Explainable Machine Learning Assists in Revealing Associations Between Polysomnographic Biomarkers and Incident Type 2 Diabetes in Men

Introduction

Obstructive sleep apnoea (OSA) and type 2 diabetes (T2D) show bidirectional relationships that potentially exacerbate each other’s effects on overall health.^1,2 Although the pathophysiological mechanisms linking OSA and diabetes are not fully understood, chronic intermittent hypoxia, inflammation, sleep fragmentation and increased sympathetic drive could potentially all contribute to dysregulation of glucose metabolism.^1,3,4

The relationship between OSA severity and T2D has been studied extensively and multiple OSA severity markers derived from polysomnography have been found to be associated with prevalence and incidence of T2D. For example, the apnoea hypopnea index (AHI), a frequency count of respiratory airflow reduction (hypopnea) or complete cessation (apnoea) due to upper airway collapse during sleep and the most common index for quantifying OSA severity, has shown modest associations of moderate-severe with incident diabetes.^1,5–8 However, we and others have reported no association of OSA severity based on the AHI with diabetes incidence.^9–11 Other markers related to OSA severity and polysomnographic measures also shown to be associated with T2D include oxygen saturation metrics,^5,9,10,12 AHI in rapid eye movement (REM),¹³ time spent in REM sleep¹⁴ and total sleep time.⁵ AHI in rapid eye movement (REM) sleep has been associated with glycaemic control in diabetes¹³ and insulin resistance.¹⁵ Previous studies only have focused on the association between limited markers of OSA severity and type 2 diabetes, with no systematic search or cross-validation of the best biomarkers. Additionally, traditional statistical models constrained by underlying assumptions regarding linear relationships, distribution shapes and model structures may mask potentially complex effects of biomarkers that may interact with other variables. For example, relationships between biomarkers and multiple outcomes may not be linear and may be co-dependent on other variables such as gender, body habitus and potentially a wide range of inter-related variables. Applying arbitrary cutoffs for biomarkers to help address potential nonlinear effects is unlikely to fully resolve this issue. These gaps in the literature underscore the need for new approaches that can systematically test for more complex non-linear and co-dependent relationships between variables to help reveal and rank the most important biomarkers to overcome the key limitations with traditional statistical models.

Explainable machine learning (ML) has the potential to systematically discover novel biomarkers, as it has demonstrated its power in drug discovery,¹⁶ critical illness,¹⁷ mental health¹⁸ and hypertension in OSA.¹⁹ In contrast to traditional ML approaches where the accuracy of prediction is the main priority, explainable ML focuses on both the accuracy and the transparency of how the model predicts the outcome. This transparency is completed by comprehensively looking into individual predictions (local explanations) to understand the global behavior of the ML model.²⁰ This state-of-the-art explainable ML goes further than traditional ML by revealing not only the importance of the biomarkers but also the nature and direction of the relationship between the biomarker and the outcome. This makes it particularly suitable for medical applications where a greater understanding of causal relationships and predictive power are typically of primary interest.

This study aims to investigate whether state-of-the-art explainable ML models can provide deeper insights into the relationship between polysomnographic features and the development of type 2 diabetes (T2D). The focus is threefold: first, to rank the importance of individual biomarkers; second, to move beyond ranking and explore how explainable ML can uncover novel relationships between sleep physiology and T2D risk; and third, to support hypothesis testing by offering interpretable insights into potential risk factors. The overall study design is shown in Figure 1.

Figure 1 Schematic demonstrating data-driven approaches for exploring the association between sleep biomarkers and type 2 diabetes. (a) Flow diagram of the study and follow-up assessments. (b) Biomarkers and features from different domains including sleep, breathing disorders, and demographics were extracted to input into the XGBoost machine learning model. (c) The trained model was passed through the AI Explainer framework for estimating the relative importance of biomarkers based on Shapley values. (d) The least important biomarker was removed from the data, and this process repeated until the top 15 most important biomarkers were found (d and e). (e) The output of the explainable ML model was used to reveal novel biomarkers (f), assist hypothesis testing (g), and gain insights into risk factors and assist treatment approaches (h).

