Tesi di Laurea: “Development of Machine Learning Algorithms Using Anthropometric Features for Obstructive Sleep Apnea Screening and Severity Stratification” (Sintesi)

Introduction

Obstructive Sleep Apnoea (OSA) is a prevalent, yet underdiagnosed condition, affecting 9-38% of adults worldwide^1, 2. Although more prevalent in men, it remains clinically relevant in both sexes, especially among obese and senior individuals. Moderate-to-severe OSA, defined by an apnea-hypopnea index (AHI) ≥15, affects 9-27% of adults, and up to 49% of geriatric populations.

Despite its prevalence, diagnostic and therapeutic gaps remain substantial³. Only 4-12% of individuals with OSA receive a formal diagnosis, and less than 6% receive treatment, reflecting critical inefficiencies in screening and diagnostic pathways³. The economic burden of untreated OSA is substantial, with mean annual healthcare costs of €2,500 per moderate-to-severe patient and an estimated €9 billion annual societal loss in quality-adjusted life years³. Full-night Polysomnography (PSG) represents the diagnostic gold standard, providing an accurate assessment of sleep architecture. However, its cost, infrastructure requirements, and availability restrict its accessibility. To overcome these constraints, Home-based Sleep Apnoea Testing (HSAT) offers a more accessible alternative. Among HSAT devices, the WatchPAT^® ONE (Itamar Medical Ltd., Caesarea, Israel) estimates AHI using Peripheral Arterial Tonometry combined with oximetry, actigraphy, and body position analysis. Its cloud-based data transmission allows reliable at-home recordings with minimal patient supervision. Despite the advancements offered by HSAT, the diagnostic gap remains underfilled. Screening questionnaires, such as the Epworth Sleepiness Scale (ESS), STOP-BANG, and Berlin questionnaires, are widely used but show limited specificity, resulting in high false-positive rates and unnecessary referrals⁴. Their reliance on subjective reporting reduces accuracy in women and older adults, in whom classic symptoms are less pronounced. Artificial Intelligence (AI), particularly machine learning (ML), has emerged as a powerful tool in clinical research for detecting complex patterns in large or incomplete datasets, enhancing predictive accuracy and diagnostic efficiency⁵.​ Our systematic review of AI applications in OSA screening and diagnostic settings highlighted that anthropometric indices, such as body mass index (BMI), neck circumference (NC), and waist circumference (WC), are among the most informative predictors⁶. These easily retrievable parameters are particularly meaningful in remote or resource-limited settings, where they can serve as cost-effective tools for preliminary assessment or even self-evaluation. However, most studies are limited to binary classification tasks, such as discriminating between healthy and affected individuals. Moreover, class imbalance, i.e., the disproportionate representation of outcome categories, is seldom addressed in current literature, as many rely on pre-balanced datasets. This practice can bias performance estimates, since an artificially balanced class distribution does not mirror the true prevalence of each class in the target population⁷. In light of this rationale, the current study aims to develop ML-based models using anthropometric variables for:
1. OSA Screening (binary task, AHI < 5 vs AHI ≥ 5);
2. OSA Severity Stratification (ternary task, AHI 5-15 vs AHI 15-30 vs AHI ≥ 30).

To address the class imbalance observed in the population distribution, we evaluated three commonly adopted approaches, random undersampling, random oversampling, and SMOTE, within our ML pipeline.

Methods
Study design and population

This prospective observational study was conducted at Fondazione Policlinico Universitario Campus Bio-Medico di Roma between October 2024 and February 2025. The study was conducted in accordance with the Declaration of Helsinki. This observational study, involving no deviation from standard clinical practice and no experimental intervention, was deemed exempt from formal Institutional Review Board approval. Written informed consent was obtained from all participants. Inclusion criteria were age ≥ 18 years and clinical suspicion of OSA based on symptoms or physician referral.

