Content uploaded by D.K. Sreekantha Karanam Desai
Author content
All content in this area was uploaded by D.K. Sreekantha Karanam Desai on Dec 15, 2020
Content may be subject to copyright.
Predicting difficulties in Mask Ventilation using
Machine Learning techniques
D.K.Sreekantha
Department of CSE
NMAM Institute of Technology
Nitte, India
sreekantha@nitte.edu.in
Sripada G. Mehandale
Department of Anesthesiology & Critical Care
K.S.Hegde Medical Academy
Mangalore, India
dr.sgmehandale@gmail.com
Roline Stapny Saldanha
Department of CSE
NMAM Institute of Technology
Nitte, India
rolinestapny@gmail.com
Rodrigues Rhea Carmel Glen
Department of CSE
NMAM Institute of Technology
Nitte, India
rhea.carmel@gmail.com
Jotsna Gowda Krishnappa
Department of CSE
NMAM Institute of Technology
Nitte, India
gowdajotsna@gmail.com
Prajna M K
Department of CSE
NMAM Institute of Technology
Nitte, India
prajnamk05.pmkmk@gmail.com
Abstract—The oxygen is to be supplied constantly through
the mask to a patient in the Operation Theater (OT) when
performing an operation. Any interruption in oxygen or air
supply to the patient may lead to severe bodily damage or even
death of the patient. The mask ventilation can be categorized into
3 levels namely easy, difficult and impossible mask ventilation.
An expert anesthesiologist can accurately predict the difficulties
in mask ventilation. Currently, expert anesthesiologists use their
experience to manually analyze the patient features and predict
the difficulties in mask ventilation. So authors have implemented
a software solution by applying machine learning algorithms
to predict the difficulties in mask ventilation. Authors have
identified twelve physical parameters of the patient that are
significant in predicting the difficulties in mask ventilation.
The representative patient data collected from hospital and the
knowledge of experienced anesthesiologist is used to synthesize
the data set. The data set is mined using machine learning
algorithms namely Logistic Regression, Random Forest, Support
Vector Machine, Naive Bayes and k-Nearest Neighbors. Logistic
Regression algorithm is proved to be better in predicting the
difficulties in Mask Ventilation.
Keywords—Mask ventilation, Machine Learning, Data Syn-
thesis, Rulebase, Risk parameters
I.
INTRODUCTION
Mask ventilation is a technique to provide constant supply
of oxygen to the patient in operation theater when performing
an operation. A suitable mask is placed on patients mouth
and nose with a tight seal. A constant flow of oxygen is
supplied to the mask using a bag that is attached to the mask.
The air should constantly and easily flow into the patient
and the patients chest should rise and fall with each cycle
of ventilation.
According to American Society of Anesthesiologist (ASA),
difficult mask ventilation may be due to improper mask seal,
gas leakage beyond limits or opposition to the ingress or egress
of gas [1]. The difficult mask ventilation is a serious problem
and can cause major complications like hypoxic brain injury
or myocardial ischaemia, injuries to the nose, mouth, eyes
and even the death of the patient [2]. The difficulty in mask
ventilation may also injure patients teeth, lips, soft palate,
uvula and nerves due to the excess force on the mask[2].
Ensuring constant supply of oxygen to the patient is the
primary duty of anesthesiologist. Anesthesiologist usually
performs the physical examination of the patient to anticipate
any difficulty in mask ventilation. Expert anesthesiologist can
predict the difficulties in mask ventilation easily through his
past experiences. The difficult mask ventilation occurs only
in less than 5% of the patients. Employing a full time expert
anesthesiologist will be very expensive for the hospital
management. Therefore, a system which can predict the
difficulties in mask ventilation is of vital importance. This
system can help novice anesthesiologist to accurately predict
the difficulties in mask ventilation. These systems can
prevent major airway injuries in the operating patient.
So authors are motivated to design an expert system
software to help novice doctors to predict the difficulties in
mask ventilation. In this work, authors have designed a system
which classifies mask ventilation into 3 types, namely easy
mask ventilation, difficult mask ventilation and impossible
mask ventilation. Authors have identified twelve physical
parameters of patients that cause difficulties in mask
ventilation through Literature survey and discussions with the
expert anaesthesiologist. Authors have used the patient data
from hospital, the rulebase, and the knowledge of expert
anesthesiologist to simulate the dataset for experimentation.