Methods

Data Collection

Approval for the conduct of these Men Androgens Inflammation Lifestyles Environment and Stress (MAILES) study assessments was provided by the Central Adelaide Local Health Network Human Research Ethics Committee, with approval numbers 02305H, 2010054, HREC/17/TQEH/74, HREC/15/TQEH/127 and H-2020-109. This study complied with the Declarations of Helsinki. All participants provided signed consent for participation in all stages of data collection. All data used in this report was de-identified. Detailed data collection methods for the MAILES baseline have previously been reported.²¹ Briefly, participants underwent clinical assessment (including anthropometry, sphygmomanometry, a fasting blood sample was drawn) and completed questionnaires regarding sociodemographics, doctor-diagnosed medical conditions, chronic disease risk factors, and medication use, The polysomnography (PSG) sub-study methodology [8-channel in-home unattended polysomnography (Embletta X100, Embla Systems, Colorado) measuring EEG, EOG, EMG, nasal pressure, thoracic and abdominal effort, oximetry, body position, and limb movements] is previously described.²² Briefly, the final number of participants included in this study was 536 (52 with incident diabetes vs 484 free of diabetes at the follow-up assessment, Figure 1a).

Outcomes and Biomarkers

As previously described,¹⁰ participants with pre-existing diabetes identified during the 2007–10 clinical assessments were excluded from the present analyses based on criteria including self-reported doctor-diagnosed diabetes, fasting plasma glucose ≥7.0 mmol/L, glycated haemoglobin (HbA1c) ≥6.5%, or use of diabetes medications. In those free of diabetes at baseline, participation in the three follow-up assessments after the PSG sub-study was used to ascertain incident diabetes as illustrated in Figure 1a. Incident cases were identified through one or more of the following measures:

1. Postal survey in 2015–16: self-report of doctor-diagnosed diabetes or use of diabetes medications.
2. Clinical follow-up from June 2018 to March 2020: self-report of doctor-diagnosed diabetes, use of diabetes medications, fasting plasma glucose ≥7.0 mmol/L or non-fasting plasma glucose ≥11.0 mmol/L, and glycated haemoglobin (HbA1c) ≥6.5%.
3. Postal survey in October 2020 to January 2021: self-report of doctor-diagnosed diabetes or use of diabetes medications.

The study identified T2D cases using a combination of standard clinical diagnostics and self-reported data. Of the 52 incident cases of diabetes, 18 (35%) were identified with an elevated plasma glucose or HbA1c at the second follow-up in 2018–20. Self-reported doctor diagnosed diabetes has been shown to be valid²³ and captures diabetes that is controlled through diet or medication use.

No data cleaning procedures were applied to this dataset for the current study. A full list of biomarkers was shown in Supplementary Table S1.

Machine Learning Model

We used the XGBoost machine learning model²⁴ to predict the incidence of type 2 diabetes (Figure 1b and c). XGBoost was chosen because it outperforms other ML models, particularly for tabular data, and requires less tuning.²⁵ This is important as it significantly reduces the computational time required for iterative feature selection in this study to address multicollinearity. Model hyperparameters, such as maximum depth and learning rate, were determined through five-fold cross-validation. The final values selected were a max_depth of 16 and a learning_rate of 0.23. Additionally, to address the imbalance in the number of incident type 2 diabetes cases (52 vs 484), we applied a weight to the positive class using scale_pos_weight = 9. Imbalanced data were handled by assigning a higher weight to the minority (positive) class in the loss function.²⁶ This ensures that misclassified rare events are penalized more heavily, helping the model focus on them. This approach does not require a resampling method, making it suitable for the XGBoost model.²⁷

Local Explanation and Shapley Values

In our study, the output of the XGBoost model represents the log odds of a participant developing type 2 diabetes. While the XGBoost model provides a final output, its decision-making process is opaque. To gain insight into how each biomarker contributes to the model output, we used TreeExplainer²⁰ to compute Shapley values, a parameter from game theory used to infer the contribution of each feature to the outcome (Figure 1c). These Shapley values were calculated for each participant based on their biomarker values. Shapley values assess all possible combinations of biomarkers and quantify each biomarker’s marginal contribution by measuring how the prediction changes when a feature is included or excluded. Since computing Shapley values is computationally expensive, we used TreeExplainer to approximate them,²⁰ making the approach feasible for datasets with a large number of features and complex models. Since the output of the XGBoost model is in log odds, the Shapley values in our study represent the change in log odds of incident diabetes.