Anthropometric and clinical information were collected during the initial consultation, including age, sex, weight, height, BMI, NC, and WC. Medical history and comorbidities were confirmed from electronic health records when available. Each participant completed a short questionnaire addressing five sleep-related symptoms (daytime sleepiness, mood changes, nocturnal awakenings, nycturia, and difficulties driving caused by drowsiness). Snoring intensity was investigated with a 10-point visual analog scale (VAS), and excessive daytime sleepiness was quantified using the Epworth Sleepiness Scale (ESS). Smoking status was classified as never, former, or current smoker. The final dataset included 30 variables. OSA diagnosis and severity were established using the WatchPAT^® ONE device (Itamar Medical Ltd., Caesarea, Israel), with OSA status defined by an AHI ≥ 5 events/h, and severity was classified as mild (AHI 5–15), moderate (AHI 15–30), or severe (AHI ≥ 30).

Dataset structure and preprocessing

To ensure robust model validation, we applied the stratified k-fold cross-validation, with k=10, which preserves the original class distribution among the folds. For each iteration, data were split into training (80%), validation (10%), and test (10%) subsets. Normalization was performed using z-score standardization based on training data only, and the same transformation was applied to validation and test sets to avoid data leakage.

The methodological framework adopted in this consists of two sequential stages, In Step 1, the optimal data balancing method is identified based on validation performance across classifiers and 10 folds of the cross-validation, while Step 2 involves the selection of the best-performing combination of feature selection strategy and classifier based on the validation set, as illustrated in Fig. 1.

 

Fig. 1. Schematic representation of the proposed method. Step 1, describing the balancing method selection process, and Step 2, illustrating the subsequent selection of the best-performing combination of feature set, feature ranking-selection method, and ML model. This process was repeated for each k-fold of the cross-validation.

 

Machine learning design

Two predictive tasks were defined:
1. binary classification, to screen for OSA presence (AHI < 5 vs ≥ 5 events·h⁻¹);
2. ternary classification, to predict OSA severity (mild, moderate, severe).

To account for any potential class imbalance, three balancing methods were tested: undersampling, which reduces the majority class to match the size of the minority class; oversampling, which duplicates instances from the minority class to achieve balance; and SMOTE (Synthetic Minority Over-sampling Technique)⁸, which generates synthetic examples by interpolating between existing minority class instances. These balancing methods were selected because they are among the most commonly adopted resampling strategies in supervised learning for medical/clinical datasets, including in the field of OSA research, and thus allow our results to remain comparable with existing literature (e.g.,^9-11). We evaluated five supervised classifiers from distinct algorithmic families and commonly used in OSA screening and diagnostic research⁶: random forest (RF), an ensemble tree-based method; logistic regression (LR), a generalized linear model; support vector machine (SVM), a maximum-margin kernel-based classifier; extreme gradient boosting (XGB), a boosting-based ensemble; and k-nearest neighbours (k-NN), a distance-based non-parametric method. Their hyperparameters were optimized via grid search within a cross-validation framework⁶.

Feature selection

To identify the smallest and most informative predictor subset, recursive feature elimination (RFE) was combined with three ranking criteria: RIDGE regression, LASSO regression, and GINI impurity. We selected RFE because it is among the most commonly used wrapper-based strategies in recent OSA and biomedical prediction studies and has demonstrated strong performance in detecting compact, high-value feature sets⁶. Feature subsets of 30, 20, 10, 9, and 5 variables were tested sequentially. The optimal combination of classifier, balancing method, and feature-selection strategy was determined based on validation performance.

Model evaluation

Model discrimination was quantified using the area under the curve (AUC) as the primary metric, in accordance with current literature. Sensitivity, specificity, precision, F1 score, and accuracy were also computed. For the binary task, paired t-tests were used to compare AUCs across models since performance differences were normally distributed. For the multiclass task, the Wilcoxon signed-rank test was adopted instead, given the non-normal distribution of the differences and the ordinal nature of the outcomes. A two-tailed p ≤ 0.05 was considered statistically significant.

