The determination of the optimal time of the patients in weaning trial process from mechanical ventilation, between patients capable of maintaining spontaneous breathing and patients that fail to maintain spontaneous breathing, is a very important task in intensive care unit. Wavelet Transform (WT) and Neural Networks (NN) techniques were applied in order to develop a classifier for the study of patients on weaning trial process. The respiratory pattern of each patient was characterized through different time series. Genetic Algorithms (GA) and Forward Selection were used as feature selection techniques. A classification performance of 77.00± 0.06% of well classified patients, was obtained using a NN and GA combination, with only 6 variables of the 14 initials.
PAPER ICIMTH 2014 (GRECIA)
Neural Network Classification of Patients on Weaning Trials Using Respiratory Patterns
1. Patients Classification on Weaning
Trials Using Neural Networks and
Wavelet Transform
Carlos Arizmendia,1
, Juan Viviescas a
Hernando Gonzáleza
, Beatriz Giraldob
a
Control & Mecatrónica Research Group
Universidad Autónoma de Bucaramanga, Bucaramanga, Colombia
b
Dept.of ESAII, Universitat Politècnica de Catalunya (UPC),
Institut de Bioenginyeria de Catalunya (IBEC) and
CIBER de Bioingeniería, Biomateriales y Nanomedicina, Barcelona, Spain.
Abstract. The determination of the optimal time of the patients in weaning trial process from mechanical
ventilation, between patients capable of maintaining spontaneous breathing and patients that fail to maintain
spontaneous breathing, is a very important task in intensive care unit. Wavelet Transform (WT) and Neural
Networks (NN) techniques were applied in order to develop a classifier for the study of patients on weaning
trial process. The respiratory pattern of each patient was characterized through different time series. Genetic
Algorithms (GA) and Forward Selection were used as feature selection techniques. A classification
performance of 77.00± 0.06% of well classified patients, was obtained using a NN and GA combination,
with only 6 variables of the 14 initials.
Keywords. Weaning process, Neural Networks, Discrete Wavelet Transform,
Genetic Algorithms.
Introduction
A classifier is a system that assigns categories or classes to the data presented. The
classifier does not have to know the structure of the data, to recognize and classify data
or objects belonging to a given category, being necessary to learn from the experience,
the essence of that category [1]. The NN is a useful parametric technique for modeling
data density. The NN model consists of a network with an input layer with as many
nodes as inputs have, a hidden layer with variable number of nodes that depend on the
characteristics of the problem, and an output layer with as many nodes as possible
outputs [2].
Weaning trials process of patients in intensive care units, is the process of
discontinuing mechanical ventilation and a complex clinical procedure. Despite
advances in mechanical ventilation and respiratory support, the science of determining
if the patient is ready for extubation is still very imprecise [3]. When mechanical
ventilation is discontinued, up to 25% of patients have respiratory distress severe
enough to require reinstitution of ventilator support. A failed weaning trial is
1
Corresponding Author: Carlos Arizmendi, is professor of the Mecatronics Faculty in the
Universidad Autónoma de Bucaramanga, Avenida 42 No 48-11, Bucaramanga, Colombia, Email:
carizmendi@unab.edu.co.
2. discomforting for the patient, may induce cardiopulmonary distress and carries a higher
mortality rate.
The respiratory model that describes the mechanical function of the pulmonary
system can be characterized by the following time series: inspiratory time (TI),
expiratory time (TE), breath duration (TTot), tidal volume (VT), inspiratory fraction
(TI / TTot), average flow inspired (VT / TI), and the frequency tidal volume ratio (f /
VT) [3]. The aim of the present work is the analysis with neural networks and feature
selections techniques of the respiratory pattern variability in patients during weaning
trials, in order to find differences between patients capable of maintaining spontaneous
breathing and patients that fail to maintain spontaneous breathing.
