Novel Non-Invasive Deep Learning Model Based on Tongue Images for Early Differentiation of Ischemic and Hemorrhagic Stroke

Novel Non-Invasive Deep Learning Model Based on Tongue Images for Early Differentiation of Ischemic and Hemorrhagic Stroke | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Novel Non-Invasive Deep Learning Model Based on Tongue Images for Early Differentiation of Ischemic and Hemorrhagic Stroke Yuwei Pan, Xiguang Tian, Sande Gao, Haoran Cen, Zihao Huang, Yixuan Li, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8586557/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Stroke, the second leading cause of death globally, demands rapid differentiation between ischemic stroke (IS) and intracerebral hemorrhage (ICH) to guide optimal therapeutic strategies. Current diagnostic modalities (computed tomography [CT] or magnetic resonance imaging [MRI]) are hindered by limited accessibility and prolonged turnaround times, particularly in resource-constrained settings. Herein, we propose StrokeNet, a deep learning model integrating a modified ResNet-18 backbone with wavelet-based hierarchical attention (HWAttention) modules, for the early differentiation of stroke subtypes via tongue images, which serve as an underutilized non-invasive diagnostic tool in modern stroke care. A single-center cross-sectional study was conducted, enrolling 201 acute stroke patients (144 IS, 57 ICH). Tongue regions were accurately segmented using the YOLOv5 object detection model, and six-channel composite images (combining full facial and cropped tongue RGB features) were constructed as model inputs. On the internal validation set, StrokeNet achieved a classification accuracy of 82.93%, an area under the receiver operating characteristic curve (AUC) of 0.8463 (95% CI:0.7125–0.9799), a sensitivity of 89.47% (for IS), a specificity of 68.00% (for ICH), and an F1-score of 0.8793. The model outperformed mainstream baseline architectures (EfficientNet-B2, ResNet-18) across key metrics, with ablation experiments confirming that the HWAttention module and six-channel input design synergistically enhanced discriminative feature capture. Clinical risk factor analysis identified age > 60 years as an independent predictor of ICH (adjusted OR = 2.326, 95% CI: 1.243–4.352, p = 0.008). Subgroup analysis validated the model’s robustness across age, gender, comorbidity status and neurological deficit severity. To our knowledge, this study is the first to leverage artificial intelligence (AI)-driven tongue imaging for stroke subtype differentiation. StrokeNet offers non-invasiveness, rapid diagnostic turnaround and low cost, establishing a novel paradigm to optimize emergency triage in resource-limited healthcare settings and address global disparities in stroke care. Biological sciences/Computational biology and bioinformatics Health sciences/Health care Health sciences/Medical research Health sciences/Neurology Stroke Ischemic stroke Intracerebral Hemorrhagic stroke Tongue diagnosis Deep learning Attention mechanism Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Background Stroke is the second leading cause of mortality worldwide, responsible for approximately 13.7 million incident cases and 6.2 million deaths annually. Among these, the acute-phase mortality rate of intracerebral hemorrhage (ICH) reaches as high as 40%–50%, significantly exceeding that of ischemic stroke (IS). Given that the two conditions require diametrically opposed management strategies, early and accurate differentiation is critical in determining patient outcomes[ 1 , 2 ]. Intravenous thrombolysis and endovascular therapy for ischemic stroke must be administered within the golden time window (4.5 hours), whereas inadvertent thrombolysis in ICH patients can lead to hematoma expansion and a marked increase in mortality[ 3 ]. Although intravenous and endovascular thrombolytic therapies have substantially improved outcomes in IS patients, nearly 40% of patients experience treatment delays in clinical practice due to the inability to obtain timely neuroimaging[ 4 ]. Currently, computed tomography (CT), the primary imaging modality used in clinical practice, demonstrates an average sensitivity of 66% and specificity of 87% for diagnosing ischemic stroke within 6 hours of onset. Moreover, it is prone to false negatives in cases involving small infarcts[ 5 ]. Although magnetic resonance imaging (MRI) can improve diagnostic accuracy, the average examination time exceeds 45 minutes. Furthermore, even in countries with relatively ample medical resources, the availability of MRI in primary hospitals is consistently lower[ 6 ]. Consequently, the median time to initial treatment for stroke patients is as long as 144 minutes, with only 9.3% of patients receiving treatment within the first third (90 minutes) of the golden therapeutic window. Treatment delay is directly associated with unfavorable outcomes[ 7 ]. These technical limitations in neuroimaging contribute to the underutilization of the golden time window in acute stroke management, substantially increasing the risk of disability among patients. Tongue diagnosis in Traditional Chinese Medicine (TCM) offers an innovative approach for the differential diagnosis of stroke. TCM theory posits that the tongue reflects systemic physiological and pathological changes, a notion that aligns with modern medical mechanisms including inflammatory responses and microcirculatory dysfunction. Clinical evidence indicates that ischemic stroke patients often present with pale-red tongue with white and greasy coating, whereas those with hemorrhagic stroke typically exhibit dark-purplish tongue with yellow and greasy coating[ 8 ], supporting an objective correlation between tongue manifestations and stroke subtypes. However, conventional tongue diagnosis relies heavily on practitioners' subjective experience, leading to suboptimal inter-observer consistency. Advances in deep learning have effectively addressed this limitation, with mechanisms such as attention networks and multi-scale feature extraction having been successfully applied in modern tongue image analysis[ 9 , 10 ]. Existing studies have demonstrated the diagnostic utility of tongue imaging in both neoplastic (e.g., gastric, pancreatic, hepatocellular, and colorectal cancers) and non-neoplastic conditions (e.g., gastritis, rheumatoid arthritis, diabetes, and chronic hepatitis B)[ 11 – 14 ]. Nevertheless, most existing models focus on digestive diseases, and a dedicated framework for stroke-specific differential diagnosis remains notably lacking. Moreover, no study to date has employed AI-driven tongue analysis for early discrimination between IS and ICH. This study introduces for the first time the StrokeNet model, which integrates residual networks with hierarchical attention mechanisms. In a cross-sectional cohort of 201 patients, it achieved a classification accuracy of 82.93% in distinguishing early cerebral infarction from intracerebral hemorrhage. To our knowledge, this is the first research to apply artificial intelligence (AI)-driven deep learning to evaluate the value of tongue imaging in the early differential diagnosis of these conditions. A comprehensive search of the PubMed database (up to June 2025, without language restrictions), using keywords such as "stroke" and "tongue diagnosis," identified no prior publications in this domain. Notably, the StrokeNet model demonstrates strong clinical feasibility, particularly in resource-limited settings. It requires an average diagnosis time of only 43 seconds and incurs minimal cost, substantially outperforming conventional imaging in both time efficiency and accessibility. This model fills a critical gap as a rapid tool for stroke subtype discrimination and offers a new paradigm for optimizing emergency triage workflows. It holds particular promise for primary care institutions and clinical settings in developing countries, with the potential to enhance the utilization of the golden treatment window and reduce the risks of misdiagnosis and inappropriate treatment. 2 Materials and Methods 2.1 Study Population and Ethical Statements From November 2023 to December 2024, consecutive patients with acute stroke aged 18–80 years were recruited from the Department of Neurology, Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University. All participants or their legal guardians (for patients with impaired consciousness) signed a written informed consent form, and the study was conducted in accordance with the Declaration of Helsinki (revised 2013). The ethical review of this study was carried out and approved by the Institutional Review Board (IRB) of Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University (IRB number: 2023NK004), with the clinical trial registration number ChiCTR2300077147. All source codes and data analyzed in this study can be obtained from the corresponding author upon reasonable request. The diagnostic criteria for acute stroke refer to the Chinese Guidelines for the Diagnosis and Treatment of Acute Ischemic Stroke and Chinese Guidelines for the Diagnosis and Treatment of Intracerebral Hemorrhage: intracerebral hemorrhage (ICH) is defined as hyperdense lesions on cranial computed tomography (CT) or hypointense signals on T2* weighted magnetic resonance imaging (MRI); ischemic stroke (IS) is defined as early ischemic signs on CT or hyperintense signals on diffusion-weighted MRI, with symptom onset time ≤ 24 hours and first-episode stroke. Initially, 256 patients were screened, with the following inclusion and exclusion criteria: Inclusion Criteria: (1) Aged 18–80 years; (2) Meeting the above diagnostic criteria for acute stroke; (3) Glasgow Coma Scale (GCS) score ≥ 12, enabling cooperation with tongue image acquisition; (4) No history of tongue reconstruction surgery or severe maxillofacial trauma. Exclusion Criteria: (1) Presence of severe oral diseases (e.g., oral cancer, extensive ulcers) or active tongue lesions; (2) Comorbidity with systemic diseases affecting tongue morphology (e.g., malignant tumors, systemic lupus erythematosus, Sjögren's syndrome); (3) Use of medications known to alter tongue appearance within 2 weeks prior to enrollment (e.g., antibiotics, antifungals, chemotherapeutic agents); (4) Pregnancy or lactation period; (5) Impaired tongue image quality (tongue coverage ≥ 10% by teeth/lips, motion blur, or other quality defects). A total of 55 patients were excluded based on the above criteria, and 201 eligible patients were finally included in the analysis. 2.2 Data Collection Trained research physicians conducted interviews and took tongue photographs of participants following a standardized collection process. The interviews gathered baseline data on general conditions (age, gender), medical history (hypertension, diabetes mellitus), and clinical manifestations (onset time, neurological deficit symptoms). Tongue photographs were collected using a YZAI-02 tongue diagnosis instrument (Fig. 1 ) either before breakfast or 2 hours after breakfast. The specific steps for image collection are as follows: (1) Power on the instrument after inspection and adjust the camera parameters. (2) Disinfect the areas of the instrument that may come into direct contact with the participant using 75% alcohol. (3) Instruct the patient to stick out their tongue flatly. (4) Turn on the built-in ring light source and complete the image capture. (5) Following the evaluation of the acquired tongue image, image acquisition is finalized if the image meets predefined quality criteria; otherwise, repeated image capture is performed until the established quality requirements are satisfied. Qualification criteria for photo quality: no problems such as occlusion, blurring, fogging, overexposure, or underexposure; the tongue should be relaxed and flattened with no twisting or tension; there should be no foreign objects, staining, or other conditions affecting the appearance of the tongue surface. 2.3 Data preprocessing All tongue images utilized in this study were acquired using standardized equipment and manually labeled as ischemic stroke (IS) or intracerebral hemorrhage (ICH), corresponding to labels 0 and 1, respectively. The original images included the entire facial region, with the tongue accounting for only a small proportion. To isolate the tongue region and eliminate background noise, we employed a YOLOv5s object detection model trained on a self-constructed dataset. This model automatically identified and localized the tongue region, providing bounding boxes for direct cropping. No image resizing or margin extension was required for the cropped regions. Samples with failed tongue detection were marked as abnormal and excluded from subsequent training. Each valid sample consisted of a complete facial image and the corresponding cropped tongue region. RGB channels were extracted from both images and concatenated along the channel dimension to form a six-channel composite image, enabling the model to jointly process global and local features. To preserve anatomical details, the images retained their original resolution without resizing. Prior to model training, all images were normalized using channel-wise means and standard deviations of [0.5, 0.5, 0.5]. To enhance model generalization, data augmentation strategies were implemented during training, including random horizontal flipping, random cropping, and color jittering. These augmentation techniques were applied to the complete facial images and cropped tongue images separately before channel concatenation. The final six-channel images featured a unified structure and consistent labeling, which could be directly used as input for downstream classification models. 2.4 StrokeNet-Based Classification Model We proposed a deep convolutional neural network (DCNN) for differentiating ischemic stroke (IS) from intracerebral hemorrhage (ICH) based on tongue images. The network backbone was adapted from ResNet-18 and modified to accommodate six-channel input (Fig. 2 A). This model accepts RGB channels from both the original full facial images and cropped tongue regions as inputs, enabling it to learn joint representations of global and local features. Specifically, the first convolutional layer of ResNet-18 was modified to accept six input channels instead of the conventional three. To enhance the model’s capability to capture fine-grained features such as tongue texture and spatial variations, we incorporated HWAttention modules at six insertion points: before and after the 1st, 2nd, and 3rd residual blocks, respectively. Each HWAttention module integrates discrete wavelet transform (DWT), inverse wavelet transform (IWT), and a dual-channel spatial attention mechanism. These modules assist the model in dynamically emphasizing discriminative features in both spatial and frequency domains (Fig. 2 B). The final model architecture comprises a global average pooling layer followed by a fully connected layer that outputs binary prediction outcomes (IS or ICH) (Fig. 2 A). Due to the limited sample size, to prevent overfitting, we frozen the first two residual blocks of ResNet-18 and only trained the deeper layers along with the classifier. Weighted cross-entropy loss was employed during training to address class imbalance. The stochastic gradient descent (SGD) optimizer was utilized with an initial learning rate of 0.001, which was decayed every two epochs via a StepLR scheduler. The training process lasted for 100 epochs with a batch size of 4, and model checkpoints were saved after each epoch for performance evaluation. 