Software and environment

All analyses were performed in Python 3.9 (scikit-learn 1.2, XGBoost 1.7). The computational pipeline was executed on a standard desktop workstation (Apple M1, 16 GB RAM).

Results

The final cohort included 254 patients, 200 men (78.7%) and 54 women (21.3%), with a mean age of 55.34 ± 12.42 years and a BMI of 27.28 ± 3.76 kg/m². The mean AHI for the entire cohort was 23.4 ± 17.9 events/h, with men exhibiting higher AHI values than women (24.9 ± 18.2 vs 17.2 ± 15.3). 50 participants (19.7%) were classified as non-OSA (AHI < 5 events/h), 45 (17.7%) as mild OSA (5–15), 78 (30.7%) as moderate (15–30), and 81 (31.9%) as severe (≥ 30). A full list of the 30 anthropometric variables is illustrated in Tab. 1.

 

Tab. 1. List of employed features, stratified by non-OSA vs OSA patients.

 

Handling of missing data

Missing values affected weight, height, WC, and NC. Weight and height measurements displayed missingness rates below 9% across all AHI categories. WC data missingness ranged from 19.23% in moderate OSA to 46.67% in mild OSA. Likewise, NC measurements were missing in 55.13-73.3% of cases across severity groups. Data was complete in non-OSA individuals. Missing data were imputed using a k-nearest neighbors approach (with k = 5) within the training set only. All other variables were complete. Although NC and WC exhibited high missingness, both variables are well-established predictors of OSA severity and represent key anthropometric markers in clinical practice. Removing them would have resulted in information loss and reduced model interpretability. For this reason, instead of discarding these features or excluding patients, we applied k-nearest neighbors imputation (k = 5), which has been shown to preserve multivariate relationships and outperform mean or regression-based imputation in clinical datasets with moderate-to-high missingness. The imputation was performed strictly within the training folds to avoid data leakage.

Data balancing

Since our dataset was imbalanced, with an underrepresentation of non-OSA (19.7%) and mild OSA (17.7%) cases, three balancing techniques were applied across all classifiers.

In the binary task (OSA vs non-OSA), oversampling displayed the highest discriminative power. RF achieved a mean validation AUC of 95.44% ± 4.92%, followed by LR (95.07% ± 5.17%). SMOTE achieved intermediate performances (RF 95.37% ± 4.47%), and undersampling yielded the lowest AUCs (RF 94.55% ± 5.28%), as illustrated in Tab. 2.

 

Tab. 2. Performance assessment for the binary task.

 

Likewise, oversampling offered the best balance in the ternary task (mild vs moderate vs severe OSA) too, with a mean validation AUC of 66.62% ± 7.73% for RF, and 65.59% ± 7.68% for LR. Undersampling produced slightly lower AUCs (RF 65.33% ± 6.85%), and SMOTE yielded the lowest AUCs on average (RF 59.87% ± 7.68%), as illustrated in Tab. 3.

 

Tab. 3. Performance assessment for the ternary task.

 

Oversampling was therefore selected as the preferred balancing method for the subsequent feature selection analysis.

Feature selection

RFE with RIDGE, LASSO, and GINI ranking was applied to identify optimal combinations of predictors and features.

In the binary task, RIDGE regression yielded the most consistent ranking and minimal AUC variation across folds, where the RF classifier maintained high discrimination from the full 30-variable set (95.44% ± 4.92%) to the peak 9-feature set (95.33% ± 4.83%), to a compact 5-feature subset (94.33% ± 5.66%), as shown in Fig. 2.

Statistical comparison via paired t-test between the peak 9-feature and more parsimonious 5-feature sets revealed no significant differences (p = 0.80), supporting the selection of the smaller feature set as an ideal compromise between computational efficiency and predictive power. On the independent test set, the RIDGE-RF combination achieved an AUC of 92.01% ± 7.94%. Comparable results were observed for other classifiers (from RF 92.07% ± 7.75% to k-NN 91.59% ± 7.98%).

