We consecutively enrolled adults with suspected OSA who underwent in-laboratory PSG at the Sleep Center of Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, between June 2022 and July 2024. During the same period, we additionally recruited community-based participants who underwent HSAT. Inclusion criteria were age ≥18 years, prominent snoring, and provision of informed consent. Exclusion criteria included: (1) chronic diseases that may contribute to hypoxemia, such as heart failure, chronic obstructive pulmonary disease, chronic kidney disease; (2) chronic use of medications affecting sleep, including sedative-hypnotics, anxiolytics, antidepressants, and antipsychotics; (3) other concurrent sleep disorders, such as upper airway resistance syndrome, restless legs syndrome, or hypersomnia; (4) prior treatment for OSA; and (5) incomplete data. The participant flowchart is shown in Figure 1.

The study protocol complied with the Declaration of Helsinki and was approved by the Ruijin Hospital Ethics Committee (approval number: 2018-107). Due to its retrospective design, all data were fully deidentified, and the ethics committee waived the requirement for informed consent. All data were securely stored in accordance with institutional research data management standards, and no compensation was provided to participants.

Participants abstained from sedatives, alcohol, and caffeinated beverages for at least 24 hours prior to the study. In-laboratory PSG was performed using the Alice 6 system (Philips Respironics, Murrysville, PA, USA) with standard monitoring including electroencephalography, submental electromyography, bilateral electrooculography, electrocardiography, pulse oximetry, oronasal airflow via thermistor and pressure transducer, thoracoabdominal effort, snoring, and body position. HSAT used the Alice NightOne device (Philips Respironics, Murrysville, PA, USA) to record nasal airflow, respiratory effort, and fingertip SpO₂. Recordings with more than 4 hours of analyzable data following manual review were considered valid. Two certified sleep specialists independently scored PSG and HSAT data according to the AASM scoring manual [3]: apnea was defined as ≥ 90% airflow reduction for ≥ 10 seconds, and hypopnea as ≥ 30% airflow reduction for ≥ 10 seconds accompanied by ≥ 4% SpO₂ desaturation. AHI was calculated as the total number of apneas and hypopneas per hour of sleep, and OSA was defined as AHI ≥ 5 events/hour.

All analyses were conducted using Python (version 3.11; Python Software Foundation). Continuous variables are presented as median and IQR, and categorical variables as frequency and percentage. The Anderson–Darling test was used to assess normality. Group differences were evaluated with the Kruskal–Wallis H test, followed by Dunn’s post-hoc test (significance threshold P<.05). To examine linear and nonlinear associations between continuous predictors and the binary outcome, restricted cubic spline (RCS) regression was fitted within an LR framework. Likelihood-ratio tests compared RCS models against linear specifications, and spline curves were used to visualize dose-response relationships. To further interpret model predictions, Shapley additive explanations (SHAP) were used to quantify the contribution of each feature. Finally, stratified analyses by sex and age were conducted for these oximetry parameters.

Among 4156 screened participants, 2641 were included in the final analysis: 2195 undergoing PSG and comprised the internal development cohort, and 446 undergoing HSAT and formed the external validation cohort (Figure 1). The internal cohort consisted of 943 non-OSA and 1252 OSA participants. Compared with the non-OSA group, the OSA group was significantly older, a higher male proportion, experienced more frequent hypoxic episodes, and had longer hypoxic durations. The external cohort comprised 76 non-OSA and 370 OSA participants. These OSA patients displayed higher AHI, ODI, ST90, T90, and HB values alongside lower oxygen saturation, yet they were younger than non-OSA participants, with no significant between-group difference in sex distribution. Demographic and clinical characteristics are summarized in Table 2. Violin plots (Figure 3) revealed a higher median age (60.0 vs 45.0 years) and more severe nocturnal hypoxemia in the external validation cohort, underscoring distinct disease severity and physiological profiles between the 2 cohorts. These differences provide a robust foundation for validating the generalizability of the multi-parameter oximetry model across diverse clinical scenarios.

We evaluated the predictive performance of 8 OSA-related oximetry parameters using 6 ML algorithms. Given the class imbalance between OSA and non-OSA groups in the internal cohort, the F₁-score was selected as the primary metric to balance precision and recall, with AUC used to assess overall discriminative performance [43]. Substantial heterogeneity was observed in model performance, with F₁-scores ranging from 0.5332 to 0.9269 and AUC values from 0.5660 to 0.9808. Notably, ODI and HB exhibited the strongest discriminative ability. Figure 4A summarizes the top 4 single-parameter oximetry models ranked by F₁-score. The SVM model achieved optimal performance for ODI (F₁-score = 0.9269, AUC = 0.9712), and the LightGBM model performed best for HB (F₁-score = 0.9043, AUC = 0.9590). By contrast, MeanSpO₂ and TotalPower showed comparatively weaker discriminative capacity, with F₁-score of 0.7073 (LR model) and 0.6713 (CatBoost model), respectively.

