Artificial Intelligence, Difficult Endotracheal Intubation
Conditions
Brief summary
The assessment and management of difficult airway is of critical importance. Unsuccessful airway management leads to serious mortality and morbidity. From the beginning of the pre-anesthesia examination, 3% to 13% of patients who are considered suitable for routine airway management may be difficult to intubate. Airway assessment issues include risk assessment and airway examination (bedside and forward) to estimate the risk of difficult airway or aspiration. Airway examination aims to determine the presence of upper airway pathologies or anatomical anomalies. Some physical characteristics are associated with difficult airways and unsuccessful intubation. Examples of these are; limited neck movement, snoring, short sternomental distance, neck circumference thickness, etc. Physical characteristics can be measured with a meter or more detailed upper airway ultrasonographic measurements. In this study, researchers aimed to evaluate the anthropometric and ultrasonographic measurement values of patients who underwent preoperative airway assessment and to see the predictability of difficult intubation with artificial intelligence-supported decision support programs.
Detailed description
Difficult intubation, particularly unpredictable difficult intubation, is a challenging scenario for every anesthesiologist. Patients who are initially assessed as suitable for routine airway management may present as difficult to intubate in 5% to 22% of cases. Accurate evaluation and management of difficult airways are crucial, as failure in airway management can lead to serious morbidity and mortality. Airway assessment helps identify predictable difficult airways, but it does not exclude patients with normal clinical evaluations who may still experience unpredictable difficult intubation. The primary goal of airway examination is to detect upper airway pathologies or anatomical anomalies. Several physical characteristics are associated with difficult airways and failed intubation, including limited neck mobility, snoring, a short sternomental distance, and increased neck circumference. Common airway assessment tools, such as the Mallampati classification and the upper lip bite test, require patient cooperation, which limits their applicability in sedated, trauma, or unresponsive patients. The Cormack-Lehane classification, used during direct laryngoscopy, is invasive and does not allow for pre-procedural preparation. In this context, non-invasive, bedside, rapid, and accessible ultrasonographic assessments and anthropometric measurements have gained importance in predicting difficult airways. With technological advancements, decision-support systems and artificial intelligence (AI)-assisted applications are increasingly used to prevent adverse outcomes. Successful airway management is particularly critical in high-risk patients, where rapid decision-making is essential. Easily accessible, bedside, non-invasive ultrasonographic measurements, integrated with AI-based learning programs, have the potential to predict difficult intubation in advance. This enables early preparation, timely interventions, and the reduction of life-threatening risks. In this study, researchers aimed to predict difficult intubation preoperatively using non-invasive anthropometric and ultrasonographic upper airway measurements, combined with AI-assisted decision-support programs, without requiring any invasive procedures. Our hypothesis is that preoperative airway assessment through anthropometric and ultrasonographic measurements, supported by AI-based decision-support programs, can accurately predict difficult intubation and facilitate early preparation
Interventions
OTHERThyromental distance
Distance between the chin and thyroid cartilage with a tape measure when the patient is in a neutral position
OTHERNeck circumference
Measurement of neck circumference with a tape measure when the patient is in a neutral position
OTHERMouth opening distance
Distance between the upper and lower teeth at the point where the mouth opening is maximum when the patient is in a neutral position.
