Neuroscience-guided bilateral thalamic MRI deep learning for obstructive sleep apnea diagnosis: a dual-center translational study Authors: Chong Tao, Saiyue Yu, Lin Han, Qiyuan Cai, Lingzi Meng, Yan Jiang, Lin Wang Published: 2026 ABSTRACT BackgroundObstructive sleep apnea (OSA) remains substantially underdiagnosed, and recurrent intermittent hypoxia (IH) is a key pathophysiological feature of OSA and may affect thalamic function. We aimed to develop a neuroscience-guided MRI-based deep learning model for auxiliary OSA diagnosis.MethodsMale C57BL/6J mice were exposed to IH, and thalamic activation was assessed using immunofluorescence staining and in vivo calcium imaging. Clinical brain MRI and polysomnography (PSG) data were retrospectively collected from 760 adults across two centers. Center 1 was divided into training and validation cohorts, and center 2 served as an independent external test set. Based on the experimental findings, bilateral thalamic regions were segmented, and a dual-branch 3D CNN with Transformer-based feature fusion was developed.ResultsIH induced prominent thalamic activation in mice, and central lateral thalamic nucleus (CL) showed the strongest relative response among the examined thalamic subregions. Compared with the control group, the IH group showed a marked increase in c-Fos-positive cells in the CL (62.50 ± 11.67 vs. 9.67 ± 4.89, n = 6, P < 0.0001). Further calcium signal recordings showed that calcium activity in the CL was significantly increased within the 1-minute time window after stimulation. In the human MRI cohort, the bilateral thalamic deep learning model achieved AUCs of 0.944 (95% CI, 0.926–0.960), 0.928 (95% CI, 0.893–0.957), and 0.911 (95% CI, 0.856–0.957) in the training, validation, and external test sets, respectively, outperforming the unilateral models (P < 0.001).ConclusionsIH-induced thalamic activation, with CL serving as a representative highly responsive thalamic subregion, provided experimental support for thalamic ROI selection. The bilateral thalamic MRI model provides a biologically grounded and externally validated imaging-based approach that may assist OSA identification.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12888-026-08472-5. Introduction Obstructive sleep apnea (OSA) is a common yet underdiagnosed sleep disorder characterized by recurrent upper airway collapse during sleep, which leads to intermittent hypoxia, sleep fragmentation, sympathetic overactivation and multisystem injury [1–3]. Previous studies have shown that OSA is closely associated with a range of adverse outcomes, including hypertension, coronary heart disease, stroke, metabolic disorders, and cognitive decline [4–6]. Despite its high prevalence and multisystem involvement, the clinical diagnosis rate of OSA remains low. Previous studies have reported that among patients with moderate-to-severe sleep apnea identified by polysomnography (PSG), more than 80% had not received a clinical diagnosis before screening [7, 8]. Early diagnosis and treatment of OSA may relieve excessive daytime sleepiness, fatigue, and impaired attention, thereby improving quality of life [9]. At present, PSG remains the gold standard for the diagnosis of OSA [10]. However, its widespread use in routine clinical practice is limited by several factors. First, PSG requires specialized equipment, a standardized sleep-monitoring environment, and trained technical personnel, and therefore is available in only a limited number of hospitals on a routine basis. Second, PSG is relatively costly and cumbersome, and patients are typically required to wear multiple sensors during the examination, which may cause discomfort and reduce acceptance. Together, these limitations may delay standardized diagnosis in patients with OSA and consequently hinder early detection and intervention. Compared with PSG, magnetic resonance imaging (MRI) is more accessible in clinical practice. Many patients have already undergone brain MRI during routine health examinations, neurological assessments, or screening for other conditions. Therefore, brain MRI acquired for other clinical purposes may provide an opportunity to develop models that assist OSA risk identification [11]. Such an approach may help prioritize individuals for formal PSG evaluation. From a pathophysiological perspective, one of the hallmark features of OSA is recurrent intermittent hypoxia (IH) during sleep [2, 12, 13]. IH can induce structural and functional abnormalities in the central nervous system [14, 15]. Previous studies have shown that the central medial thalamus region containing the central lateral thalamic nucleus (CL) is an important node in the regulation of arousal and consciousness [16]. The thalamus lies between the brainstem arousal system, hypothalamic arousal network, and cerebral cortex, and serves as a key hub for sleep-wake transitions, thalamocortical rhythms, and the integration of sensory and visceral information. The thalamus plays a critical role in sleep-wake transitions, and the central thalamus contributes to forebrain arousal regulation through its widespread interactions with cortical and subcortical networks [17]. Although the involvement of the thalamus in sleep–wake regulation has long been recognized, the key question in the context of OSA is not whether the thalamus participates in sleep, but whether recurrent IH and microarousals preferentially engage thalamic-related circuits and generate imaging features that can be captured by clinical MRI. However, direct translational studies linking hypoxia-driven thalamic activation, sleep–wake regulation, and deep learning-based clinical modeling remain limited. Therefore, we proposed a novel deep learning-based approach. We first experimentally verified the activation of the thalamus and its subregions under IH conditions in a mice model. In the human study, routine clinical MRI could not reliably and stably segment small thalamic nuclei such as the CL. Therefore, we did not equate the animal CL findings with changes in a single nucleus on human MRI; instead, we used the whole bilateral thalamus as a reproducible clinical imaging ROI. We then systematically evaluated a dual-branch three-dimensional convolutional neural network (3D CNN) incorporating a Transformer-based feature fusion module. This design reflects a cross-scale translational logic: animal experiments provide neuroscience evidence that the thalamus is involved in OSA-related IH responses, whereas the human MRI model tests whether thalamic imaging features can support OSA identification. Thus, this study aimed to develop a neuroscience-guided bilateral thalamic MRI deep learning model and evaluate its diagnostic value for OSA in a dual-center cohort. Materials and methods Animals Male C57BL/6J mice, aged 8–12 weeks and weighing 18–20 g, were purchased from Jinan Pengyue Laboratory Animal Technology Co., Ltd. After arrival, the mice were allowed to acclimate to the animal facility for approximately 1 week. During this acclimation period, the mice gained weight, and male mice weighing approximately 22–25 g were selected for stereotaxic viral injection, optical fiber implantation, and subsequent calcium imaging experiments. This weight range was considered more suitable for stereotaxic surgery and stable optical fiber implantation. All animals were housed in a specific pathogen-free facility under controlled environmental conditions, including a temperature of 22 ± 1 °C, a humidity of 55 ± 5%, and a 12-h light/12-h dark cycle (lights on at 7:00 a.m.). Cages were changed once weekly. The mice had free access to standard laboratory chow and sterilized water. All experimental procedures were approved by the Animal Ethics Committee of the Affiliated Hospital of Qingdao University (approval No. QYFY WZLL 42353). Intermittent hypoxia The mice were placed in a hypoxia chamber connected to a KY-2 F oxygen controller (Hangzhou Peijia Instrument Equipment, China). To minimize the potential influence of environmental stress on the experimental results, the mice, together with their home cages, were placed into the chamber. During a 3-day acclimation period, the device was turned on at 7:00 a.m. and supplied with air at a normal oxygen concentration, and then turned off at 7:00 p.m [18, 19]. Beginning at 7:00 a.m. on day 4, the IH protocol was initiated as follows: the fraction of inspired oxygen decreased from 21% to 6% over 30 s, remained at 6% for the next 30 s, then returned to 21% over 30 s and was maintained at 21% for another 30 s. Each cycle lasted 120 s [20]. Immunohistochemistry Mice were anesthetized by intraperitoneal injection of sodium pentobarbital, transcardially perfused with 20 mL of 0.9% saline, and then perfused with 4% paraformaldehyde. The brains were removed and fixed in 4% paraformaldehyde at 4 °C for 12 h, followed by dehydration in 30% sucrose solution at 4 °C until the tissue sank. Brain sections were cut into four series at a thickness of 30 μm using a cryostat (CM1950, Leica, Germany) and collected in PBS. For c-Fos staining, the sections were rinsed three times in PBS for 5 min each to remove residual OCT embedding medium. The sections were then blocked with immunostaining blocking solution for 1.5 h on a shaker at room temperature, followed by incubation with the primary antibody at 4 °C for 24 h on a shaker (guinea pig anti-c-Fos, 1:1000; 226308, Synaptic Systems). After three washes in PBS (5 min each), the sections were incubated with the secondary antibody for 2 h at room temperature in the dark (Alexa Fluor 488 donkey anti-guinea pig IgG, 1:500; 706-545-14, Jackson ImmunoResearch). After rinsing in PBS, the sections were mounted with a DAPI-containing mounting medium (ab104139, Abcam). Fluorescence images were acquired using an Olympus VS120 microscope (Japan) [21]. Stereotaxic surgeries and viral injections Anesthesia was induced using an anesthesia machine (RWD Life Technology, China), with 4% isoflurane for induction and 2% isoflurane for maintenance. After anesthesia, the mice were secured in a stereotaxic apparatus (RWD Life Technology, China). The scalp was shaved and disinfected with povidone-iodine. Erythromycin ophthalmic ointment was then applied to protect the eyes. A midline scalp incision was made, the connective tissue was removed, and a cranial hole was drilled above the target brain region. A glass micropipette filled with mineral oil and mounted on a microsyringe injector (RWD Life Technology, China) was used to deliver 50 nL of virus at a rate of 10 nL/min, and the pipette was left in place for 10 min after injection [22]. In vivo optical fiber photometer recording To evaluate dynamic functional activity in the thalamic CL under IH conditions, the calcium indicator virus rAAV-hSyn-GCaMP6f (serotype: AAV2/9; titer: ≥ 2.00 × 10^12 VG/mL) was stereotactically injected into the CL region. All viruses were packaged by Brain VTA (Brain VTA Co. Ltd., Wuhan, China) and stored at − 80 °C before use. Under isoflurane anesthesia, the mice head was fixed in a stereotaxic apparatus. After routine disinfection, a midline scalp incision was made to expose the skull, and a burr hole was drilled above the target brain region. The coordinates for viral injection were as follows: anteroposterior (AP), − 1.30 mm; mediolateral (ML), 0.75 mm; and dorsoventral (DV), − 3.25 mm. The coordinates for optical fiber implantation were AP, − 1.30 mm; ML, 0.75 mm; and DV, − 3.15 mm. A glass micropipette was used to inject 50 nL of virus at a rate of 10 nL/min, and the pipette was left in place for 10 min after injection to minimize reflux. After viral injection, a 3-week expression period was allowed. Once stable GCaMP6f expression was achieved, calcium signals were recorded and analyzed under IH conditions. Study population and data sources This was a retrospective study. Adult patients from two centers who underwent both standard PSG and MRI during the same period were consecutively enrolled. The inclusion criteria were as follows: age 18–75 years; an interval of no more than 12 months between PSG and MRI [23]; and clear MRI images without obvious motion artifacts or susceptibility artifacts, allowing reliable thalamic segmentation. The exclusion criteria were as follows: known intracranial structural lesions affecting thalamic volume or signal, including tumors, cerebral infarction, intracerebral hemorrhage, and multiple sclerosis; a history of cranial surgery or radiotherapy; poor MRI image quality resulting in unreliable segmentation; and missing PSG data or technical quality not meeting the standards of the American Academy of Sleep Medicine (AASM). Patients with an interval of more than 12 months between PSG and MRI were also excluded. The diagnosis of OSA was based on the AASM guidelines, with the apnea–hypopnea index (AHI) measured by PSG used as the reference standard. This study was conducted in accordance with the Declaration of Helsinki and was approved by the Medical Ethics Committee of the Affiliated Hospital of Qingdao University (approval No. QYFY WZLL 42352). Written informed consent was obtained from all participants for the use of their clinical and imaging data for research purposes. All clinical, polysomnographic, and MRI data were anonymized before analysis. Patient identifiers were removed prior to model training and statistical analysis to protect participant privacy. MRI acquisition Clinical MRI data were collected from two centers. The center 1 was acquired using a Philips Medical Systems scanner, whereas the center 2 was acquired using a Siemens scanner. Detailed acquisition parameters are provided in Table 4. The original images were exported in DICOM format and converted to NIfTI format using dcm2niix for subsequent processing. Image preprocessing and spatial normalization All MRI images first underwent format harmonization and spatial normalization, as detailed below. FSL tools were used to unify the orientation matrices and coordinate systems of all NIfTI files to the RAS standard, thereby ensuring consistency across subsequent multistep processing. All modal images were then resampled to an isotropic voxel size of 1.0 × 1.0 × 1.0 mm3 using linear interpolation (trilinear interpolation) to reduce resolution differences caused by varying scanning parameters. In addition, the N4 bias field correction algorithm in ANTs (N4ITKBiasFieldCorrection) was applied to T1-weighted images to correct intensity nonuniformity and improve the stability of subsequent segmentation. Finally, HD-BET was used for skull stripping to remove non-brain tissues and reduce interference from extracranial signals in subsequent segmentation. Automatic bilateral thalamic segmentation and ROI extraction A pretrained automatic segmentation model based on the 3D U-Net architecture was used to perform three-dimensional volumetric segmentation of the bilateral thalami on the preprocessed T1-weighted images, generating