Five classification techniques namely Logistic Regression,
Random Forest, Support Vector Machine (SVM), K Nearest
neighbor and Naive Bayes classifier are applied on the dataset
and their performance is compared.
II.
LITERATURE SURVEY
Authors have carried out an exhaustive review of literature
to understand the past work in predicting the difficulties in
mask ventilation (DMV). The mask ventilation is successful
only when a proper amount of gas flows through patients
lungs using a face mask. No proper seal, higher resistance
to air flow in lung path, a reduced compliance of the lungs
leading to increased distal pressure causes DMV [3]. Various
studies have been conducted to identify the parameters of the
patient that cause DMV. In study conducted on 1399 patients,
out of which 124 patients had DMV. Comparison between
patients with and without difficulties in mask ventilation
was done to identify risk factors. seven risk factors were
identified using univariate and multivariate analysis [4]. The
risk parameters were Age (>=47 years), neck circumference
(>=40 cm), body mass index (BMI) (>=35kg/m
2
), history of
difficulty in intubation, presence of hair on face, short neck
and obstructive type of sleep apnea. However, no patient had
more than 4 of these risk parameters.
110 mask ventilation attempts were analyzed in a hospital
in Taiwan [5]. The statistical analysis was carried out to assess
the risk parameters. Age(>=65 years), lack of teeth, sunken
cheeks, double chin and thick short neck were found to be
associated with DMV. Absence of teeth, sunken cheeks and a
double chin were found to independently identify DMV. DMV
was reported to occur in 5% of the patients and Age >=55
years, BMI >= 26 kg/m
2
, absence of teeth, presence of beard,
snoring history were independent risk parameters[6]. Any two
of these risk parameters best indicates DMV. Failure to bring
lower
incisors in front of upper incisors was also found to
predict DMV [7].
Mallampathi grade of III or IV and severely limited
mandibular protrusion were found to cause DMV [8]. DMV
were found to be more common in infants and children as they
develop hypoxemia more commonly as compared to adults [9].
Stiff lungs was found to be useful in predicting DMV [10].
Univariate analysis and multivariate analysis was conducted on
500 adult patients[11]. Body mass Index, age, Mallampati
score grading, macroglossia, absence of teeth, presence of
beard,
waist-hip ratio, short neck, double chin and previous
history of snoring were statistically significant in predicting
DMV using univariate analysis. BMI>26kg/m
2
, limited
mandibular protrusion, double chin, small neck, high
Mallampati score, age > 55 years and weight were
independent risk factors for DMV in multivariate analysis.
Impossible mask ventilation is rare and it can be found once in
690 cases. Neck Radiation, male gender, sleep apnea,
Mallampati score of III or IV and presence of beard are found
be the risk factors for Impossible mask ventilation [12].
Currently expert anesthesiologist manually evaluate the
various patient parameters and carry out some manual test
such mallampati score grading, Upper lip bite test etc and
anticipate the difficulties in mask ventilation. Sometimes the
patient physical parameters are scored and the values of the
scores were used to categorize Mask Ventilation. In this paper,
a novel approach of using Machine learning algorithms to
predict the difficulties in mask ventilation is done.
Fig. 1. Flow of steps involved to build mask ventilation difficulty prediction
system.
III.
PROPOSED METHODOLOGY
Fig. 3 outlines the steps that are followed in this paper to
build a software system to predict difficulties in Mask
ventilation. Initially, the existing literature is perused and the
mask ventilation difficulties are studied. Discussions with the
expert anesthesiologists are done to carefully understand the
problem. Two methods are employed to prepare the dataset
for experimentation. In the first method, the data of patients
with difficulties in mask ventilation is collected from the
hospital. Patients with difficulties in Mask ventilation are rare,
hence authors received very small amount of data from the
hospital. The collected data ware carefully analyzed and used
to simulate a big data set. In the second method, accurate rule
base was defined and implemented to synthesize data. Finally
Various Machine Learning are used for dataset classification
and the best algorithm is selected. The outcomes are verified
by expert anesthesiologist and the model is fine tuned.
A.
Identification of Risk parameters
Risk parameters associated with difficulties in Mask
Ventilation were identified by conducting extensive literature
review and discussions with expert anesthesiologists.