We ranked the importance of each feature by calculating the mean absolute value of the Shapley values for each biomarker. Higher values indicate more important biomarkers (Figure 1d and e). Additionally, by examining individual Shapley values for each feature, we can clearly examine the contribution of each biomarker to the final prediction for a specific participant (Figure 1f–h).

Iterative Feature Selection

We used an iterative process to rank the importance of biomarkers. In the iterative feature selection process [1], the XGBoost model was initially trained on all 299 features (N), and feature importance was ranked using Shapley values. The least important feature was then removed, and the model was retrained with the remaining N-1 features. This process was repeated iteratively, removing the least important feature at each step and retraining model, until only 15 features remained, ranked from highest to lowest importance. The issue with correlated features is that they tend to underestimate each other’s contribution when included together in a machine learning model. If all correlated features are removed at once, it can result in the complete elimination of important variables. In contrast, the iterative process eliminates only one correlated feature at a time and reassesses the importance of the remaining features by retraining the model. This allows for a systematic reassessment of the contribution of each feature, reducing the impact of multicollinearity, particularly in high-dimensional datasets.²⁸

Statistical Analysis

Traditional statistical models with adjusted factors were used to demonstrate the utility of explainable ML in assisting hypothesis testing. The adjusted model also aimed to help further test the reliability of the ML model in extracting independent predictors of T2D. We utilized logistic regression to examine the relationship between novel biomarkers and the incidence of diabetes. Our model included covariates of age, glucose concentration, testosterone levels, and waist circumference at baseline. To automatically address potential nonlinear relationships between these covariates and the outcome, we employed a generalized additive model (GAM),²⁹ where the covariates are modeled as smooth functions. However, the independent variables in our study were assumed to be linear. The detailed equations of the models are as follows:

Model 1

Model 2

Model 3

Model 4

In which, is a logit function, is the intercept and is a spline function to account for nonlinear relationships. Ndesat is the continuous number of nonsupine desaturation events ≥3%. Ndesat_4categories is four quartile categories. Ndesat_2categories is a categorical variable determined by applying the explainable ML threshold. Similarly, HR_2categores and Hypopnea duration_2categories are categorical variables determined using application of each explainable ML threshold.

Results

Participants in the longitudinal study, who underwent 8 channel home-based polysomnography in 2010–11 (n=824) and were free of diabetes at baseline (n=682) were eligible for inclusion in the analysis. Of these, 78.6% (n=536) completed at least one follow-up assessment, between 2015 and 2021. A total of 536 participants (52 diabetes cases) were included in the analysis (Figure 1a). Demographic, health, and polysomnographic characteristics of the sample overall are shown in Table 1. The cohort comprised of middle aged and older men (aged 41.6–87.2 years) of who 82% reported post high school education. Participants had high rates of overweight or obesity (79.2% with BMI ≥25.0), hypertension and poor sleep quality. At least mild OSA (AHI≥10/hr) was present in 49.8%, and severe OSA, (AHI ≥30/hr) was present in 9.6%.

Table 1 Demographic, Health, and Polysomnographic Characteristics of the Sample Overall and in Relation to OSA Severity

Discovery of Novel Biomarkers

As expected, the strongest biomarker associated with the incidence of T2D was baseline plasma glucose concentration (Figure 2a). Among the top five most predictive biomarkers, apart from glucose and waist circumference, we identified three novel biomarkers including the total number of 3% desaturations in non-supine sleep, mean heart rate (HR) in the supine position, and mean duration of hypopnea. In the top 15 biomarkers, we also identified baseline glycated haemoglobin (HbA1c), and testosterone level. Other novel biomarkers not previously reported in the literature also emerged across various domains, including descriptive measures of sleep (ie, total sleep time, sleep on the left side), sleep breathing disorder (snoring), and oxygen saturation metrics (ie, mean SaO₂ levels in REM sleep and wake). Traditional AHI was ranked in 81st. The full list of biomarkers ranked from most to least important is shown in Supplementary Table S1.