a left thalamic mask (mask_L) and a right thalamic mask (mask_R), respectively. After segmentation, all masks underwent quality control review. Two neuroradiologists with 10 years of experience in brain MRI interpretation visually inspected the segmentation results. Cases with obvious boundary deviations were corrected semi-automatically, and the final results were determined by consensus. After the bilateral thalamic masks were obtained, three-dimensional morphological closing was applied to each mask using a spherical structuring element with a radius of 2 voxels to fill internal holes, and isolated small connected components of fewer than 200 voxels were removed to reduce segmentation noise. Subsequently, a region-of-interest image patch was extracted using the three-dimensional axis-aligned bounding box of the final mask as the center, with an additional 10 mm margin in each direction to preserve the boundaries of the peripheral thalamic nuclei and the surrounding microenvironmental information; regions extending beyond the image boundary were padded with zeros. The left and right thalamic image patches were cropped independently, and the target input size after cropping was 48 × 64 × 64 voxels (corresponding to D × H × W). If the bounding box with the added margin exceeded the target size, center cropping was performed around the mask centroid; if it was smaller than the target size, symmetric zero-padding was applied to ensure identical input sizes for the two branches. Intensity preprocessing was performed independently within each patient’s image patches. First, voxel intensities were clipped to the 1st–99th percentiles to remove extreme intensity outliers, followed by z-score normalization (mean = 0, variance = 1) to reduce intensity drift introduced by different scanners and acquisition parameters. Dataset splitting and class balancing strategy Data from center 1 were randomly split into the training and validation sets at a ratio of 7:3, whereas data from center 2 were used as an independent test set (Table 1). To address class imbalance arising from differences in the numbers of OSA-positive and OSA-negative samples, a positive class weight (pos_weight) was applied to the minority class in the loss function. Its value was determined according to the ratio of negative to positive samples in the training set, thereby increasing the gradient contribution of the minority class during training and reducing the model’s bias toward the majority class. Table 1Distribution of obstructive sleep apnea in datasetsCenter 1Center 2Training sets Normal196- OSA258- Total454-Val sets Normal85- OSA110- Total195-Test sets Normal-47 OSA-64 Total-111 Table 1Distribution of obstructive sleep apnea in datasetsCenter 1Center 2Training sets Normal196- OSA258- Total454-Val sets Normal85- OSA110- Total195-Test sets Normal-47 OSA-64 Total-111 Distribution of obstructive sleep apnea in datasets Data augmentation during training To effectively expand the training sample size and improve model generalizability, synchronized three-dimensional random augmentation was applied to the left and right thalamic image patches during training, meaning that the same random parameters were simultaneously applied to both branches. The main augmentation operations were as follows: (1) random flipping, with three-dimensional flipping along each axis performed with a probability of 0.5; (2) random rotation within a range of ± 10° along each axis; (3) random scaling, with scaling factors uniformly sampled from 0.9 to 1.1; (4) random translation within ± 8 voxels along each axis; (5) mild elastic deformation, implemented using a B-spline deformation field with a control point spacing of 20 voxels and a displacement standard deviation of 2 voxels to simulate soft-tissue morphological variation; and (6) intensity perturbation, including Gaussian noise addition (σ = 0.01), random contrast scaling (factor 0.9–1.1), and gamma transformation (γ∈[0.9, 1.1]). All of the above augmentation operations were applied only to the training set, whereas no augmentation was performed on the validation or test sets. Deep learning model architecture Overall architecture We proposed a dual-branch three-dimensional convolutional neural network (dual-branch 3D CNN) based on the bilateral thalami to jointly leverage the morphological and signal features of the left and right thalami for sleep apnea prediction. The network consisted of three components: a left thalamic encoding branch, a right thalamic encoding branch, and a feature fusion and classification module. The two encoding branches shared the same architecture but had independent parameters, allowing the model to learn side-specific representations of the left and right thalami separately and to capture complementary and asymmetric features through the subsequent fusion layer (Fig. 1). Fig. 1The overall workflow of deep learning models The overall workflow of deep learning models Encoder: 3D ResNet18 The encoder of each branch used a three-dimensional residual network (3D ResNet18) as the backbone. In terms of topology, 3D ResNet18 followed the standard ResNet18 architecture. The input first passed through a 3D convolutional layer with a stride of 2 (kernel size, 7 × 7 × 7; 64 output channels) and a 3D max-pooling layer for initial downsampling, followed by four stages of residual basic blocks for multi-scale feature extraction, with output channels of 64, 128, 256, and 512, respectively. Each residual basic block consisted of two 3D convolutional layers (kernel size, 3 × 3 × 3), each followed by batch normalization and a ReLU activation function. When the number of channels changed, a shortcut connection with a 1 × 1 × 1 3D convolution was used for dimensional matching. In each stage, the first residual block performed spatial downsampling through a 3D convolution with a stride of 2, thereby progressively enlarging the receptive field and reducing the feature-map resolution. All convolutional layers were followed by BN and ReLU to stabilize gradient propagation during training. To fully exploit the spatial prior information provided by the thalamic segmentation masks, mask-weighted global average pooling was performed on the high-level feature maps output from the fourth stage of the backbone network. The corresponding thalamic mask for each side was downsampled to the same spatial resolution as that feature map, and weighted global average pooling was performed only over voxels covered by the mask, thereby compressing the high-level feature map of each branch into a fixed-dimensional feature vector. The resulting vectors were then projected to a unified 256-dimensional space through a fully connected projection layer (Linear, output dimension 256, followed by BN and ReLU), denoted as F_L ∈ R^256 for the left thalamic feature and F_R ∈ R^256 for the right thalamic feature. Considering the sample size and the risk of overfitting, dropout (p = 0.5) was added after the projection layer, and weight decay was incorporated into the optimizer for regularization. Bilateral feature fusion module After the left and right thalamic feature vectors were obtained, a late-fusion strategy was adopted, and a Transformer encoder was introduced to explicitly model the symmetry and complementarity between the bilateral thalami. Specifically, the features from the two branches were treated as a token sequence of length 2, X = [F_L, F_R] ∈ R^(2 × 256). Before input