Anthropometric and US-Guided Difficult Intubation Prediction With ML Models: Complete Statistical Analysis

1. Trial at a Glance

This page covers the registry results for Anthropometric and US-Guided Difficult Intubation Prediction With ML Models (NCT06904586), a registered phase not stated study. The primary outcome is Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations.

2. Clinical Question

The study addresses a prediction question: whether information collected during preoperative airway assessment can be used by machine-learning methods to identify patients labeled as having difficult intubation. The registry describes both physical measurements and ultrasonographic measurements, with the data labeled according to expert assessment. Its stated aim is to assess the predictability of difficult intubation using artificial-intelligence-supported decision-support programs.

The listed measures span external anthropometry and ultrasound distances to anatomical landmarks. The endpoint focuses specifically on support vector machine accuracy, although the registry description says that eight algorithms were applied. The primary question is therefore narrower than the full modeling exercise: it concerns the accuracy of the support vector machine classification, not the comparative benefit of a clinical intervention.

The registry does not report the final performance estimate, the confidence interval limits, or the number of participants retained for the final evaluation. It also does not report the difficult-intubation label counts or a detailed definition of the expert assessment. Consequently, the stated question and analysis approach can be described, but the observed predictive performance cannot be interpreted from the registry record.

3. Trial Design

This is a prospective observational study, not a randomized clinical trial. The record lists a total enrollment of 329 and a completed status. The recorded start date is 2024-03-01, and the primary completion date is 2024-12-03. Duzce University is listed as the sponsor. These record details describe the study's registration and conduct period, but do not establish how many participants contributed complete records to the final model evaluation.

The study population is described as ASA I–III patients over the age of 18 who meet the inclusion criteria for general anesthesia. The arm title instead says patients between the ages of 18 and 20. The registry record does not explain this difference, and it does not provide further eligibility criteria or participant characteristics. It would therefore be inappropriate to infer a more specific age range or participant profile.

Preoperative assessment included anthropometric and ultrasonographic measurements. The record lists thyromental distance, neck circumference, mouth opening distance, jawbone-to-hyoid distance with the neck neutral and extended, skin-to-trachea distance, skin-to-epiglottis distance, and skin-to-anterior-commissure-of-vocal-cord distance. The stated approximate collection time was 20 minutes for each participant's measurements; the machine-learning estimate was described as taking approximately 1 min for each participant.

The analytic description says the data were labeled using expert assessment of difficult intubation, and that eight machine-learning algorithms were applied. It describes selecting 15 influential features from an original 30 parameters, excluding incomplete records, and dividing the dataset into training and test sets in an 80% and 20% split. Hyperparameter tuning was described as using 5-fold cross-validation. Final performance was to be evaluated on the independent test set. The registry record does not report the number of records in each split, whether the split was stratified, or whether there were repeated partitions or an external validation cohort.

No treatment allocation or control comparison is described. The analysis is a model-development and test-set performance exercise, rather than a randomized comparison. The record lists one outcome measure and no analyses, so the registered entry does not contain a numerical analysis result to interpret.

4. Participant Flow and Accounting

The registry reports how participants moved through each stage of the study. All counts below are exactly as posted; missing counts are not inferred from other stages.

Overall Study

5. Baseline Characteristics

The table reproduces the baseline characteristics posted in the registry, which describe the participants analysed at baseline and help show whether the randomized groups were comparable.

6. Primary Endpoint

The registry lists the primary outcome as Support Vector Machine Algorithm Percentage of Accuracy in Predicted Difficult Intubations, measured over Taking ultrasonographic and anthropometric measurements of each patient took approximately 20 minutes. Machine learning estimates for each patient are approximately 1 min..

The registered primary outcome is the support vector machine algorithm's percentage accuracy in predicted difficult intubations. Accuracy describes the share of evaluated cases for which a model's classification agrees with the reference label. In this study, the labels are described as being based on expert assessment of difficult intubation. The outcome is therefore intended to summarize agreement between model predictions and those labels on the test data.

Accuracy is a familiar overall measure, but its interpretation depends on the composition of the cases and on the consequences of different kinds of classification errors. A model can make correct classifications overall while still missing a meaningful share of one class, or while incorrectly flagging cases in another. Other measures can describe these error patterns separately. The registry record does not identify additional registered primary performance measures, and the numerical distribution of the labels is not reported.

