|Year : 2023 | Volume
| Issue : 1 | Page : 4-9
Predictive machine learning algorithms in anticipating problems with airway management
Muthapillai Senthilnathan, Pankaj Kundra
Department of Anaesthesiology and Critical Care, Jawaharlal Institute of Postgraduate Medical Education and Research, Puducherry, India
|Date of Submission||14-Feb-2023|
|Date of Acceptance||20-Feb-2023|
|Date of Web Publication||20-Apr-2023|
Dr. Pankaj Kundra
Department of Anaesthesiology and Critical Care, Jawaharlal Institute of Postgraduate Medical Education and Research, Second Floor, Institute Block, Puducherry - 605 006
Source of Support: None, Conflict of Interest: None
Machine learning is artificial intelligence (AI) which can predict the output variable with the fed input features. This allows computers to learn from experience without being programmed. The outcome variable in machine learning algorithm may be continuous variable or categorical variable. Supervised machine learning is commonly applied artificial intelligence (AI) in medical field. Decision tree, gradient boost machine (GBM) learning, extreme GBM (XGBM), Support vector machine, K nearest neighbour and multi-layer perceptron are few machine learning algorithms which are being utilised to address the classification and regression problems. Though the incidence of difficult intubation (DI) is rare, occurrence of such event in an unanticipated situation can result in development of arrhythmias due to desaturation and cardiac arrest if not intervened on time. It is preferred to choose the physical parameters that can predict the difficult airway more accurately in clinical scenario and train the algorithm rather than including all the non-specific parameters. Body mass index (BMI) [>30 kg.m-2: anticipated difficult mask ventilation (DMV), direct laryngoscopy (DL) and DI], inter-insicor distance (IID) (<2 cm: anticipated DL), modified Mallampati (MMP) (Grade 1 and 2: Ease of intubation; Grade 3 and 4: anticipated DI), temporomandibular distance (TMD) (<6.5 cm - anticipated DI), restriction of neck extension (if present: anticipated DL and DI), receded mandible (if present: anticipated DL and DI), and poor submandibular space compliance (if present: anticipated DL and DI) parameters which are used to predict DA by clinical assessment, can be used to feed to train the machine learning algorithm. Despite using these sophisticated tools, extubation may fail and patients require reintubation in ICU. It is very challenging to assess the lung compliance in spontaneously breathing patients as compliance will be overestimated due to generation of negative pressure. Cause for which patient has been placed on mechanical ventilation is resolved/resolving, BMI (>30 kg.m-2), intact sensorium (absence of delirium), absence of consolidation, absence of copious secretions, oxygenation status (PaO2/FiO2: >250), ventilation status (paCO2: 30-45 mmHg), measure of work of breathing (respiratory rate, rapid shallow breathing index), heart rate and blood pressure during spontaneous breathing trial (SBT) and diaphragmatic thickness fraction can be used as input features to predict the success of extubation in critically ill patients. With widespread utility of applications in medical fraternity, applications for prediction of difficult airway (or for weaning success) can be programmed which can be accessed by the clinicians to predict DA, thereby all the preparations for managing DA may be done to prevent adverse consequences of unanticipated difficult airway.
Keywords: Difficult airway, extubation failure, machine learning, predictive models
|How to cite this article:|
Senthilnathan M, Kundra P. Predictive machine learning algorithms in anticipating problems with airway management. Airway 2023;6:4-9
|How to cite this URL:|
Senthilnathan M, Kundra P. Predictive machine learning algorithms in anticipating problems with airway management. Airway [serial online] 2023 [cited 2023 Jun 7];6:4-9. Available from: https://www.arwy.org/text.asp?2023/6/1/4/374375
| Introduction|| |
Machine learning is artificial intelligence (AI) that can predict the output variable with the fed input features. This technique allows computers to learn from experience without being programmed. Widely used machine learning techniques include supervised, unsupervised, semi-supervised and reinforcement learning., Amongst these methods, the medical field commonly uses supervised and unsupervised methods of AI. As mentioned above, an appropriate type of algorithm should be chosen depending on the available type/types and category/categories of the input data. The significant difference between statistical analysis and AI is that the ultimate target will be to understand the relationships among the studied variables in statistical analysis, and the goal is to accurately predict the outcome variable in AI. The outcome variable in the machine learning algorithm may be continuous (obesity body mass index [BMI] and diabetes mellitus [blood sugar level]; the relationship between continuous variables) or categorical (relationship between smoking [yes or no] and development of peripheral vascular disease [yes or no]).
