Abstract
Objectives This paper attempts to use machine-learning (ML) algorithms to predict the presence of sleep-disordered breathing (SDB) in a patient based on their body habitus, craniofacial anatomy, and social history. Materials and methods Data from a group of 69 adult patients who attended a dental clinic for oral surgeries and dental procedures in the last 10 years was used to train machine-learning models to predict whether a subject is likely to have SDB based on input information such as age, gender, smoking history, body mass index (BMI), oropharyngeal airway (Mallampati assessment), forward head posture (FHP), facial skeletal pattern, and sleep quality. Logistic Regression (LR), K-nearest Neighbours (kNN), Support Vector Machine (SVM) and Naïve Bayes (NB) were selected as these are the most frequently used supervised machine-learning models for classification of outcomes. The data was split into two sets for machine training (80% of total records) and the remaining was used for testing (validation). Results Initial analysis of collected data showed overweight BMI (at 25 or above), periorbital hyperchromia (dark circle eyes), nasal deviation, micrognathia, convex facial skeletal pattern (class 2) and Mallampati class 2 or above have positive correlations with SDB. Logistic Regression was found to be the best performer amongst the four models used with an accuracy of 86%, F1 score of 88% and area under the ROC curve (AUC) of 93%. LR also had 100% specificity and 77.8% sensitivity. Support Vector Machine was the second-best performer with an accuracy of 79%, F1 score of 82% and AUC of 93%. k-Nearest Neighbours and Naïve Bayes performed reasonably well with F1 scores of 71% and 67%, respectively. Conclusions This study demonstrated the feasibility of using simple machine-learning models as a credible predictor of sleep-disordered breathing in patients with structural risk factors for sleep apnoea such as craniofacial anomalies, neck posture and soft tissue airway obstruction. By utilising higher-level machine-learning algorithms, it is possible to incorporate a broader range of risk factors, including non-structural features like respiratory diseases, asthma, medication use, and more, into the prediction model.