2.5 Statistical Analysis All statistical analyses were performed using R statistical software (version 4.2.1; R Foundation for Statistical Computing, Vienna, Austria). For data conforming to a normal distribution, descriptive statistics were presented as mean ± standard deviation (SD), and intergroup differences were compared using the independent samples t-test (for two-group comparisons). Prior to t-test analysis, Levene's test was conducted to verify the homogeneity of variances: if variances were homogeneous, the standard t-test was applied; if not, Welch's t-test was used as an alternative. For data that did not follow a normal distribution, descriptive statistics were expressed as interquartile range (IQR, P25 ~ P75), and intergroup comparisons were performed using the Mann–Whitney U test (for two-group nonparametric comparisons). Additionally, binary logistic regression analysis was employed to explore potential risk factors for ischemic stroke (IS), with intracerebral hemorrhage (ICH) as the reference group. Results were presented as adjusted odds ratios (OR) with 95% confidence interval (95% CI). A two-tailed p -value < 0.05 was considered statistically significant for all analyses. 2.5 Evaluation Metrics To comprehensively evaluate the model's performance, we calculated standard classification metrics: accuracy (ACC), sensitivity (Recall), specificity, precision, F1-score, and the area under the receiver operating characteristic curve (AUC). We also leveraged the components of the confusion matrix, comprising true positive (TP), false positive (FP), true negative (TN), and false negative (FN), to conduct a more granular performance analysis. The following formulas were used: $$\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{F}\text{P}+\text{T}\text{N}+\text{F}\text{N}}$$ $$\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\frac{\text{T}\text{P}+\text{T}\text{N}}{\text{T}\text{P}+\text{F}\text{P}+\text{T}\text{N}+\text{F}\text{N}}$$ $$\:\text{S}\text{p}\text{e}\text{c}\text{i}\text{f}\text{i}\text{c}\text{i}\text{t}\text{y}=\frac{\text{T}\text{N}}{\text{T}\text{N}+\text{F}\text{P}}$$ $$\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{\text{T}\text{P}}{\text{T}\text{P}+\text{F}\text{P}}$$ $$\:\text{F}1=\frac{2\times\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\times\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}$$ AUC was used to assess model robustness across classification thresholds, with values closer to 1 indicating better discriminative power. Class-wise metrics were also reported to evaluate model performance on both mild and severe cases under data imbalance.All metrics were computed on the validation set using predictions from the best-performing model checkpoint (based on highest ACC). 3 Results 3.1 Patient recruitment and data analysis The patient recruitment and screening procedures in this study strictly adhered to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement (Figure 3). A total of 256 consecutive patients with suspected acute stroke were screened from the Department of Neurology, Guangzhou Medical University Affiliated Traditional Chinese Medicine Hospital, between November 2023 and December 2024. All participants completed standardized structured interviews and tongue image acquisition. According to predefined inclusion and exclusion criteria, a stepwise screening was performed, resulting in the exclusion of 55 patients (21.48%). The exclusion reasons included severe oral diseases (n=8), systemic diseases affecting tongue morphology (n=12), use of tongue image-interfering medications within 2 weeks prior to enrollment (n=10), and inadequate tongue image quality (n=25). Finally, 201 eligible patients were included for subsequent analyses. Among the included patients, 144 (71.64%) were diagnosed with ischemic stroke (IS) and 57 (28.36%) with intracerebral hemorrhage (ICH). A stratified random sampling method was used to partition the dataset into a training set and an internal validation set at an 8:2 ratio, ensuring balanced distribution of key baseline characteristics (e.g., stroke subtype, age, gender) between the two sets: the training set included 160 patients (115 IS, 45 ICH; 79.60%), and the internal validation set included 41 patients (29 IS, 12 ICH; 20.40%). Table 1 presents the detailed baseline clinical characteristics of IS and ICH patients in the training and internal validation sets, including demographic data (age, gender), past medical history (hypertension, diabetes mellitus), vital signs (pulse rate, blood pressure), and neurological function scores [National Institutes of Health Stroke Scale (NIHSS), modified Rankin Scale (mRS)]. Intergroup comparisons were performed using appropriate statistical tests: independent samples t-test or Mann–Whitney U test for continuous variables, and chi-square test or Fisher’s exact test for categorical variables. No statistically significant differences in baseline characteristics were observed between the training set and the validation set (all p >0.05), confirming the rationality of dataset partitioning and laying a foundation for the reliability of model training and validation. To identify potential risk factors for ICH, binary logistic regression analysis was conducted with stroke subtype (IS as the reference group) as the dependent variable, and age, gender, history of hypertension, history of diabetes mellitus, pulse rate, blood pressure, and neurological function scores as independent variables (Table 2). The results showed that age> 60 years (adjusted OR = 2.326, 95%CI: 1.243–4.352, p = 0.008) was an independent risk factor for ICH ( p <0.05). Table1.Baseline clinical characteristics of IS and ICH patients in the training and internal validation datasets (mean±SD or n, %). ICH (n = 57) IS (n = 144) NIHSS score, n (%) ≤15 57 (28.9%) 136 (67.2%) >15 0 (0%) 8 (4%) mRS score, n (%) 0-2 36 (18.4%) 104 (51.2%) 3-5 21 (10.4%) 40 (19.9%) Age, n (%) ≤60 28 (14.4%) 44 (21.4%) >60 29 (14.4%) 100 (49.8%) Gender, n (%) Male 36 (18.4%) 100 (49.3%) Female 21 (10.4%) 44 (21.9%) Diabetes, n (%) 0 44 (22.4%) 108 (53.2%) 1 13 (6.5%) 36 (17.9%) Hypertension, n (%) 1 37 (18.9%) 86 (42.3%) 0 20 (10%) 58 (28.9%) Pulse Rate, n (%) ≤90 52 (26.4%) 118 (58.2%) >90 5 (2.5%) 26 (12.9%) Systolic blood pressure, n (%) >140 33 (16.9%) 89 (43.8%) ≤140 24 (11.9%) 55 (27.4%) Diastolic blood pressure, n (%) >90 20 (10.4%) 42 (20.4%) ≤90 37 (18.4%) 102 (50.7%) Table2. Binary logistic regression analysis of risk factors for ICH (reference group: IS) Variables OR(95% CI) P-value mRS score (3-5) 0.684 (0.358 – 1.308) 0.251 Age>60 2.326 (1.243 – 4.352) 0.008 Female 0.783 (0.412 – 1.489) 0.456 Diabetes 1.165 (0.565 – 2.401) 0.680 Hypertension 1.296 (0.686 – 2.449) 0.424 Hyperlipidemia 1.528 (0.410 – 5.689) 0.528 Pulse Rate>90 2.356 (0.857 – 6.471) 0.097 Systolic blood pressure≤140 0.885 (0.475 – 1.649) 0.701 Diastolic blood pressure≤90 1.412 (0.740 – 2.696) 0.296 3.2 Classification Performance Analysis The discriminative performance of the model was comprehensively evaluated using multiple core metrics and visualization curves (Figure 4B, 4C, 4D). On the internal validation set, the StrokeNet model achieved an accuracy of 82.93% (34/41) for differentiating ischemic stroke (IS) from intracerebral hemorrhage (ICH), with an area under the receiver operating characteristic curve (AUC) of 0.8463 (95% CI: 0.7125–0.9799), a precision of 86.44%, a sensitivity of 89.47% (for IS), a specificity of 68.00% (for ICH), and an F1-score of 0.8793. These metrics collectively demonstrate the model’s robust classification efficacy, particularly its high sensitivity in ischemic stroke (IS) identification, which aligns with the core clinical requirement of minimizing missed diagnoses. The confusion matrix further quantified the model’s classification outcomes (Figure 4B). Among the validation set, the model correctly identified 26 IS patients (true positives [TP]) and 8 ICH patients (true negatives [TN]). Only 3 IS cases were misclassified as ICH (false negatives [FN]), and 4 ICH cases were misclassified as IS (false positives [FP]). The false negative rate (missed diagnosis rate) was merely 10.53%, and the false positive rate (misdiagnosis rate) was 32.00%. These results indicate that the model can effectively reduce the risk of missed IS diagnoses in emergency clinical settings, providing actionable support for timely decision-making within the thrombolytic therapy window. 3.3 Baseline Model Comparison and Ablation Analysis To verify the stability and specificity of tongue image features in stroke subtype differentiation while eliminating confounding effects from model architectural discrepancies, two classical deep learning models widely used in clinical research were selected as baseline controls: ResNet-18 and EfficientNet-B2. All models were trained with identical hyperparameters, including the following settings, optimizer: Adam; learning rate: 1×10⁻⁴; batch size: 16; training epochs: 200; loss function: cross-entropy loss, to ensure fairness in comparative analysis. Visualization of the three models’ performance is presented in Figure 5. The PR curve (Figure 5C) shows that StrokeNet achieved a PR value of 0.924, which was higher than that of ResNet-18 (0.888) and EfficientNet-B2 (0.896). Similarly, the ROC curve (Figure 5D) demonstrates that StrokeNet yielded an AUC of 0.846, outperforming ResNet-18 (0.752) and EfficientNet-B2 (0.801). Table 3 provides a quantitative comparison of comprehensive performance metrics across the three models. StrokeNet outperformed the two baseline models in accuracy, recall, F1-score, and AUC. Only the precision of StrokeNet (0.86) was slightly lower than that of ResNet-18. These results confirm that the integrated strategies of wavelet attention mechanism and global-local feature representation in StrokeNet enable more precise capture of key tongue image features associated with stroke subtypes (e.g., tongue color, coating thickness, tongue shape), thereby significantly enhancing the superiority and stability of discriminative performance. Tabel 3.Quantitative comparison of classification performance among different deep learning models. Algorithm AUC ACC SPE Precision SEN F1-Score EfficientNet-B2 0.8400 0.7317 0.48 0.7869 0.8421 0.8136 ResNet-18 0.7500 0.7073 0.56 0.9000 0.7719 0.7857 StrokeNet 0.8463 0.8292 0.68 0.8644 0.8947 0.8793 To evaluate the applicability and robustness of the StrokeNet model across diverse populations, subgroup stratification was performed based on age, gender, past medical history, admission blood pressure, admission pulse rate, and neurological deficit severity (mRS score). The AUC was used to quantify the model’s performance in each subgroup. As illustrated in Figure 6, subgroup-specific performance analysis yielded the following key findings. In terms of age, the model exhibited superior diagnostic performance in the elderly population (≥ 65 years old) compared to younger individuals. Regarding gender, the model demonstrated better discriminative ability in male patients than in female patients. With respect to risk factors, the model maintained relatively stable performance across subgroups stratified by hypertension or diabetes mellitus status. For admission mRS scores, the model’s performance remained consistent between patients with mRS ≤ 2 and those with mRS > 3, without significant discrepancies. 3.3 Model Robustness and Subgroup Performance To further assess the robustness and generalization ability of the StrokeNet model, we conducted ablation studies and comparative analyses. Specifically, we evaluated the performance degradation induced by removing the HWAttention modules from the network. The ablated model exhibited noticeable declines in AUC and recall, confirming the critical role of frequency-enhanced attention in capturing discriminative features associated with stroke subtype differentiation. Additionally, we compared the proposed six-channel input design (integrating full facial images and cropped tongue regions) with conventional three-channel inputs. The six-channel configuration achieved superior classification performance, demonstrating that the integration of global and local visual information yields more comprehensive feature representations and strengthens the model’s decision-making robustness. Collectively, these experimental results validate that the proposed StrokeNet model not only attains high classification accuracy but also retains stable performance across different input configurations, highlighting its promising potential for clinical translation. 