Figure 2 Explainable ML reveals novel biomarkers associated with diabetes. (a) Top 15 strongest biomarkers associated with incident T2D. The ranking is based on the mean absolute Shapley values (see Methods for details). Colors indicate the values of biomarkers categorized into four quartiles. A positive change in log odds indicates a higher risk of T2D. (b) The relationship between the top 15 biomarkers and the risk of T2D. The change in log odds is indicated on the y-axis, while the change in biomarker values is shown on the x-axis. The black dashed lines represent smooth curves across the data points to illustrate the overall shape of the relationship. Dot points indicate data from 536 participants.

Relationship Between Top Biomarkers and Diabetes

To explore the direction and shape of relationships between the top biomarkers associated with T2D risk, we used partial dependence plots, where the relative change in log odds of incident diabetes is plotted against the range of values of a given biomarker (Figure 2b). Among the top five biomarkers, for example, we observed a sigmoid-shaped relationship between glucose concentration levels, mean HR in the supine position, and mean hypopnea duration. The number of non-supine SaO₂ desaturations and waist circumference exhibited more distinct step changes indicative of a strong threshold of risk. As shown in Figure 2b, specifically, by utilizing the zero crossing of log odds change, we determined the threshold for non-supine SaO₂ desaturation as 19 events and a male waist circumference of 96.4 cm. Incident diabetes was associated with shorter hypopnea duration and lower SaO₂ levels in wake or REM sleep while a typical U-shaped relationship with total sleep time was also evident (Figure 2b). Higher testosterone levels were associated with lower T2D risk compared to lower testosterone levels.

Explainable Machine Learning to Assist Incident T2D Prediction

We further investigated if explainable ML could be useful to predict future incident T2D. Specifically, we tested the statistical significance of novel biomarkers to predict incident T2D onset using a logistic regression model adjusted for age, serum glucose and testosterone concentrations and waist circumference measured at baseline (see Methods for details). Model 1, which used the continuous value of non-supine SaO₂ desaturation events at baseline to predict future incident T2D, failed to reject the null hypothesis (Table 2). This is expected as the relationship between the biomarkers and risk is nonlinear as observed in Figure 2b. To help deal with nonlinear relationships, we applied a traditional approach to categorize continuous values into quartiles (Model 2). However, model 2 also failed to reject the null hypothesis (Table 2). We further performed Model 3, where non-supine desaturation events were categorized based on the threshold identified by explainable ML (≥19 events) as shown in Figure 2b. The model showed a strong association between the number of desaturation events and incident T2D (Odds ratio = 2.41 [95% CI 1.20 to 4.82], P-value = 0.013, Table 1). We finally added mean HR in REM sleep and mean hypopnea duration as categorized using the identified thresholds in Figure 2b, which all showed better model fits and strong associations of these novel biomarkers with incident diabetes (Model 4, Table 2).

Table 2 Explainable Machine Learning Assisting Hypothesis Testing Association Between Diabetes and OSA Biomarkers

Precision Medicine Approach to Risk Factor Assessment

We further explored the detailed output of explainable ML by examining two typical scenarios where participants initially exhibited similar glucose concentration levels at baseline but subsequently showed different outcomes at follow-up assessments. In Figure 3a and b, we considered two typical participants who had normal serum glucose concentration levels at baseline (4.8 mmol/L). However, other risk factors were associated with the outcome. Figure 3a shows biomarkers that were strongly associated with a very low risk of T2D, while an unfavorable range of PSG and an inflammatory biomarker (tumor necrosis factor alpha) placed the participant in Figure 3b at high risk of T2D, despite a healthy waist circumference (88.3 cm).

Figure 3 Comparison of typical scenarios leading to different outcomes at follow-up assessment. (a and b) Comparison of two participants with serum low glucose concentrations at baseline. (c and d) Comparison of two participants with high serum glucose concentrations at baseline. Red and blue colors indicate increased or decreased risk, respectively. The baseline risk for all participants was calculated as the mean of the model output, indicated by green dashed lines (log odds = −0.193), while the final risk outcome is indicated by red dashed lines after adding the additive contributions of all other risk factors.