into the Transformer encoder, learnable side embeddings were added to distinguish the left from the right side and to enable cross-side information interaction. The Transformer encoder comprised two layers with four-head multi-head self-attention (64 dimensions per head), and the hidden dimension of the feed-forward network was 512. Dropout (p = 0.5) was applied to both the attention and feed-forward layers, and residual connections together with LayerNorm were used to stabilize training. The output token sequence of the encoder was averaged along the token dimension to obtain a fused representation G ∈ R^256, which was then passed through a fusion MLP for dimensionality reduction (Linear 256→128, ReLU, Dropout 0.5), yielding a shared fused representation H∈R^128. Classification head and output A binary classification head (Linear 128→64, ReLU, Dropout 0.3, Linear 64→1) was attached after the shared fused representation H to output the OSA prediction logit z_OSA. After Sigmoid activation, the predicted probability was obtained as p_OSA = σ(z_OSA); samples with p_OSA > 0.5 were classified as OSA-positive (OSA+). Model training strategy A weighted binary cross-entropy loss with a positive class weight was used as the loss function: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{gathered}\mathcal{L}{\text{ }} = {\text{ }} - [{w^ + }{\text{ }}\cdot\cdot\cdot{\text{ }}y{\text{ }}\cdot\cdot\cdot{\text{ }}log(p\_OSA){\text{ }} \hfill \\\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, + {\text{ }}\left( {1{\text{ }} - {\text{ }}y} \right){\text{ }}\cdot\cdot\cdot{\text{ }}log(1{\text{ }} - {\text{ }}p\_OSA)] \hfill \\ \end{gathered} $$\end{document} where y is the ground-truth label (0 or 1), and w⁺ denotes the positive class weight, which was determined according to the ratio of negative to positive samples in the training set. The model was optimized using AdamW, with an initial learning rate of 1 × 10⁻⁴ and a weight decay of 1 × 10⁻². The mini-batch size was set to 4, and training was performed for 200 epochs. A cosine annealing schedule was used for learning rate adjustment, and early stopping was implemented based on the validation loss with a patience of 30 epochs. To further improve training stability, gradient clipping (maximum norm threshold = 5) and automatic mixed precision (AMP) training were applied to accelerate computation and reduce GPU memory consumption. The model was implemented in the PyTorch framework (version 2.0) and trained on an RTX 4060 Ti GPU. Model evaluation and statistical analysis Model performance was evaluated on the independent test set using the following metrics: area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The optimal classification threshold was determined based on the Youden index in the validation set and then fixed for application to the test set. The 95% confidence interval (CI) for the AUC was calculated using 1,000 bootstrap resamples. Differences in AUC between models, including unilateral branch models and the bilateral fusion model, were statistically compared using the DeLong test. All statistical tests were two-sided, and P < 0.05 was considered statistically significant. In addition, saliency maps were generated using the Shapley value method to improve understanding of the model decision-making process. Derived from the Shapley value concept in game theory, this method quantifies the importance of each image region by calculating its marginal contribution to the prediction outcome. Red regions indicate a greater positive contribution to the prediction, whereas blue regions indicate an inhibitory effect on the prediction. This visualization approach may help clinicians better understand the imaging features emphasized by the model, thereby enhancing its clinical interpretability and credibility. For animal experiments, data are presented as mean ± SEM unless otherwise stated. Each dot in the bar graphs represents one animal. For c-Fos quantification, values from multiple matched sections within the same brain region were first averaged for each animal, and the animal was used as the unit of analysis. For comparisons between the IH group and the Air control group within each brain region, two-sided Welch’s t tests were used because this method does not assume equal variances between groups. Significance annotations in the c-Fos bar graphs were based on the raw P values from Welch’s t tests. For calcium imaging, the stimulus-evoked peak ΔF/F values within the 1-min post-stimulation window were compared between groups using a two-sided Welch’s t test. Statistical significance was defined as P < 0.05. The significance levels were indicated as follows: *P < 0.05, **P < 0.01, and ***P < 0.001. Statistical analyses were performed using GraphPad Prism and Python. Results Intermittent hypoxia induced prominent thalamic activation To identify the brain regions activated by IH (Fig. 2a), we first assessed neuronal activation using c-Fos immunofluorescence staining (Fig. 2b). We then compared the acute brain-wide expression pattern of c-Fos following IH exposure. IH induced neuronal activation across multiple arousal, cognition, and hypoxia-related brain regions. Among the examined macroscopic brain regions, the whole thalamus showed the most prominent c-Fos response, with greater relative fold change and effect size than the mPFC, hypothalamus, and DG. Further analysis of thalamic subregions showed that the CL, CM, and PVT were all significantly activated, with the CL showing the highest relative fold change and largest effect size (Fig. S1). Therefore, the CL may serve as a representative thalamic nucleus for IH induced thalamic arousal-related responses. Compared with the Air control group, the IH group showed a significant increase in c-Fos-positive cells in the CL (62.50 ± 11.67 vs. 9.67 ± 4.89, n = 6, P < 0.0001) (Fig. 2c), indicating that neurons in this region were activated after IH exposure. To further investigate the neuronal response of the CL to IH stimulation, we recorded calcium signal changes in the thalamic CL of mice under IH using fiber photometry (Fig. 2d). Compared with baseline, calcium signals in the CL changed immediately after IH stimulation, whereas no obvious changes were observed in the control group (Fig. 2e). Quantification of the stimulus-evoked peak ΔF/F (%) showed a significant increase in the CL compared with the control group (10.54 ± 1.47 vs. 2.86 ± 6.39, n = 6, P = 0.002) (Fig. 2f). These findings indicate that intermittent hypoxia activated the thalamus. Fig. 2Intermittent hypoxia activates the thalamic CL region. (a) Schematic illustration of the intermittent hypoxia experimental protocol. (b) Immunofluorescence staining showing c-Fos expression in the CL after intermittent hypoxia stimulation. (c) Quantification of the number of c-Fos-positive cells in the CL. (d) Schematic of fiber photometry recording and the site of viral injection. (e) Average ΔF/F signal heatmaps (upper panel) and calcium transient traces (lower panel) after intermittent hypoxia and normoxic stimulation (n = 6). (f) Quantification of stimulus-evoked peak ΔF/F (%) (n = 6) Intermittent hypoxia activates the thalamic CL region. (a) Schematic illustration of the intermittent hypoxia experimental protocol. (b) Immunofluorescence staining showing c-Fos expression in the CL after intermittent hypoxia stimulation. (c) Quantification of the number of c-Fos-positive cells in the CL. (d) Schematic of