4 Discussion A critical unmet need in the emergency triage of acute stroke lies in the imbalance between the urgency of the therapeutic window and the accessibility of diagnostic tools, which remains a global healthcare challenge[ 15 ]. Despite significant advancements in neuroimaging techniques and blood-based biomarker assays, delays in stroke subtype differentiation persist as a key determinant of poor patient outcomes in real-world clinical settings. Notably, the World Health Organization (WHO) has highlighted that unequal access to medical equipment is particularly prevalent in low- and middle-income countries (LMICs), with inadequate availability of essential diagnostic tools, especially in rural and remote regions. This resource gap directly translates to substantial treatment delays[ 16 ]. Specifically, a retrospective study involving 588 patients with acute ischemic stroke demonstrated that the mean delay in sequential CT imaging (non-contrast CT followed by CTA/CTP) was 53.7 minutes[ 17 ]. Notably, for patients with large vessel occlusion stroke (LVOS) transferred from non-endovascular stroke centers (non-ESCs) to endovascular stroke centers (ESCs) for endovascular treatment (EVT), two key factors were linked to prolonged door-in-door-out (DIDO) times: presentation to a non-stroke certified non-ESC was associated with 20 additional minutes at the transferring hospital, and the acquisition of vascular imaging added another 16 minutes[ 18 ]. This primarily attributed to the lack of integrated imaging equipment and standardized protocols in non-ESCs, which necessitate repeated patient transfers or prolonged waits for equipment availability. Blood-based biomarkers, as valuable supplements to neuroimaging, have emerged as a novel technical avenue for stroke subtype differentiation. Glial fibrillary acidic protein (GFAP) stands as the most well-validated specific biomarker: a meta-analysis encompassing 12 studies demonstrated that serum GFAP exhibited a sensitivity of 81.1% and a specificity of 97% for differentiating intracerebral hemorrhage (ICH) from ischemic stroke (IS) within 1–6 hours of symptom onset, with significantly higher GFAP levels observed in ICH patients compared to IS patients[ 19 ]. S100 calcium-binding protein β (S100β) also showed promising discriminative efficacy: at a cutoff value of 67 pg/ml, it achieved an AUC of 0.903 for distinguishing the two subtypes, with a high sensitivity of 95.7% and a relatively lower specificity of 70.4%, which renders it a viable complementary biomarker to GFAP[ 20 ]. Notably, N-terminal pro-B-type natriuretic peptide (NT-proBNP) performed exceptionally well in combined detection strategies: when used in conjunction with retinol-binding protein 4 (RBP-4), it identified 29.7% of IS patients with 100% specificity; further integration with GFAP elevated this identification rate to 51.5%[ 21 ]. Furthermore, systematic reviews and meta-analyses have confirmed that additional biomarkers, such as brain natriuretic peptide (BNP), matrix metalloproteinase-9 (MMP-9), and D-dimer, possess moderate discriminative potential[ 21 ]. Multimarker combinations, including 2GFAP + D-dimer + S100β and RBP-4 + NT-proBNP + GFAP, have emerged as particularly promising approaches, optimizing sensitivity while maintaining high specificity (98%–100%) [ 22 ]. Nevertheless, biomarker-based detection still harbors non-negligible limitations. First, no single or combined biomarker has yet obtained approval for routine clinical application, necessitating further validation through large-scale prospective multicenter studies to confirm reliability and generalizability. Second, detection costs for stroke subtype-discriminating biomarkers remain prohibitive given that a single GFAP assay incurs an approximate cost of $ 150–200, while assays for alternative biomarkers including miR-124-3p and metabolomic indicators demand advanced specialized laboratory infrastructure, a requirement that renders such testing modalities inaccessible in resource-limited healthcare settings. Third, turnaround times for results (1–2 hours) may still delay critical decisions within the golden therapeutic window. Fourth, certain biomarkers have inherent detection contraindications. For example, metabolomic profiles are prone to interference in patients with severe hepatic or renal dysfunction, potentially leading to false-positive results. Additional limitations of current diagnostic tools include inadequate coverage of special populations and poor adaptability to diverse clinical scenarios. Magnetic resonance imaging (MRI) is absolutely contraindicated in patients with severe renal impairment or metallic implants (e.g., cardiac pacemakers). Conversely, computed tomography (CT) entails radiation exposure risks with the mean effective dose for a single non-contrast head CT is approximately 2.7 mSv[ 23 ], which pose potential hazards to patients requiring repeated imaging. In settings lacking imaging equipment including emergency scenes and remote areas, healthcare providers are precluded from conducting timely stroke subtype differentiations. As emphasized in authoritative guidelines, the development of non-invasive, point-of-care rapid differentiation tools represents a critical unmet need in stroke care, which would substantially improve global equity in stroke diagnosis and treatment. This context underscores the clear clinical positioning and guideline alignment of the AI-based tongue image differentiation protocol proposed in the present study. Tongue diagnosis, a core component of inspection diagnosis in traditional Chinese medicine (TCM), is characterized by non-invasiveness, convenience, and cost-effectiveness, serving as a pivotal diagnostic tool in TCM clinical practice[ 24 ]. In recent years, with advancements in modern science and technology, significant progress has been made in the objectification of tongue diagnosis—laying a scientific foundation for its integration into modern medical practice[ 24 ]. The tongue, featuring a thin, delicate mucosa and abundant vascularization, acts as an intuitive window to reflect systemic microcirculatory status[ 25 ]. For ischemic stroke (IS) patients, insufficient cerebral blood perfusion is often accompanied by systemic microcirculatory disorders, manifesting as pale red tongue color and white greasy coating; this phenomenon is associated with oxidative stress responses and inflammatory factor release induced by cerebral tissue hypoxia in infarcted regions[ 26 ]. In contrast, intracerebral hemorrhage (ICH) patients, due to cerebrovascular rupture and hemorrhage, exhibit dark purple tongue color and yellow greasy coating, which are closely linked to elevated levels of hemoglobin metabolites released by erythrocyte rupture and vascular endothelial injury markers in the bloodstream [ 24 , 25 ]. However, the clinical application of traditional tongue diagnosis is limited by its subjectivity and lack of standardization. Moreover, existing AI models for tongue image analysis primarily concentrate on digestive system diseases, and no specific tongue image-based AI tools have been developed for stroke subtype differentiation, which serves as the core innovative entry point of the present study. By integrating TCM tongue diagnosis theory with deep learning technology, we innovatively proposed the StrokeNet model for stroke subtype differentiation based on tongue images, with three core advantages: First, during data preprocessing, the YOLOv5 object detection model was employed to accurately segment the tongue region, effectively eliminating interference from irrelevant tissues (e.g., teeth, lips). Additionally, standardized illumination and image augmentation techniques were utilized to mitigate biases arising from variations in data acquisition environments, providing a high-quality data foundation for feature extraction[ 27 ]. Second, to address the high-dimensional and non-linear characteristics of tongue images, and in contrast to previous studies that relied primarily on linear regression or shallow feature extraction techniques and thus failed to capture the complex mapping relationships between tongue features, we designed the HWAttention module that integrates discrete wavelet transform (DWT), inverse wavelet transform (IWT), and a dual-channel spatial attention mechanism. This module enables simultaneous capture of key features in both the spatial domain (e.g., tongue shape, coating distribution) and frequency domain (e.g., tongue color gradients, texture details), significantly enhancing the ability to discriminate stroke subtype-specific tongue image differences. Third, by optimizing the input layer structure, we expanded the traditional three-channel image to a six-channel input, realizing efficient fusion of multi-modal information. This not only alleviates the problem of reduced specificity caused by feature aliasing in conventional convolutional neural networks (CNNs) but also avoids the computational burden associated with a sharp increase in parameters[ 28 ]. Performance validation results demonstrated that the StrokeNet model achieved an accuracy of 82.93%, an AUC of 0.8463, and a sensitivity of 89.47% on the internal validation set, while simultaneously outperforming mainstream baseline models including ResNet-18 (AUC = 0.752) and EfficientNet-B2 (AUC = 0.801). Notably, its high sensitivity aligns with the core clinical requirement of minimizing missed diagnoses for acute stroke triage. Subgroup analysis further confirmed the clinical generalizability of the StrokeNet model: it maintained stable performance across subgroups stratified by age, gender and exhibited no significant differences in subgroups with or without hypertension/diabetes mellitus, as well as across different neurological deficit severities.This finding is particularly notable, given that age > 60 years was identified as an independent risk factor for intracerebral hemorrhage (ICH) in our regression analysis and that stroke patients with comorbid hypertension or diabetes mellitus often present with more complex clinical conditions, both of which are factors that tend to interfere with traditional stroke subtype differentiation methods.The stability of the StrokeNet model stems not only from the precise capture of core tongue image features by the HWAttention module, but also from the inherent nature of tongue images as a window reflecting systemic physiological and pathological states. Tongue manifestations mirror holistic changes in microcirculation, which are less susceptible to interference from individual underlying diseases. Compared with existing diagnostic methods, the core advantages of the StrokeNet model lie in its non-invasiveness, low cost, and operational simplicity. It eliminates the need for expensive imaging or laboratory equipment, requiring only standardized tongue image acquisition via a dedicated device to complete differentiation within 1 minute. This makes it particularly suitable for resource-limited settings such as primary healthcare facilities and emergency on-site triage, offering a new paradigm for addressing the global challenge of difficult and delayed stroke diagnosis. Nevertheless, the present study has several limitations: First, it was a single-center cross-sectional study with a relatively limited sample size (n = 201), and all included patients were recruited from the Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University, which may introduce selection bias. The external validity of the model thus requires further verification in multi-center, large-sample cohorts. Second, no independent external validation set was incorporated, and the generalization ability of the model needs to be tested using datasets from diverse geographical regions and populations. Third, potential confounding factors affecting tongue manifestations, including medication use (e.g., antiplatelet agents, antihypertensive drugs), diet, and oral hygiene status, were not accounted for in the present study. Future studies should refine inclusion/exclusion criteria and incorporate confounding factor adjustment to enhance result reliability. Fourth, the model was developed based solely on tongue image data, without integrating clinical parameters (e.g., age, pulse rate) or biomarkers. Constructing a multi-modal fusion model that combines tongue images with clinical and laboratory data holds promise for further improving discriminative performance. Future research directions can focus on three key areas: First, conducting multi-center, large-sample prospective studies to enroll patients from different regions and healthcare institutions, establishing a standardized dataset with an independent external validation set to optimize model parameters and validate its real-world applicability. Second, exploring the association between tongue image features and stroke prognosis, developing a comprehensive AI tool integrating early subtype differentiation, prognosis prediction, and treatment guidance to further enhance clinical utility. Third, investigating the molecular biological mechanisms underlying the correlation between tongue manifestations and stroke subtypes, providing novel experimental evidence for the integration of traditional TCM theory with modern medicine. 4 Conclusion This study for the first time integrated deep learning technology with TCM tongue diagnosis to develop the StrokeNet model for early differentiation of acute stroke subtypes. Characterized by non-invasiveness, low cost, rapidity, and high sensitivity, the model holds broad application prospects in primary healthcare and emergency settings. The findings not only validate the practical value of tongue images in stroke subtype differentiation but also provide a successful example of the interdisciplinary integration of traditional medical knowledge and modern artificial intelligence technology. This work is expected to offer new insights and tools for optimizing the global stroke diagnosis and treatment system. Declarations Ethical Statements All participants or their legal guardians (for patients with impaired consciousness) signed a written informed consent form, and the study was conducted in accordance with the Declaration of Helsinki (revised 2013). The ethical review of this study was carried out and approved by the Institutional Review Board (IRB) of Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University (IRB number: 2023NK004), with the clinical trial registration number ChiCTR2300077147. All source codes and data analyzed in this study can be obtained from the corresponding author upon reasonable request. Data Availability Statement All data supporting the findings of this study are not publicly available due to ethical restrictions and the terms of the informed consent obtained from participants, which prohibit the unrestricted sharing of sensitive clinical information. However, de-identified data can be requested from the first author, Dr. Yuwei Pan, via email at [email protected] for non-commercial research purposes. Requests will be reviewed by the Institutional Review Board (IRB) of the Affiliated TCM Hospital of Guangzhou Medical University to ensure compliance with ethical guidelines and participant privacy protections. Conflict of Interest The authors declare no conflicts of interest. Acknowledgement Funding by Guangzhou Science and Technology Bureau Municipal University Institute Joint Project (Grant Number:2025A03J3431), Young Science and Technology Talents Fund of The Affiliated TCM Hospital of Guangzhou Medical University (Grant Number:2023RC15), Youth Talent Projects in Guangzhou Medical University (Grant Number:2024SRP211) Author Contributions Yuwei Pan, Shaopeng Liu contributed substantially to the design of work, data analysis and experimental procedure. Haoran Cen assisted in modeling. Xiguang Tian, Sande Gao, Zihao Huang, Yixuan Li, Wenjian Du, Nan Zhang, Guodong Li assisted with the data collection. References GBD 2021 Stroke Risk Factor Collaborators. Global, regional, and national burden of stroke and its risk factors, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet Neurol. 