=Paper=
{{Paper
|id=Vol-2142/paper4
|storemode=property
|title=Predicting ICU mortality by supervised bidirectional LSTM networks
|pdfUrl=https://ceur-ws.org/Vol-2142/paper4.pdf
|volume=Vol-2142
|authors=Yao Zhu,Xiaoliang Fan,Jinzhun Wu,Xiao Liu,Jia Shi,Cheng Wang
|dblpUrl=https://dblp.org/rec/conf/ijcai/ZhuFWLSW18
}}
==Predicting ICU mortality by supervised bidirectional LSTM networks==
https://ceur-ws.org/Vol-2142/paper4.pdf
Predicting ICU Mortality by Supervised Bidirectional
LSTM Networks
Yao Zhu1,4, Xiaoliang Fan1,4,*, Jinzhun Wu2,4, Xiao Liu3, Jia Shi5, Cheng Wang1
1
Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University,
Xiamen, China
2
The First Affiliated Hospital, Xiamen University, Xiamen, China
3
School of Information Technology, Deakin University, Melbourne, Australia
4
Digital Fujian Institute of Healthcare & Biomedical Big Data Research, Xiamen
University, Xiamen, China
5
School of Information Science and Engineering, Lanzhou University, Lanzhou, China
csyzhu@stu.xmu.edu.cn, xfb_fxl@xm.gov.cn, 1923731201@qq.com, xiao.liu@deakin.edu.au,
shij2016@lzu.edu.cn, cwang@xmu.edu.cn
Abstract. Mortality prediction in the Intensive Care Unit (ICU) is considered as
one of critical steps for the treatment of patients in serious condition. It is a big
challenge to model time-series variables for mortality prediction in ICU, because
physiological variables such as heart rate and blood pressure are sampled with
inconsistent time frequencies. In addition, it is difficult to capture the timing
changes of clinical data and to interpret the prediction result of ICU mortality. To
deal with these challenges, in this paper, we propose a novel ICU mortality
prediction algorithm combining bidirectional LSTM (Long Short-Term Memory)
model with supervised learning. First, we preprocess 37 time-series variables
related to patients’ signs. Second, we construct the Bidirectional LSTM model
with supervision technique to accurately reflect significant changes in patients’
signs. Finally, we train and evaluate our model using a real-world dataset
containing 4,000 ICU patients. Experimental results show that our proposed
method can significantly outperform many baseline methods.
Keywords: Deep learning, ICU, mortality prediction, LSTM.
1 Introduction
Intensive Care Unit (ICU) is a rescue center for critically ill patients, which provides
the upmost service to reduce “dead in bed” events. For example, in the United States1,
5.7 million people are admitted to the ICU each year, and 2.3 million will require a
mechanical ventilator to help them breathe. Mortality prediction in ICU is considered
as one of critical steps for the treatment of patients with serious condition in ICU. The
major objectives of mortality prediction in ICU are: 1) assess and monitor the severity
of patients’ illness continuously based on their physical condition; and 2) determine
those patients of the highest risk to be treated with the upmost treatments, interventions
and resources. Therefore, accurate ICU mortality prediction can not only give clinicians
* Correspondence author.
1
Critical Care Statistics, http://www.sccm.org/Communications/Pages/CriticalCareStats.aspx
a better and earlier sense of which patients are likely to get worse, but also facilitates
the efficient allocation of hospital resources.
Many machine learning methods [4-7] have been employed to optimize the
prediction model with the availability of high-quality ICU datasets. In fact, the mortally
ill patients in ICU could generate a large amount of time-series data, such as heart rate,
blood pressure, temperature, Glasgow Coma Scale (GCS), and so on. Recently, deep
neural networks such as RNNs [8], LSTM [9], have been applied to process continuous
data sequences such as time-series variables in ICU.
Inspired by studies on ICU mortality prediction with deep neural networks (DNNs)
[10, 11], we aim to make use of the strong learning ability of DNNs to capture the
fluctuations in time-series variables which could adequately reflect changes in patients’
illness states. However, it is non-trivial to capture the fluctuations in time-series
variables in ICU, and we need to address the following two challenges:
—To model time-series variables in ICU, which is challenging because
physiological variables are usually sampled with inconsistent time patterns. For
instance, heart rate variables are collected every 20 minutes, while urine data is sampled
in a 5-hour interval.