In the ternary task, GINI impurity displayed the most stable feature rankings, where the GINI-SVM combination displayed the best AUCs, from the full 30-feature set (65.93% ± 8.19%), to a compact 10-feature subset (70.28% ± 8.16%), as shown in Fig. 3.

 

Fig. 2. Performance variations (expressed in AUC%) of the binary task observed with decreasing feature subsets, using RIDGE as a feature ranking-selection method.		Fig. 3. Performance variations (expressed in AUC%) of the ternary task observed with decreasing feature subsets, using GINI impurity as a feature ranking-selection method.

 

Statistical comparison via the Wilcoxon signed-rank test carried out between the optimal 20-feature and more parsimonious 10-feature sets revealed no significant difference (p = 0.30), supporting the clinical adoption of the reduced feature set. AUCs of 62.90% ± 10.27% (SVM), 62.90% ± 10.27% (LR), 64.71% ± 9.53% (RF), 64.07% ± 6.44% (k-NN), and 63.13% ± 10.64% (XGB) were achieved on the independent test set. Although lower than binary performance, these models still provided meaningful clinical separation between mild and severe cases using purely anthropometric and questionnaire-derived data.

Feature importance

For each task, we analyzed the percentage of times each feature was selected across the folds of the stratified cross-validation for the best-performing combination of feature selection and classifier.

Tab. 4 reports the feature selection frequencies, expressed as percentages, observed in the final test set evaluation for the binary and ternary tasks.

 

Tab. 4. Feature selection percentages for the test set, using RIDGE for feature selection and RF as ML model in the binary task and GINI Impurity for feature selection and SVM as ML model in the ternary task (only the top 11 features are reported in the table, as the remaining features exhibited selection frequencies < 20%).

 

Therefore, the final 5-feature set included the 5 features with the highest selection frequency, namely: weight, NC, WC, Friedman tongue position (FTP), and nocturnal awakenings, while the optimal 10-feature subset comprised ESS, BMI, NC, Mallampati score (MS), FTP, snoring VAS, age, WC, height, and weight.

Discussion

The management of class imbalance represents a fundamental challenge in developing clinical ML models¹². In our study, oversampling outperformed undersampling and SMOTE in both OSA screening and severity stratification.

For binary OSA vs non-OSA classification, oversampling maximized recall, a critical metric for minimizing false negatives in a screening context. While undersampling marginally improved specificity, this entailed discarding valuable data, representing a suboptimal clinical trade-off. The most robust classifiers consistently performed better with oversampling, as its simplicity makes it ideal for implementation in clinical environments where reliability is key. To balance performance with clinical practicality, we employed an RFE process. For binary classification, a compact 5-feature set including body weight, NC, WC, FTP, and nocturnal awakenings proved to be ideal. The clinical rationale behind these features is well-established. NC is a powerful metric due to its direct correlation to upper airway collapsibility caused by peri-pharyngeal fat, and remains a cornerstone of established tools such as the STOP-BANG questionnaire^{6, 13}. Meta-analytic data reinforce NC as an independent predictor of OSA, with pooled odds ratios of 1.8-3.2 for moderate-to-severe OSA per 1 cm increase in NC¹⁴. FTP provides a standardized assessment of the oropharyngeal anatomy, with higher grades (III-IV) strongly correlating with increased pharyngeal collapsibility and AHI values, even more than the MS¹⁵. Nocturnal awakenings constitute the hallmark arousal response to apnoeic events and correlate with sympathetic surges resulting in transient increases in heart rate and blood pressure. Over time, this heightened sympathetic response contributes to the development of cardiovascular comorbidities associated with OSA^16, 17. Body weight and WC reflect visceral fat accumulation, a key pathogenetic driver of OSA. WC often outperforms the predictive power of indicators of general adiposity, such as BMI alone, with meta-analyses reporting adjusted odds ratios of 2.1–3.5 for moderate-to-severe OSA per 10 cm WC increase, independent of BMI¹⁸. However, body measurements present practical challenges in busy real-world environments, as the correct acquisition of these measurements requires patient cooperation and precise tape placement. This is reflected in our dataset, where WC values exhibited considerable rates of missingness.