2024;23(10):973-1003. O'Donnell MJ, Xavier D, Liu L, et al. Risk factors for ischaemic and intracerebral haemorrhagic stroke in 22 countries (the INTERSTROKE study): a case-control study. Lancet. 2010;376(9735):112-123. Hacke W, Kaste M, Bluhmki E, et al. Thrombolysis with alteplase 3 to 4.5 hours after acute ischemic stroke. N Engl J Med. 2008;359(13):1317-1329. Mowla A, Doyle J, Lail NS, et al. Delays in door-to-needle time for acute ischemic stroke in the emergency department: A comprehensive stroke center experience. J Neurol Sci. 2017;376:102-105. Wardlaw JM, Mielke O. Early signs of brain infarction at CT: observer reliability and outcome after thrombolytic treatment--systematic review. Radiology. 2005;235(2):444-453. Burdorf BT. Comparing magnetic resonance imaging and computed tomography machine accessibility among urban and rural county hospitals. J Public Health Res. 2021;11(1):2527. Saver JL, Fonarow GC, Smith EE, et al. Time to treatment with intravenous tissue plasminogen activator and outcome from acute ischemic stroke. JAMA. 2013;309(23):2480-2488. Xu L, Sun Y, Wang HB. 基于文献探讨舌诊在中风病健康管理中的应用价值[Article in Chinese].中国中医基础医学杂志（Journal of Basic Chinese Medicine）. 2019;25(3):326-329. Liu Q, Li Y, Yang P, et al. A survey of artificial intelligence in tongue image for disease diagnosis and syndrome differentiation. Digit Health. 2023;9:20552076231191044. Zhou J, Li S, Wang X, et al. Weakly Supervised Deep Learning for Tooth-Marked Tongue Recognition. Front Physiol. 2022;13:847267. Li J, Yuan P, Hu X, et al. A tongue features fusion approach to predicting prediabetes and diabetes with machine learning. J Biomed Inform. 2021;115:103693. Ma C, Zhang P, Du S, Li Y, Li S. Construction of Tongue Image-Based Machine Learning Model for Screening Patients with Gastric Precancerous Lesions. J Pers Med. 2023;13(2):271. Jiang T, Guo XJ, Tu LP, et al. Application of computer tongue image analysis technology in the diagnosis of NAFLD. Comput Biol Med. 2021;135:104622. Yuan L, Yang L, Zhang S, et al. Development of a tongue image-based machine learning tool for the diagnosis of gastric cancer: a prospective multicentre clinical cohort study. EClinicalMedicine. 2023;57:101834. Meretoja A, Keshtkaran M, Saver JL, et al. Stroke thrombolysis: save a minute, save a day. Stroke. 2014;45(4):1053-1058. Katz JM, Wang JJ, Boltyenkov AT, et al. Rescan Time Delays in Ischemic Stroke Imaging: A Retrospective Observation and Analysis of Causes and Clinical Impact. AJNR Am J Neuroradiol. 2021;42(10):1798-1806. Katz JM, Wang JJ, Boltyenkov AT, et al. Rescan Time Delays in Ischemic Stroke Imaging: A Retrospective Observation and Analysis of Causes and Clinical Impact. AJNR Am J Neuroradiol. 2021;42(10):1798-1806. Kuc A, Isenberg DL, Kraus CK, et al. Factors associated with door-in-door-out times in large vessel occlusion stroke patients undergoing endovascular therapy. Am J Emerg Med. 2023;69:87-91. Zhang J, Zhang CH, Lin XL, Zhang Q, Wang J, Shi SL. Serum glial fibrillary acidic protein as a biomarker for differentiating intracerebral hemorrhage and ischemic stroke in patients with symptoms of acute stroke: a systematic review and meta-analysis. Neurol Sci. 2013;34(11):1887-1892. Zhou S, Bao J, Wang Y, Pan S. S100β as a biomarker for differential diagnosis of intracerebral hemorrhage and ischemic stroke. Neurol Res. 2016;38(4):327-332. Bustamante A, Penalba A, Orset C, et al. Blood biomarkers to differentiate ischemic and hemorrhagic strokes. Neurology. 2021;96(15):e1928-e1939. Misra S, Montaner J, Ramiro L, et al. Blood biomarkers for the diagnosis and differentiation of stroke: A systematic review and meta-analysis. Int J Stroke. 2020;15(7):704-721. Mnyusiwalla A, Aviv RI, Symons SP. Radiation dose from multidetector row CT imaging for acute stroke. Neuroradiology. 2009;51(10):635-640. Ma YJ, Li T, Tang YP. 中医舌诊在中风病中的应用研究进展[Article in Chinese].大众科技(Popular Science ＆Technology). 2020;22(12):64-66. Zhang JH, Zhao F, Hui Z, et al. 中医舌诊在脑卒中患者临床应用[Article in Chinese]. 吉林中医药（Jilin Journal of Traditional Chinese Medicine). 2018;38(4):425-428. Liu H, Zhang P, Huang Y, et al. Research on multi-label recognition of tongue features in stroke patients based on deep learning. Sci Rep. 2024;14(1):32144. Ma Y, Liu J, Liu Y, et al. Structure and Illumination Constrained GAN for Medical Image Enhancement. IEEE Trans Med Imaging. 2021;40(12):3955-3967. H. Michaeli, T. Michaeli and D. Soudry, "Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations," 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 2023, pp. 16333-16342 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8586557","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":602985891,"identity":"045064d6-c66a-4f3e-91ba-5ac389e8754d","order_by":0,"name":"Yuwei Pan","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yuwei","middleName":"","lastName":"Pan","suffix":""},{"id":602985892,"identity":"325ae969-9499-4ea3-bb8e-11804cf495ec","order_by":1,"name":"Xiguang Tian","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xiguang","middleName":"","lastName":"Tian","suffix":""},{"id":602985893,"identity":"aab8208c-adec-4d24-8bcc-9e48719a5f54","order_by":2,"name":"Sande Gao","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Sande","middleName":"","lastName":"Gao","suffix":""},{"id":602985894,"identity":"98e452b3-1976-41a3-ad1f-7adc02f9b9aa","order_by":3,"name":"Haoran Cen","email":"","orcid":"","institution":"School of Computer Science, Guangdong Polytechnic Normal University","correspondingAuthor":false,"prefix":"","firstName":"Haoran","middleName":"","lastName":"Cen","suffix":""},{"id":602985896,"identity":"e9f910d4-3f45-4b35-b61c-0e1216a8aa21","order_by":4,"name":"Zihao Huang","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zihao","middleName":"","lastName":"Huang","suffix":""},{"id":602985898,"identity":"73cf2071-140d-4f84-a5fa-a184d1035ff1","order_by":5,"name":"Yixuan Li","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yixuan","middleName":"","lastName":"Li","suffix":""},{"id":602985900,"identity":"562c422f-4774-4f95-a687-a2c970238d47","order_by":6,"name":"Wenjian Du","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wenjian","middleName":"","lastName":"Du","suffix":""},{"id":602985902,"identity":"0ed38743-8747-4e55-aaaa-dd3c6f5c4ad4","order_by":7,"name":"Nan Zhang","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Nan","middleName":"","lastName":"Zhang","suffix":""},{"id":602985903,"identity":"ef582721-9f83-48ab-8549-f6f3fcc29174","order_by":8,"name":"Guodong Li","email":"","orcid":"","institution":"The Affiliated Traditional Chinese Medicine Hospital, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Guodong","middleName":"","lastName":"Li","suffix":""},{"id":602985906,"identity":"c140fe10-2d2a-4d4d-aadb-fa4c8446eb65","order_by":9,"name":"Shaopeng Liu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAxklEQVRIiWNgGAWjYBACNobDBx98/CPBY9/eQKQWPsZjyYYzG2zkDHgOEKlFjvmMmTRnQ5qxgUQCsQ5jO2MgzbjjcOJ2yccbbzDU2EQT1sJzrMC48MzhxJ2z04otGI6l5TYQ1CJxeEPyDLbDiQ23c8wkGBsOE6FF/oHBYR6QlptniNXCcMSwmbcN6P0bPERrOZbMOOOMjZxkD9AvCcT4Rb7h8PEfHyokePjZD2+88aHGhrAWZEB81CBpIVXHKBgFo2AUjAwAACBbRAeFyNJKAAAAAElFTkSuQmCC","orcid":"","institution":"School of Computer Science, Guangdong Polytechnic Normal University","correspondingAuthor":true,"prefix":"","firstName":"Shaopeng","middleName":"","lastName":"Liu","suffix":""}],"badges":[],"createdAt":"2026-01-13 02:58:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8586557/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8586557/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104343445,"identity":"7b746da2-c80c-4621-ba70-8de472037520","added_by":"auto","created_at":"2026-03-10 17:11:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":377879,"visible":true,"origin":"","legend":"\u003cp\u003eThe tongue diagnosis instrument. 1: lens hood, 2: LED light resource, 3: high-definition camera. (A)Front view of the device. (B)Side view of the device. (C)Back view of the device. (D)Image acquisition process.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/6c9a671ac2844f9568d677b9.png"},{"id":104343444,"identity":"bfb1db37-14c1-4eb3-bdc0-72948def081d","added_by":"auto","created_at":"2026-03-10 17:11:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":114329,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of the StrokeNet for classifying ischemic stroke patient and hemorrhagic stroke patient based on tongue images.(A)The architecture of the StrokeNet model. (B)Structural diagram of the HWA.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/0073db44f6a08b6ecf068519.png"},{"id":104343449,"identity":"9f5267c9-f234-40d0-a382-32ef32fddbd0","added_by":"auto","created_at":"2026-03-10 17:11:14","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":241063,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram of patient recruitment and screening\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/2f2b08a20462d77aec69107a.png"},{"id":104405435,"identity":"b8e72310-bf5d-40a2-a89a-9c33ea748065","added_by":"auto","created_at":"2026-03-11 12:22:53","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":110478,"visible":true,"origin":"","legend":"\u003cp\u003eClassification performance of the proposed StrokeNet model on the internal validation dataset. (A)Training and validation loss curves of the StrokeNet model. (B)Confusion matrix of the StrokeNet model. (C)Precision-Recall (PR) curve of the StrokeNet model. (D)Receiver Operating Characteristic (ROC) curve of the StrokeNet model.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/5c80ec977cf029b091d67f66.png"},{"id":104343450,"identity":"9b427fdd-f7a3-40b8-8ecf-0c3b5e6b1628","added_by":"auto","created_at":"2026-03-10 17:11:14","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":190113,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance comparison of different deep learning models. (A) Architecture diagram of the ResNet-18 model. (B) Architecture diagram of the EfficientNet-B2 model. (C) PR curves of three models. (D) ROC curves of three models .\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/4fd1f5418d088ac1dff42f46.png"},{"id":104405427,"identity":"ed841b4b-67a4-4875-96c9-d725461b615f","added_by":"auto","created_at":"2026-03-11 12:22:52","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":116243,"visible":true,"origin":"","legend":"\u003cp\u003eAnalysis of the discriminative performance of the StrokeNet model across different population subgroups. (A) Age subgroup (≥60 years). (B) Age subgroup (\u0026lt;60 years). (C) Gender subgroup (male). (D) Gender subgroup (female). (E) Hypertension history subgroup (with hypertension). (F) Hypertension history subgroup (without hypertension). (G) Diabetes mellitus history subgroup (with diabetes mellitus). (H) Diabetes mellitus history subgroup (without diabetes mellitus). (I) mRS score ≤ 2; (J) mRS score \u0026gt;3.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/5a21da8a355c70e429867db9.png"},{"id":105199363,"identity":"e128a120-e3f5-4070-a9fb-62fe8fbb289f","added_by":"auto","created_at":"2026-03-23 10:57:51","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1718884,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8586557/v1/5081bf14-2e0d-4c8b-8703-724e0cee1319.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Novel Non-Invasive Deep Learning Model Based on Tongue Images for Early Differentiation of Ischemic and Hemorrhagic Stroke","fulltext":[{"header":"1 Background","content":"\u003cp\u003eStroke is the second leading cause of mortality worldwide, responsible for approximately 13.7\u0026nbsp;million incident cases and 6.2\u0026nbsp;million deaths annually. Among these, the acute-phase mortality rate of intracerebral hemorrhage (ICH) reaches as high as 40%\u0026ndash;50%, significantly exceeding that of ischemic stroke (IS). Given that the two conditions require diametrically opposed management strategies, early and accurate differentiation is critical in determining patient outcomes[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Intravenous thrombolysis and endovascular therapy for ischemic stroke must be administered within the golden time window (4.5 hours), whereas inadvertent thrombolysis in ICH patients can lead to hematoma expansion and a marked increase in mortality[\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Although intravenous and endovascular thrombolytic therapies have substantially improved outcomes in IS patients, nearly 40% of patients experience treatment delays in clinical practice due to the inability to obtain timely neuroimaging[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCurrently, computed tomography (CT), the primary imaging modality used in clinical practice, demonstrates an average sensitivity of 66% and specificity of 87% for diagnosing ischemic stroke within 6 hours of onset. Moreover, it is prone to false negatives in cases involving small infarcts[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Although magnetic resonance imaging (MRI) can improve diagnostic accuracy, the average examination time exceeds 45 minutes. Furthermore, even in countries with relatively ample medical resources, the availability of MRI in primary hospitals is consistently lower[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Consequently, the median time to initial treatment for stroke patients is as long as 144 minutes, with only 9.3% of patients receiving treatment within the first third (90 minutes) of the golden therapeutic window. Treatment delay is directly associated with unfavorable outcomes[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. These technical limitations in neuroimaging contribute to the underutilization of the golden time window in acute stroke management, substantially increasing the risk of disability among patients.\u003c/p\u003e \u003cp\u003eTongue diagnosis in Traditional Chinese Medicine (TCM) offers an innovative approach for the differential diagnosis of stroke. TCM theory posits that the tongue reflects systemic physiological and pathological changes, a notion that aligns with modern medical mechanisms including inflammatory responses and microcirculatory dysfunction. Clinical evidence indicates that ischemic stroke patients often present with pale-red tongue with white and greasy coating, whereas those with hemorrhagic stroke typically exhibit dark-purplish tongue with yellow and greasy coating[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], supporting an objective correlation between tongue manifestations and stroke subtypes.