Multi-channel ConvNet Approach to Predict the Risk of in-Hospital

Mortality for ICU Patients

Fabien Viton, Mahmoud Elbattah, Jean-Luc Guérin and Gilles Dequen

Laboratoire MIS, Université de Picardie Jules Verne, Amiens, France

Keywords: ConvNet, Deep Learning, Multivariate Time Series, in-Hospital Mortality.

Abstract: The healthcare arena has been undergoing impressive transformations thanks to advances in the capacity to

capture, store, process, and learn from data. This paper re-visits the problem of predicting the risk of in-

hospital mortality based on Time Series (TS) records emanating from ICU monitoring devices. The problem

basically represents an application of multi-variate TS classification. Our approach is based on utilizing

multiple channels of Convolutional Neural Networks (ConvNets) in parallel. The key idea is to disaggregate

multi-variate TS into separate channels, where a ConvNet is used to extract features from each univariate TS

individually. Subsequently, the features extracted are concatenated altogether into a single vector that can be

fed into a standard MLP classification module. The approach was experimented using a dataset extracted from

the MIMIC-III database, which included about 13K ICU-related records. Our experimental results show a

promising accuracy of classification that is competitive to the state-of-the-art.

1 INTRODUCTION

Healthcare services are delivered in data-rich

environments where a wealth of data is created at

multiple levels of operational and medical records. In

view of that, data analytics is increasingly becoming

a key enabling thrust to leverage of such massive data

amounts. An important part of the analytics

capabilities is based on Time Series (TS) data.

Applications of TS analytics are essential in a wide

diversity of domains, especially healthcare where the

use of temporal data is ubiquitous.

However, the multi-dimensionality of TS data

brings up further challenges regarding the extraction

and selection of features. In this respect, Deep

Learning (LeCun, Bengio, and Hinton 2015) could

present as an appropriate approach. Deep Learning

allows for learning hierarchical feature

representations from raw data automatically. Deep

architectures of Convolutional Neural Networks

(ConvNets) (LeCun et al. 1989; LeCun et al. 1998)

have been successfully implemented in complex

tasks. Examples include Computer Vision and

Machine Translation (e.g. Krizhevsky, Sutskever,

and Hinton 2012; Gehring et al. 2017).

Correspondingly, Deep Learning has also been

considered as an attractive path for tackling tough TS

problems. Particularly, in the case of multiple

variates, complex relationships, and large amounts of

data. There has been a growing interest over the past

few years in this regard (e.g. Karim et al. 2019), as an

alternative approach to avoid the need for developing

conventional hand-crafted features.

In this context, this study approaches a multi-

variate TS problem. The task is to predict the risk of

in-hospital mortality among ICU patients. The

problem under consideration represents a typical

application of multi-variate TS classification. Using a

multi-channel architecture, our approach utilizes

multiple ConvNets to extract features from each

univariate TS individually. The experiments used a

dataset extracted from the MIMIC-III database,

which provides a freely accessible repository of ICU

records (Johnson et al. 2016).

The study attempts to make contributions in two

aspects. On one hand, the study is conceived to

contribute to the ongoing efforts towards availing of

Deep Learning methods for multi-variate TS

problems. While from a practical standpoint, the

performance of channel-wise ConvNet architectures

is explored with respect to the problem of predicting

in-hospital mortality.

Viton, F., Elbattah, M., Guérin, J. and Dequen, G.

Multi-channel ConvNet Approach to Predict the Risk of in-Hospital Mortality for ICU Patients.

DOI: 10.5220/0009891900980102

In Proceedings of the 1st International Conference on Deep Learning Theory and Applications (DeLTA 2020), pages 98-102

ISBN: 978-989-758-441-1

 2020 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reserved

2 RELATED WORK

The problem of TS classification was notably

identified as one of the key challenges in Data Mining

research (Yang and Wu, 2006). Distance-based

methods such as Dynamic Time Warping (DTW)

have been long recognized as the most performing

technique in this respect (Berndt and Clifford, 1994).

The development of feature-based similarity

measures was also explored (Fulcher, 2018).

However, the intensive process of pre-processing and

feature extraction was generally considered as a

limiting factor. While the analysis and forecasting of

TS have been dominated by regression-based

modelling methods such as Auto Regressive

Integrated Moving Average (ARIMA).

With recent advances, Machine Learning (ML)

has increasingly come into prominence, especially for

complex multivariate TS problems. Various

implementations of ConvNet and Recurrent Neural

Network (RNN) architectures were successfully

applied for TS classification tasks. For instance, a

ConvNet-based framework was proposed by (Cui,

Chen, and Chen, 2016). ConvNets were exploited to

automatically extract features through a sequence of

convolution and pooling operations. The extracted

features could represent the internal structure of the

input TS. Furthermore, (Wang, Yan, and Oates, 2017)

demonstrated that ConvNet models could provide

better performance over traditional DTW methods.