In the second scenario as shown in Figure 3c and d, high baseline glucose concentrations were observed in two participants. The risk of T2D for the participant in Figure 3c was associated with a range of sleep biomarkers. This participant was free of T2D at the follow-up assessment despite a high waist circumference (104.7 cm). Conversely, the participant in Figure 3d, with an unfavorable range of risk factors, combined with a high fasting glucose concentration, placed them into a high-risk category consistent with their T2D diagnosis at follow-up. For this participant, several biomarkers associated with elevated T2D risk could potentially be targeted to reduce risk.

Discussion

This study applied an explainable ML approach to explore associations between OSA biomarkers and incident T2D over around 8 years. This approach detected both well established and novel biomarkers associated with increased incident T2D risk. We further demonstrate that the explainable ML approach substantially benefits T2D risk factor identification and cut-off selection otherwise missed by our more traditional statistical modelling approaches¹⁰ due to non-linear relationships and sharper transition points than reliably discernible via inter-quartile ranges. This more detailed risk assessment approach may have important clinical applications to improve T2D risk identification and management and help to identify potential underlying mechanisms and more effective risk-modifying interventions suggested by this approach. The novelty of our analytical approach lies in our use of ML not just for prediction tasks, but to gain insights into the relationship between high-dimensional polysomnographic features and incident T2D. This opens new opportunities for exploring these novel biomarkers and potentially advancing personalized medicine approaches.

Fasting serum glucose, HbA1c,³⁰ serum testosterone^31,32 and waist circumference³³ are established biomarkers and were identified in the top 15 biomarkers, supporting the validity of this approach. We also observed mean hypopnea duration to be associated with incident diabetes, consistent with recent findings that shorter apnoea-hypopnea event duration was associated with all-cause mortality risk in the Sleep Heart Health Study.³⁴ The presence of hypopnea index in REM sleep in the top 15 biomarkers is similar to findings of the DREAM Study of US Veterans.³⁵ Ding et al reported that a cluster analysis derived PSG phenotype characterised by frequent events in REM sleep with hypoxia was associated with incident diabetes independent of age, sex, race, fasting glucose, however, the association did not persist with additional adjustment for BMI and BMI change. Our finding that mean HR in the supine position predicted incident diabetes is also consistent with the findings of⁵ who showed an association between incident diabetes and mean total sleep time and heart rate. Even with established biomarkers associated with T2D we were able to more comprehensively evaluate the nature of the relationships via explainable ML. For example, we identified a sharp step-like transition point in the relationship between the odds of incident T2D and waist circumference. Based on the profile of the relationship, we identified that participants with a waist circumference above 96 cm were at significantly higher risk of incident T2D compared to those with a smaller waist circumference. This is comparable with previous research that identified a waist circumference above 94 cm for men as indicative of an increased risk of developing T2D³³ and guideline waist circumference values of ≥94 cm and 102 cm for men based on increased risk and substantially increased risk of chronic conditions respectively (WHO 2000). Another example was fasting serum glucose concentrations which showed a sigmoidal relationship with T2D risk. Our results identified baseline glucose levels above 5.45 mmol/L as being associated with significantly increased risk of incident T2D at follow-up, consistent with the threshold typically used to define normal fasting plasma glucose levels as less than 5.55 nmol/L.³⁶ This finding is striking because explainable ML not only confirmed known biomarkers linked to T2D but also identified critical thresholds for evaluating increased risk. While well-controlled studies are required to confirm the utility of identified novel biomarkers and the identified thresholds of elevated risk, a greater understanding of the strength and directions of individual and multiple biomarker relationships with T2D could be valuable for clinicians and patients to target preventative interventions to help reduce overall risks.