The outcome specifies a 95% confidence interval as its measure of dispersion. In general, a confidence interval communicates uncertainty around an estimated performance value: a narrower interval indicates greater precision under the method used to construct it, while a wider interval indicates less precision. The registry record does not report the interval limits or explain the method used to calculate the interval. It also does not report the accuracy estimate itself. Thus, the endpoint's definition is available, but neither the magnitude nor the precision of the observed performance can be assessed from the registry record.

The stated timeframe describes assessment and prediction duration, not a follow-up period for an outcome. The measurements took approximately 20 minutes per participant, and machine-learning estimates approximately 1 min per participant, according to the record. These timing details do not indicate how long the model was evaluated or how difficult intubation was adjudicated.

7. Why No Re-analysis Is Shown

8. Statistical Interpretation

9. Statistical Methods Explained

The statistical approach described in the registry record is a supervised classification workflow. In general, supervised learning uses cases with known labels to train a model to assign labels to other cases. Here, difficult-intubation labels were based on expert assessment, and the intended task was to classify those labels using anthropometric and ultrasonographic measurements. The record identifies the model inputs and general workflow, but does not report the final numerical performance.

Training and test data

The record describes splitting the dataset into training and test sets in an 80% and 20% allocation. In general, the training data are used to fit model parameters, while the held-out test data are reserved for evaluating performance after model development. A test set can provide a less biased estimate of performance than evaluating the model on the same data used to fit it, provided the split and analysis are carried out without information from the test set influencing model choices.

The test set is described as independent, but it is drawn from the study dataset rather than identified as data from a separate external population. It therefore represents an internal holdout evaluation as described, not evidence of performance across other settings. The registry record does not explain whether the division was random or stratified, how reproducibility of the split was handled, or the participant counts in the two sets. Such details affect how reproducible and representative the estimate may be.

Cross-validation and tuning

Hyperparameter tuning is described as using 5-fold cross-validation. In general, cross-validation repeatedly divides training data into subsets so that model settings can be assessed on data not used to fit that particular iteration. The process can help select settings while making better use of the training data than relying on one development split alone. It does not remove the need for a separate final evaluation, and the record says that final performance was assessed on the test set.

Overfitting occurs when a model adapts too closely to patterns in its development data, including random quirks that may not recur in new data. Cross-validation can help identify settings that generalize better within the development data, but it does not guarantee generalization. The registry does not provide the specific tuning ranges, selection criteria, or implementation details. It also does not clarify whether each preprocessing and feature-selection step was performed within the cross-validation process.

Feature selection and preprocessing

The described analysis began with 30 parameters and selected 15 influential features using feature-extraction methods and literature relevance. Feature selection can reduce the number of inputs and focus modeling on variables considered informative. If feature selection uses information from the test data, however, the test estimate can become optimistic because the held-out data have indirectly influenced model development. The registry record does not state how this selection was separated from the final test evaluation, so that aspect cannot be assessed.

The record says preprocessing included handling missing values and that incomplete records were excluded. Excluding incomplete records is one approach to missing data, but it can reduce the analyzed sample and can affect which participant profiles are represented if incompleteness is not random. In general, missingness patterns and the amount of missing data help determine how an exclusion approach should be interpreted. The number of records excluded, the variables with missing values, and any comparison of included and excluded cases are not reported in the registry record.

Accuracy and uncertainty

Accuracy is the proportion of evaluated classifications that match the reference labels, commonly expressed as a percentage. It offers a compact measure of overall agreement, but does not identify whether errors are false negatives or false positives. For a clinical prediction task, those error types can have different implications. In general, sensitivity, specificity, predictive values, and a confusion matrix can add detail, but the registry record does not report these measures.

A 95% confidence interval is specified for the primary accuracy outcome. In general, a confidence interval describes uncertainty in an estimate under the assumptions of its calculation and sampling process. Its width can be affected by the amount of evaluable data and the distribution of outcome labels. The actual interval and its calculation method are not reported, so precision cannot be judged. The record also gives no class counts, so accuracy cannot be interpreted in light of the balance between difficult and non-difficult cases.

Multiple models and interpretation

The description names eight algorithms, while the registered primary outcome refers to support vector machine accuracy. Testing or comparing several algorithms can raise multiplicity and model-selection concerns, particularly if the best-performing result is selected after seeing test-set performance. The registry record does not explain whether the test set was used only once, how model comparisons were conducted, or whether any adjustment for multiple testing was planned. These are unresolved reporting details, not evidence that a particular analytic practice occurred.