Further efforts concentrated on employing Deep

Learning potentials for complex TS problems that

involve large-scale datasets and multiple variables.

For example, a ConvNet-based feature extractor was

developed for multivariate TS classification (Zheng

et al. 2016). (Purushotham et al. 2017) provided an

exhaustive evaluation of Deep Learning against other

ML models based on the MIMIC dataset. They

demonstrated that Deep learning consistently

outperformed other approaches, especially in the case

of large multi-variate TS data.

Other studies experimented Long Short-Term

Memory (LSTM) models. For example, (Siami-

Namini, Tavakoli, and Namin, 2018) reported that

LSTM outperformed traditional algorithms including

ARIMA. Another study explored the use of bi-

directional LSTMs, which provided a better

performance as well (Siami-Namini, Tavakoli, and

Namin, 2019). A detailed presentation of such efforts

would go beyond the scope of this study, but (Fawaz

et al. 2019) provides a comprehensive review of the

state-of-the-art Deep Learning implementations for

TS classification.

3 DATA DESCRIPTION

The study used a dataset extracted from the MIMIC-

III database (Johnson et al. 2016). The MIMIC

database provides a rich repository of ICU admissions

to the Beth Israel Deaconess Medical Center in

Boston between 2001 and 2012. It is considered to be

one of the largest databases of its kind publicly

available. It has been utilized in plentiful studies (e.g.

Desautels et al. 2016; Komorowski et al. 2018).

The dataset comprised more than 13K patient

records related to a variety of ICU admissions

including cardiac, medical, surgical, trauma, and

others. The TS variables described the patient status

over the 48-hour timespan after admission (e.g. heart

rate, blood pressure, temperature, etc.). Specifically,

the dataset included 17 temporal measurements,

which represented the typical readings used during

ICU monitoring (Silva et al. 2012). As such, a

(17x48) matrix could describe the development of

each ICU patient. The dataset variables are listed in

Table1 below.

For each case, a binary label corresponded to the

outcome (i.e. in-hospital mortality). The mortality

rate among patients was about 9%. To establish a

benchmark for comparison, the dataset was prepared

following the set of procedures provided by

(Harutyunyan et al. 2019). Under normal conditions,

some variables could suffer from missing values (e.g.

blood or urine samples). To fill in missing values, we

applied values from the previous time point. All

variables were normalized with zero mean and unit

standard deviation.

Table 1: Dataset variables.

Variables

Heart Rate

Respiratory Rate

Capillary Refill Rate

Systolic Blood Pressure

Diastolic Blood Pressure

Mean Blood Pressure

Fraction Inspired Oxygen (FiO

)

Oxygen Saturation (SaO

)

Temperature

Glucose

Glascow Coma Scale Eye Opening

Glascow Coma Scale Motor Response

Glascow Coma Scale Verbal Response

Glascow Coma Scale Total

Height

Weight

Multi-channel ConvNet Approach to Predict the Risk of in-Hospital Mortality for ICU Patients

Figure 1: Approach overview.

4 APPROACH OVERVIEW

The approach is based on the multi-channel ConvNet

architecture proposed by (Zheng et al. 2014). The

architecture includes a combination of unsupervised

and supervised learning over two stages as follows.

4.1 Feature Extraction

Learning representations is a fundamental question

for ML. In this respect, ConvNets introduced a potent

mechanism to automatically learn complex patterns

from raw data. ConvNets were cleverly designed to

deal with the spatial or temporal variability

underlying data. Likewise, ConvNets were utilized in

our case to learn features from the TS data, and hence

eliminating the need to develop hand-crafted features.

Initially, the multi-variate TS is separated into

univariate channels. For each channel, hierarchical

features are extracted through operations of

convolution and pooling. The convolutional layers

extract temporal patterns by applying 1-D filters over

TS sequences. The convolutions are followed by a

ReLU activation layer, which introduces the non-

linearity into the learning process. Subsequently, a

global average pooling is applied, which computes

the mean value of filters across the time dimension.

The operations conducted on each TS channel are

described in the equations below.

= Wi ⁎ TS

+ b (1)

= ReLU(y

) (2)

feat

= GlobalAveragePooling(h

) (3)

Where y

is the output filter and ⁎ is the convolution

operation, and feat

is the output feature map.