—To incorporate the learning ability of deep neural networks into the interpretability
of the prediction result of ICU mortality, which is challenging as deep learning
approaches are known to have difficulties in inherently modeling the causation that
could provide straightforward decision supports for clinicians.
To address these challenges, in this paper we propose a novel ICU mortality
prediction method, named BiLSTM-ST, which combines a bidirectional LSTM (Long
Short-Term Memory) model with supervised learning technique. First, based on the
experience of the clinicians and the occurrence frequency of time-series variables, we
choose the available general descriptors and time-series variables as the input of each
patient. And the time-series variables are preprocessed via data simplification, data
completion and data normalization. Second, we construct the Bidirectional LSTM
model with supervision technique to accurately reflect significant changes in patients’
signs. Finally, we train and evaluate the proposed model on a public ICU dataset with
4,000 ICU patients. Experimental results show that our proposed Bidirectional LSTM
model with supervised learning can significantly outperform seven baseline methods.
2 Related Works
High-quality ICU datasets are gradually becoming open and available for research
purpose, which enables an increasing number of works on mortality prediction in ICUs.
To assess the mortality risk of patients in ICUs, there are several classic scoring
methods such as SAPS II [1] and APACHE IV [2]. SAPS stands for Simplified Acute
Physiology Score while APACHE stands for Acute Physiology And Chronic Health
Evaluation, and they are both applied within 24 hours of admission of a patient to the
ICU. However, it is well known that these scores are limited to a few fixed indicators,
which could make it very difficult to accurately reflect the dynamic evolution of
patients’ signs in ICU [3].
Machine learning technology has been widely utilized in statistical analysis. The
work in [4] achieves a high score in Mortality Prediction Challenge of Physionet by
using Support Vector Machine (SVM) classifiers. In addition, both general descriptors
and aggregated variables are used as features for the model. Later, the authors in [6]
come up with the CHISQ Classification Algorithm which is designed to address the
imbalance problem in the binary classification. However, the inconsistence in sampling
frequencies for time-series variables is known to cause great difficulties in statistical
analysis, as the time gaps could not be easily characterized to provide valuable decision
supports for clinicians.
More recently, deep neural networks are utilized to learn from the sequence data and
able to grasp the long-term dependence of time series data [15-16]. Since a large part
of the medical data is of time-series type, LSTM networks have been applied to learn
to classify diagnoses based on patient’s Electronic Health Record in pediatric intensive
care unit (PICU) [12]. A bidirectional LSTM model with attention mechanism to
predict mortality outcomes in ICUs is proposed showing competitive results on 2012
PhysioNet datasets [9]. The work in [13] combines LSTM and latent topic modeling
for mortality prediction so that it can not only predict but also interpret the predictive
results on mortality.
Furthermore, the work in [14] proposed a full-time supervision based bidirectional
RNN method, called FTS-BRNN, for QA tasks. Compared with our work, the major
difference is that we directly modify the LSTM neural network by adding supervision
technique, and we design the particular Loss function method to better apply to clinical
data.
3 Preliminary and Datasets
3.1 Basic Notations
In this work, we attempt to predict the mortality of patients in the ICU using deep
learning methods. The data we possess includes up to 37 time-series variables of each
patient recorded during the first 48 hours of their stay in the ICU, including such as
heart rate, blood pressure, weight, etc. We are taking advantage of these time-series
data to explore the patterns of the patient's physical condition changes, in order to
achieve the high accuracy of mortality prediction.
The mortality prediction in the ICU essentially evaluates the risk of the death based
on the specific patient's current physiological condition so that the doctor can take an
appropriate care. Given the clinical experience of doctors, patients’ physiological states
during the first 48 hours in the ICU could profoundly influence his or her clinical trends
within the next 30 days. Thus, as Equation (1) shows, we divided the predictions into
two categories: 1) death in ICU within 30 days, and 2) death in ICU exceeding 30 days
or survival in ICU. In other words, the first type of patients requires more monitoring
and care services than those of the second type.