Beyond the linear association of MinSpO₂, all other oximetry parameters exhibited nonlinear relationships with OSA risk (P<.001), accounting for the heterogeneous predictive performance across indicators. Strong predictors, including ODI, HB, T90, and ST90, exhibited steep dose-response curves with pronounced threshold effects (Figure 5). For instance, ODI showed a rapid risk escalation at lower values followed by a plateau, thereby providing distinct decision boundaries that enhanced the model’s discriminative ability and optimized F₁-scores. Conversely, weaker predictors exhibited contrasting profiles: MeanSpO₂ showed shallow gradients within the clinically critical 88%-92% range, resulting in classification ambiguity, whereas TotalPower displayed marked variability with widened 95% CIs at higher values, indicating substantial noise that limited predictive utility (Figure 5). Given the complex nonlinear patterns of most key predictors, traditional linear regression models fail to capture these critical features.

We constructed and evaluated multi-parameter oximetry models, including 28 dual-, 56 triple-, and 70 quadruple-parameter combinations, with top-performing models illustrated in Figure 4B-D. Among the dual-parameter models, the CatBoost-trained ODI-HB model achieved optimal performance (F₁-score = 0.9472, AUC = 0.9865; Table 3, Figure 4B). The ODI-HB-MinSpO₂ model performed best in the triple-parameter category (F₁-score = 0.9496, AUC = 0.9869; Table 3, Figure 4C), whereas the quadruple-parameter ODI-HB-MinSpO₂-ST90 model attained the highest overall discriminative ability (F₁-score = 0.9516, AUC = 0.9879), significantly outperforming single-parameter oximetry models (Table 3, Figure 6). CatBoost demonstrated consistent superiority across all evaluation metrics (Table 3). Notably, adding 5 or more oximetry parameters yielded only marginal gains, underscoring the importance of selecting informative and complementary features rather than an indiscriminate increase in input dimensionality.

Subgroup analyses revealed significant performance variations across demographics. In the male subgroup, the optimal model (ODI-HB-MinSpO₂-ST90) achieved an F₁-score of 0.9460 and an AUC of 0.9853, with CatBoost outperforming other algorithms (Table 4, Figure 7A). In the female subgroup, the best-performing combination was ODI-HB-MinSpO₂-MeanSpO₂ (F₁-score = 0.9543, AUC = 0.9919; Table 4, Figure 7B), suggesting sex-specific differences in OSA-related hypoxic patterns. In the age-stratified analysis, the older subgroup demonstrated superior overall performance (F₁-score = 0.9398-0.9701, AUC = 0.9913-0.9933) with ODI-HB-MinSpO₂-ST90 as the optimal model, whereas the younger subgroup exhibited stable but slightly lower performance (F₁-score = 0.9163-0.9467, AUC = 0.9774-0.9863), favoring ODI-HB-MinSpO₂-MeanSpO₂ (Figure 7C-D). Across all subgroups, CatBoost maintained consistently superior classification performance (Table 4).

To elucidate the predictive mechanisms of the optimal 4-parameter model, we integrated SHAP analysis with normalized feature importance scores. SHAP values quantified each feature’s marginal contribution and revealed nonlinear relationships between oximetry parameters and OSA risk (Figure 8A), while normalized scores reflected relative contribution weights (Figure 8B). In the internal cohort, ODI, HB, and MinSpO₂ emerged as the top 3 predictors, with importance scores of 0.437, 0.320, and 0.137, respectively (Figure 8B). Subgroup analyses revealed heterogeneous contribution patterns across sex and age strata (Figure 8C-J). Notably, male and older subgroups showed consistent dominance of ODI, HB, MinSpO₂, and ST90 (Figure 8C, D, G, H), with ODI exhibiting the highest contribution in the older subgroup (importance score: 0.511). Conversely, younger and female subgroups were characterized by ODI-HB-MinSpO₂-MeanSpO₂, where MeanSpO₂ replaced ST90 as a stronger predictor (Figure 8E,F,I,J). Particularly in females, MeanSpO₂ surpassed MinSpO₂ in contribution strength (Figure 8F).