4.2 MLP Classifier

The output of each ConvNet channel is a feature map

that can be regarded as a compressed representation

of the input TS. The feature maps are subsequently

concatenated to jointly form a single feature map.

Eventually, the aggregated feature map is fed to a

conventional MLP classifier, which would be trained

to perform the classification task.

Supervised training is basically performed at this

stage. The filter coefficients output from ConvNet

channels are updated simultaneously during the MLP

learning process. The classifier may include a single

layer or multiple hidden layers, which would perform

further non-linear transformation of the feature map.

Figure 1 sketches the approach architecture.

5 EMPIRICAL EXPERIMENTS

As alluded earlier, the goal was to predict the risk of

in-hospital mortality based on the initial 48h interval

of ICU monitoring. The MIMIC dataset was

randomly divided into 75% train and 25% test

portions. The hyperparameters (e.g. filter size) of

ConvNet channels were decided empirically.

Specifically, we could achieve the highest accuracy

with filter size= 8, and number of filters=8. Given 17

TS variables, the model included 136 convolutional

filters (i.e. 17*8), and the output layer was composed

of 136 weights. The model was trained using Adam

optimizer (Kingma and Ba, 2014).

DeLTA 2020 - 1st International Conference on Deep Learning Theory and Applications

100

Figure 2: Model loss in training and validation sets.

Figure 3: ROC curve.

Figure 4: Precision-Recall curve.

Various structures of MLP were experimented for

training the model. It turned out that the best

performance could be achieved using 3 fully

connected layers. Specifically, the hidden layers

consisted of 64, 32 and 16 neurons, respectively.

Further, the dropout technique was applied to help

reduce the model over-fitting (Srivastava et al, 2014).

Figure 2 plots the model loss in training and

validation over 10 epochs with 20% of the dataset

used for validation.

Figure 3 examines the classification accuracy

based on the Receiver Operating Characteristics

(ROC) curve. The ROC curve plots the relationship

between the true positive rate and the false positive

rate across a full range of possible thresholds. The

model could achieve a very good accuracy (AUC-

ROC≈0.85). Figure 4 plots the Precision-Recall

curve, which is particularly important in the case of

imbalanced datasets (AUC-PR≈0.60).

Overall, the model could largely provide

comparable performance to the literature.

Furthermore, we could achieve a relatively higher

AUC-PR compared to the work conducted by

(Harutyunyan et al. 2019), which did not include the

multi-channel ConvNet approach. The experiments

were implemented using the Keras library (Chollet,

2015) with the TensorFlow (Abadi et al. 2016)

backend. The model implementation is shared on the

GitHub repository (Elbattah, 2020).

6 CONCLUSIONS

The multi-channel ConvNet approach could yield

promising results applied to the problem of predicting

in-hospital mortality. Despite using a relatively

simple ConvNet architecture, the accuracy achieved

is competitive to the state-of-the-art. It is conceived

that further improvements could be realized by

applying more sophisticated architectures.

Our future work aims to explore further

interesting prospects. On one hand, we are concerned

with the explainability of predictions. Analyzing and

visualizing the output of ConvNet channels may be

employed to bring insights into the most influential

Multi-channel ConvNet Approach to Predict the Risk of in-Hospital Mortality for ICU Patients

101

variables on the predicted outcome. On the other

hand, with the mounting successful applications of

Transfer Learning, we endeavor to explore that path

as well. Transfer Learning methods might be a key to

improve the performance by fine-tuning pre-trained

models rather than training from scratch.

REFERENCES

Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean,

J., ... & Kudlur, M. (2016). Tensorflow: A system for

large-scale machine learning. In Proceedings of the

12th USENIX Symposium on Operating Systems Design

and Implementation (pp. 265-283).

Berndt, D. J., & Clifford, J. (1994). Using dynamic time

warping to find patterns in time series. In Proceedings

of KDD workshop (Vol. 10, No. 16, pp. 359-370).

Chollet, F., (2015). Keras. https://github.com/fchollet/keras

Cui, Z., Chen, W., & Chen, Y. (2016). Multi-scale

convolutional neural networks for time series

classification. arXiv preprint arXiv:1603.06995.

Desautels, T., Calvert, J., Hoffman, J., Jay, M., Kerem, Y.,

Shieh, L., ... & Wales, D. J. (2016). Prediction of sepsis

in the intensive care unit with minimal electronic health

record data: a machine learning approach. JMIR

Medical Informatics, 4(3), e28.

Elbattah, M., 2020. https://github.com/Mahmoud-

Elbattah/DeLTA2020.

Fawaz, H. I., Forestier, G., Weber, J., Idoumghar, L., &

Muller, P. A. (2019). Deep learning for time series

classification: a review. Data Mining and Knowledge

Discovery, 33(4), 917-963.