fiber photometry recording and the site of viral injection. (e) Average ΔF/F signal heatmaps (upper panel) and calcium transient traces (lower panel) after intermittent hypoxia and normoxic stimulation (n = 6). (f) Quantification of stimulus-evoked peak ΔF/F (%) (n = 6) Baseline characteristics A total of 760 patients from two centers were included. Data from center 1 were randomly split into the training set (n = 454) and validation set (n = 195) at a ratio of 7:3, whereas data from center 2 were used as the external test set (n = 111). According to the AHI, participants were classified into the normal group (AHI < 5) and the OSA group (AHI ≥ 5) (Table 1). The two cohorts showed no statistically significant differences in age, sex, or body mass index (BMI) (Table 2). Further stratification by AHI severity showed that mild, moderate, and severe OSA in Center 1 included 159, 96, and 113 cases, accounting for 43.2%, 26.1%, and 30.7% of patients with OSA, respectively. In Center 2, mild, moderate, and severe OSA included 22, 19, and 23 cases, accounting for 34.4%, 29.7%, and 35.9% of patients with OSA, respectively. The distribution of OSA severity did not differ significantly between the two centers (Table 3). In addition, differences in MRI scanner vendors and acquisition parameters provided a relatively stringent test of the model’s cross-center generalizability (Table 4). Table 2Baseline demographic and clinical characteristics in the two centersCenter1NormalCenter1OSAP valueCenter2NormalCenter2OSAP valueNumber281368-4764-Age, years55.0 ± 14.356.2 ± 12.50.24853.5 ± 15.754.1 ± 11.50.820Male sex, n (%)209 (74.4%)281 (76.4%)0.56133(70.2%)47 (73.4%)0.708BMI, kg/m²28.9 ± 3.929.2 ± 4.10.37328.6 ± 3.629.2 ± 3.70.433Note: Data are presented as mean ± standard deviation or number (percentage), as appropriate. The normal group was defined as participants with an AHI < 5 events/h, and the OSA group was defined as participants with an AHI ≥ 5 events/h. Continuous variables were compared using the independent-samples Student’s t test when the assumption of equal variances was satisfied and Welch’s t test when variances were unequal. Categorical variables were compared using the Pearson χ² test. All statistical tests were two-sided, and P < 0.05 was considered statistically significant Table 2Baseline demographic and clinical characteristics in the two centersCenter1NormalCenter1OSAP valueCenter2NormalCenter2OSAP valueNumber281368-4764-Age, years55.0 ± 14.356.2 ± 12.50.24853.5 ± 15.754.1 ± 11.50.820Male sex, n (%)209 (74.4%)281 (76.4%)0.56133(70.2%)47 (73.4%)0.708BMI, kg/m²28.9 ± 3.929.2 ± 4.10.37328.6 ± 3.629.2 ± 3.70.433Note: Data are presented as mean ± standard deviation or number (percentage), as appropriate. The normal group was defined as participants with an AHI < 5 events/h, and the OSA group was defined as participants with an AHI ≥ 5 events/h. Continuous variables were compared using the independent-samples Student’s t test when the assumption of equal variances was satisfied and Welch’s t test when variances were unequal. Categorical variables were compared using the Pearson χ² test. All statistical tests were two-sided, and P < 0.05 was considered statistically significant Baseline demographic and clinical characteristics in the two centers Note: Data are presented as mean ± standard deviation or number (percentage), as appropriate. The normal group was defined as participants with an AHI < 5 events/h, and the OSA group was defined as participants with an AHI ≥ 5 events/h. Continuous variables were compared using the independent-samples Student’s t test when the assumption of equal variances was satisfied and Welch’s t test when variances were unequal. Categorical variables were compared using the Pearson χ² test. All statistical tests were two-sided, and P < 0.05 was considered statistically significant Table 3Distribution of obstructive sleep apnea severity across the two centersCenter 1n(% of total cohort)Center1% of OSA casesCenter2n(% of total cohort)Center2% of OSA casesNormal281(43.3%)-47(42.3%)-Mild OSA159(24.5%)(43.2%)22(19.8%)(34.4%)Moderate OSA96(14.8%)(26.1%)19(17.1%)(29.7%)Severe OSA113(17.4%)(30.7%)23(20.7%)(35.9%)Total OSA368(56.7%)(100.0%)64(57.7%)(100.0%)Total participate649(100.0%)-111(100.0%)-Notes: OSA severity was classified according to the apnea–hypopnea index as follows: normal, AHI < 5 events/h; mild OSA, 5 ≤ AHI < 15 events/h; moderate OSA, 15 ≤ AHI < 30 events/h; and severe OSA, AHI ≥ 30 events/h. Percentages in the “% of total cohort” columns were calculated using all participants in each center as the denominator. Percentages in the “% of OSA cases” columns were calculated using participants with OSA in the corresponding center as the denominator. The distribution of mild, moderate, and severe OSA among participants with OSA did not differ significantly between the two centers (Pearson χ²=1.754, df = 2, P = 0.416) Table 3Distribution of obstructive sleep apnea severity across the two centersCenter 1n(% of total cohort)Center1% of OSA casesCenter2n(% of total cohort)Center2% of OSA casesNormal281(43.3%)-47(42.3%)-Mild OSA159(24.5%)(43.2%)22(19.8%)(34.4%)Moderate OSA96(14.8%)(26.1%)19(17.1%)(29.7%)Severe OSA113(17.4%)(30.7%)23(20.7%)(35.9%)Total OSA368(56.7%)(100.0%)64(57.7%)(100.0%)Total participate649(100.0%)-111(100.0%)-Notes: OSA severity was classified according to the apnea–hypopnea index as follows: normal, AHI < 5 events/h; mild OSA, 5 ≤ AHI < 15 events/h; moderate OSA, 15 ≤ AHI < 30 events/h; and severe OSA, AHI ≥ 30 events/h. Percentages in the “% of total cohort” columns were calculated using all participants in each center as the denominator. Percentages in the “% of OSA cases” columns were calculated using participants with OSA in the corresponding center as the denominator. The distribution of mild, moderate, and severe OSA among participants with OSA did not differ significantly between the two centers (Pearson χ²=1.754, df = 2, P = 0.416) Distribution of obstructive sleep apnea severity across the two centers Notes: OSA severity was classified according to the apnea–hypopnea index as follows: normal, AHI < 5 events/h; mild OSA, 5 ≤ AHI < 15 events/h; moderate OSA, 15 ≤ AHI < 30 events/h; and severe OSA, AHI ≥ 30 events/h. Percentages in the “% of total cohort” columns were calculated using all participants in each center as the denominator. Percentages in the “% of OSA cases” columns were calculated using participants with OSA in the corresponding center as the denominator. The distribution of mild, moderate, and severe OSA among participants with OSA did not differ significantly between the two centers (Pearson χ²=1.754, df = 2, P = 0.416) Table 4Demographics and MRI characteristicsCenter 1(n = 649)Center 2(n = 111)Age, yr55.7 ± 13.353.8 ± 13.4Male sex, n(%)490(75.5%)80(72.1%)BMI, kg/m229.0 ± 4.028.9 ± 3.7ScannerPhilips Medical SystemsSIEMENSSlice thickness (mm)5.55Number of slices1824Section gap (mm)7.56Matrix336 × 336288 × 288TR (ms)550.0008220TE (ms)152.46Flip angle (°)6970Field strength3.0T3.0T Table 4Demographics and MRI characteristicsCenter 1(n = 649)Center 2(n = 111)Age, yr55.7 ± 13.353.8 ± 13.4Male sex, n(%)490(75.5%)80(72.1%)BMI, kg/m229.0 ± 4.028.9 ± 3.7ScannerPhilips Medical SystemsSIEMENSSlice thickness (mm)5.55Number of slices1824Section gap (mm)7.56Matrix336 × 336288 × 288TR (ms)550.0008220TE (ms)152.46Flip angle (°)6970Field strength3.0T3.0T Demographics and MRI characteristics Deep learning model