To assess generalizability, we tested model performance on an independent external cohort. The CatBoost algorithm demonstrated robust generalizability, achieving an F₁-score of 0.9667 with single-parameter configurations and maintaining high performance as oximetry parameter complexity increased (Table 5). Specifically, the optimal 4-parameter oximetry model (ODI-HB-MinSpO₂-ST90) achieved an F₁-score of 0.9838 and an AUC of 0.9881 (Table 5, Figure 9D), suggesting that the model captures shared OSA pathophysiological features rather than overfitting the internal cohort. Subgroup analysis further confirmed the robustness of sex- and age-stratified models in the external cohort (Table 6). Sex-optimized models achieved F₁-scores of 0.9848 (male subgroup: ODI-HB-MinSpO₂-ST90) and 0.9799 (female subgroup: ODI-HB-MinSpO₂-MeanSpO₂), with AUCs exceeding 0.98 (Figure 9E, F). Age-stratified models similarly achieved F₁-scores exceeding 0.98 across subgroups (Figure 9G, H). These results validate the excellent generalizability of the CatBoost-based oximetry model for diverse OSA screening applications (Table 6).

This study presents the first comprehensive evaluation of multidimensional oximetric parameters for OSA screening. Using 6 ML algorithms, we developed and rigorously validated an integrated multi-parameter model that overcomes the inherent limitations of conventional single- or dual-parameter approaches [44,45]. Through a novel algorithm-parameter matching framework, we established a CatBoost model combining ODI, HB, MinSpO₂, and ST90. This model showed robust performance in external validation, providing a streamlined and high-precision tool for OSA screening. Furthermore, our findings elucidate the heterogeneous contributions of oximetric parameters across sex- and age-specific subgroups, addressing a critical gap in population-specific research [46] and providing the foundation for personalized risk stratification.

We used a large internal development set derived from PSG and an independent external validation set derived from HSAT, which included older patients with severe nocturnal hypoxemia, representing distinct clinical phenotypes. This integration of community and clinical data improves model generalizability and may streamline OSA diagnosis [4,23]. Consistent with previous reports, HSAT showed high diagnostic accuracy and strong correlation with PSG in older patients with severe OSA [47]. As expected, the OSA group was predominantly male and exhibited worse oximetry profiles [1,22,23]. Single-parameter oximetry models demonstrated marked variability in predictive performance, with ODI and HB emerging as the strongest predictors. ODI quantifies hourly desaturation frequency and correlates strongly with AHI, acting as an independent predictor of OSA severity regardless of sleep stage or body position [48,49]. Although ODI >20 events/h showed high sensitivity (96.6%) for severe OSA, its AUC for mild disease was only 0.62, suggesting limited standalone utility [50,51]. In contrast, HB integrates desaturation depth and duration, better capturing the cumulative physiological burden of intermittent hypoxemia [44,46,52]. Notably, RCS analysis revealed that the MeanSpO₂ risk curves plateaued within the 88%-92% range, indicating that mean values fail to capture transient desaturation events and lack diagnostic sensitivity without complementary parameters [53]. TotalPower reflects global signal fluctuations without a specific mechanistic link to respiratory events [54]. These nonlinear parameter-risk relationships explain the suboptimal performance of linear models such as LR [10,23,44,55]. The steep threshold effects for ODI and HB suggest that even mild OSA can trigger substantial risk escalation, supporting the hypothesis of a critical threshold for hypoxic exposure [56,57].

We further quantified interactions among multiple oximetric indices. While ODI tracks event frequency, it neglects desaturation depth and duration, failing to capture physiological nuances of OSA heterogeneity [10,57]. Our 2-parameter ODI-HB CatBoost model achieved an F₁-score of 0.9472. By characterizing both temporal and intensity dimensions, this synergy explains why ODI, as the strongest predictor, requires HB integration to improve performance [57,58]. The superior performance of our primary model, ODI-HB-MinSpO₂-ST90 (F₁-score = 0.9516, AUC = 0.9879), stems from the inherent complementarity of these parameters. MinSpO₂ captures severe hypoxic nadirs [16], while ST90 quantifies nocturnal hypoxemia duration [59]. Together, this 4-parameter ensemble facilitates comprehensive multidimensional phenotyping of OSA, encompassing the frequency, depth, duration, and cumulative burden of hypoxic events [57,60]. In terms of benchmarking, our model outperformed the multidimensional oximetry approach proposed by Kong et al [8] (AUC=0.939) and the least squares boosting model (AUC=0.889-0.924) developed by Gutiérrez-Tobal et al [23]. Moreover, integrated models incorporating demographics, questionnaires, and facial photography achieve AUCs ranging from 0.88 to 0.89 [1,61], whereas our oximetry-only approach achieved superior accuracy without auxiliary clinical data. This suggests multidimensional oximetry indices serve as effective PSG surrogates, encapsulating more direct pathophysiological information than traditional clinical markers [4].