Finally, model discrimination and clinical usefulness are distinct concepts. A classifier may agree with labels on an evaluation dataset without demonstrating that its use improves decisions or outcomes. In general, assessment of clinical utility would require evidence about how predictions are used and their consequences. Such an evaluation is not described in the registry record. The posted entry supports a description of the modeling plan, but not a conclusion about real-world utility or the relative performance of the algorithms.

10. Multiplicity, Missing Data and Analysis Populations

The registry record lists one outcome measure, support vector machine accuracy, and zero analyses. Its methods description nevertheless names eight machine-learning algorithms. These registry entries do not establish which algorithm results were formally compared, whether one model was designated in advance as the principal model beyond the named primary outcome, or whether any multiplicity procedure was planned. Adjustment for multiple comparisons is not reported in the registry record.

Multiplicity matters in general when many models, outcomes, or analysis choices are examined: selecting a favorable result from many possibilities can make the selected performance appear more convincing than it would under a single prespecified test. The record does not say whether model comparisons were inferential or descriptive, whether the test set was used for model selection, or whether the reported support vector machine estimate was intended as a sole confirmatory result. No claim about adjustment or its absence in the actual analysis can be made beyond noting that it is not reported in the registry record.

For missing data, the description says that incomplete records were excluded after preprocessing addressed missing values. The number excluded, the variables affected, the reasons for missingness, and the final evaluable sample are not reported. In general, excluding incomplete cases can reduce precision and may affect representativeness if included and excluded participants differ. The record does not describe imputation or sensitivity analyses.

The description distinguishes training and test data and reports 5-fold cross-validation for hyperparameter tuning. The participant counts in those sets and the precise analysis populations are not reported. There is also no reported account of whether all enrolled participants were eligible for model analysis or how exclusions were handled across the workflow. These omissions limit assessment of the reproducibility and robustness of the performance estimate.

11. Safety Results

The registry reports serious adverse events as affected participants divided by participants at risk.

Deaths

Other (non-serious) adverse events

Participants with at least one non-serious event at or above the registry's 0% reporting threshold.

The registry record reports one group with 329 participants at risk for serious adverse events and zero with serious adverse events. It also reports 329 at risk for other adverse events and zero with other adverse events, and lists zero deaths. These are the recorded event counts for the group described as patients receiving general anesthesia.

These counts are not a comparison between randomized treatment groups. The study is observational and evaluates airway measurements and prediction methods; the registry does not describe a control group for safety comparison. The event figures should therefore be read as counts reported for the study group, not as evidence of a relative safety difference.

Zero recorded events does not establish that an event is impossible, nor does it show that all possible events were captured. In general, the interpretation of safety counts depends on how events were defined, monitored, and attributed, as well as the completeness of reporting. The registry record does not provide further event definitions, monitoring details, or a narrative explaining the zero counts. It also does not describe the duration of safety follow-up. The figures communicate what was recorded in the registry, not a broader estimate of risk beyond the study context.

12. What the Result Does, and Does Not, Mean

The registry record establishes that the study was designed to evaluate prediction of difficult intubation from preoperative anthropometric and ultrasonographic measurements. It describes a dataset labeled by expert assessment, a training and test split, cross-validation for tuning, and a primary accuracy outcome for a support vector machine. It does not report the observed accuracy or its 95% confidence interval. Therefore, the record does not show how accurately the model classified the evaluated cases or how precise that estimate was.

Even if an accuracy result were available, it would describe agreement with the study's reference labels in the evaluated dataset. It would not alone prove that the model can predict performance in a different population or clinical setting. The registry does not describe external validation. Nor would accuracy by itself establish that a model improves decisions, outcomes, or the process of airway management. Those are separate questions requiring evidence beyond a classification metric.

The study is observational, so it does not estimate the causal effect of a treatment. Its purpose is prediction rather than a comparison of therapies. The listed measurements and algorithms may be evaluated as candidate predictors, but the entry does not report results for the other algorithms, additional performance metrics, or a comparative ranking.

The enrollment count is 329, while incomplete records were excluded according to the analysis description. The final number analyzed is not reported, so enrollment cannot be treated as the denominator for the accuracy estimate. The outcome-label distribution, the number of difficult-intubation cases, and detailed error counts are also absent. These missing quantities prevent a fuller account of the model's behavior.