1 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 𝑑𝑎𝑦𝑠 ≤ 30
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 = (1)
0 𝑠𝑢𝑟𝑣𝑖𝑣𝑎𝑙 days > 30
3.2 Datasets
The ICU data we employed are from the PhysioNet/Computing in Cardiology
Challenge 20122. There are 4,000 patients and records for each patient consist of two
parts: general descriptors (as shown in italic fonts in Table 1) and time-series data. First,
general descriptors contain patients’ basic information, including recordID, age, gender,
height, and ICU type. Second, time-series data is composed of 37 variables that reflect
the patient's physiological state. Each variable has an associated time-stamp indicating
the observed time of the variable. In addition, the dataset provides the number of days
the patient has stayed in ICUs and the fact of death or survival for patients. Thus, we
label patients who died in the ICU after more than 30 days or survival in ICU as 0,
while patients who died in ICU within 30 days as 1.
Table 1. Data samples of the PhysioNet dataset
Time Parameter Value Time Parameter Value
00:00 RecordID 132548 01:39 GCS 15
00:00 Age 68 01:39 NIDiasABP 80
00:00 Gender 0 01:39 NIMAP 112.7
00:00 Height 162.6 01:39 NISysABP 178
00:00 ICUType 3 02:14 MAP 144
00:09 GCS 15 19:09 Urine 18
00:09 NIDiasABP 79 20:09 Weight 87
00:09 NIMAP 112 25:09 Temp 36.7
00:09 NISysABP 178 47:09 HR 60
3.3 Observations
3.3.1 General Descriptors
We perform the basic statistics on general descriptors including patients’ basic
information. The preliminary observations are described as follows: 1) 12.88% of
patients survived for less than 30 days; 2) male patients are slightly more than female
patients; 3) there are four different types of ICU: coronary care unit (14.43%), cardiac
surgery recovery unit (21.85%), medical ICU (37.02%), and surgical ICU (26.70%);
and 4) the elderly (age over 60) take the highest proportion (61.98%).
3.3.2 Time-series Data
In Fig. 1, we further illustrate the 48-hour time-series data between two patients,
patient_a was dead in ICU within 30 days, while patient_b survived in the ICU. We
selected two high-frequency variables, HR and GCS, for comparison. heart rate (HR)
is the most basic sign of patients and GCS reflects the patient’s degree of coma.
2
PhysioNet/Computing in Cardiology Challenge 2012,
https://www.physionet.org/challenge/2012/
For example, the fluctuating range of patient_a's heart rate (HR) is relatively larger,
showing an obvious downward trend at the same time, while patient_b's heart rate has
fluctuated within the normal range. In addition, GCS of patient_a has a big drop to
below 10 at 18:00, whereas GCS of patient_b remains stable within 48 hours. The
dashed box marks the period during which both HR and GCS indicators of patient_a
change drastically.
(a) patient_a (dead in ICU within 30 days) (b) patient_b (survived in ICU)
Figure 1. The comparison between two patients.
4 The Proposed Method
4.1 Data Preprocessing
There are some general descriptors for each patient and 37 time-series variables in the
dataset. For the general descriptors, we chose the Age and ICUType as part of input for
the individual patient. It's worth mentioning that the ICUType is represented by the one-
hot code. For the 37 time-series variables, Cholesterol, TropI and TropT even have no
record in the dataset. Thus, we utilized 34 time-series variables for patients ultimately.
To sum up, the input of each patient consists of two general descriptors and thirty-four
time-series variables.
We recorded 34 time-series variables about the vital signs of the patients with a one-
hour time interval, that is to say, there are at most forty-eight records in 48 hours for
the individual variable. Specifically, three data preprocessing steps are conducted as
follow: 1) data simplification. For multiple records of the same variable in an hour,
we use the average value as there are very small changes; 2) data completion. For
missing records in an hour, there are two circumstances. For the first circumstance, if
there is no such data for a period of 48 hours, the variable is assigned with the average
value of the variable in the same type of ICU the patient belongs to. For the second
circumstance, if the record is only missing occasionally within 48 hours, the missing
values are replaced by the neighboring records; 3) data normalization. The mean and
standard deviation are used for normalization, so that the time-series information for
each patient is represented by a 34×48 matrix.