| Supervised Machine Learning|| |
In supervised machine learning, there will be a range of various input variables with which the machine predicts the known outcome. Regression algorithms are a type of supervised learning. There is a gross difference in the clinical application of regression in statistics and AI. In statistics, linear regression can refer to continuous outcome variables, and logistic regression can refer to binary outcome data. The toxic dose of a drug amongst diverse population or complications associated with high blood sugar levels can be predicted with machine learning. The dataset can be divided into two parts: the training data part and testing data domain [Figure 1]. The randomly chosen input variables and their outcome data will be assigned to each domain. The machine will be learning with these data and is called training the algorithm. Once training is over, the machine should be able to predict the outcome accurately with the new input data. In a model with overfitting, many data will be classified appropriately, but it has less sensitivity (high specificity). There will be importance to individual data rather than the problem in the model with the underfitting of the data. Most data are classified appropriately with few errors in a model with an appropriate fit.
|Figure 1: Performance of supervised machine learning technique in medical field|
Click here to view
| Unsupervised Machine Learning|| |
There is no predefined outcome in unsupervised machine learning. It is explorative where the machine makes clusters within the data (that have similar characteristics), and the variables that do not have similarities with others will be excluded from the clusters. As this method decreases the number of variables within the dataset, it is also known as the dimension reduction technique. This method of AI can be used when the input variables are too many and to compress them to fewer input characteristics to predict the outcome accurately.
| Machine Learning Models that are Useful in the Fields of Anaesthesia and Intensive Care Unit|| |
The decision tree is a supervised machine learning technique for analysing classification-related and regression problems. The outcome variable is categorical in the classification tree, and the outcome variable is continuous in the regression tree. The decision tree is known to be affected by overfitting. Random forest (RF) can be used to overcome this overfitting as it works by the Ensemble AI model. Boosting AI is a subtype of ensemble AI used to build robust classifiers from weak classifiers. Gradient boosting machine (GBM) learning is used when the decision tree handles weak classifiers. Extreme GBM (XGBM) is a gradient boosting method wherein the decision trees are added sequentially, which leads to learning from the error of the previous tree, thereby improving the accuracy of prediction. Light GBM (LGBM) handles lots of datasets with little complexity, but it cannot be used when the number of data is lesser.
Support vector machine (SVM) may deal with classification and regression problems. If input features are related to various classes, then a decision plane is needed to separate them., SVM requires a larger dataset to get trained.
K nearest neighbour (KNN) algorithm deals with classification problems. With the fed dataset, KNN groups the data to classify the problems. Irrelevant input features may result in a worsening of the accuracy of prediction by the KNN model.
K means clustering algorithm is a subtype of unsupervised learning. When the number of input variables is high, it resolves the clustering problems more effectively.
The backpropagation algorithm is an effective method for processing the gradient in the neural network. This method of AI is utilised in deep learning. This method uses a neural network which has specific applications. Multi-layer perceptron is a subtype of feed-forward neural network., It has three layers, namely input, output and hidden layers. The input features will be fed to the input layer, and the output layer predicts classification problems. Numerous hidden layers will be there between the input and output layers. The data flow from the input towards the output layer.
| Machine Learning Algorithm in the Prediction of the Difficult Airway|| |
Difficult Airway (DA) can result in catastrophic desaturation and haemodynamic perturbance, especially whenever cannot intubate and difficult mask ventilation (DMV) situation exist. Although the incidence of difficult intubation (DI) is rare, the occurrence of such an event in an unanticipated situation can result in the development of arrhythmias due to desaturation and cardiac arrest if not intervened on time. Direct laryngoscopy (DL) and intubation are performed in the operating room (OR), emergency department and wards. OR is a relatively controlled environment, the patient's airway will be evaluated before being taken up for surgery, and intubation will be performed in optimal conditions. Despite preanaesthetic airway evaluation with various physical parameters, optimal intubating conditions and availability of sophisticated video-laryngoscopes, there are occurrences of DL and DI. Incidence of DA ranges between up to 20% depending upon the criteria to define DMV, DL and DI were used. There are physical airway examination tools that can predict DMV, DL and DI. These parameters include but are not limited to inter-incisor distance (IID), modified Mallampati (MMP) score, range of neck movements, edentulousness, the recession of mandible, thyromental distance (TMD), temporomandibular joint movement and compliance of the submandibular space. None of these parameters is highly sensitive or specific. Hence, a combination of these parameters is utilised to improve the sensitivity in predicting DMV, DL and DI. LEMON score, the 3-3-2 rule, and Benumof's 11 parameters scale are a few combination tools that comprised a few of the above-mentioned DA prediction parameters. These DA prediction tools are also not highly sensitive. Prediction of DA by machine learning algorithms is one promising tool that can successfully predict DA, DL and DI. Physical examination DA predictors such as limited IID, higher grade on MMP score (Grade 3 and 4), restricted neck extension, edentulousness, receded mandible, poor submandibular compliance, previous history of DL or DI from the anaesthesia record, lesions in the airway (lingual thyroid), features of the compromised airway (carcinoma of the larynx, laryngeal papillomatosis), imaging DA predictors (X-ray; Computed tomography, magnetic resonance imaging [performed for the diagnostic purpose]) and findings of ultrasonography (USG) imaging of the airway can be collected and utilised as input features for training the algorithm.