TABLE I
DMV
RISK PARAMETERS AND THEIR VALUES
Sl.No Parameter Min
Value
Max
Value
Typical
Value
DMV
1
A
g
e 50 110 51-110 51- 110
2
Mallampati
score
1
4
1 or 2 3 or 4
3
Mandibular
protrusion
1
3
3
2,3
4
BMI 15 60 <26 >= 26
5
Neck cir-
cumference
20 54 <40 >42
6
Gende
r
M or F M or F M or F
M
7
Neck Radia-
tio
n
Y or N Y or N
N
Y
8
Bear
d
P or A P or A
A
P
9
Stiff Lun
g
s P or A P or A
A
P
10 History of
snoring
P or A P or A
A
P
11 Teeth P or A P or A
P
A
12 Macro
g
lossia P or A P or A
A
P
Twelve risk parameters are found to effectively predict the
difficulties in mask ventilation. They are Age, Gender, Body
Mass Index (BMI), Beard, Mallampati score, Neck radiation,
Upper lip byte test (ULBT), Absence of Teeth, Neck
circumference, Macroglossia, History of snoring and stiff
lungs.
B.
Description of Risk Parameters
1)
Mallampati score: This score is given based on the
visual assessment of the distance from the ceiling of the mouth
to the tongue base. The score is noted by asking the patient
to protrude his tongue as much as possible while in the sitting
posture. There are 4 possible mallampati scores. The possible
value of Mallampati score are class I, II, III and IV .
2)
Neck Radiation: Neck Radiation refers to any radiation
given to the patients. In Cancer patients neck radiation may
be seen. The possible values are true and false.
3)
ULBT (upper lip bite test): This test is done to check
the mandibular movement of the patient. The output of test
is either grade 1, grade 2 or grade 3.Grade 1 - The lower
incisors can bite the upper lip above vermilion line. Grade 2 -
The lower incisors can bite the upper lip below vermilion line
Grade 3 - The lower incisors cannot bite the upper lip.
4)
Macroglossia: It is the medical term for an unusual
enlargement of tongue when compared to other parts in the
mouth.
5)
Stiff lungs: The walls of the air sacs in the patients lungs
may get inflamed. Overtime, the walls become scarred making
the lungs stiff. This condition is called as Stiff Lungs.
The Table I shows the values of the twelve parameters along
with the values that causes Difficulties in mask ventilation.
This table was created under the guidance of expert
anesthesiologist.
IV.
PREPARING THE DATA
The Difficulties in Mask Ventilation (DMV) is rare, but
when it is present causes serious damage to the patient.
Fig. 2. Rules for identifying Difficulties in Mask ventilation.
TABLE II
S
UBSET OF THE
D
ATASET
Age BMI Brd MS NR ULBT Tth NC Ma HS SL Tgt
9
20
0101 1
20
0
000
94 15
0302 1
36
0
001
108 25
0101 0
20
0
001
54 24
0101 1
32
1
011
106 27
1101 1
50
1
112
104 15
0301 0
38
0
001
66 40
1201 1
32
0
001
64 41
0103 1
32
0
011
102 26
1202 1
51
1
112
51 41
0101 1
28
0
101
There are no standard datasets for DMV which could be used
in this work. The authors have procured a small amount of
data from hospital. The risk parameters associated with
difficulties in mask ventilation are studied. The study showed
that different combinations of risk parameters caused the
difficulties in mask ventilation. The proper combination of the
risk parameters that result in difficulties in mask ventilation
were studied from the expert anesthesiologists of KS Medical
Academy, Mangalore and written in the form of rules.The Fig.
2 shows few rules from the rulebase which was created under
the guidance of expert anesthesiologist. Around Fifty rules are
present in the rule base. Authors have implemented the rules
using python program and synthesized dataset for
experimentation. The dataset contains six lakh patient records
and each record has twelve features (risk parameters). Table II
shows few patient records present in the dataset.
V.
PREPROCESSING
The dataset has 6,04,121 records, out of this 5,02,448 are
representing difficult mask ventilation, 16,984 are representing
Fig. 3. Plot showing distribution of dataset.
Fig. 4. Dataset distribution based on Age.
easy mask ventilation and 84,688 are representing impossible
mask ventilation. The Fig. 3 describes the dataset used in
this work. The dataset is visualized by plotting graphs for
every parameter. The graphs in Fig. 4, 5 and 6 are plotted
using Jupyter notebook tools. The red colour portion in the
graph corresponds to the records of difficult mask ventilation
in the dataset. The blue colour portion in the graph
corresponds to the data of impossible mask ventilation. The
pink portion in the graph corresponds to the records of easy
mask ventilation. The x - axis values indicates the possible
values of the parameters and y - axis represents the frequency
of parameter values present in the dataset.