\u003c/p\u003e \u003cp\u003eHowever, conventional tongue diagnosis relies heavily on practitioners' subjective experience, leading to suboptimal inter-observer consistency. Advances in deep learning have effectively addressed this limitation, with mechanisms such as attention networks and multi-scale feature extraction having been successfully applied in modern tongue image analysis[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Existing studies have demonstrated the diagnostic utility of tongue imaging in both neoplastic (e.g., gastric, pancreatic, hepatocellular, and colorectal cancers) and non-neoplastic conditions (e.g., gastritis, rheumatoid arthritis, diabetes, and chronic hepatitis B)[\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Nevertheless, most existing models focus on digestive diseases, and a dedicated framework for stroke-specific differential diagnosis remains notably lacking. Moreover, no study to date has employed AI-driven tongue analysis for early discrimination between IS and ICH.\u003c/p\u003e \u003cp\u003eThis study introduces for the first time the StrokeNet model, which integrates residual networks with hierarchical attention mechanisms. In a cross-sectional cohort of 201 patients, it achieved a classification accuracy of 82.93% in distinguishing early cerebral infarction from intracerebral hemorrhage. To our knowledge, this is the first research to apply artificial intelligence (AI)-driven deep learning to evaluate the value of tongue imaging in the early differential diagnosis of these conditions. A comprehensive search of the PubMed database (up to June 2025, without language restrictions), using keywords such as \"stroke\" and \"tongue diagnosis,\" identified no prior publications in this domain.\u003c/p\u003e \u003cp\u003eNotably, the StrokeNet model demonstrates strong clinical feasibility, particularly in resource-limited settings. It requires an average diagnosis time of only 43 seconds and incurs minimal cost, substantially outperforming conventional imaging in both time efficiency and accessibility. This model fills a critical gap as a rapid tool for stroke subtype discrimination and offers a new paradigm for optimizing emergency triage workflows. It holds particular promise for primary care institutions and clinical settings in developing countries, with the potential to enhance the utilization of the golden treatment window and reduce the risks of misdiagnosis and inappropriate treatment.\u003c/p\u003e"},{"header":"2 Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study Population and Ethical Statements\u003c/h2\u003e \u003cp\u003eFrom November 2023 to December 2024, consecutive patients with acute stroke aged 18\u0026ndash;80 years were recruited from the Department of Neurology, Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University. All participants or their legal guardians (for patients with impaired consciousness) signed a written informed consent form, and the study was conducted in accordance with the Declaration of Helsinki (revised 2013). The ethical review of this study was carried out and approved by the Institutional Review Board (IRB) of Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University (IRB number: 2023NK004), with the clinical trial registration number ChiCTR2300077147. All source codes and data analyzed in this study can be obtained from the corresponding author upon reasonable request.\u003c/p\u003e \u003cp\u003e The diagnostic criteria for acute stroke refer to the Chinese Guidelines for the Diagnosis and Treatment of Acute Ischemic Stroke and Chinese Guidelines for the Diagnosis and Treatment of Intracerebral Hemorrhage: intracerebral hemorrhage (ICH) is defined as hyperdense lesions on cranial computed tomography (CT) or hypointense signals on T2* weighted magnetic resonance imaging (MRI); ischemic stroke (IS) is defined as early ischemic signs on CT or hyperintense signals on diffusion-weighted MRI, with symptom onset time\u0026thinsp;\u0026le;\u0026thinsp;24 hours and first-episode stroke. Initially, 256 patients were screened, with the following inclusion and exclusion criteria:\u003c/p\u003e \u003cp\u003eInclusion Criteria: (1) Aged 18\u0026ndash;80 years; (2) Meeting the above diagnostic criteria for acute stroke; (3) Glasgow Coma Scale (GCS) score\u0026thinsp;\u0026ge;\u0026thinsp;12, enabling cooperation with tongue image acquisition; (4) No history of tongue reconstruction surgery or severe maxillofacial trauma.\u003c/p\u003e \u003cp\u003eExclusion Criteria: (1) Presence of severe oral diseases (e.g., oral cancer, extensive ulcers) or active tongue lesions; (2) Comorbidity with systemic diseases affecting tongue morphology (e.g., malignant tumors, systemic lupus erythematosus, Sj\u0026ouml;gren's syndrome); (3) Use of medications known to alter tongue appearance within 2 weeks prior to enrollment (e.g., antibiotics, antifungals, chemotherapeutic agents); (4) Pregnancy or lactation period; (5) Impaired tongue image quality (tongue coverage\u0026thinsp;\u0026ge;\u0026thinsp;10% by teeth/lips, motion blur, or other quality defects). A total of 55 patients were excluded based on the above criteria, and 201 eligible patients were finally included in the analysis.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Data Collection\u003c/h2\u003e \u003cp\u003e Trained research physicians conducted interviews and took tongue photographs of participants following a standardized collection process. The interviews gathered baseline data on general conditions (age, gender), medical history (hypertension, diabetes mellitus), and clinical manifestations (onset time, neurological deficit symptoms).\u003c/p\u003e \u003cp\u003eTongue photographs were collected using a YZAI-02 tongue diagnosis instrument (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) either before breakfast or 2 hours after breakfast. The specific steps for image collection are as follows: (1) Power on the instrument after inspection and adjust the camera parameters. (2) Disinfect the areas of the instrument that may come into direct contact with the participant using 75% alcohol. (3) Instruct the patient to stick out their tongue flatly. (4) Turn on the built-in ring light source and complete the image capture. (5) Following the evaluation of the acquired tongue image, image acquisition is finalized if the image meets predefined quality criteria; otherwise, repeated image capture is performed until the established quality requirements are satisfied.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eQualification criteria for photo quality: no problems such as occlusion, blurring, fogging, overexposure, or underexposure; the tongue should be relaxed and flattened with no twisting or tension; there should be no foreign objects, staining, or other conditions affecting the appearance of the tongue surface.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Data preprocessing\u003c/h2\u003e \u003cp\u003eAll tongue images utilized in this study were acquired using standardized equipment and manually labeled as ischemic stroke (IS) or intracerebral hemorrhage (ICH), corresponding to labels 0 and 1, respectively. The original images included the entire facial region, with the tongue accounting for only a small proportion. To isolate the tongue region and eliminate background noise, we employed a YOLOv5s object detection model trained on a self-constructed dataset. This model automatically identified and localized the tongue region, providing bounding boxes for direct cropping. No image resizing or margin extension was required for the cropped regions. Samples with failed tongue detection were marked as abnormal and excluded from subsequent training.\u003c/p\u003e \u003cp\u003eEach valid sample consisted of a complete facial image and the corresponding cropped tongue region. RGB channels were extracted from both images and concatenated along the channel dimension to form a six-channel composite image, enabling the model to jointly process global and local features. To preserve anatomical details, the images retained their original resolution without resizing.\u003c/p\u003e \u003cp\u003ePrior to model training, all images were normalized using channel-wise means and standard deviations of [0.5, 0.5, 0.5]. To enhance model generalization, data augmentation strategies were implemented during training, including random horizontal flipping, random cropping, and color jittering. These augmentation techniques were applied to the complete facial images and cropped tongue images separately before channel concatenation. The final six-channel images featured a unified structure and consistent labeling, which could be directly used as input for downstream classification models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 StrokeNet-Based Classification Model\u003c/h2\u003e \u003cp\u003eWe proposed a deep convolutional neural network (DCNN) for differentiating ischemic stroke (IS) from intracerebral hemorrhage (ICH) based on tongue images. The network backbone was adapted from ResNet-18 and modified to accommodate six-channel input (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). This model accepts RGB channels from both the original full facial images and cropped tongue regions as inputs, enabling it to learn joint representations of global and local features. Specifically, the first convolutional layer of ResNet-18 was modified to accept six input channels instead of the conventional three.\u003c/p\u003e \u003cp\u003eTo enhance the model\u0026rsquo;s capability to capture fine-grained features such as tongue texture and spatial variations, we incorporated HWAttention modules at six insertion points: before and after the 1st, 2nd, and 3rd residual blocks, respectively. Each HWAttention module integrates discrete wavelet transform (DWT), inverse wavelet transform (IWT), and a dual-channel spatial attention mechanism. These modules assist the model in dynamically emphasizing discriminative features in both spatial and frequency domains (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB).\u003c/p\u003e \u003cp\u003eThe final model architecture comprises a global average pooling layer followed by a fully connected layer that outputs binary prediction outcomes (IS or ICH) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). Due to the limited sample size, to prevent overfitting, we frozen the first two residual blocks of ResNet-18 and only trained the deeper layers along with the classifier. Weighted cross-entropy loss was employed during training to address class imbalance. The stochastic gradient descent (SGD) optimizer was utilized with an initial learning rate of 0.001, which was decayed every two epochs via a StepLR scheduler. The training process lasted for 100 epochs with a batch size of 4, and model checkpoints were saved after each epoch for performance evaluation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Statistical Analysis\u003c/h2\u003e \u003cp\u003eAll statistical analyses were performed using R statistical software (version 4.2.1; R Foundation for Statistical Computing, Vienna, Austria). For data conforming to a normal distribution, descriptive statistics were presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (SD), and intergroup differences were compared using the independent samples t-test (for two-group comparisons). Prior to t-test analysis, Levene's test was conducted to verify the homogeneity of variances: if variances were homogeneous, the standard t-test was applied; if not, Welch's t-test was used as an alternative. For data that did not follow a normal distribution, descriptive statistics were expressed as interquartile range (IQR, P25\u0026thinsp;~\u0026thinsp;P75), and intergroup comparisons were performed using the Mann\u0026ndash;Whitney U test (for two-group nonparametric comparisons).\u003c/p\u003e \u003cp\u003eAdditionally, binary logistic regression analysis was employed to explore potential risk factors for ischemic stroke (IS), with intracerebral hemorrhage (ICH) as the reference group. Results were presented as adjusted odds ratios (OR) with 95% confidence interval (95% CI). A two-tailed \u003cem\u003ep\u003c/em\u003e-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant for all analyses.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eTo comprehensively evaluate the model's performance, we calculated standard classification metrics: accuracy (ACC), sensitivity (Recall), specificity, precision, F1-score, and the area under the receiver operating characteristic curve (AUC). We also leveraged the components of the confusion matrix, comprising true positive (TP), false positive (FP), true negative (TN), and false negative (FN), to conduct a more granular performance analysis.\u003c/p\u003e \u003cp\u003eThe following formulas were used:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{c}\\text{c}\\text{u}\\text{r}\\text{a}\\text{c}\\text{y}=\\frac{\\text{T}\\text{P}+\\text{T}\\text{N}}{\\text{T}\\text{P}+\\text{F}\\text{P}+\\text{T}\\text{N}+\\text{F}\\text{N}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{c}\\text{c}\\text{u}\\text{r}\\text{a}\\text{c}\\text{y}=\\frac{\\text{T}\\text{P}+\\text{T}\\text{N}}{\\text{T}\\text{P}+\\text{F}\\text{P}+\\text{T}\\text{N}+\\text{F}\\text{N}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\text{S}\\text{p}\\text{e}\\text{c}\\text{i}\\text{f}\\text{i}\\text{c}\\text{i}\\text{t}\\text{y}=\\frac{\\text{T}\\text{N}}{\\text{T}\\text{N}+\\text{F}\\text{P}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}=\\frac{\\text{T}\\text{P}}{\\text{T}\\text{P}+\\text{F}\\text{P}}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:\\text{F}1=\\frac{2\\times\\:\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}\\times\\:\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}{\\text{P}\\text{r}\\text{e}\\text{c}\\text{i}\\text{s}\\text{i}\\text{o}\\text{n}+\\text{R}\\text{e}\\text{c}\\text{a}\\text{l}\\text{l}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eAUC was used to assess model robustness across classification thresholds, with values closer to 1 indicating better discriminative power. Class-wise metrics were also reported to evaluate model performance on both mild and severe cases under data imbalance.All metrics were computed on the validation set using predictions from the best-performing model checkpoint (based on highest ACC).