Table 1 Related work

1

Predictive Machine Learning Algorithms in Anticipating Problems with Airway Management [14]

The study aims at the use of artificial intelligence (ML) methods to foresee airway management

obstacles. The methods used are both supervised and unsupervised machine learning. Some relevant models include Decision Trees, Random Forests, SVM, and Multi-layer Perceptron Neural

Networks. The models are employed to tackle classification and regression

issues. For instance, the training of algorithms with clinical data via supervised learning models to predict outcomes, such as difficult airway (DA), difficult mask ventilation (DMV), and difficult intubation (DI) using input

features like BMI, inter-incisor distance, Mallampati score, and neck extension limitation are the algorithms’ tasks.

The set of features the algorithms were trained on includes the physical parameters such as BMI (> 30 kg/m²), inter-incisor distance (IID) (< 2 cm), Modified Mallampati (MMP) scores (Grade 3 and 4 indicate anticipated DI), thyromental distance (TMD) (< 6.5 cm), restricted neck extension, receded mandible, and poor submandibular compliance. These variables serve the purpose of predicting airway difficulties in clinical assessments.

The ways to evaluate the models are accuracy, sensitivity, specificity, and precision, and the study shows that many models such as GBM and XGBM usually do better than simpler models because of their error reduction and their prediction-improving feature by sequential learning. Finally, GBM and Logistic Regression were the best models in terms of offering both high accuracy and good discrimination. Nonetheless, the model selection is the primary healthcare concern and the amount of training data.

2

Development of a machine learning algorithm to predict intubation among hospitalized patients with COVID-19 [15].

The purpose of this study was to create a machine learning algorithm that would forecast intubation in COVID-19 hospitalized patients. Using data from 4,087 patients admitted between February and April 2020, a retrospective cohort design was employed. The method turned patient demographic, vital, and lab data into time-series data for model training using a random forest classifier. The model updated every 12 h and generated forecasts based on 24-hour windows.

Physical parameters like BMI (> 30 kg/m2), inter-incisor distance (IID) (< 2 cm), Modified Mallampati (MMP) scores (Grades 3 and 4 indicate anticipated DI), thyromental distance (TMD) (< 6.5 cm), restricted neck extension, receding mandible, and poor submandibular compliance are among the features that the algorithms were trained on. In clinical examinations, these indicators are used to predict airway issues.

Among the variables were:

1. Vitals: pulse, oxygen saturation, systolic and diastolic blood pressure, and respiratory rate.

2. Laboratory data: levels of D-Dimer, creatinine, C-reactive protein, platelet count, white blood cell count, and arterial O2 and CO2.

3. Comorbidities include diabetes, chronic obstructive lung disease, liver disease, renal disease, and hypertension.

The performance of the model was assessed by the study using the following metrics:

1. Area Under the Curve (AUC): Evaluate how well the model distinguishes between positive and negative instances.

2. Area Under the Precision-Recall Curve (AUPRC): This measures how well the model can detect real positives while maintaining sensitivity.

3. Kaplan-Meier Survival Analysis: To assess the rates of intubation-free patients who were informed by the model vs. those who were not.

The model’s AUC is 0.84, whereas the ROX index is 0.64.

AUPRC: The 0.30 score of the model beat the 0.13 score of the ROX index.

The model-identified patients had a noticeably increased risk of intubation during their hospital stay.

3

Predicting the need for intubation in the first 24 h after critical care admission using machine learning approaches [16].

Within 24 h after ICU admission, the study creates a machine-learning model to forecast the necessity for intubation. With information from two sizable databases, MIMIC-III and eICU-CRD, containing more than 17,000 critically ill patients, it employs a retrospective cohort design. Two machine learning models, Random Forest (RF) and Logistic Regression (LR), were used to complete the prediction challenge. Autoencoders (AEs) were used to impute missing data, and models trained on 60% of the available data were tested on the 40% that remained.

The following are the main variables in the model:

Demographics: Medical specialization, age, and gender.