It is preferred to choose the physical parameters that can predict the DA more accurately in the clinical scenario and train the algorithm rather than including all the non-specific parameters. Some studies analysed the accuracy of predicting DA with a few parameters such as TMD, MMP class and USG airway findings. They found that algorithms predicted the DA accurately. The algorithm needs to be trained with higher numbers of patients' data as the machine learns by experience with the fed data. Hence, more training with the data will result in an accurate prediction of the DA.
BMI (>30 kg/m2: Anticipated DMV, DL and DI), IID (<2 cm: Anticipated DL), MMP (Grades 1 and 2: Ease of intubation; Grades 3 and 4: Anticipated DI), TMD (<6.5 cm-anticipated DI), restriction of neck extension (if present: Anticipated DL and DI), receded mandible (if present: Anticipated DL and DI) and poor submandibular space compliance (if present: Anticipated DL and DI) parameters which are used to predict DA by clinical assessment, can be used to feed to train the machine learning algorithm. The input features should be clearly defined, and their relationship with the outcome variable (DMV, DL or DI) should be clearly defined so that the algorithm can predict the outcome with the fed data accurately. A supervised type of machine learning algorithm would be appropriate to predict DA. Eighty per cent of the patient's data can be provided to train the algorithm, and 20% of the patients' data can be used to test the algorithm. A stepwise approach to train and test the data to predict DA is depicted in [Figure 2]. Health-care centres can prepare the appropriate machine learning algorithms to expect DA with standard input features with respect to their centre that results in DMV, DL and DI. Yamanaka et al. performed a trial in which various machine learning approaches in predicting DA and first-attempt intubation success in emergency department was studied. A total of 10,741 patients were recruited for this study, and 543 (5.1%) patients had DA. Seventy-one per cent of patients had first-attempt intubation success. Machine learning models such as logistic regression, gradient boost, and ensemble models had higher predictive accuracy for DA and models such as logistic regression, gradient boost and multi-layer perceptron had higher discriminatory capacity in predicting first-attempt intubation success. Commonly tools used to evaluate the prediction capacity of the algorithm to predict DA include accuracy, sensitivity, specificity and precision. Accuracy is the ratio between appropriate predictions to the total number of predictions (True Positives + True Negatives/All the predictions). Transparent reporting of multivariable prediction model of individual prognosis or diagnosis and Prediction model risk of bias assessment tool statements are available for improvised reporting and appraisal of machine learning prediction models for the medical field.,
|Figure 2: Simplified stepwise approach machine learning algorithm to predict DA. DA: Difficult airway|
Click here to view
| Machine Learning Algorithms in Predicting Successful Liberation from Mechanical Ventilation in Intensive Care Unit|| |
Critically ill patients are prone to develop diaphragmatic and skeletal muscle weakness that can make a failure of liberation from mechanical ventilation. Reintubation in critically ill patients increases morbidity. Many tools and indices such as rapid shallow breathing index (RSBI = VT/respiratory rate [RR]), CORE index (lung compliance, oxygenation, respiratory rate (RR) and Effort), integrated weaning index (Static compliance*SaO2/[RR/VT]), occlusion pressure at 100 m.sec (P0.1) and diaphragmatic thickness fraction are available to anticipate the success of extubation in the intensive care unit (ICU). Despite using these sophisticated tools, extubation attempts fail many times and require reintubation in ICU. It is very challenging to assess lung compliance in spontaneously breathing patients as compliance will be overestimated due to the generation of negative pressure. Few patients may require reintubation in whom reintubation has not been expected. The cause for which the patient has been placed on mechanical ventilation is resolved/resolving, BMI (>30 kg/m2), intact sensorium (absence of delirium), absence of consolidation, absence of copious secretions, oxygenation status (paO2/FiO2: >250), ventilation status (paCO2: 30–45 mmHg), a measure of work of breathing (RR, RSBI), heart rate and blood pressure during spontaneous breathing trial and diaphragmatic thickness fraction can be used as input features to predict the success of extubation in critically ill patients. Liu et al. performed a study to predict optimal timing for weaning from mechanical ventilation with a two-staged AI prediction approach. This study used machine learning algorithms such as logistic regression, RF, SVM, KNN, LGBM, XGBM and multi-layer perceptron. A total of 5873 patients were recruited for this study. The authors found that these AI prediction models could accurately predict the weaning success in ICU patients.