\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Results","content":"\u003cp\u003e3.1 Patient recruitment and data analysis\u003c/p\u003e\n\u003cp\u003eThe patient recruitment and screening procedures in this study strictly adhered to the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) statement (Figure 3). A total of 256 consecutive patients with suspected acute stroke were screened from the Department of Neurology, Guangzhou Medical University Affiliated Traditional Chinese Medicine Hospital, between November 2023 and December 2024. All participants completed standardized structured interviews and tongue image acquisition. According to predefined inclusion and exclusion criteria, a stepwise screening was performed, resulting in the exclusion of 55 patients (21.48%). The exclusion reasons included severe oral diseases (n=8), systemic diseases affecting tongue morphology (n=12), use of tongue image-interfering medications within 2 weeks prior to enrollment (n=10), and inadequate tongue image quality (n=25). Finally, 201 eligible patients were included for subsequent analyses.\u003c/p\u003e\n\u003cp\u003eAmong the included patients, 144 (71.64%) were diagnosed with ischemic stroke (IS) and 57 (28.36%) with intracerebral hemorrhage (ICH). A stratified random sampling method was used to partition the dataset into a training set and an internal validation set at an 8:2 ratio, ensuring balanced distribution of key baseline characteristics (e.g., stroke subtype, age, gender) between the two sets: the training set included 160 patients (115 IS, 45 ICH; 79.60%), and the internal validation set included 41 patients (29 IS, 12 ICH; 20.40%).\u003c/p\u003e\n\u003cp\u003eTable 1 presents the detailed baseline clinical characteristics of IS and ICH patients in the training and internal validation sets, including demographic data (age, gender), past medical history (hypertension, diabetes mellitus), vital signs (pulse rate, blood pressure), and neurological function scores [National Institutes of Health Stroke Scale (NIHSS), modified Rankin Scale (mRS)]. Intergroup comparisons were performed using appropriate statistical tests: independent samples t-test or Mann\u0026ndash;Whitney U test for continuous variables, and chi-square test or Fisher\u0026rsquo;s exact test for categorical variables. No statistically significant differences in baseline characteristics were observed between the training set and the validation set (all \u003cem\u003ep\u003c/em\u003e\u0026gt;0.05), confirming the rationality of dataset partitioning and laying a foundation for the reliability of model training and validation.\u003c/p\u003e\n\u003cp\u003eTo identify potential risk factors for ICH, binary logistic regression analysis was conducted with stroke subtype (IS as the reference group) as the dependent variable, and age, gender, history of hypertension, history of diabetes mellitus, pulse rate, blood pressure, and neurological function scores as independent variables (Table 2). The results showed that age\u0026gt; 60 years (adjusted OR = 2.326, 95%CI: 1.243\u0026ndash;4.352, \u003cem\u003ep\u003c/em\u003e = 0.008) was an independent risk factor for ICH (\u003cem\u003ep\u003c/em\u003e\u0026lt;0.05).\u003c/p\u003e\n\u003cp\u003eTable1.Baseline clinical characteristics of IS and ICH patients in the training and internal validation datasets (mean\u0026plusmn;SD or n, %).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"99%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eICH (n = 57)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eIS (n = 144)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eNIHSS score, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026le;15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e57 (28.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e136 (67.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026gt;15\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e8 (4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003emRS score, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0-2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e36 (18.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e104 (51.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e3-5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e21 (10.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e40 (19.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eAge, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026le;60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e28 (14.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e44 (21.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026gt;60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e29 (14.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e100 (49.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eGender, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e36 (18.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e100 (49.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e21 (10.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e44 (21.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eDiabetes, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e44 (22.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e108 (53.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e13 (6.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e36 (17.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eHypertension, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e37 (18.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e86 (42.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e20 (10%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e58 (28.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003ePulse Rate, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026le;90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e52 (26.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e118 (58.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026gt;90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e5 (2.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e26 (12.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eSystolic blood pressure, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026gt;140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e33 (16.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e89 (43.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026le;140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e24 (11.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e55 (27.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003eDiastolic blood pressure, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026gt;90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e20 (10.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e42 (20.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u0026le;90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e37 (18.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e102 (50.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTable2. Binary logistic regression analysis of risk factors for ICH (reference group: IS)\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"99%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eVariables\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003eOR(95% CI)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003eP-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003emRS score (3-5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e0.684 (0.358 \u0026ndash; 1.308)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.251\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eAge\u0026gt;60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e2.326 (1.243 \u0026ndash; 4.352)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e0.783 (0.412 \u0026ndash; 1.489)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.456\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eDiabetes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e1.165 (0.565 \u0026ndash; 2.401)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.680\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eHypertension\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e1.296 (0.686 \u0026ndash; 2.449)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.424\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eHyperlipidemia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e1.528 (0.410 \u0026ndash; 5.689)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.528\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003ePulse Rate\u0026gt;90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e2.356 (0.857 \u0026ndash; 6.471)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.097\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eSystolic blood pressure\u0026le;140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e0.885 (0.475 \u0026ndash; 1.649)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.701\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 43px;\"\u003e\n \u003cp\u003eDiastolic blood pressure\u0026le;90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 36px;\"\u003e\n \u003cp\u003e1.412 (0.740 \u0026ndash; 2.696)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 20px;\"\u003e\n \u003cp\u003e0.296\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e3.2 Classification Performance Analysis\u003c/p\u003e\n\u003cp\u003eThe discriminative performance of the model was comprehensively evaluated using multiple core metrics and visualization curves (Figure 4B, 4C, 4D). On the internal validation set, the StrokeNet model achieved an accuracy of 82.93% (34/41) for differentiating ischemic stroke (IS) from intracerebral hemorrhage (ICH), with an area under the receiver operating characteristic curve (AUC) of 0.8463 (95% CI: 0.7125\u0026ndash;0.9799), a precision of 86.44%, a sensitivity of 89.47% (for IS), a specificity of 68.00% (for ICH), and an F1-score of 0.8793. These metrics collectively demonstrate the model\u0026rsquo;s robust classification efficacy, particularly its high sensitivity in ischemic stroke (IS) identification, which aligns with the core clinical requirement of minimizing missed diagnoses.\u003c/p\u003e\n\u003cp\u003eThe confusion matrix further quantified the model\u0026rsquo;s classification outcomes (Figure 4B). Among the validation set, the model correctly identified 26 IS patients (true positives [TP]) and 8 ICH patients (true negatives [TN]). Only 3 IS cases were misclassified as ICH (false negatives [FN]), and 4 ICH cases were misclassified as IS (false positives [FP]). The false negative rate (missed diagnosis rate) was merely 10.53%, and the false positive rate (misdiagnosis rate) was 32.00%. These results indicate that the model can effectively reduce the risk of missed IS diagnoses in emergency clinical settings, providing actionable support for timely decision-making within the thrombolytic therapy window.\u003c/p\u003e\n\u003cp\u003e3.3 Baseline Model Comparison and Ablation Analysis\u003c/p\u003e\n\u003cp\u003eTo verify the stability and specificity of tongue image features in stroke subtype differentiation while eliminating confounding effects from model architectural discrepancies, two classical deep learning models widely used in clinical research were selected as baseline controls: ResNet-18 and EfficientNet-B2. All models were trained with identical hyperparameters, including the following settings, optimizer: Adam; learning rate: 1\u0026times;10⁻⁴; batch size: 16; training epochs: 200; loss function: cross-entropy loss, to ensure fairness in comparative analysis.\u003c/p\u003e\n\u003cp\u003eVisualization of the three models\u0026rsquo; performance is presented in Figure 5. The PR curve (Figure 5C) shows that StrokeNet achieved a PR value of 0.924, which was higher than that of ResNet-18 (0.888) and EfficientNet-B2 (0.896). Similarly, the ROC curve (Figure 5D) demonstrates that StrokeNet yielded an AUC of 0.846, outperforming ResNet-18 (0.752) and EfficientNet-B2 (0.801).\u003c/p\u003e\n\u003cp\u003eTable 3 provides a quantitative comparison of comprehensive performance metrics across the three models. StrokeNet outperformed the two baseline models in accuracy, recall, F1-score, and AUC. Only the precision of StrokeNet (0.86) was slightly lower than that of ResNet-18. These results confirm that the integrated strategies of wavelet attention mechanism and global-local feature representation in StrokeNet enable more precise capture of key tongue image features associated with stroke subtypes (e.g., tongue color, coating thickness, tongue shape), thereby significantly enhancing the superiority and stability of discriminative performance.\u003c/p\u003e\n\u003cp\u003eTabel 3.Quantitative comparison of classification performance among different deep learning models.\u003c/p\u003e\n\u003cdiv align=\"center\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eAlgorithm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eACC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003eSPE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eSEN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eF1-Score\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eEfficientNet-B2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8400\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.7317\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7869\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.8421\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.8136\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eResNet-18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.7500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.7073\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0.56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.9000\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.7719\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.7857\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21px;\"\u003e\n \u003cp\u003eStrokeNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8463\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.8292\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 10px;\"\u003e\n \u003cp\u003e0.68\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14px;\"\u003e\n \u003cp\u003e0.8644\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.8947\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.8793\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTo evaluate the applicability and robustness of the StrokeNet model across diverse populations, subgroup stratification was performed based on age, gender, past medical history, admission blood pressure, admission pulse rate, and neurological deficit severity (mRS score). The AUC was used to quantify the model\u0026rsquo;s performance in each subgroup.