Furthermore, we labeled all patients as {0, 1}, according to Equation (1). Since the
information provided in the dataset includes the patient's survival days, we could screen
cases by the 30-day threshold. Thus, we obtained the label for each patient.
4.2 Algorithm
In order to realize this classification model for mortality prediction, we try to utilize
deep learning method to construct the model. Specifically, LSTM recurrent neural
network is employed in this paper as it is more suitable for time-series data. Therefore,
we start from the basic LSTM classification model and then propose the Bidirectional
LSTM model with continuous improvement. At the same time, supervised learning
technique is applied to these neural network structures to effectively improve the
accuracy of the prediction results. Finally, during the model training phase, we find that
the imbalance between positive and negative samples significantly affects the
classification result. To deal with this problem, we use the up-sampling method to
balance the training data and obtain a more accurate model.
4.2.1 Bidirectional LSTM
In recurrent neural networks, Long Short-Term Memory (LSTM) is a relatively more
efficient structure. Compared with the basic RNN, LSTM are able to make better use
of long-term dependence among the data. The following Equation (2) shows how the
LSTM structure works in a looping module. Specifically, s>? represents the state of
Cell at time t, which is related to a?> (the input at time t) and s>?@A (the state at time t-
1). For each patient, his or her physical condition at time t is affected by the previous
illness state and the specific vital signs of the subject at this time. Of course, in order to
deal with long-term dependencies, b?C means the forget gates’ reservations about the
state of the Cell at the last moment, but b?D decides to what extent the input information
is received. The absolute value of the disease index and its changes reflect the patient's
physiological status. Clearly, the abnormally fluctuating time-series data in the model
require more attention than the normal stable values. From this point of view, memory
and forgetting mechanisms are very important to the prediction model.
𝑠EF = 𝑏HF ∗ 𝑠EF@A + 𝑏KF ∗ 𝑔 𝑎EF (2)
For our scenario, each patient has 48 hours of time-series variables, which can be
captured by the LSTM network. Furthermore, the introduction of Bidirectional LSTM
can help us make better use of both past and future time-series data at each time step.
In fact, BiLSTM consists of forward and backward LSTMs, which allows for a more
complete understanding of the characteristics of all time-series data. In another word,
BiLSTM could avoid the blindness that propagation in unidirectional may cause. As a
result, BiLSTM is employed to process the input data to reflect the fluctuation of
patients’ illness condition as it is more capable of sensitively capturing the changes of
patient's physical signs than unidirectional LSTM models.
4.2.2 Supervision Technique
In a simplest Bidirectional LSTM recurrent neural network, there is only one output per
recurrent module of the hidden layer. When the gradients back propagate, the entire
module calculates the loss value only at the last step. In contrast, a recurrent module
can actually be unfolded into multiple modules, each part of which corresponds to a
certain moment of the input. Thus, every step can generate output respectively. As for
the calculation of the loss value, we apply a supervision technique. For each generated
output, as shown in Equation (3) and (4), we compute the loss with the label value, and
the ℒ (loss at last) is weighted by all N losses through all N steps. N is the total number
of steps, and 𝑔K are linear interpolation between 0 and 1. In this way, during the
training period of the model, the output of all steps will affect the final parameters so
that the model can detect subtle changes of the vital signs more sensitively.
K
𝑔K = 𝑖 = 1,2 … , 𝑁 (3)
N
N
ℒ= KTA 𝑔K × 𝐿𝑜𝑠𝑠K (4)
The supervision technique here has some advantages over the attention mechanism.
First, the attention mechanism only concerned with global dependencies, but in fact, it
is the local state that affects the outcome for the mortality prediction. And grabbing
characteristics of local state is exactly what the supervision technique is good at. Second,
the supervision technique shows the ordered incremental importance for all time steps,
which is consistent with the trend of patients’ states. However, the attention on time
steps is unordered according to the attention mechanism. Third, attention mechanism
tends to converge slower and need a larger amount of computation when dealing with
superabundant steps, but supervision technique improves the loss value in recurrent
module to make the model effectively handle the long sequence problem. For example,
in our experiment, model with attention mechanism converge slower than the
supervised LSTM networks when the number of total steps equals to 48.