The Fig. 4 shows the plot of the dataset based on Age. The
red portion is present above the age 50 years which indicates
that people older than 50 years have higher chance of difficult
mask ventilation.
Fig. 5. Dataset distribution based on Beard.
Fig. 6. Dataset distribution based on Neck Circumference.
The plot of the dataset based on patient beard parameter in
shown in Fig. 5. The easy mask ventilation is only present in
patients without beard. The Fig. 6 is the plot of the dataset
based on Neck circumference. A higher neck circumference
causes difficulties in mask ventilation. In patients with neck
circumference greater than 49cm, mask ventilation becomes
impossible to carry out. Therefore, very high Neck
circumference is an independent predictor for impossible
mask ventilation. Similar plots are plotted for other
parameters. The plots indicate that high values of
mallampati grade, ULBT, BMI, Age, No teeth, presence of
macroglossia, previous snoring history and presence of stiff
lungs have contributed to difficulties in mask ventilation.
VI.
CLASSIFICATION OF DATASET USIN G MACHINE
LEARNING ALGORITHMS
The dataset is classified using five Machine learning
algorithms namely Logistic Regression, Random Forest,
Support vector Machine, Naive bayes classifier and K Nearest
Neighbor. Since the twelve parameters present in the dataset
were carefully selected based on literature survey and expert
anesthesiologist guidance, all twelve parameters are
considered as features for classification.
The twelve parameters Age, Gender, BMI, Beard,
Mallampati score, Neck radiation, Upper lip byte test ULBT,
Presence of teeth, Neck circumference, Macroglossia, History
of snoring and stiff lungs are considered as input features for
machine learning Algorithm. The entire dataset is divided into
training dataset and testing dataset. Randomly selected 70% of
the dataset is considered as training dataset and remaining
30% is considered as testing data.
The experiments are carried out using Python programming
in Jupyter Notebook Application. The machine learning
algorithms are invoked from the scikit-learn library for
classifying the dataset. Authors have carried out the
experiments in two ways, Simple Validation and 5-Fold-
Cross validation. In simple validation, the model is trained
using 70% of the dataset and remaining 30% of the data is
used to test the model. This method is less efficient because
many variations are not captured effectively which creates an
under fitted, over fitted or biased model. The 5-fold-cross
Validation is more stable method. This method captures
variations across the dataset and better approximation is
obtained since it performs training and testing on every part of
the dataset. In the 5 fold cross validation, 4 folds are used to
train the model and left out fold is used to test the model.
VII.
RESULTS DISCUSSIONS
The results of machine learning models are shown in Table
III.
Logistic Regression model has given an accuracy of
98.3% during simple validation method and an accuracy of
98.1% during cross validation method. Random forest model
and support vector Machine showed similar simple validation
accuracy whereas their cross validation score slightly differed.
Naive Bayes model showed the minimum accuracy among all
Fig. 7. Screenshot of Classification Algorithm accuracy
TABLE III
R
ESULTS FROM
M
ACHINE
L
EARNING
A
LGORITHMS
Fig. 8. Model output for 25 test cases
the models. K nearest neighbour model has shown an accuracy
of 98%. However, K nearest neighbour model took higher
time to execute and hence is computationally expensive. Fig.
7 displays the output of the machine learning models during
experimentation.
The Fig. 8 shows the output of the models for 25 input
test
values. The title of the graphs displays the name of the
algorithm used. The green point indicates the actual output
that is expected. The red point indicates the output of model.
If the expected output and obtained output overlaps the red
point is hidden under the green point. In the figures, a red
dot
indicates that obtained values does not match with the
expected value. The algorithm with many red dots has high
error. Logistic Regression classifier has shown the least error.
In the past no attempts were made to predict the difficulties in
Mask ventilation using Machine learning algorithms. Hence,
the current work stands as a baseline work for future studies.
The results of the experiments on the dataset indicate that the
identified twelve risk parameters are accurate in predicting
difficulties in mask ventilation. Logistic Regression is the
efficient algorithm for predicting the difficulties in Mask
Ventilation.