\u003c/p\u003e\n\u003cp\u003eAs illustrated in Figure 6, subgroup-specific performance analysis yielded the following key findings. In terms of age, the model exhibited superior diagnostic performance in the elderly population (\u0026ge; 65 years old) compared to younger individuals. Regarding gender, the model demonstrated better discriminative ability in male patients than in female patients. With respect to risk factors, the model maintained relatively stable performance across subgroups stratified by hypertension or diabetes mellitus status. For admission mRS scores, the model\u0026rsquo;s performance remained consistent between patients with mRS \u0026le; 2 and those with mRS \u0026gt; 3, without significant discrepancies.\u003c/p\u003e\n\u003cp\u003e3.3 Model Robustness and Subgroup Performance\u003c/p\u003e\n\u003cp\u003eTo further assess the robustness and generalization ability of the StrokeNet model, we conducted ablation studies and comparative analyses. Specifically, we evaluated the performance degradation induced by removing the HWAttention modules from the network. The ablated model exhibited noticeable declines in AUC and recall, confirming the critical role of frequency-enhanced attention in capturing discriminative features associated with stroke subtype differentiation.\u003c/p\u003e\n\u003cp\u003eAdditionally, we compared the proposed six-channel input design (integrating full facial images and cropped tongue regions) with conventional three-channel inputs. The six-channel configuration achieved superior classification performance, demonstrating that the integration of global and local visual information yields more comprehensive feature representations and strengthens the model\u0026rsquo;s decision-making robustness.\u003c/p\u003e\n\u003cp\u003eCollectively, these experimental results validate that the proposed StrokeNet model not only attains high classification accuracy but also retains stable performance across different input configurations, highlighting its promising potential for clinical translation.\u003c/p\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eA critical unmet need in the emergency triage of acute stroke lies in the imbalance between the urgency of the therapeutic window and the accessibility of diagnostic tools, which remains a global healthcare challenge[\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. Despite significant advancements in neuroimaging techniques and blood-based biomarker assays, delays in stroke subtype differentiation persist as a key determinant of poor patient outcomes in real-world clinical settings. Notably, the World Health Organization (WHO) has highlighted that unequal access to medical equipment is particularly prevalent in low- and middle-income countries (LMICs), with inadequate availability of essential diagnostic tools, especially in rural and remote regions. This resource gap directly translates to substantial treatment delays[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSpecifically, a retrospective study involving 588 patients with acute ischemic stroke demonstrated that the mean delay in sequential CT imaging (non-contrast CT followed by CTA/CTP) was 53.7 minutes[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Notably, for patients with large vessel occlusion stroke (LVOS) transferred from non-endovascular stroke centers (non-ESCs) to endovascular stroke centers (ESCs) for endovascular treatment (EVT), two key factors were linked to prolonged door-in-door-out (DIDO) times: presentation to a non-stroke certified non-ESC was associated with 20 additional minutes at the transferring hospital, and the acquisition of vascular imaging added another 16 minutes[\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. This primarily attributed to the lack of integrated imaging equipment and standardized protocols in non-ESCs, which necessitate repeated patient transfers or prolonged waits for equipment availability.\u003c/p\u003e \u003cp\u003eBlood-based biomarkers, as valuable supplements to neuroimaging, have emerged as a novel technical avenue for stroke subtype differentiation. Glial fibrillary acidic protein (GFAP) stands as the most well-validated specific biomarker: a meta-analysis encompassing 12 studies demonstrated that serum GFAP exhibited a sensitivity of 81.1% and a specificity of 97% for differentiating intracerebral hemorrhage (ICH) from ischemic stroke (IS) within 1\u0026ndash;6 hours of symptom onset, with significantly higher GFAP levels observed in ICH patients compared to IS patients[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. S100 calcium-binding protein β (S100β) also showed promising discriminative efficacy: at a cutoff value of 67 pg/ml, it achieved an AUC of 0.903 for distinguishing the two subtypes, with a high sensitivity of 95.7% and a relatively lower specificity of 70.4%, which renders it a viable complementary biomarker to GFAP[\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eNotably, N-terminal pro-B-type natriuretic peptide (NT-proBNP) performed exceptionally well in combined detection strategies: when used in conjunction with retinol-binding protein 4 (RBP-4), it identified 29.7% of IS patients with 100% specificity; further integration with GFAP elevated this identification rate to 51.5%[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Furthermore, systematic reviews and meta-analyses have confirmed that additional biomarkers, such as brain natriuretic peptide (BNP), matrix metalloproteinase-9 (MMP-9), and D-dimer, possess moderate discriminative potential[\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Multimarker combinations, including 2GFAP\u0026thinsp;+\u0026thinsp;D-dimer\u0026thinsp;+\u0026thinsp;S100β and RBP-4\u0026thinsp;+\u0026thinsp;NT-proBNP\u0026thinsp;+\u0026thinsp;GFAP, have emerged as particularly promising approaches, optimizing sensitivity while maintaining high specificity (98%\u0026ndash;100%) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eNevertheless, biomarker-based detection still harbors non-negligible limitations. First, no single or combined biomarker has yet obtained approval for routine clinical application, necessitating further validation through large-scale prospective multicenter studies to confirm reliability and generalizability. Second, detection costs for stroke subtype-discriminating biomarkers remain prohibitive given that a single GFAP assay incurs an approximate cost of \u003cspan\u003e$\u003c/span\u003e150\u0026ndash;200, while assays for alternative biomarkers including miR-124-3p and metabolomic indicators demand advanced specialized laboratory infrastructure, a requirement that renders such testing modalities inaccessible in resource-limited healthcare settings. Third, turnaround times for results (1\u0026ndash;2 hours) may still delay critical decisions within the golden therapeutic window. Fourth, certain biomarkers have inherent detection contraindications. For example, metabolomic profiles are prone to interference in patients with severe hepatic or renal dysfunction, potentially leading to false-positive results.\u003c/p\u003e \u003cp\u003eAdditional limitations of current diagnostic tools include inadequate coverage of special populations and poor adaptability to diverse clinical scenarios. Magnetic resonance imaging (MRI) is absolutely contraindicated in patients with severe renal impairment or metallic implants (e.g., cardiac pacemakers). Conversely, computed tomography (CT) entails radiation exposure risks with the mean effective dose for a single non-contrast head CT is approximately 2.7 mSv[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], which pose potential hazards to patients requiring repeated imaging. In settings lacking imaging equipment including emergency scenes and remote areas, healthcare providers are precluded from conducting timely stroke subtype differentiations.\u003c/p\u003e \u003cp\u003e As emphasized in authoritative guidelines, the development of non-invasive, point-of-care rapid differentiation tools represents a critical unmet need in stroke care, which would substantially improve global equity in stroke diagnosis and treatment. This context underscores the clear clinical positioning and guideline alignment of the AI-based tongue image differentiation protocol proposed in the present study.\u003c/p\u003e \u003cp\u003eTongue diagnosis, a core component of inspection diagnosis in traditional Chinese medicine (TCM), is characterized by non-invasiveness, convenience, and cost-effectiveness, serving as a pivotal diagnostic tool in TCM clinical practice[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. In recent years, with advancements in modern science and technology, significant progress has been made in the objectification of tongue diagnosis\u0026mdash;laying a scientific foundation for its integration into modern medical practice[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. The tongue, featuring a thin, delicate mucosa and abundant vascularization, acts as an intuitive window to reflect systemic microcirculatory status[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. For ischemic stroke (IS) patients, insufficient cerebral blood perfusion is often accompanied by systemic microcirculatory disorders, manifesting as pale red tongue color and white greasy coating; this phenomenon is associated with oxidative stress responses and inflammatory factor release induced by cerebral tissue hypoxia in infarcted regions[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. In contrast, intracerebral hemorrhage (ICH) patients, due to cerebrovascular rupture and hemorrhage, exhibit dark purple tongue color and yellow greasy coating, which are closely linked to elevated levels of hemoglobin metabolites released by erythrocyte rupture and vascular endothelial injury markers in the bloodstream [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eHowever, the clinical application of traditional tongue diagnosis is limited by its subjectivity and lack of standardization. Moreover, existing AI models for tongue image analysis primarily concentrate on digestive system diseases, and no specific tongue image-based AI tools have been developed for stroke subtype differentiation, which serves as the core innovative entry point of the present study.\u003c/p\u003e \u003cp\u003eBy integrating TCM tongue diagnosis theory with deep learning technology, we innovatively proposed the StrokeNet model for stroke subtype differentiation based on tongue images, with three core advantages: First, during data preprocessing, the YOLOv5 object detection model was employed to accurately segment the tongue region, effectively eliminating interference from irrelevant tissues (e.g., teeth, lips). Additionally, standardized illumination and image augmentation techniques were utilized to mitigate biases arising from variations in data acquisition environments, providing a high-quality data foundation for feature extraction[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. Second, to address the high-dimensional and non-linear characteristics of tongue images, and in contrast to previous studies that relied primarily on linear regression or shallow feature extraction techniques and thus failed to capture the complex mapping relationships between tongue features, we designed the HWAttention module that integrates discrete wavelet transform (DWT), inverse wavelet transform (IWT), and a dual-channel spatial attention mechanism. This module enables simultaneous capture of key features in both the spatial domain (e.g., tongue shape, coating distribution) and frequency domain (e.g., tongue color gradients, texture details), significantly enhancing the ability to discriminate stroke subtype-specific tongue image differences. Third, by optimizing the input layer structure, we expanded the traditional three-channel image to a six-channel input, realizing efficient fusion of multi-modal information. This not only alleviates the problem of reduced specificity caused by feature aliasing in conventional convolutional neural networks (CNNs) but also avoids the computational burden associated with a sharp increase in parameters[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e \u003cp\u003ePerformance validation results demonstrated that the StrokeNet model achieved an accuracy of 82.93%, an AUC of 0.8463, and a sensitivity of 89.47% on the internal validation set, while simultaneously outperforming mainstream baseline models including ResNet-18 (AUC\u0026thinsp;=\u0026thinsp;0.752) and EfficientNet-B2 (AUC\u0026thinsp;=\u0026thinsp;0.801). Notably, its high sensitivity aligns with the core clinical requirement of minimizing missed diagnoses for acute stroke triage.\u003c/p\u003e \u003cp\u003eSubgroup analysis further confirmed the clinical generalizability of the StrokeNet model: it maintained stable performance across subgroups stratified by age, gender and exhibited no significant differences in subgroups with or without hypertension/diabetes mellitus, as well as across different neurological deficit severities.This finding is particularly notable, given that age\u0026thinsp;\u0026gt;\u0026thinsp;60 years was identified as an independent risk factor for intracerebral hemorrhage (ICH) in our regression analysis and that stroke patients with comorbid hypertension or diabetes mellitus often present with more complex clinical conditions, both of which are factors that tend to interfere with traditional stroke subtype differentiation methods.The stability of the StrokeNet model stems not only from the precise capture of core tongue image features by the HWAttention module, but also from the inherent nature of tongue images as a window reflecting systemic physiological and pathological states. Tongue manifestations mirror holistic changes in microcirculation, which are less susceptible to interference from individual underlying diseases.\u003c/p\u003e \u003cp\u003eCompared with existing diagnostic methods, the core advantages of the StrokeNet model lie in its non-invasiveness, low cost, and operational simplicity. It eliminates the need for expensive imaging or laboratory equipment, requiring only standardized tongue image acquisition via a dedicated device to complete differentiation within 1 minute. This makes it particularly suitable for resource-limited settings such as primary healthcare facilities and emergency on-site triage, offering a new paradigm for addressing the global challenge of difficult and delayed stroke diagnosis.