Figure 2. BiLSTM model with supervised learning
Fig. 2 depicts the structure of the supervision technique applied to the Bidirectional
LSTM network. The figure shows the LSTM structure in both forward and backward
directions after the hidden layer is expanded, where Label is the expected output of the
module. In addition, dotted arrows represent inputs to the forward LSTM process, while
solid arrows represent inputs to the backward LSTM process. The small orange circle
is the cell of the forward LSTM network and the small yellow circle is the cell of the
backward LSTM network.
4.2.3 Samples Balance
In the process of scanning the datasets with 4,000 patients, we find that samples with
positive and negative labels are uneven. For example, there are 3,485 patients with the
label 0, while only 515 patients are labeled as 1. If trained with the unbalanced data
directly, it will be difficult for the model to capture features of patients with a label of
1. Thus, we augment the sample with the up-sampling method to reduce the gap
between the positive and negative samples. Finally, with a total of 4,200 patients for
training, there are 2,786 negative samples and 1,414 positive samples so that a
reasonable balance is achieved.
5 Experiments and Evaluations
5.1 Experimental Settings
We perform the experiment on 4,000 cases of the Physionet dataset, with 3,200 patients
as the training set and the remaining 800 as the test set. After data preprocessing, for
the training set, we up-sampled the records for balance, expanding the training set to
4,200. We apply the supervision technique to three different types of recurrent neural
networks including GRU, LSTM and Bidirectional LSTM, so as to produce a variety
of models for mortality prediction for comparison purpose. There are 128 cells per
hidden layer in our final model, and the softmax function is utilized to classify the
output into two categories. We make parameters optimization with 5-fold cross-
validation by taking 20% of the training set for parameter validation.
The algorithms are developed using Python 3.6.1, TensorFlow 1.2.1 and Keras 2.1.1.
Meanwhile, the experiment is conducted on 14 CPU cores (Intel(R) Xeon(R) CPU E5-
2683 v3 @ 2.00 GHz), with two GPUs (GeForce GTX TITAN X).
5.2 Evaluation Metrics
As the prediction of mortality is a binary classification problem, we choose Precision,
Recall, F1, and AUC to evaluate our model and compare with baselines. For a binary
classification problem, we usually take the class of interest as the positive class and the
others as negative class. Thus, there are four cases where the classifier is predicted
correctly or incorrectly on the dataset, as shown in Table 2.
TABLE 2. Confusion Matrix.
Outcome
1 0
Predicted
1 TP FP
0 FN TN
Precision equals TP divided by TP plus FP, and Recall equals TP divided by TP
plus FN. Besides, F1 is the harmonic mean of Precision and Recall, which is defined
as:
U A A
= + (5)
VA WXYEKZK[\ ]YE^__
In addition, AUC characterizes the ability of the classifier to rank positive samples
in front of negative samples. The greater the value of AUC is, the better the effect of
classification achieves.
5.3 Baselines
We compare our BiLSTM-ST model with the following six baselines:
● CNN [17]. Convolutional Neural Network, one of the most popular models of deep
learning but it lacks handling of time series data.
● LSTM [18]. Long short-term memory (LSTM) network is a special kind of RNN. It
is explicitly designed to avoid the long-term dependency problem and it is capable
of remembering information for long periods of time.
● BiLSTM [19]. Bidirectional LSTM network has both forward and backward LSTM
structure and tt has a better overall understanding of time series data.
● GRU [20]. Gated Recurrent Unit, is a variant of LSTM that maintains the effect of
LSTM while making the structure simpler.
● GRU-ST. Gated Recurrent Unit (GRU) network with supervision technique,
improve the prediction of the model to a certain extent.
● LSTM-ST. Long short-term memory (LSTM) network with supervision technique.
● BiLSTM (attention) [9]. Bidirectional LSTM network with attention mechanism.
The attention mechanism is to weight importance of each time step and it might
capture the human decision making implicitly.
5.4 Results Summary
As shown in Table 3, we employ four indicators Precision, Recall, F1 and AUC to
compare different models for mortality prediction.