VIII. SUMMARY AND CONCLUSIONS
Authors have studied the Difficult Mask Ventilation (DMV)
problem from several research papers. Researchers also visited
the hospital several times to gather information and to interact
with expert anesthesiologist. Twelve risk parameters leading
to difficulties in mask ventilation were identified. An accurate
dataset was prepared and processed using the conditions that
cause difficulties in mask ventilation. The dataset was analyzed
using graphs plotted for different parameters. The machine
learning model was built using Jupyter notebook tool. Five
different classification techniques namely Logistic Regression
Sl.
No
Classifier SimpleValidation
Accuracy (%)
5-Fold Cross
Validation
Accuracy
(%)
1
Lo
g
isticRe
g
ression 98.8 98.1
2
RandomForest 96.0 96.2
3
Su
pp
ortVectorMachine 96.0 94.1
4
Naive Ba
y
es 88.9 89.1
5
K Nearest 98.4 98.5
classifier, Random Forest classifier, Support Vector Machine,
K nearest neighbour classifier and Naive Bayes classifier are
used for developing the system for predicting Difficulties
in Mask Ventilation. Logistic Regression has reported with
performance of 98% accuracy. This work can be further
extended for other difficulties in airway management.
REFERENCES
[1] Apfelbaum, J.L., Hagberg, C.A., Caplan, R.A., Blitt, C.D., Connis, R.T.,
Nickinovich, D.G., Benumof, J.L., Berry, F.A., Bode, R.H., Cheney, F.W.
and Guidry, O.F., “Practice guidelines for management of the difficult
airwayan updated report by the American Society of Anesthesiologists
task force on management of the difficult airway”, Anesthesiology: The
Journal of the American Society of Anesthesiologists 118.2, 2013, pp.
251–270.
[2] Jonathan Holland and Will Donaldson. “Difficult Mask Ventilation”,
AnesthesiaTutorial of the Week 321, 2015, pp. 1–7
[3] Saddawi-Konefka, D., Hung, S.L., Kacmarek, R.M. and Jiang, Y,
“Optimizing mask ventilation: literature review and development of a
conceptual framework”. Respiratory care 60.12, 2015, pp. 1834–1840.
[4] Cattano D, Killoran PV, Cai C, Katsiampoura AD, Corso RM, Hagberg
CA, “Difficult mask ventilation in general surgical population:
observation of risk factors and predictors”. F1000Research 3, 2014.
[5] Lee, S.Y., Chien, D.K., Huang, M.Y., Huang, C.H., Shih, S.C., Wu,
K.M., Ku, H.C. and Chang, W.H. “Patient-specific factors associated with
difficult mask ventilation in the emergency department”, International
Journal of Gerontology 11.4 , 2017, pp. 263–266.
[6] Langeron, O., Masso, E., Huraux, C., Guggiari, M., Bianchi, A., Coriat,
P. and Riou, B, “Prediction of difficult mask ventilation”.
Anesthesiology: The Journal of the American Society of
Anesthesiologists 92.5, 2000, pp. 1229–1236.
[7] Singh, G., Gautam, P.L., Kaul, T.K. and Luthra, N., “Prediction of
difficult mask ventilation”. Canadian Journal of Anesthesia/Journal
canadien d’anesthe´sie 53, 2006, pp. 26409–26409.
[8] Kheterpal, S., Han, R., Tremper, K.K., Shanks, A., Tait, A.R., O’Reilly,
M. and Ludwig, “Incidence and predictors of difficult and impossible
mask ventilation”, Anesthesiology: The Journal of the American Society
of Anesthesiologists 105.5, 2006, pp. 885–891.
[9] Mohammad El-Orbany and Harvey J Woehlck. “Difficult mask
ventilation”, Anesthesia & analgesia 109.6, 2009, pp. 1870–1880.
[10] Evelyn Wong and Yih-Yng Ng. “The difficult airway in the emergency
department”, International journal of emergency medicine 1.2, 2008, p.
107.
[11] P Gautam, TK Gaul, and N Luthra. “Prediction of difficult mask
ventilation”. In: European Journal of Anaesthesiology (EJA) 22.8, 2005,
pp. 638–640.
[12] Kheterpal, S., Martin, L., Shanks, A.M. and Tremper, K.K, “Prediction
and outcomes of impossible mask ventilationa review of 50,000
anesthetics”. Anesthesiology: The Journal of the American Society of
Anesthesiologists 110.4, 2009, pp. 891–897.