1. Data Base
Respiratory flow signals were measured in 66 patients on weaning trials from
mechanical ventilation (WEANDB database). These patients were recorded in the
Departments of Intensive Care Medicine at Santa Creu i Sant Pau Hospital, Barcelona,
Spain, and Getafe Hospital, Getafe, Spain, according to the protocols approved by the
local ethic committees. The patients were submitted under T-tube test, disconnected
from the ventilator and maintained spontaneous breathing through an endotracheal tube
during 30 min. If the patients maintained the spontaneous breathing with normality
they were extubated, if not, they were reconnected. The patients were classified into
two groups: Successful Group (GS), 33 patients with successful weaning trials; and
Failed Group (GF), 33 patients that failed to maintain spontaneous breathing.
2. Methodology
Several studies for the analysis of respiratory signals, in patients during the process
extubation have been successful performed, demonstrating the relevance of the
breathing cycle duration (TTot) and the rapid shallow index (F / VT) series, as relevant
variable extubation [4,5]. In [6] illustrates the Tobin index (F / VT) as one of the most
used in the clinic for good predictive value and good reproducibility, without requiring
patient cooperation or additional instrumentation. For the respiratory TTot and F / VT
time series, for each patient and condition, the statistics: mean, Standard Deviation
(SD), mode, Intercualtil Range (IR), Skewness (SK), variance and kurtosis were
computed. Table 1 illustrates the computed statistics for each time series.
Table 1. Number of inputs and their computed statistics variables for each time series.
Input Variable Input Variable
1 Mean TTot 8 Mean F / VT
2 SD TTot 9 SD F / VT
3 Mode TTot 10 Mode F / VT
4 IR TTot 11 IR F / VT
5 SK TTot 12 SK F / VT
6 Variance TTot 13 Variance F / VT
7 Kurtosis TTot 14 Kurtosis F / VT
3. 2.1 Neural Networks Classification and Forward Feature Selection
A Neural Network classifier was performed with an exhaustive search of the
optimal number of neurons in the hidden layer, varying two neurons for each iteration.
In order to reduce the dimensionality of the system, Forward Feature Selection method
was used [7], in both cases, 5-fold Cross Validation method, 60% of the patterns were
used for the training stage, 20% for validation and 20% for the test. The classification
rates were computed as the mean ± standard deviation of 150 runs of the output system.
To avoid the overfitting, the NN was trained with the Bayesian Regularization and
Early Stoping methods, Early Stopping method was implemented using the functions
of training: Variable Learning Rate (traingdx), Resilient Backpropagation (trainrp),
Levenberg-Marquardt Backpropagation (trainlm) (Matlab-Trademark®).
Genetic Algorithm is an adaptive method based on the evolution theory to solve
search and optimization problems [2]. The configuration of the algorithm is a binary
type population where each bit represents a variable, where is made an exhaustive
search for the variables that results with the best classification results. Table 2
illustrates the combinations of the variables with the best classification results, after
implementing NN with GA and forward Selection methods.
Table2. Best classification results using GA and forward selection method.
2.2 Wavelet Transform and neural network classification
The WT performs the decomposition of a signal generated from two basic functions
through combinations of scaling and translations of these [8]. In this case the two base
functions are called Wavelet and scaling function, respectively. It was implemented the
Biorthogonal family in the time series TTot and F/VT, using the entropy criterion in
order to determine the maximum number of decomposition levels, obtaining the detail
and approximation coefficients. For each decomposition level the U _Mann Whitney
test was performed, and the decomposition levels with the less p_value was taken. Once
the best p_value were determined for the decomposition levels, the index equation (1)
was implemented in order to determine the wavelet decomposition coefficients which
allows maximum separation between classes with a smaller number of decomposition
levels. The Biorthogonal 3.3 wavelet family obtained the best index result (0.1534).