The appropriate conclusion from the posted record is limited: a prospective observational prediction study was registered and completed, and its modeling workflow is described. The registry record does not contain the primary numerical result needed to judge predictive performance. That absence should not be interpreted as evidence either for or against the ability of the model to classify difficult intubation.

13. Why This Trial Matters Statistically

This study raises statistical questions that are important whenever machine-learning predictions are developed from clinical measurements. First, the prediction target must be defined clearly. The record says that labels were based on expert assessment of difficult intubation, but it does not give a detailed label definition or adjudication procedure. The quality and consistency of reference labels affect what a model is learning to predict.

Second, an evaluation design must separate model development from performance assessment. The described training and test split and 5-fold cross-validation are relevant safeguards against evaluating a fitted model only on the same data used to build it. Their value depends on how feature selection, preprocessing, tuning, and test-set evaluation were kept separate. The registry record does not provide enough implementation detail to assess those boundaries fully.

Third, a single overall accuracy measure has limits. Without the number of cases in each label class and information about the errors, it is difficult to understand what the model gets wrong and for whom. A confidence interval is also needed to judge the precision of the estimate. Although the registry identifies a 95% confidence interval as part of the primary outcome, neither it nor the accuracy result is reported.

Finally, several algorithms are listed, but the registered entry provides no performance results or comparison plan. Transparent reporting of the selected model, the evaluation population, exclusions, and performance uncertainty would help readers distinguish a promising development result from evidence of reliable prediction. The record is statistically relevant because it describes an intended workflow, but the absent numerical results prevent evaluation of its predictive claims.

14. Limitations

• The posted summary statistics are the only data these results rest on; patient-level data, full protocols and statistical analysis plans are not available.
• Recalculated results are simple, unadjusted calculations from posted summaries. They do not replace the sponsor's prespecified analysis.

The main limitation for interpretation is that the registry record contains no numerical result for the primary accuracy outcome. The specified 95% confidence interval is also not reported. As a result, neither performance magnitude nor precision can be assessed. The record lists zero analyses, so the entry does not provide a result-level account of the modeling exercise.

The participant description is not fully consistent: the arm title refers to patients between the ages of 18 and 20, while the arm description specifies ASA I–III patients over the age of 18. The registry record does not clarify the intended age range. It also does not report detailed inclusion and exclusion criteria or baseline characteristics, limiting assessment of the study population.

The record identifies expert assessment as the basis for difficult-intubation labels but does not report the detailed criteria, assessor process, or any agreement assessment. The meaning and consistency of the prediction target therefore cannot be fully evaluated. The outcome-label counts and final analysis denominator are also absent.

Although the workflow describes a test set and cross-validation, important implementation details are not reported, including participant counts by split, the exact feature-selection procedure, and how preprocessing and feature selection were separated from test evaluation. No external validation cohort is described. The record also does not report class-specific performance measures or the handling of missingness beyond excluding incomplete records.

Finally, the registry does not report model-comparison results, a multiplicity approach, or a clinical utility evaluation. These omissions limit conclusions about model ranking, robustness, generalizability, and implications for decision-making. They do not establish that the analyses were conducted incorrectly; they identify information not reported in the registry record.

15. Related Tutorials

16. Related Statistical Calculators

17. Sources

• ClinicalTrials.gov: Anthropometric and US-Guided Difficult Intubation Prediction With ML Models — NCT06904586.
• Registry record last updated 2025-05-31; status last verified 2025-05.
• Results first posted 2025-05-31.
• Data retrieved from ClinicalTrials.gov on 2026-11-09.
• Group sizes, means, event counts and sponsor-reported p-values are reproduced as posted. Effect estimates, confidence intervals and p-values not attributed to the sponsor are independently calculated from the posted summary statistics.

18. Record Summary

The registry describes a completed prospective observational study of anthropometric and ultrasonographic measurements for predicting difficult intubation. Enrollment was 329. The recorded study dates are 2024-03-01 to 2024-12-03, and Duzce University is listed as sponsor.

The primary outcome is support vector machine percentage accuracy with a 95% confidence interval. The analysis description names eight algorithms, selection of 15 features from 30 parameters, exclusion of incomplete records, an 80% training and 20% test split, and 5-fold cross-validation for tuning. The final model performance was described as evaluated on an independent test set.

The registry record does not report the numerical accuracy, confidence interval, final analysis sample, or algorithm comparison results. It reports zero serious adverse events, zero other adverse events, and zero deaths among the recorded group. The available entry supports a description of the study design and intended workflow, but not a quantitative conclusion about predictive performance.