Furthermore, we demonstrated proof-of-concept to support the utility of multivariable risk prediction models guided by explainable ML. Epidemiological studies commonly categorize continuous variables using quartiles to help account for potential non-normal or non-linear relationships. However, our observation of a sharp sigmoidal relationship between non-supine desaturation events and increased odds of incident T2D at a cut-off ≥19 events identified by explainable ML was not discernible between quartiles. While nonlinear models are advocated for use instead of categorizing variables,³⁷ they may not model nonsmoothed relationships well.²⁹ Data-driven optimal cut-points could potentially be bias towards fitting relationships that may not necessarily exist.³⁸ Thus, as with any approach, data-driven methods require independent validation to confirm the presence of reliably reproducible relationships. Nevertheless, explainable ML clearly has significantly improved power to detect complex multivariable effects not easily discernible through more traditional methods, but still requires further testing of novel biomarkers on independent data.³⁸

Our study has limitations that could be addressed in future studies. The thresholds established using data from middle-aged and older men, may not generalize to women or younger populations. This is especially relevant for variables such as waist circumference, which likely require sex-specific cut-off points. Whether the specific cut-off threshold for non-supine desaturation events applies to other demographic groups is also unknown. Men with OSA show worse beta-cell function compared to women,³⁹ further supporting sex-specific differences in type 2 diabetes risk for which different thresholds may well need to be applied. Overfitting issues might result in identifying irrelevant biomarkers. To mitigate this, we used Shapley values for feature ranking, which provides a more interpretable and robust alternative to permutation-based methods. The model’s correct identification of known T2D biomarkers further supports its reliability and reduces overfitting risk. Although strategies to reduce overfitting were applied, a limitation of our study is the lack of independent datasets for validation. Thus, further cross-sectional and longitudinal validation studies remain needed in independent and more diverse cohorts, including women and young men, to determine the reliability of the identified biomarkers and thresholds.

High ranking novel biomarkers identified through this approach are hypothesis generating and clearly warrant further investigation. However, as with more traditional statistical methods, associations may not be causal and require cautious speculation and interpretation surrounding potential mechanisms. As reviewed by Landry et al,⁴⁰ patients with supine predominant or supine isolated OSA are younger and have a lower BMI and AHI compared to non-positional OSA patients. The association of 3% desaturations in non-supine sleep with incident diabetes may reflect in part the diabetes risk associated with older age, obesity and more frequent respiratory events. Given high rates of supine predominant OSA (50–60%) and supine isolated OSA (20–30%) in clinical sleep populations and epidemiological studies,⁴⁰ the persistence of OSA in the non-supine position may reflect a more severe OSA phenotype.

Our finding that mean HR in the supine position predicted incident diabetes is also consistent with the findings of Kendzerska et al⁵ who showed an association between incident diabetes and mean total sleep time and heart rate. Continuous sympathetic stimuli occurring at the end of respiratory events promotes increased the sympathetic drive and increased heart rate⁴¹ has been linked to insulin resistance.⁴² Higher incident diabetes risk with lower mean hypopnea duration in this study is consistent with increased all-cause mortality risk with shorter apnoea-hypopnea event duration in the Sleep Heart Health Study,³⁴ which the authors speculated may indicate a predisposition to augmented autonomic nervous system responses or increased ventilatory instability. The Multi-Ethnic Study of Atherosclerosis study also showed shorter hypopnea and apnoea events to be associated with lower circulatory delay, lower arousal threshold⁴³ so multiple potential mechanisms could potentially help explain novel associations.

It was of interest to observe a low ranking of AHI (81st) in the Shapley values. Associations of varying strength between incident diabetes and OSA severity assessed by the AHI have been reported.¹⁰ Previous studies have also reported no association of diabetes incidence with AHI including the Wisconsin Sleep Cohort Study^1,19 a study of US veterans.⁹ Nocturnal hypoxemia metrics have been consistently described as predictors of diabetes. It is also worth noting that AHI ranked only 81st in importance in explaining incident diabetes risk. This is perhaps not surprising given that AHI is a somewhat arbitrary and highly variable manually scored composite outcome of apnoea, hypopnea, desaturation and arousal events. Thus, the measure itself is problematic for identifying underlying pathophysiological mechanisms, and previous studies typically show only weak and inconsistent associations between AHI a range of clinical endpoints. Systematic ranking of all available candidate biomarkers head-to-head without the need a-prior assumptions, linear model constraints or overly restricted variable choices is a comprehensive and efficient approach to potential biomarker investigation.

In conclusion, explainable ML not only supports the relevance of established biomarkers but also identifies new biomarkers for T2D, which could play a key role in guiding future hypothesis testing and the development of more reliable and clinically applicable biomarkers. Although further validation is needed, these proof-of-concept findings underscore the value of explainable ML in prospective data analysis.