we use his/her most popular health status as the current status, regardless of time variations. The second baseline is a multi-task logistic regression (LR). To predict information at Vt+1, we feed the health records at Vt to a logistic regression model with multiple softmax classi ers. This can be viewed as a simpli ed model of Figure 1 without using RNNs and attention mechanism to learn latent states. This model only considers the e ect from the previous one time step, rather than long time history. Diagnosis Prediction Table 2 shows the accuracy of the proposed approaches in comparison with baselines on the two datasets. For each patient in the testing set, we predict the health conditions for the subsequent visits using his/her historical health records. For the SOF dataset, we predict the probability of BMD states of normal, osteopenia and osteoporosis for di erent measurements such as hip and femoral neck. For the BloodTest dataset, we predict the probability of each blood analyte falling into normal, low abnormal and high abnormal. The results are averaged over 5 random trials of 5-fold cross validation. Avg.# Correct represents the average number of correctly predicted claims of 5 random trials. Accuracy represents the ratio between correctly predicted claims and total number of claims to be predicted. For the two datasets, RNNl, RNNg and RNNc can clearly outperform plain RNN. Since the prediction of RNN mostly depends on recent visits, it may not memorize all the past information. Through attention-mechanism, RNNl, RNNg and RNNc can fully take all the previous visit information into consideration, assign di erent attention scores for past visits, and achieve better performance compared to RNN.
Visit Interpretation The attention mechanism can be used to understand the importance of historical visits to the current visit. As an example, here we analyze the concatenation-based attention mechanism on the SOF dataset. Figure 2 shows a case study for predicting the diagnoses in the sixth visit through the previous ve visits.
For chronic diseases, the last visit is often the most important since patients' health conditions change slowly. As in the gure, for the rst, fourth and fth patients, the importance of visit increases with time going on. However, this is not always the case due to the complexity of disease progression and impact from risk factors. Table 3 shows the variation of bone mineral density (BMD) diagnoses and attention scores of di erent visits of the second patient. In each visit, there are ve di erent BMD diagnoses, and the values in the table indicate the severity of bone density loss. Although V4 and V5 are closer to V6 in terms of time, V2 and V3 have the same condition as V6. Thus health records of V2 and V3 are more important to V6. We can see that the attention mechanism correctly assigns larger weights to V2 and V3. As for the BloodTest dataset, using attention mechanism to memorize all the past information is also important. An abnormal blood analyte can temporarily turn into normality via medicine, but it may fall back after some time. Therefore, interpreting visit importance through the attention mechanism can help to better monitor disease progression.
In diagnosis prediction, making decisions using very recent record is usually not enough, and it is important to lookup long term health information. To understand the relationship between the length of patient medical history and the prediction performance, we select 1,000 patients from the BloodTest dataset with more than seven visits. Table 4 shows the accuracy of RNNl in predicting the diagnoses from V2 to V7. We can see that with the number of visit increasing, the performance can often improve. We believe that it is due to the fact that RNN is able to learn better estimates of patient information as it memorizes longer health records. Acknowledgement This work was supported in part by NSF IIS-1218393 and IIS-1514204, and by SISTABENE POR project as PIHGIS POR project.