performance Because the thalamus is a paired structure and the left and right sides are not fully symmetrical in anatomical connectivity and functional organization, we developed three deep learning models (Left thalamus, Right thalamus, and Bilateral thalami) to evaluate the unilateral characteristics of thalamic imaging information and the value of bilateral fusion. Specifically, the unilateral models were used to determine whether either thalamus had independent diagnostic value and to explore potential laterality-related differences in OSA-related thalamic imaging features, whereas the bilateral model was used to assess whether bilateral feature fusion could provide additional diagnostic gain. the unilateral thalamic models showed moderate discriminative ability in both the validation and test sets. In the training set, the AUCs of the left and right thalamic models were 0.851 (95% CI, 0.817–0.883) and 0.846 (95% CI, 0.810–0.878), respectively (Fig. 3a). In the validation set, the corresponding AUCs were 0.821 (95% CI, 0.760–0.876) and 0.798 (95% CI, 0.731–0.851), respectively (Fig. 3b). In the external test set, the AUCs were 0.790 (95% CI, 0.703–0.870) and 0.819 (95% CI, 0.740–0.887), respectively (Fig. 3c). In contrast, the bilateral thalamic model substantially improved discriminative performance, achieving AUCs of 0.944 (95% CI, 0.926–0.960), 0.928 (95% CI, 0.893–0.957), and 0.911 (95% CI, 0.856–0.957) in the training, validation, and external test sets, respectively (Fig. 3a–c). The corresponding accuracies were 0.850 (95% CI, 0.817–0.881), 0.836 (95% CI, 0.785–0.887), and 0.829 (95% CI, 0.757–0.892), respectively. The corresponding sensitivities were 0.833 (95% CI, 0.787–0.876), 0.855 (95% CI, 0.791–0.918), and 0.813 (95% CI, 0.719–0.906), respectively. The corresponding specificities were 0.872 (95% CI, 0.821–0.913), 0.812 (95% CI, 0.729–0.894), and 0.851 (95% CI, 0.745–0.936), respectively. The corresponding F1 scores were 0.863 (95% CI, 0.832–0.892), 0.855 (95% CI, 0.806–0.901), and 0.846 (95% CI, 0.780–0.905), respectively (Table 5). Fig. 3Comparison of the diagnostic performance of different thalamic models in the training, validation, and test sets. (a) ROC curves of the left thalamic, right thalamic, and bilateral thalamic models in the training set. (b) ROC curves of the left thalamic, right thalamic, and bilateral thalamic models in the validation set. (c) ROC curves of the left thalamic, right thalamic, and bilateral thalamic models in the test set. (d–f) Heatmaps of pairwise DeLong test comparisons among different models, showing the significance of AUC differences in the test, validation, and training sets, respectively. (g–i) Bubble plots of pairwise DeLong test comparisons among different models, showing the distribution of P values in the test, validation, and training sets, respectively. (j–l) Decision curve analysis of the three models in the training, validation, and test sets, used to evaluate the clinical net benefit across different threshold probabilities. (m–o) Calibration curves of the three models in the training, validation, and test sets, used to assess the agreement between predicted probabilities and observed event rates. Overall, the bilateral thalamic model demonstrated superior discrimination, calibration, and clinical utility across all datasets Comparison of the diagnostic performance of different thalamic models in the training, validation, and test sets. (a) ROC curves of the left thalamic, right thalamic, and bilateral thalamic models in the training set. (b) ROC curves of the left thalamic, right thalamic, and bilateral thalamic models in the validation set. (c) ROC curves of the left thalamic, right thalamic, and bilateral thalamic models in the test set. (d–f) Heatmaps of pairwise DeLong test comparisons among different models, showing the significance of AUC differences in the test, validation, and training sets, respectively. (g–i) Bubble plots of pairwise DeLong test comparisons among different models, showing the distribution of P values in the test, validation, and training sets, respectively. (j–l) Decision curve analysis of the three models in the training, validation, and test sets, used to evaluate the clinical net benefit across different threshold probabilities. (m–o) Calibration curves of the three models in the training, validation, and test sets, used to assess the agreement between predicted probabilities and observed event rates. Overall, the bilateral thalamic model demonstrated superior discrimination, calibration, and clinical utility across all datasets Table 5Performance of the deep learning modelsDatasetModelACC(95% CI)Sensitivity(95% CI)Specificity(95% CI)F1(95% CI)NPV(95% CI)PPV(95% CI)AUC(95% CI)TrainLeft_thalamus0.744(0.703–0.784)0.721(0.667–0.779)0.776(0.714–0.832)0.762(0.720–0.801)0.678(0.632–0.727)0.808(0.766–0.851)0.851(0.817–0.883)Right_thalamus0.749(0.707–0.789)0.764(0.706–0.814)0.730(0.668–0.791)0.776(0.737–0.812)0.700(0.652–0.749)0.788(0.749–0.830)0.846(0.810–0.878)Bilateral_thalami0.850(0.817–0.881)0.833(0.787–0.876)0.872(0.821–0.913)0.863(0.832–0.892)0.799(0.756–0.841)0.895(0.862–0.928)0.944(0.926–0.960)ValLeft_thalamus0.733(0.672–0.795)0.718(0.636–0.810)0.753(0.659–0.836)0.752(0.691–0.812)0.673(0.608–0.750)0.790(0.729–0.856)0.821(0.760–0.876)Right_thalamus0.713(0.651–0.769)0.691(0.600-0.773)0.741(0.647–0.824)0.731(0.667–0.787)0.649(0.584–0.716)0.775(0.714–0.843)0.798(0.731–0.851)Bilateral_thalami0.836(0.785–0.887)0.855(0.791–0.918)0.812(0.729–0.894)0.855(0.806–0.901)0.811(0.742–0.886)0.854(0.798–0.911)0.928(0.893–0.957)TestLeft_thalamus0.730(0.649–0.811)0.750(0.641–0.859)0.702(0.574–0.830)0.762(0.683–0.836)0.673(0.577–0.783)0.774(0.698–0.867)0.790(0.703–0.870)Right_thalamus0.739(0.658–0.811)0.719(0.609–0.828)0.766(0.638–0.872)0.760(0.679–0.835)0.666(0.583–0.767)0.807(0.724–0.893)0.819(0.740–0.887)Bilateral_thalami0.829(0.757–0.892)0.813(0.719–0.906)0.851(0.745–0.936)0.846(0.780–0.905)0.769(0.684–0.860)0.881(0.813–0.948)0.911(0.856–0.957) Table 5Performance of the deep learning modelsDatasetModelACC(95% CI)Sensitivity(95% CI)Specificity(95% CI)F1(95% CI)NPV(95% CI)PPV(95% CI)AUC(95% CI)TrainLeft_thalamus0.744(0.703–0.784)0.721(0.667–0.779)0.776(0.714–0.832)0.762(0.720–0.801)0.678(0.632–0.727)0.808(0.766–0.851)0.851(0.817–0.883)Right_thalamus0.749(0.707–0.789)0.764(0.706–0.814)0.730(0.668–0.791)0.776(0.737–0.812)0.700(0.652–0.749)0.788(0.749–0.830)0.846(0.810–0.878)Bilateral_thalami0.850(0.817–0.881)0.833(0.787–0.876)0.872(0.821–0.913)0.863(0.832–0.892)0.799(0.756–0.841)0.895(0.862–0.928)0.944(0.926–0.960)ValLeft_thalamus0.733(0.672–0.795)0.718(0.636–0.810)0.753(0.659–0.836)0.752(0.691–0.812)0.673(0.608–0.750)0.790(0.729–0.856)0.821(0.760–0.876)Right_thalamus0.713(0.651–0.769)0.691(0.600-0.773)0.741(0.647–0.824)0.731(0.667–0.787)0.649(0.584–0.716)0.775(0.714–0.843)0.798(0.731–0.851)Bilateral_thalami0.836(0.785–0.887)0.855(0.791–0.918)0.812(0.729–0.894)0.855(0.806–0.901)0.811(0.742–0.886)0.854(0.798–0.911)0.928(0.893–0.957)TestLeft_thalamus0.730(0.649–0.811)0.750(0.641–0.859)0.702(0.574–0.830)0.762(0.683–0.836)0.673(0.577–0.783)0.774(0.698–0.867)0.790(0.703–0.870)Right_thalamus0.739(0.658–0.811)0.719(0.609–0.828)0.766(0.638–0.872)0.760(0.679–0.835)0.666(0.583–0.767)0.807(0.724–0.893)0.819(0.740–0.887)Bilateral_thalami0.829(0.757–0.892)0.813(0.719–0.906)0.851(0.745–0.936)0.846(0.780–0.905)0.769(0.684–0.860)0.881(0.813–0.948)0.911(0.856–0.957) Performance of the deep learning