\u003c/p\u003e \u003cp\u003eNevertheless, the present study has several limitations: First, it was a single-center cross-sectional study with a relatively limited sample size (n\u0026thinsp;=\u0026thinsp;201), and all included patients were recruited from the Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University, which may introduce selection bias. The external validity of the model thus requires further verification in multi-center, large-sample cohorts. Second, no independent external validation set was incorporated, and the generalization ability of the model needs to be tested using datasets from diverse geographical regions and populations. Third, potential confounding factors affecting tongue manifestations, including medication use (e.g., antiplatelet agents, antihypertensive drugs), diet, and oral hygiene status, were not accounted for in the present study. Future studies should refine inclusion/exclusion criteria and incorporate confounding factor adjustment to enhance result reliability. Fourth, the model was developed based solely on tongue image data, without integrating clinical parameters (e.g., age, pulse rate) or biomarkers. Constructing a multi-modal fusion model that combines tongue images with clinical and laboratory data holds promise for further improving discriminative performance.\u003c/p\u003e \u003cp\u003eFuture research directions can focus on three key areas: First, conducting multi-center, large-sample prospective studies to enroll patients from different regions and healthcare institutions, establishing a standardized dataset with an independent external validation set to optimize model parameters and validate its real-world applicability. Second, exploring the association between tongue image features and stroke prognosis, developing a comprehensive AI tool integrating early subtype differentiation, prognosis prediction, and treatment guidance to further enhance clinical utility. Third, investigating the molecular biological mechanisms underlying the correlation between tongue manifestations and stroke subtypes, providing novel experimental evidence for the integration of traditional TCM theory with modern medicine.\u003c/p\u003e"},{"header":"4 Conclusion","content":"\u003cp\u003eThis study for the first time integrated deep learning technology with TCM tongue diagnosis to develop the StrokeNet model for early differentiation of acute stroke subtypes. Characterized by non-invasiveness, low cost, rapidity, and high sensitivity, the model holds broad application prospects in primary healthcare and emergency settings. The findings not only validate the practical value of tongue images in stroke subtype differentiation but also provide a successful example of the interdisciplinary integration of traditional medical knowledge and modern artificial intelligence technology. This work is expected to offer new insights and tools for optimizing the global stroke diagnosis and treatment system.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthical Statements\u003c/p\u003e\n\u003cp\u003eAll participants or their legal guardians (for patients with impaired consciousness) signed a written informed consent form, and the study was conducted in accordance with the Declaration of Helsinki (revised 2013). The ethical review of this study was carried out and approved by the Institutional Review Board (IRB) of Affiliated Traditional Chinese Medicine Hospital of Guangzhou Medical University (IRB number: 2023NK004), with the clinical trial registration number ChiCTR2300077147. All source codes and data analyzed in this study can be obtained from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003eData Availability Statement\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll data supporting the findings of this study are not publicly available due to ethical restrictions and the terms of the informed consent obtained from participants, which prohibit the unrestricted sharing of sensitive clinical information. However, de-identified data can be requested from the first author, Dr. Yuwei Pan, via email at [email protected] for non-commercial research purposes. Requests will be reviewed by the Institutional Review Board (IRB) of the Affiliated TCM Hospital of Guangzhou Medical University to ensure compliance with ethical guidelines and participant privacy protections.\u003c/p\u003e\n\u003cp\u003eConflict of Interest\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflicts of interest.\u003c/p\u003e\n\u003cp\u003eAcknowledgement\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Funding by Guangzhou Science and Technology Bureau Municipal University Institute Joint Project (Grant Number:2025A03J3431), Young Science and Technology Talents Fund of The Affiliated TCM Hospital of Guangzhou Medical University (Grant Number:2023RC15), Youth Talent Projects in Guangzhou Medical University (Grant Number:2024SRP211)\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAuthor Contributions\u003c/p\u003e\n\u003cp\u003eYuwei Pan, Shaopeng Liu contributed substantially to the design of work, data analysis and experimental procedure. Haoran Cen assisted in modeling. Xiguang Tian, Sande Gao, Zihao Huang, Yixuan Li, Wenjian Du, Nan Zhang, Guodong Li assisted with the data collection.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eGBD 2021 Stroke Risk Factor Collaborators. Global, regional, and national burden of stroke and its risk factors, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet Neurol. 2024;23(10):973-1003.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eO\u0026apos;Donnell MJ, Xavier D, Liu L, et al. Risk factors for ischaemic and intracerebral haemorrhagic stroke in 22 countries (the INTERSTROKE study): a case-control study. Lancet. 2010;376(9735):112-123.\u003c/li\u003e\n \u003cli\u003eHacke W, Kaste M, Bluhmki E, et al. Thrombolysis with alteplase 3 to 4.5 hours after acute ischemic stroke. N Engl J Med. 2008;359(13):1317-1329.\u003c/li\u003e\n \u003cli\u003eMowla A, Doyle J, Lail NS, et al. Delays in door-to-needle time for acute ischemic stroke in the emergency department: A comprehensive stroke center experience. J Neurol Sci. 2017;376:102-105.\u003c/li\u003e\n \u003cli\u003eWardlaw JM, Mielke O. Early signs of brain infarction at CT: observer reliability and outcome after thrombolytic treatment--systematic review. Radiology. 2005;235(2):444-453.\u003c/li\u003e\n \u003cli\u003eBurdorf BT. Comparing magnetic resonance imaging and computed tomography machine accessibility among urban and rural county hospitals. J Public Health Res. 2021;11(1):2527.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eSaver JL, Fonarow GC, Smith EE, et al. Time to treatment with intravenous tissue plasminogen activator and outcome from acute ischemic stroke. JAMA. 2013;309(23):2480-2488.\u003c/li\u003e\n \u003cli\u003eXu L, Sun Y, Wang HB.\u0026nbsp;基于文献探讨舌诊在中风病健康管理中的应用价值[Article in Chinese].中国中医基础医学杂志（Journal of Basic Chinese Medicine）. 2019;25(3):326-329.\u003c/li\u003e\n \u003cli\u003eLiu Q, Li Y, Yang P, et al. A survey of artificial intelligence in tongue image for disease diagnosis and syndrome differentiation. Digit Health. 2023;9:20552076231191044.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eZhou J, Li S, Wang X, et al. Weakly Supervised Deep Learning for Tooth-Marked Tongue Recognition. Front Physiol. 2022;13:847267.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eLi J, Yuan P, Hu X, et al. A tongue features fusion approach to predicting prediabetes and diabetes with machine learning. J Biomed Inform. 2021;115:103693.\u003c/li\u003e\n \u003cli\u003eMa C, Zhang P, Du S, Li Y, Li S. Construction of Tongue Image-Based Machine Learning Model for Screening Patients with Gastric Precancerous Lesions. J Pers Med. 2023;13(2):271.\u003c/li\u003e\n \u003cli\u003eJiang T, Guo XJ, Tu LP, et al. Application of computer tongue image analysis technology in the diagnosis of NAFLD. Comput Biol Med. 2021;135:104622.\u003c/li\u003e\n \u003cli\u003eYuan L, Yang L, Zhang S, et al. Development of a tongue image-based machine learning tool for the diagnosis of gastric cancer: a prospective multicentre clinical cohort study. EClinicalMedicine. 2023;57:101834.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMeretoja A, Keshtkaran M, Saver JL, et al. Stroke thrombolysis: save a minute, save a day. Stroke. 2014;45(4):1053-1058.\u003c/li\u003e\n \u003cli\u003eKatz JM, Wang JJ, Boltyenkov AT, et al. Rescan Time Delays in Ischemic Stroke Imaging: A Retrospective Observation and Analysis of Causes and Clinical Impact. AJNR Am J Neuroradiol. 2021;42(10):1798-1806.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eKatz JM, Wang JJ, Boltyenkov AT, et al. Rescan Time Delays in Ischemic Stroke Imaging: A Retrospective Observation and Analysis of Causes and Clinical Impact. AJNR Am J Neuroradiol. 2021;42(10):1798-1806.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eKuc A, Isenberg DL, Kraus CK, et al. Factors associated with door-in-door-out times in large vessel occlusion stroke patients undergoing endovascular therapy. Am J Emerg Med. 2023;69:87-91.\u003c/li\u003e\n \u003cli\u003eZhang J, Zhang CH, Lin XL, Zhang Q, Wang J, Shi SL. Serum glial fibrillary acidic protein as a biomarker for differentiating intracerebral hemorrhage and ischemic stroke in patients with symptoms of acute stroke: a systematic review and meta-analysis. Neurol Sci. 2013;34(11):1887-1892.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eZhou S, Bao J, Wang Y, Pan S. S100\u0026beta; as a biomarker for differential diagnosis of intracerebral hemorrhage and ischemic stroke. Neurol Res. 2016;38(4):327-332.\u003c/li\u003e\n \u003cli\u003eBustamante A, Penalba A, Orset C, et al. Blood biomarkers to differentiate ischemic and hemorrhagic strokes. Neurology. 2021;96(15):e1928-e1939.\u003c/li\u003e\n \u003cli\u003eMisra S, Montaner J, Ramiro L, et al. Blood biomarkers for the diagnosis and differentiation of stroke: A systematic review and meta-analysis. Int J Stroke. 2020;15(7):704-721.\u003c/li\u003e\n \u003cli\u003eMnyusiwalla A, Aviv RI, Symons SP. Radiation dose from multidetector row CT imaging for acute stroke. Neuroradiology. 2009;51(10):635-640.\u003c/li\u003e\n \u003cli\u003eMa YJ, Li T, Tang YP.\u0026nbsp;中医舌诊在中风病中的应用研究进展[Article in Chinese].大众科技(Popular Science\u0026nbsp;＆Technology). 2020;22(12):64-66.\u003c/li\u003e\n \u003cli\u003eZhang JH, Zhao F, Hui Z, et al.\u0026nbsp;中医舌诊在脑卒中患者临床应用[Article in Chinese].\u0026nbsp;吉林中医药（Jilin Journal of Traditional Chinese Medicine). 2018;38(4):425-428.\u003c/li\u003e\n \u003cli\u003eLiu H, Zhang P, Huang Y, et al. Research on multi-label recognition of tongue features in stroke patients based on deep learning. Sci Rep. 2024;14(1):32144.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMa Y, Liu J, Liu Y, et al. Structure and Illumination Constrained GAN for Medical Image Enhancement. IEEE Trans Med Imaging. 2021;40(12):3955-3967.\u003c/li\u003e\n \u003cli\u003eH. Michaeli, T. Michaeli and D. Soudry, \u0026quot;Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations,\u0026quot; 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 2023, pp. 16333-16342\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Stroke, Ischemic stroke, Intracerebral Hemorrhagic stroke, Tongue diagnosis, Deep learning, Attention mechanism","lastPublishedDoi":"10.21203/rs.3.rs-8586557/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8586557/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eStroke, the second leading cause of death globally, demands rapid differentiation between ischemic stroke (IS) and intracerebral hemorrhage (ICH) to guide optimal therapeutic strategies. Current diagnostic modalities (computed tomography [CT] or magnetic resonance imaging [MRI]) are hindered by limited accessibility and prolonged turnaround times, particularly in resource-constrained settings. Herein, we propose StrokeNet, a deep learning model integrating a modified ResNet-18 backbone with wavelet-based hierarchical attention (HWAttention) modules, for the early differentiation of stroke subtypes via tongue images, which serve as an underutilized non-invasive diagnostic tool in modern stroke care.\u003c/p\u003e\n\u003cp\u003eA single-center cross-sectional study was conducted, enrolling 201 acute stroke patients (144 IS, 57 ICH). Tongue regions were accurately segmented using the YOLOv5 object detection model, and six-channel composite images (combining full facial and cropped tongue RGB features) were constructed as model inputs. On the internal validation set, StrokeNet achieved a classification accuracy of 82.93%, an area under the receiver operating characteristic curve (AUC) of 0.8463 (95% CI:0.7125–0.9799), a sensitivity of 89.47% (for IS), a specificity of 68.00% (for ICH), and an F1-score of 0.8793. The model outperformed mainstream baseline architectures (EfficientNet-B2, ResNet-18) across key metrics, with ablation experiments confirming that the HWAttention module and six-channel input design synergistically enhanced discriminative feature capture. Clinical risk factor analysis identified age \u0026gt; 60 years as an independent predictor of ICH (adjusted OR = 2.326, 95% CI: 1.243–4.352, \u003cem\u003ep\u003c/em\u003e = 0.008).\u003c/p\u003e\n\u003cp\u003eSubgroup analysis validated the model’s robustness across age, gender, comorbidity status and neurological deficit severity. To our knowledge, this study is the first to leverage artificial intelligence (AI)-driven tongue imaging for stroke subtype differentiation. StrokeNet offers non-invasiveness, rapid diagnostic turnaround and low cost, establishing a novel paradigm to optimize emergency triage in resource-limited healthcare settings and address global disparities in stroke care.\u003c/p\u003e","manuscriptTitle":"Novel Non-Invasive Deep Learning Model Based on Tongue Images for Early Differentiation of Ischemic and Hemorrhagic Stroke","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-10 17:11:09","doi":"10.21203/rs.3.rs-8586557/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0a2b101a-a6c4-48ac-b4be-aa0394871654","owner":[],"postedDate":"March 10th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":64157536,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":64157537,"name":"Health sciences/Health care"},{"id":64157538,"name":"Health sciences/Medical research"},{"id":64157539,"name":"Health sciences/Neurology"}],"tags":[],"updatedAt":"2026-03-23T10:57:06+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-10 17:11:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8586557","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8586557","identity":"rs-8586557","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}