Comparison with Baselines. We can see clearly from Table 3 that the Bidirectional
LSTM model with supervised learning technique (BiLSTM-ST) outperforms baselines
in general. Specifically, the prediction effect of recurrent neural networks (e.g., GRU,
LSTM, BiLSTM) is obviously better than that of convolution neural network (CNN)
because recurrent neural networks can better deal with the time series data. Furthermore,
the effectiveness of both supervised learning technique and attention mechanism
indicates that it is helpful to combine such techniques in recurrent neural networks for
ICU mortality prediction.
Comparison with BiLSTM (attention). BiLSTM-ST performs better than BiLSTM
(attention). The reason is that in contrast to the attention mechanism, the supervision
technique could supervise each step in reasonable order so that the model can learn
features and predict more effectively. Meanwhile, the attention mechanism lacks of
attention to local features during the time steps. In addition, the supervision technique
is more suitable for mortality prediction because it can better deal with the long
sequence problem like the time series data of 48 hours.
TABLE 3. Comparison among different methods.
Model Precision Recall F1 AUC
CNN 0.573 0.529 0.550 0.717
GRU 0.691 0.650 0.670 0.804
LSTM 0.638 0.733 0.682 0.809
BiLSTM 0.712 0.723 0.717 0.825
GRU-ST 0.789 0.670 0.724 0.822
LSTM-ST 0.826 0.685 0.749 0.832
BiLSTM (attention) 0.798 0.738 0.767 0.838
BiLSTM-ST 0.848 0.745 0.793 0.869
BiLSTM-ST(24-hour) 0.839 0.731 0.781 0.831
BiLSTM-ST(36-hour) 0.845 0.747 0.793 0.868
Figure 3. Comparison of BiLSTM-ST with data in varying length
Variants of BiLSTM-ST. In addition to utilizing all 48 hours of data (BiLSTM-ST),
we also selected the sequence data of the first 24-hour and first 36-hour after the patient
was admitted into the ICU, and the mortality prediction is carried out with the BiLSTM-
ST model. As shown in Fig. 3, BiLSTM-ST(24-hour) is relatively weak compared to
BiLSTM-ST(36-hour) and BiLSTM-ST(48-hour), which indicates that the data within
24 hours is not enough to accurately reflect the severity of the patient's illness condition
for the following 30 days. Nevertheless, the prediction accuracy of BiLSTM-ST(36-
hour) and BiLSTM-ST(48-hour) is almost identical. Specifically, BiLSTM-ST(36-hour)
even exceeds BiLSTM-ST(48-hour) on Recall. As a result, we could use 36-hour of
time-series data instead of 48-hour so as to give clinicians an earlier sense of which
patients will require critical targeted treatments.
6 Conclusion
In this paper, we proposed a novel mortality prediction algorithm in Intensive Care Unit
(ICU). We trained and tested our model with a real-world dataset of 4,000 patients to
accurately reflect significant changes in patients’ vital signs. Our experimental results
demonstrated that the bidirectional LSTM model with supervised learning mechanism
outperforms all other baseline methods. The significance of our novel prediction model
includes: 1) by effectively capturing fluctuations in time-series variables, it could give
clinicians an earlier sense of the patient’s mortality status; and 2) it could be used to
help hospitals to allocate ICU resources more efficiently.
In the future, our work can be extended, for example: 1) more sophisticated data
preprocessing steps will be conducted to capture the characteristics of the time-series
data; and 2) more extensive ICU datasets will be employed to evaluate and improve our
model.
Acknowledgments
The work is supported by grants from the Natural Science Foundation of China
(61300232); the China Postdoc Foundation (2015M580564); and Fundamental
Research Funds for the Central Universities (lzujbky-2016-br04).