models 0.744 (0.703–0.784) 0.721 (0.667–0.779) 0.776 (0.714–0.832) 0.762 (0.720–0.801) 0.678 (0.632–0.727) 0.808 (0.766–0.851) 0.851 (0.817–0.883) 0.749 (0.707–0.789) 0.764 (0.706–0.814) 0.730 (0.668–0.791) 0.776 (0.737–0.812) 0.700 (0.652–0.749) 0.788 (0.749–0.830) 0.846 (0.810–0.878) 0.850 (0.817–0.881) 0.833 (0.787–0.876) 0.872 (0.821–0.913) 0.863 (0.832–0.892) 0.799 (0.756–0.841) 0.895 (0.862–0.928) 0.944 (0.926–0.960) 0.733 (0.672–0.795) 0.718 (0.636–0.810) 0.753 (0.659–0.836) 0.752 (0.691–0.812) 0.673 (0.608–0.750) 0.790 (0.729–0.856) 0.821 (0.760–0.876) 0.713 (0.651–0.769) 0.691 (0.600-0.773) 0.741 (0.647–0.824) 0.731 (0.667–0.787) 0.649 (0.584–0.716) 0.775 (0.714–0.843) 0.798 (0.731–0.851) 0.836 (0.785–0.887) 0.855 (0.791–0.918) 0.812 (0.729–0.894) 0.855 (0.806–0.901) 0.811 (0.742–0.886) 0.854 (0.798–0.911) 0.928 (0.893–0.957) 0.730 (0.649–0.811) 0.750 (0.641–0.859) 0.702 (0.574–0.830) 0.762 (0.683–0.836) 0.673 (0.577–0.783) 0.774 (0.698–0.867) 0.790 (0.703–0.870) 0.739 (0.658–0.811) 0.719 (0.609–0.828) 0.766 (0.638–0.872) 0.760 (0.679–0.835) 0.666 (0.583–0.767) 0.807 (0.724–0.893) 0.819 (0.740–0.887) 0.829 (0.757–0.892) 0.813 (0.719–0.906) 0.851 (0.745–0.936) 0.846 (0.780–0.905) 0.769 (0.684–0.860) 0.881 (0.813–0.948) 0.911 (0.856–0.957) To compare differences in AUC among the deep learning models, we performed the DeLong test. The results showed no significant difference between the left and right thalamic models in the training, validation, or test sets, whereas the bilateral thalamic model consistently achieved significantly higher AUCs than either unilateral model (P < 0.001), indicating superior discriminative performance (Fig. 3d–i). Decision curve analysis showed that, across the training, validation, and external test sets, the bilateral thalamic model yielded a higher net benefit over a wide range of threshold probabilities and overall outperformed the “treat all” and “treat none” strategies. Compared with the left and right unilateral thalamic models, the bilateral thalamic model showed greater clinical utility across most clinically relevant threshold ranges, suggesting better potential applicability in real-world settings (Fig. 3j-l).Calibration curve analysis showed that all three deep learning models demonstrated a reasonable agreement between predicted probabilities and observed positive rates in the training, validation, and external test sets. Notably, the calibration curve of the bilateral thalamic model was closer to the ideal reference line across multiple datasets (Fig. 3m-o). To further evaluate model performance at the individual-sample level, we plotted the distributions of prediction scores for the left thalamic, right thalamic, and bilateral thalamic models in the training and test sets. The results showed more substantial overlap between samples in the intermediate score range for the left and right unilateral thalamic models, whereas the bilateral thalamic model exhibited a clearer separation between negative and positive samples overall, with negative samples clustering in the lower score range and positive samples in the higher score range, and only limited overlap near the decision boundary. In the external test set, the bilateral thalamic model likewise showed a favorable separation trend, with most negative samples distributed in the negative prediction range and positive samples in the positive prediction range. By contrast, the left and right unilateral thalamic models still showed more samples located near the decision boundary in the test set, further supporting the greater robustness and generalizability of the bilateral thalamic deep learning model (Fig. 4a–f). These findings were consistent with the results of the ROC analysis and DeLong test for the bilateral thalamic model. Fig. 4Distribution of sample prediction scores and interpretability analysis of different thalamic models. (a) Distribution of sample prediction scores for the right thalamic model in the training set. (b) Distribution of sample prediction scores for the left thalamic model in the training set. (c) Distribution of sample prediction scores for the bilateral thalamic model in the training set. (d) Distribution of sample prediction scores for the right thalamic model in the test set. (e) Distribution of sample prediction scores for the left thalamic model in the test set. (f) Distribution of sample prediction scores for the bilateral thalamic model in the test set. (g–j) Original images of representative samples and the corresponding SHAP heatmaps, where (g) shows a positive sample from the right thalamic model, (h) shows a negative sample from the right thalamic model, (i) shows a negative sample from the left thalamic model, and (j) shows a positive sample from the left thalamic model. Red regions indicate a greater positive contribution to the positive prediction, whereas blue regions indicate a negative or inhibitory contribution to the prediction Distribution of sample prediction scores and interpretability analysis of different thalamic models. (a) Distribution of sample prediction scores for the right thalamic model in the training set. (b) Distribution of sample prediction scores for the left thalamic model in the training set. (c) Distribution of sample prediction scores for the bilateral thalamic model in the training set. (d) Distribution of sample prediction scores for the right thalamic model in the test set. (e) Distribution of sample prediction scores for the left thalamic model in the test set. (f) Distribution of sample prediction scores for the bilateral thalamic model in the test set. (g–j) Original images of representative samples and the corresponding SHAP heatmaps, where (g) shows a positive sample from the right thalamic model, (h) shows a negative sample from the right thalamic model, (i) shows a negative sample from the left thalamic model, and (j) shows a positive sample from the left thalamic model. Red regions indicate a greater positive contribution to the positive prediction, whereas blue regions indicate a negative or inhibitory contribution to the prediction To investigate the basis of model decisions, we performed SHAP-based interpretability analysis on representative samples. SHAP analysis further enhanced the biological interpretability of the model. The results showed that in OSA-positive cases, the central thalamic region predominantly exhibited positive contributions, whereas in non-OSA cases, the same region predominantly exhibited negative contributions (Fig. 4g–j). This class-dependent direction of contribution suggests that the model did not rely merely on diffuse and nonspecific imaging noise for classification, but may have captured discriminative imaging features located in the central thalamic region. This finding is partly consistent with the animal experiments. The animal data showed that IH induced neuronal activation in the whole thalamus and multiple thalamic subregions, with central thalamic-related nuclei such as the CL showing marked responses. The higher contribution of the central thalamic region in the human SHAP maps suggests that the clinical MRI model may have utilized distributed imaging information related to the central thalamic region. Overall, the bilateral thalamic deep learning model had stronger discriminative ability than the unilateral models. Its advantages were reflected not only in improved overall performance metrics, including AUC, accuracy, sensitivity, specificity, and