Fulcher, B. D. (2018). Feature-based time-series analysis.

In Feature Engineering for Machine Learning and

Data Analytics (pp. 87-116). CRC Press.

Gehring, J., Auli, M., Grangier, D., Yarats, D., & Dauphin,

Y. N. (2017). Convolutional sequence to sequence

learning. In Proceedings of the 34th International

Conference on Machine Learning (ICML)-Volume 70

(pp. 1243-1252).

Harutyunyan, H., Khachatrian, H., Kale, D. C., Ver Steeg,

G., & Galstyan, A. (2019). Multitask learning and

benchmarking with clinical time series data. Scientific

Data, 6(1), 1-18.

Karim, F., Majumdar, S., Darabi, H., & Harford, S. (2019).

Multivariate LSTM-FCNs for time series classification.

Neural Networks, 116, 237-245.

Kingma, D. P., and Ba, J. (2014). Adam: A method for

stochastic optimization. arXiv preprint arXiv:

1412.6980.

Komorowski, M., Celi, L. A., Badawi, O., Gordon, A. C.,

& Faisal, A. A. (2018). The artificial intelligence

clinician learns optimal treatment strategies for sepsis

in intensive care. Nature Medicine, 24(11), 1716-1720.

Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012).

ImageNet classification with deep convolutional neural

networks. In Proceedings of Advances in Neural

Information Processing Systems (pp. 1097-1105).

Kumar, M., & Singh, S. (2009). Engineering Mathematics-

III. KRISHNA Prakashan Media, Meerut, India.

LeCun, Y., Boser, B. E., Denker, J. S., Henderson, D.,

Howard, R. E., Hubbard, W. E., and Jackel, L. D.

(1989). Handwritten digit recognition with a back-

propagation network. In Proceedings of Advances in

Neural Information Processing Systems (NIPS) (pp.

396-404).

LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998).

Gradient-based learning applied to document

recognition. In Proceedings of the IEEE, 86(11), 2278-

2324.

LeCun, Y.; Bengio, Y.; and Hinton, G. (2015). Deep

learning. Nature 521(7553):436–444.

Purushotham, S., Meng, C., Che, Z., & Liu, Y. (2017).

Benchmark of deep learning models on large healthcare

MIMIC datasets. arXiv preprint arXiv:1710.08531.

Johnson, A. E., Pollard, T. J., Shen, L., Li-wei, H. L., Feng,

M., Ghassemi, M., ... & Mark, R. G. (2016). MIMIC-

III, a freely accessible critical care database. Scientific

Data, 3, 160035.

Siami-Namini, S., Tavakoli, N., & Namin, A. S. (2018,

December). A comparison of ARIMA and LSTM in

forecasting time series. In 2018 17th IEEE International

Conference on Machine Learning and Applications

(ICMLA) (pp. 1394-1401). IEEE.

Siami-Namini, S., Tavakoli, N., & Namin, A. S. (2019). A

Comparative Analysis of Forecasting Financial Time

Series Using ARIMA, LSTM, and BiLSTM. arXiv

preprint arXiv:1911.09512.

Silva, I., Moody, G., Scott, D. J., Celi, L. A., & Mark, R. G.

(2012). Predicting in-hospital mortality of ICU patients:

The PhysioNet/Computing in Cardiology Challenge

2012. In Proceedings of the 2012 Computing in

Cardiology (pp. 245-248). IEEE.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., &

Salakhutdinov, R. (2014). Dropout: a simple way to

prevent neural networks from overfitting. The Journal

of Machine Learning Research, 15(1), 1929-1958.

Wang, Z., Yan, W., & Oates, T. (2017). Time series

classification from scratch with deep neural networks:

A strong baseline. In Proceedings of the 2017

International Joint Conference on Neural Networks

(IJCNN) (pp. 1578-1585). IEEE.

Yang, Q., & Wu, X. (2006). 10 Challenging problems in

data mining research. International Journal of

Information Technology & Decision Making, 5(04),

597-604.

Zheng, Y., Liu, Q., Chen, E., Ge, Y., & Zhao, J. L. (2014,

June). Time series classification using multi-channels

deep convolutional neural networks. In International

Conference on Web-Age Information Management (pp.

298-310). Springer, Cham.

Zheng, Y., Liu, Q., Chen, E., Ge, Y., & Zhao, J. L. (2016).

Exploiting multi-channels deep convolutional neural

networks for multivariate time series classification.

Frontiers of Computer Science, 10(1), 96-112.

DeLTA 2020 - 1st International Conference on Deep Learning Theory and Applications

102