ML algorithms demonstrate variable performance in OSA screening [62]. LR yielded an AUC of approximately 0.77 [63], whereas SVM exhibited the highest accuracy exclusively for mild OSA [62]. RF reached 84.4% accuracy in predicting severe OSA but remained insensitive to complex feature interactions [62,64]. In contrast, gradient boosting frameworks excelled at handling heterogeneous interactions, class imbalances, and missing data [4,23]. By implementing ordered boosting, CatBoost effectively mitigates gradient bias, thereby enhancing model robustness and consistently outperforming established benchmarks such as XGBoost and LightGBM [4,65,66]. Our study represents the first application of CatBoost to multidimensional oximetry-based OSA screening, underscoring its capacity to resolve complex nonlinear relationships [24,25]. Notably, performance did not improve with 5 or more oximetry parameters. Indiscriminate feature addition introduces multicollinearity and overfitting without incremental gains [4]. By prioritizing core parameter selection over feature stacking, our model ensures both robustness and clinical feasibility, providing a basis for developing portable screening devices [7,23].

Given the substantial phenotypic heterogeneity of OSA, sex-specific differences have been inadequately addressed in existing models, often leading to underdiagnosis among women [67-69]. To address this, we conducted stratified analyses by sex and age. In males and older subgroups, the ODI-HB-MinSpO₂-ST90 model proved optimal, consistent with our overall findings. This result likely attributable to our predominantly male sample (median age 45 years). Conversely, the ODI-HB-MinSpO₂-MeanSpO₂ configuration demonstrated superior performance in females. Patiño et al [70] reported that despite lower AHI values, women exhibit mean SpO₂ reductions comparable to those in men, while Poka-Mayap et al [71] confirmed significantly lower mean SpO₂ levels in women with OSA, suggesting that the female OSA phenotype may be more closely linked to sustained hypoxemia. This may reflect heightened hypoxic sensitivity in women, where SpO₂ fluctuations at subclinical AHI thresholds are sufficient to induce end-organ damage [56]. Notably, in older adults, the ODI-HB-MinSpO₂-ST90 model achieved an exceptional AUC of 0.9950, surpassing that in younger participants. This enhanced accuracy likely stems from reduced hypoxic tolerance and pronounced SpO₂ variability characteristic of aging. Indeed, older patients, particularly older women, consistently exhibit lower SpO₂ levels independent of AHI burden [72].

Previous ML-based OSA diagnostic studies have focused predominantly on predictive performance, frequently neglecting to quantify parameter contributions, thereby limiting clinical trust and adoption of oximetry-based models [17,19,23]. We addressed this “black-box” limitation through SHAP values and feature importance analysis, which corroborated the central role of ODI and HB while revealing significant feature hierarchy shifts across subgroups. Notably, the importance of MeanSpO₂ increased substantially in the female subgroup, reinforcing hypotheses regarding sex-specific physiological signatures in which sustained hypoxemia may characterize the female phenotype more prominently [56,69,71]. Furthermore, our CatBoost-based model maintained exceptional performance across independent external validation sets, with sex- and age-specific models demonstrating sustained stability, attributable to CatBoost’s ability to mitigate gradient bias [73]. Despite the relatively small non-OSA control group in the external validation cohort (n = 76), the high precision maintained across populations indicates strong potential for cross-cohort generalizability [7,23,74]. This robustness suggests that ML-integrated multidimensional oximetry can streamline OSA screening protocols and reduce reliance on resource-intensive PSG [75].

This study has several limitations. First, despite the large sample size, participants were recruited from a single center and comprised individuals referred for suspected OSA. This may have resulted in a higher OSA prevalence than in the general population, thereby increasing the pretest probability and potentially overestimating diagnostic performance. Second, the lack of longitudinal follow-up precludes assessment of the model’s temporal consistency or long-term predictive capacity. Third, our cohort lacked racial and ethnic diversity, and as skin pigmentation can introduce systematic biases in SpO₂ measurements, this may limit generalizability to individuals with darker skin tones [76]. Finally, the single-center design and absence of prospective community-based validation restricts broader external validity.

In conclusion, we developed and validated a robust CatBoost-based multidimensional oximetry model that enables accurate OSA screening. Although the ODI-HB-MinSpO₂-ST90 combination demonstrated optimal performance in the general population, substituting ST90 with MeanSpO₂ proved superior for female and younger subgroups. By integrating nocturnal hypoxemia indices spanning frequency, depth, and duration, our approach overcomes the limitations of single-parameter screening and offers multidimensional physiological assessment beyond conventional AHI-centric methods. These models can be readily integrated into portable monitoring devices or wearable technologies to facilitate early OSA diagnosis. Future research should prioritize multicenter prospective trials and multi-ethnic validation studies to establish standardized protocols for personalized OSA risk stratification.