A machine learning model trained on brain structural and functional MRI data can identify young and middle-aged men with severe obstructive sleep apnea at 95% sensitivity, according to a new study in Sleep and Breathing.
Researchers from the Second Affiliated Hospital of Soochow University, in collaboration with Oxford University, developed a classifier that integrates gray matter volume measures, dynamic functional connectivity patterns, and basic demographics to flag patients with severe oxygen desaturation (oxygen desaturation index >= 30 events/hour).
What They Found
The team analyzed data from 111 young and middle-aged male patients, extracting 50 gray matter structural indices and 7 dynamic functional connectivity temporal metrics from resting-state MRI. Using a Random Forest algorithm for feature selection and a Support Vector Machine for classification, they built a model that:
- Accuracy: 70.91%
- Recall (sensitivity): 95.07%
- Precision: 72.25%
- F1-score: 81.68%
- AUC-ROC: 0.798
When feature selection was tightened to a variable importance threshold above 0.05, the model improved to 20 features and achieved an AUC of 0.813 with 97.50% recall and 73.60% accuracy, a net gain of 2.7 percentage points in accuracy.
The features selected by the model mapped onto brain regions and networks known to support memory and attention, including frontal and temporal cortical areas and default-mode network nodes. This convergence suggests the model is capturing neurobiologically meaningful patterns, the cumulative effect of chronic intermittent hypoxia on brain structure and connectivity, rather than statistical noise.
Why It Matters
Obstructive sleep apnea affects an estimated 936 million adults worldwide, with the majority undiagnosed. While polysomnography remains the gold standard, it is resource-intensive and inaccessible in many settings. A neuroimaging-based screening tool could help identify patients at highest risk of severe hypoxia-related brain injury, potentially triaging them for priority treatment.
The high sensitivity (95%) is particularly relevant for a screening tool: the model rarely misses a severe case, even if its overall accuracy is moderate. The fact that the model’s features align with known cognitive networks, OSA patients commonly report deficits in memory, attention, and executive function, adds face validity to the approach.
Limits
The study was limited to 111 patients from a single center, all male and predominantly middle-aged. The findings cannot be generalized to women, older adults, or pediatric populations without further validation. The model identifies severe OSA with severe desaturation specifically (ODI >= 30), not mild-to-moderate disease. A larger, multi-center, sex-balanced cohort would be needed before clinical deployment.
Bottom Line
Multimodal MRI analyzed by machine learning can identify severe OSA with high sensitivity. The brain regions highlighted by the model offer a window into the neurological toll of chronic hypoxia, and suggest that imaging-based screening may one day complement traditional sleep studies.
Source
Jing Wang et al. “Machine learning identification of neuroimaging signatures in young and middle-aged male patients with obstructive sleep apnea: a multimodal MRI study.” Sleep and Breathing, 2026 Jun 20;30(4):192. DOI: 10.1007/s11325-026-03740-w. PMID: 42322350.