OTHERDistance from jawbone to hyoid bone with neck in neutral position
Distance from mentum to hyoid bone with neck in neutral position by ultrasonography
OTHERDistance from jawbone to hyoid bone with neck in extension
Ultrasound measurement of distance from mentum to hyoid bone with neck in extension
OTHERDistance between skin and trachea
Ultrasound measurement of distance between skin and trachea
OTHERDistance between skin and epiglottis
Distance between skin and epiglottis measured by ultrasonography
OTHERDistance between skin and anterior commissure of vocal cord:
Distance between skin and anterior commissure of vocal cord measured by ultrasonography
OTHERDistance between skin and hyoid bone
Distance between skin and hyoid bone measured by ultrasonography
OTHERMaximum Tongue Thickness
Measurement of Maximal Tongue Thickness by Ultrasonography
Sponsors
Duzce University
Study design
Observational model
COHORT
Time perspective
PROSPECTIVE
Eligibility
Sex/Gender
ALL
Age
18 Years to 75 Years
Healthy volunteers
No
Inclusion criteria
* Patients over 18 years of age * Patients who will undergo general anesthesia
Exclusion criteria
* Pregnant women * Those with congenital and/or acquired facial deformities * Patients who have previously undergone upper neck airway surgery * Patients with head and neck tumors * Patients who will undergo thyroidectomy
Design outcomes
Primary
| Measure | Time frame | Description |
|---|---|---|
| Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | Taking ultrasonographic and anthropometric measurements of each patient took approximately 20 minutes. Machine learning estimates for each patient are approximately 1 min. | The dataset, labeled based on expert assessment of difficult intubation, was classified using eight widely accepted machine learning algorithms: logistic regression (LR) \[6\], support vector machine (SVM) \[7\], random forest (RF) \[8\], K-nearest neighbors (KNN) \[9\], Gaussian naive Bayes (GNB) \[10\], CatBoost \[11\], XGBoost \[12\], and decision tree (DT) \[13\]. From the original 30 parameters, the 15 most influential features were selected based on feature extraction methods and literature relevance. Preprocessing steps included handling missing values, with incomplete records excluded. The dataset was split into training (80%) and test (20%) sets. Models were trained on the training set, with hyperparameter tuning performed via 5-fold cross-validation to avoid overfitting. Final model performance was evaluated on the independent test set. |
Countries
Turkey (Türkiye)
Participant flow
Participants by arm
| Arm | Count |
|---|---|
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia Thyromental distance: Distance between the chin and thyroid cartilage with a tape measure when the patient is in a neutral position
Neck circumference: Measurement of neck circumference with a tape measure when the patient is in a neutral position
Mouth opening distance: Distance between the upper and lower teeth at the point where the mouth opening is maximum when the patient is in a neutral position.
Distance from jawbone to hyoid bone with neck in neutral position: Distance from mentum to hyoid bone with neck in neutral position by ultrasonography
Distance from jawbone to hyoid bone with neck in extension: Ultrasound measurement of distance from mentum to hyoid bone with neck in extension
Distance between skin and trachea: Ultrasound measurement of distance between skin and trachea
Distance between skin and epiglottis: Distance between skin and epiglottis measured by ultrasonography
Distance between skin and anterior commissure of vocal cord:: Distance between skin and anterior commissure of vocal cord measured by ultrasonography
Distance between skin and hyoid bone: Distance between skin and hyoid bone measured by ultrasonography
Maximum Tongue Thickness: Measurement of Maximal Tongue Thickness by Ultrasonography | 329 |
| Total | 329 |
Baseline characteristics
| Characteristic | Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | — |
|---|---|---|
| Age, Categorical <=18 years | 0 Participants | — |
| Age, Categorical >=65 years | 12 Participants | — |
| Age, Categorical Between 18 and 65 years | 317 Participants | — |
| Age, Continuous | 45 years STANDARD_DEVIATION 20 | — |
| Race and Ethnicity Not Collected | — | — Participants |
| Region of Enrollment Turkey | 329 Participants | — |
| Sex: Female, Male Female | 190 Participants | — |
| Sex: Female, Male Male | 139 Participants | — |
Adverse events
| Event type | EG000 affected / at risk |
|---|---|
| deaths Total, all-cause mortality | 0 / 329 |
| other Total, other adverse events | 0 / 329 |
| serious Total, serious adverse events | 0 / 329 |
Outcome results
Primary
Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations
The dataset, labeled based on expert assessment of difficult intubation, was classified using eight widely accepted machine learning algorithms: logistic regression (LR) \[6\], support vector machine (SVM) \[7\], random forest (RF) \[8\], K-nearest neighbors (KNN) \[9\], Gaussian naive Bayes (GNB) \[10\], CatBoost \[11\], XGBoost \[12\], and decision tree (DT) \[13\]. From the original 30 parameters, the 15 most influential features were selected based on feature extraction methods and literature relevance. Preprocessing steps included handling missing values, with incomplete records excluded. The dataset was split into training (80%) and test (20%) sets. Models were trained on the training set, with hyperparameter tuning performed via 5-fold cross-validation to avoid overfitting. Final model performance was evaluated on the independent test set.
Time frame: Taking ultrasonographic and anthropometric measurements of each patient took approximately 20 minutes. Machine learning estimates for each patient are approximately 1 min.
| Arm | Measure | Group | Value (NUMBER) |
|---|---|---|---|
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | support vector macine Accuracy | 89.39 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | Logistic Regression Accuracy | 77.27 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | Random Forest Accuracy | 87.88 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | Decision Tree Accuracy | 80.30 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | K-Nearest Neighbors Accuracy | 74.24 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | Gaussian Naive Bayes Accuracy | 71.21 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | CatBoost Accuracy | 83.33 percentage of estimate |
| Patients Between the Ages of 18 and 20 Who Will Receive General Anesthesia | Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations | XGBoost Accuracy | 81.82 percentage of estimate |