Classification and
feature selection
method
Training Algorithm Variables Accuracy
Neural
Networks and
Forward Selection
Traingdx 14-8-9-6-11-1 75 ± 0.3389%
Trainrp 13-6-5 73.3 ± 0.230%
Trainlm 8-4-11-14-3 73.3 ± 0.303%
Trainrb 8-13 73.3 ± 0.180%
Neural
Networks and
Genetic Algorithm
Traingdx 8-10-13 77.00 ± 0.06%
Trainrp 2-10-13 71.66± 0.07%
Trainlm 3-5-8-9-12 76.67± 0.12%
Trainrb 1-2-6-7-12-13 77.00± 0.06%
4. 𝐼𝑛𝑑𝑒𝑥 =
∑ 𝑃𝑣𝑎𝑙𝑢𝑒𝑠 𝑇𝑇𝑜𝑡 + 𝑃𝑣𝑎𝑙𝑢𝑒𝑠 𝑓/𝑉𝑇
𝑇𝑜𝑡𝑎𝑙 𝑜𝑓 𝐷𝑒𝑐𝑜𝑚𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛 𝑙𝑒𝑣𝑒𝑙𝑠 𝑏𝑦 𝐹𝑎𝑚𝑖𝑙𝑦
(1)
Once the optimal Wavelet was determined, the WT was applied in the 66 patients,
computing the 7 statistics described in Table 1, to the detail coefficients at level 3 of
the TTot series and approximation coefficients of level 2 of the F / VT series. Finally
was implemented the NN classification, with the same parameters setting
aforementioned, for the dimensionality reduction the GA and forward selection also
were implemented. Table 3 illustrates the best classification results for each feature
selection method.
Table 3. Two best classification results for the GA and feed-forward feature selection methods
3. Conclusions
NN were implemented in order to analyze the variability of the breathing pattern in
patients assisted with MV, with the aim of finding relevant variables to determine the
optimal time of extubation, obtaining encouraging results with GA feature selection
method.
Is expected to perform a more thorough study of the most relevant variables that
correspond to the best classification rate, in order to understand and explain the
physiological importance of the most variables in the process of weaning trials.
References
[1] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford, University Press, 1995.
[2] A. Celekil, et, al. Use of artificial neural networks and genetic algorithms for prediction of sort option of
an azo-metal complex dye into lentil straw, Elsevier, U.S.A, 2013.
[3] J.M. Boles, J. Bion, A. Connors, M. Herridge, B. Marsh, C. Melote, R. Pearl, H. Silverman, M. Stanchina,
A. Vieillard-Baron, and T. Welte. TASKFORCE—weaning from mechanical ventilation, European
Respiratory Journal. France, 2007.
[4] L. S. Correa, B. F. Giraldo . Multi-parameter Analysis of ECG and Respiratory Flow Signals to Identify
Success of Patients on Weaning Trials, Conference proceedings IEEE Eng Med Biol Soc.U.S.A, 2010.
[5] J.C.Hsu, Y.F.Chen, H.Hung. Lin, C.Hsiang. Li and X. Jiang, Construction of Prediction Module for
Successful Ventilator Weaning, New Trends in Applied Artificial Intelligence, Kyoto, Japan, 2007.
[6] M.J. Tobin, Advances in mechanical ventilation, New. England. Journal of Medicine, 2001.
[7] E. Romero, J.M. Sopena, G. Navarrete, R. Alquezar, Feature selection forcing overtraining may help to
improve performance, proc. Int. Joint Conference on Neural Networks, Portland, Oregon, 2003.
[8] S. Mallat, A Wavelet Tour of Signal Processing, Academic Press , USA. 1999.
Feature selection method Training
algorithm
Best variables selected Accuracy
Forward selection with
Early Stopping
Traingdx 7-8-9-6-11-1 75% ± 0.033%
Forward selection with
Bayesian Regularization
Trainrb1 8-13 73.33% ± 0.18%
Genetic algorithm with
Bayesian Regularization Trainrb1 1-2-6-7-5-13 77% ± 0.070%
GA with Early Stopping Traingdx 8- 10-13 77% ± 0.063%