F1 score, but also in better class separation, superior calibration, higher clinical net benefit, and more stable external generalizability. Bilateral thalamic feature fusion did not simply increase the amount of input information; rather, it more fully integrated potentially complementary imaging features from the two thalami, thereby enabling a more accurate and robust characterization of the OSA-related imaging phenotype. Discussion By integrating animal experiments with imaging analysis, we developed a novel diagnostic model for OSA and validated it in data from two centers. Despite differences in MRI acquisition parameters between the two centers, the deep learning model trained on the center 1 dataset showed good predictive performance in the center 2 dataset. OSA is closely related to dysfunction of the nervous system, and its recurrent intermittent hypoxia and sleep fragmentation can induce structural and functional changes in the central nervous system. In previous imaging-based auxiliary diagnostic models, the selection of the region of interest was largely based on prior literature or empirical judgment, without direct experimental evidence. To overcome this limitation, we for the first time proposed a neurobiology-driven interdisciplinary strategy: we first verified in an animal model the effect of intermittent hypoxia, a key pathophysiological process of OSA, on a specific brain region; we then used this region as a target ROI to construct a new deep learning diagnostic model. This design enabled the translation of basic neuroscience evidence into a clinical imaging model, provided a clearer biological basis for ROI selection, and strengthened the logic and persuasiveness of the model design. Our stepwise c-Fos analysis first compared macroscopic brain regions and showed that the whole thalamus had the most prominent response among the examined regions. Subregional analysis then showed activation of the CL, CM, and PVT, with the CL displaying the strongest relative response. Fiber photometry further confirmed dynamic CL activity during IH. Thus, the animal experiments support thalamic involvement and identify the CL as a representative highly responsive node, but they do not indicate that the CL is uniquely affected. Human MRI operates at a coarser spatial scale than nucleus-level animal experiments. Routine clinical acquisitions cannot reliably and reproducibly delineate small thalamic nuclei such as the CL, CM, and PVT; therefore, the whole bilateral thalamus was selected as a stable clinical ROI. Within this cross-scale framework, the animal experiments provide nucleus-level biological motivation, whereas the human model tests whether distributed whole-thalamus imaging features associated with OSA can support identification. Likewise, the central thalamic contribution observed in the SHAP maps should be interpreted as regional model attribution rather than precise subnuclear localization. The bilateral design was motivated by potential differences in left–right thalamic organization. Prior meta-analytic and intracranial studies have shown that the two thalami are not functionally identical [24, 25]. In the present study, bilateral fusion outperformed either unilateral model, suggesting that complementary imaging information was captured across sides. However, this study was not designed to establish a specific lateralized mechanism of OSA. Although external validation and decision curve analysis indicate potential value, the model should be viewed as an auxiliary approach for OSA risk identification and prioritization of PSG rather than as a replacement for PSG. Demonstrating clinical benefit will require prospective multicenter validation, comparison with established screening approaches, assessment of workflow integration, and evaluation of whether model-assisted decisions improve patient management. Our study still has several limitations. First, the animal experiments and human MRI analyses were conducted at different spatial scales; therefore, the CL calcium signal results cannot be interpreted as direct evidence that the human MRI model specifically identifies CL abnormalities. Second, OSA may affect multiple brain regions and multiple neural systems. This study can only show that the thalamus demonstrated an important response among the examined brain regions and does not exclude the possibility that other brain regions are also involved in OSA pathophysiology. Third, although external validation enhanced the reliability of the results, residual confounding and scanner-related differences cannot be completely excluded. Future studies integrating higher-resolution MRI, refined thalamic subnucleus segmentation, prospective multicenter cohorts, and more complete clinical phenotyping are needed to further validate the relationship between thalamic imaging features and OSA pathophysiology. Finally, we mainly focused on the classification of OSA and did not further investigate disease severity stratification or prognostic prediction. This is mainly because PSG remains the gold standard for OSA diagnosis and severity assessment, whereas the MRI model we developed is more suitable as an auxiliary identification tool for early screening and risk alerting in previously undiagnosed patients, rather than as a substitute for PSG in severity stratification. Prognostic prediction requires long-term follow-up and comprehensive clinical outcome data. In the future, we plan to collect additional disease-related and longitudinal follow-up data to investigate prognostic prediction in OSA. In conclusion, we developed a biologically informed and externally validated bilateral thalamic MRI model for auxiliary OSA identification. To our knowledge, this is the first OSA diagnostic model to integrate basic neuroscience evidence with deep learning. The model achieved high accuracy in datasets from two centers. This deep learning architecture showed clear advantages and may help prioritize individuals for standard PSG evaluation. Supplementary Information Below is the link to the electronic supplementary material. Figure S1: Multi-region c-Fos analysis after intermittent hypoxia exposure. (A) Quantification of c-Fos-positive cells in macroscopic brain regions, including the whole thalamus, mPFC, hypothalamus, and DG. IH induced a distributed activation pattern, with the whole thalamus showing the most prominent response among the examined macroscopic regions. (B) Quantification of c-Fos-positive cells in thalamic subregions, including CL, CM, and PVT. CL showed the strongest relative response among the examined thalamic subregions.Data are shown as mean ± SEM. Each dot represents one animal. P values were calculated using two-sided Welch’s t tests. *P < 0.05, **P < 0.01, ***P < 0.001. Figure S1: Multi-region c-Fos analysis after intermittent hypoxia exposure. (A) Quantification of c-Fos-positive cells in macroscopic brain regions, including the whole thalamus, mPFC, hypothalamus, and DG. IH induced a distributed activation pattern, with the whole thalamus showing the most prominent response among the examined macroscopic regions. (B) Quantification of c-Fos-positive cells in thalamic subregions, including CL, CM, and PVT. CL showed the strongest relative response among the examined thalamic subregions.Data are shown as mean ± SEM. Each dot represents one animal. P values were calculated using two-sided Welch’s t tests. *P < 0.05, **P < 0.01, ***P < 0.001.