=Paper=
{{Paper
|id=Vol-1149/bd2014_cai
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-1149/bd2014_cai.pdf
|volume=Vol-1149
}}
==None==
https://ceur-ws.org/Vol-1149/bd2014_cai.pdf
ABSTRACTS : scientific
Modeling of time series health data
using Dynamic Bayesian Networks: An
application to predictions of patient
outcomes after multiple surgeries
Xiongcai Caia, Oscar Perez-Conchaa, Fernando Martin-Sanchezb, Blanca Gallegoa
a
Centre for Health Informatics, Australian Institute of Health Innovation, University of New South Wales, NSW 2052
b
Health and Biomedical Informatics Centre, University of Melbourne, Victoria 3010
SUMMARY
Objective: To develop dynamic predictive models for real-time outcome predictions of hospitalised patients.
Dr Xiongcai (Peter) Cai Design: Dynamic Bayesian networks (DBNs) were built to model patient outcomes that dynamically depend
on patient’s clinical profiles, temporal patterns of ward transfers and surgery data. These models were
Research Fellow
applied to predict remaining days of hospitalisation (RDH) for patients undergoing multiple surgeries and their
Centre for Health Informatics,
performance compared against a static model based on Bayesian networks (BNs).
The University of New South Wales
Datasets: Hospital data from a Sydney metropolitan hospital.
x.cai@unsw.edu.au Results: The basic model uses static information at time of prediction. The DBN model uses static and
temporal information extracted from a series of surgeries; DBNs show a significant improvement in patient
outcome predictions with respect to the static model.
Conclusion: Time series health data can be dynamically modelled by DBNs to improve predictions of outcomes
Dr Xiongcai (Peter) Cai is a researcher with a general
for patients undergoing multiple surgeries.
background in AI with expertise in the fields of machine
learning, data mining, social network analysis and health
INTRODUCTION
informatics. He has been spending the past decade
researching and developing models to better understand Healthcare systems are under increasing pressure to identify strategies to improve the current patterns of
behaviours and patterns that relate real world human care. Although unstructured data analytics have been widely reported in big data, the importance of time
activities. series has not been fully explored yet. Real-time prediction of patient outcomes could benefit greatly from
big data in health and time series analysis. In particular, prediction of RDH1 is an important indicator to
assess healthcare delivery and hospital management. Unexpectedly long length of stay may negatively impact
patients and hospitals in a variety of ways, such as higher costs and increased exposure to adverse events.
Current methods do not allow real-time automated stratification of risk. Rapidly identifying those patients at
highest risk of extended RDH has a great potential to improve the quality of care, reduce avoidable harm and
costs. DBNs2 are specially suitable to tackle this problem, since they are probabilistic graphical models that
allow temporal order, which can better capture the dynamical nature of the healthcare delivery processes:
prognosis, treatment selection, surgery and recovery.
In this paper, we aim to develop a DBNs-based prediction model to investigate the roles of time series data for
the prediction of RDH for patients undergoing multiple surgeries.
DESCRIPTION
Medical records of patients who underwent consecutive surgeries at a Sydney metropolitan hospital between
2008-2012 were analysed. There are 5733 records in the dataset. Each admission is characterised by a set of
attributes, which include: patient’s characteristics (such as age), surgery information (main procedure, number
of procedures, length of surgery) and ward type. These attributes, together with days already in hospital,
constitute the inputs to the static BN. The outcome to be predicted is RDH. BNs3,4 are static probabilistic
graphical models that consists of nodes and arcs forming a directed acyclic graph, where nodes present
domain variables (predictors), whereas arcs represent conditional probabilistic relationships among variables.
DBNs2 represent the evolution of a system over time, allowing a fixed structure network to present variables at
multiple time points (slices), containing temporal dependencies between slices.
We learned both structures and model parameters: 1) In order to construct the static BNs (Figure 1), we
learned intra-slice structures and reinforced them with domain knowledge. These BNs will be the baselines
to compare with the DBNs; at the point of prediction, the BN will contain the information of the current
22 #bd14 | big data conference
surgery, whereas the DBN will contain the temporal information of the consecutive surgeries. 2) We then fixed the intra-slice structure and learned the inter-slice
dependencies. We computed the conditional probabilistic dependencies between any two-time slices by creating successive two-slice sequences and by learning
the DBNs (Figure 2). For testing, we unrolled the DBN to the length of test sequences (Figure 3) and input test sequences as evidences to infer RDH of patients.
In our experiments, we performed 5-fold cross validation in both the learned BNs and DBNs. RDH is discretised into 12 bins. Compared to BNs, DBNs achieved a
significant improvement in the prediction of RDH. Specifically, DBNs achieved 72.4% prediction accuracy, whereas BNs 25.8%. This implies a 180% improvement,
which might be due to the ability of DBNs to dynamically update the model using temporal information from time series data. It requires about 30 minutes to
construct the DBNs with 64-bit Windows 7 Enterprise, 2 cores of Intel® i7-3840QM CPU @ 2.80GHz and 8GB RAM.
Figure 1. Static BN. Figure 2. DBN. Figure 3. Unrolled DBN.
CONCLUSION
We developed predictive DBNs models for predicting RDH of surgical patients using patient’s trajectories and time series surgical information. Our experiments
showed that DBNs significantly outperform BNs in RDH prediction after multiple surgeries. In the future, we plan to further apply DBNs in large-scale big data
frameworks for health informatics.
Funding: This work was funded by National Health and Medical Research Council (NHMRC) Project grant 1045548 and Program Grant 568612.
Ethics approval: Ethics approval was obtained from the NSW Population and Health Services Research Ethics Committee and the NSW Human Research Ethics Committee. The corresponding author was responsible for the data analysis after the extraction. Its contents are the
responsibility of the authors and do not reflect the views of NHMRC.
REFERENCES
1. V. Liu, P. Kipnis, M. K. Gould, and G. J. Escobar, “Length of stay predictions: improvements through the use of automated laboratory and comorbidity variables,” Medical care, vol. 48, no. 8, pp. 739-744, 2010.
2. M. KP., Dynamic Bayesian networks: representation inference and learning, UC Berkeley, 2002.
3. L. PJ, v. d. G. LC, and A.-H. A., “Bayesian networks in biomedicine and health-care,” Artificial Intelligence in Medicine, vol. 30, pp. 201-214, 2004.
4. J. Pearl, Causality: Models, Reasoning, and Inference: Cambridge University Press, 2000.
3 - 4 april 2014 | melbourne 23