Vital signs include heart rate, respiration rate, systolic and diastolic blood pressure, oxygen saturation (SpO2), and Glasgow Coma Scale (GCS).

Laboratory values: HCO3, PaO2, and PaCO2 are examples of blood gas parameters.

Interventions include oxygen therapy and the use of vasopressors.

The following measures were used to assess the models’ performance:

The model’s ability to distinguish between patients who are intubated and those who are not is measured by the area under the receiver operating curve, or AUC.

Sensitivity: The model’s capacity to accurately identify intubation-required patients.

Specificity: The capacity to accurately determine which patients do not require intubation.

The likelihood that genuine predictions will be made for both intubated and non-intubated instances is shown by Positive and Negative Predictive Values (PPV, NPV).

With an AUC of 0.86 as opposed to 0.77 for Logistic Regression, the Random Forest model outperformed LR.

The Random Forest model’s sensitivity is 0.88, while its specificity is 0.66.

Over the whole range of intubation risk projections, the Random Forest model showed good calibration.

4

Machine Learning Approaches for Predicting Difficult Airway and First-Pass Success in the Emergency Department: Multicenter Prospective Observational Study [17].

The research, a multicenter prospective observational study, was carried out in 13 Japanese emergency departments (EDs). 10,741 patients who had tracheal intubations between January 2010 and December 2018 were included in the dataset. Seven machine learning models, such as XGBoost, gradient boosting, and random forest, were created utilizing regularly gathered information on patient demographics and vital signs before intubation. The capacity of these algorithms to forecast two outcomes—difficult airway and first-pass intubation success—was used to assess their effectiveness.

The following are independent variables (predictors): Glasgow Coma Scale, pre-intubation vital signs (pulse rate, systolic blood pressure, respiratory rate, oxygen saturation), patient demographics (age, sex, BMI, etc.), elements of the modified LEMON criteria, and intubation-related factors (medications, intubation techniques, devices used).

Dependent factors, or results:

A difficult airway needs to be intubated more than once.

First-pass success is the state in which intubation goes well on the first try.

With the exception of the k-point closest neighbor and multilayer perceptron, machine learning models fared better than the modified LEMON criterion for predicting problematic airways; the ensemble model had the highest c-statistic, at 0.74.

With the exception of random forest and k-point closest neighbor, machine learning models surpassed the logistic regression reference model in first-pass success prediction (the ensemble model had the highest c-statistic of 0.81).

In both scenarios, the ensemble model outperformed conventional techniques in terms of sensitivity and specificity.

5

Machine learning for the prediction of preclinical airway management in

injured patients: a Registry-based trial [18].

A retrospective study was carried out using a registry-based dataset of adult trauma patients in Germany who received emergency medical care between 2018 and 2020.

Random Forest and Naive Bayes machine learning algorithms were employed to forecast the necessity of preclinical airway control.

There were 25,556 patients in all, and 1,451 of them needed breathing assistance. Principal component analysis (PCA) was utilized to pick preprocessed attributes from the dataset that were the subject of the investigation.

Important clinical factors that were taken into account for model training were auscultation, damage patterns, oxygen therapy, and shock index.

The following are examples of independent variables or features: thoracic drainage, oxygen therapy, noninvasive breathing, injury patterns, vital signs, and the usage of specific drugs, such as catecholamines.

Dependent variable (result): Preclinical airway care is necessary.

The Glasgow Coma Scale (GCS), starting heart rate, systolic blood pressure, and respiration rate are additional significant characteristics.

Regarding performance, the Random Forest (RF) model outperformed the Naive Bayes (NB) model.

AUC-ROC (area under the receiver operating characteristic) for RF was 0.96, and its overall accuracy (97.8%) was higher than that of NB.

Additionally, the RF model was better at predicting the requirement for airway care than the NB model, as evidenced by its greater positive predictive value (PPV) of 0.85 compared to 0.46.

When it came to the precision-recall area (0.83), the RF model outperformed the NB model (0.66).