For the severity stratification task, a 10-feature set was optimal. This included essential predictors like the ESS and NC, alongside highly prevalent features such as BMI, snoring intensity, and WC.

ESS is a subjective measure of daytime sleepiness. Evidence has shown a positive correlation between ESS scores and OSA severity, where higher ESS scores are associated with increased AHI values^19, 20. While BMI is a simple mathematical calculation of weight and height (kg/m²), its clinical utility in predicting OSA extends beyond being a univariate metric. While BMI itself represents an integrated measure, modern ML approaches frequently decompose its components, leveraging height and weight independently at different algorithmic decision levels due to their distinct informative values. Snoring Intensity is not only a biomarker for disease progression, but also a significant indicator of quality of life for the patient and for their sleeping partners. Evidence suggests that untreated snoring reduces partners' sleep efficiency by 13% (equivalent to 62 minutes of lost sleep nightly) and increases their arousal index by 42%, often leading to sleep deprivation and relational strain^21, 22. This multiclass approach using simple, anthropometric features is novel, as most prior work has focused on binary classification or relied on complex physiological signals^23-25, highlighting the methodological contribution of our resource-efficient model.

The ternary task demonstrated substantially lower performance (average AUC 60-66%) than binary classification (average AUC 94-95%), reflecting the inherent difficulty of categorizing patients along a physiological continuum defined by arbitrary AHI cut-points, as opposed to making a simpler diseased/non-diseased distinction. Despite the lower absolute metrics, the preserved ability to identify severe cases supports the practical implementation of our model for triage-level decision support.

These findings should be contextualized within the study's limitations. Class imbalance itself reflects the real-world setting of a tertiary sleep center, which may limit generalizability to asymptomatic populations. Furthermore, although the WatchPAT device is validated, it is not the diagnostic gold standard, which may potentially introduce inaccuracies in borderline cases. As a single-center study, external validation is a necessary next step. Also, all features used in this study were collected at a single time point, limiting the models’ applicability to dynamic or longitudinal monitoring. The incorporation of temporal features, such as symptom progression, weight change, or follow-up diagnostics, could enhance predictive power and allow for disease trajectory modelling. Moreover, while missing data were handled using k-NN, imputation bias remains a concern, particularly for features not missing at random. Importantly, missingness is an intrinsic part of real-world practice. Some of the most predictive features in our model, such as NC and WC, were also among the most frequently missing due to their lower priority in routine workflows. While imputation helped preserve dataset integrity, it may have introduced systematic bias or reduced generalizability in low-resource or incomplete-data settings. Finally, a significant limitation common to many such models is their limited explainability. The “black box” nature of the predictions can impede clinical trust, and future work must incorporate explainable AI techniques to clarify the rationale behind each output, ensuring accountability and facilitating safer integration into clinical workflows.

Conclusions

In this observational study, ML models trained on easily obtainable anthropometric and clinical variables achieved high accuracy for OSA screening and moderate performance for severity classification. By integrating features that reflect both anatomical structure and symptomatic burden, the models reproduced key pathophysiological determinants of disease. Reliance on routine, non-invasive measurements make them suitable for deployment in primary care or telemedicine environments, where diagnostic resources are limited. Implemented as an automated prediction tool, this approach could streamline referrals and facilitate earlier identification and treatment of OSA. Broader multicentre validation and integration with simple physiological signals may further enhance precision and clinical uptake.