References
1. Legall, J.R., Lemeshow, S., Saulnier, F.: A new simplified acute physiology score (SAPS II)
based on a European/North American multicenter study. JAMA 270, 2957--2963 (1993)
2. Zimmerman, J.E., Kramer, A.A., McNair, D.S., Malila, F.M.: Acute Physiology and Chronic
Health Evaluation (APACHE) IV: hospital mortality assessment for today’s critically ill
patients. Critical care medicine 34, 1297--1310 (2006)
3. Silva, I., Moody, G., Scott, D.J., Celi, L.A., Mark, R.G.: Predicting in-hospital mortality of
icu patients: The physionet/computing in cardiology challenge 2012. Computing in
Cardiology 39(20), 245--248 (2012)
4. Citi, L., Barbieri, R.: PhysioNet 2012 Challenge: Predicting mortality of ICU patients using
a cascaded SVM-GLM paradigm. Computing in Cardiology 25(1), 257--260 (2012)
5. Luo, Y., Xin, Y., Joshi, R., Celi, L., Szolovits, P.: Predicting ICU Mortality Risk by Grouping
Temporal Trends from a Multivariate Panel of Physiologic Measurements. In: 30th AAAI
Conference on Artificial Intelligence, pp. 42--50. AAAI press, Phoenix (2016)
6. Bhattacharya, S., Rajan, V., Shrivastava, H.: ICU Mortality Prediction: A Classification
Algorithm for Imbalanced Datasets. In: 31th AAAI Conference on Artificial Intelligence, pp.
1288--1294, AAAI press, San Francisco (2017)
7. Ghassemi, M., Naumann, T., Doshi-Velez, F., Brimmer, N., Joshi, R., Rumshisky, A.,
Szolovits, P.: Unfolding physiological state: Mortality modelling in intensive care units. In:
20th ACM SIGKDD international conference on Knowledge discovery and data mining, pp.
75--84, ACM press, New York (2014)
8. Carlin, C., Van Ho, L., Ledbetter, D., Aczon, M., Wetzel, R.: Predicting Individual
Physiologically Acceptable States for Discharge from a Pediatric Intensive Care Unit. arXiv,
preprint arXiv:1712.06214. (2017)
9. Nguyen, P., Tran, T., Venkatesh, S.: Deep learning to attend to risk in ICU. In: 2nd
International Workshop on Knowledge Discovery in Healthcare Data, pp. 25--29, Morgan
Kaufmann Press, Melbourne (2017)
10. Ding, J., Kang, X., Hu, X.H., Gudivada, V.: Building A Deep Learning Classifier for
Enhancing a Biomedical Big Data Service. In: 2017 IEEE International Conference on
Services Computing, pp. 140--147, IEEE Press, Honolulu (2017)
11. Avati, A., Jung, K., Harman, S., Downing, L., Ng, A., Shah, N.H.: Improving palliative care
with deep learning. In: 2017 IEEE International Conference on Bioinformatics and
Biomedicine, pp. 311--316, IEEE Press, Kansas City (2017)
12. Lipton, Z.C., Kale, D.C., Elkan, C., Wetzel, R.: Learning to diagnose with LSTM recurrent
neural networks. arXiv, preprint arXiv:1511.03677. (2015)
13. Jo, Y., Lee, L., Palaskar, S.: Combining LSTM and Latent Topic Modeling for Mortality
Prediction. arXiv, preprint arXiv:1709.02842. (2017)
14. Xu, D., Li, W.J.: Full-Time Supervision Based Bidirectional RNN for Factoid Question
Answering. arXiv, preprint arXiv: 1606.05854. (2016)
15. Dai, A.M., Le, Q.V.: Semi-supervised sequence learning. In: 29th International Conference
on Neural Information Processing Systems, pp. 3079--3087, MIT Press, Montreal (2015)
16. Ng, Y.H., Hausknecht, M., Vijayanarasimhan, S., Vinyals, O., Monga, R., Toderici, G.:
Beyond short snippets: deep networks for video classification. Computer Vision and Pattern
Recognition, 16, 4694--4702 (2015)
17. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional
neural networks. In: 26th International Conference on Neural Information Processing
Systems, pp. 1097--1105, MIT Press, Lake Tahoe (2012)
18. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation, 9, 1735--
1780 (1997)
19. Graves, A., Fernández, S., Schmidhuber, J.: Bidirectional LSTM Networks for Improved
Phoneme Classification and Recognition. In: 15th International Conference on Artificial
Neural Networks, pp. 799--804, Springer Press, Warsaw (2005)
20. Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recurrent neural
networks on sequence modeling. arXiv, preprint arXiv:1412.3555. (2014)