| Future of Artificial Intelligence in the Prediction of Difficult Airway and Weaning Success|| |
With the appropriate input variables that can predict the target outcome; the machine can be trained with an adequate number of patients to improve its accuracy in predicting the outcome variable. Then the data can be tested by the machine to predict the outcome. DA prediction parameters by physical examination and imaging techniques can be fed into the machine to master the prediction of DA, and an application can be programmed with the training. With the widespread utility of applications in the medical fraternity [applications for calculation of acute physiology and chronic health evaluation (APACHE) score in ICU, for choosing appropriate antimicrobial, for calculation of antimicrobial use adjusted to Creatinine clearance], applications for prediction of difficult airway (or for weaning success) can be programmed. Such programmed applications can be accessed by the clinicians to predict DA, thereby all the preparations for managing DA may be done to prevent adverse consequences of unanticipated difficult airway.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Gupta R, Srivastava D, Sahu M, Tiwari S, Ambasta RK, Kumar P. Artificial intelligence to deep learning: Machine intelligence approach for drug discovery. Mol Divers 2021;25:1315-60.
Greener JG, Kandathil SM, Moffat L, Jones DT. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 2022;23:40-55.
Pfob A, Lu SC, Sidey-Gibbons C. Machine learning in medicine: A practical introduction to techniques for data pre-processing, hyperparameter tuning, and model comparison. BMC Med Res Methodol 2022;22:282.
Gyori NG, Palombo M, Clark CA, Zhang H, Alexander DC. Training data distribution significantly impacts the estimation of tissue microstructure with machine learning. Magn Reson Med 2022;87:932-47.
Kernbach JM, Staartjes VE. Foundations of machine learning-based clinical prediction modeling: Part II-Generalization and overfitting. Acta Neurochir Suppl 2022;134:15-21.
Yang L, Wu H, Jin X, Zheng P, Hu S, Xu X, et al.
Study of cardiovascular disease prediction model based on random forest in Eastern China. Sci Rep 2020;10:5245.
Dinh A, Miertschin S, Young A, Mohanty SD. A data-driven approach to predicting diabetes and cardiovascular disease with machine learning. BMC Med Inform Decis Mak 2019;19:211.
Zhang Z, Ho KM, Hong Y. Machine learning for the prediction of volume responsiveness in patients with oliguric acute kidney injury in critical care. Crit Care 2019;23:112.
Krittanawong C, Johnson KW, Rosenson RS, Wang Z, Aydar M, Baber U, et al.
Deep learning for cardiovascular medicine: A practical primer. Eur Heart J 2019;40:2058-73.
Kriegeskorte N, Golan T. Neural network models and deep learning. Curr Biol 2019;29:R231-6.
Sajayan A, Nair A, McNarry AF, Mir F, Ahmad I, El-Boghdadly K. Analysis of a national difficult airway database. Anaesthesia 2022;77:1081-8.
Yamanaka S, Goto T, Morikawa K, Watase H, Okamoto H, Hagiwara Y, et al.
Machine learning approaches for predicting difficult airway and first-pass success in the emergency department: Multicenter prospective observational study. Interact J Med Res 2022;11:e28366.
Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): The TRIPOD statement. BMJ 2015;350:g7594.
Wolff RF, Moons KG, Riley RD, Whiting PF, Westwood M, Collins GS, et al.
PROBAST: A tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med 2019;170:51-8.
Azeredo LM, Nemer SN, Barbas CS, Caldeira JB, Noé R, Guimarães BL, et al.
The integrative weaning index in elderly ICU subjects. Respir Care 2017;62:333-9.
Santangelo E, Mongodi S, Bouhemad B, Mojoli F. The weaning from mechanical ventilation: A comprehensive ultrasound approach. Curr Opin Crit Care 2022;28:322-30.
Liu CF, Hung CM, Ko SC, Cheng KC, Chao CM, Sung MI, et al.
An artificial intelligence system to predict the optimal timing for mechanical ventilation weaning for intensive care unit patients: A two-stage prediction approach. Front Med (Lausanne) 2022;9:935366.
[Figure 1], [Figure 2]