Intravitreal anti-vascular endothelial growth factor(anti-VEGF)therapy has been widely used, but the variability in its therapeutic efficacy limits individualized treatment. In recent years, the application of artificial intelligence(AI)has opened up new avenues for personalized treatment response prediction, and its core branches include machine learning(ML)and deep learning(DL). This review systematically retrieved and analyzed 41 relevant studies published up to April 2025. Comprehensive analysis reveals that AI predictive models are evolving from forecasting single endpoints(such as visual acuity or central retinal thickness)to integrating multi-dimensional endpoints(encompassing anatomical, functional, and treatment demand parameters)and generating predictive imaging outputs. In terms of technical approaches, DL models(28 studies, accounting for 68.3%)dominate this field due to their robust image interpretation capabilities, while ML models(10 studies, 24.4%)retain significant value in the analysis of structured clinical data. Cross-disease comparisons indicate that research efforts are most concentrated on age-related macular degeneration(ARMD)and diabetic macular edema(DME), with shared conceptual frameworks for model construction, yet distinct anatomical and functional indicators are prioritized for each disease. Currently, the field confronts several key challenges, including insufficient prospective clinical validation, limited model interpretability(the “black box problem”), and a scarcity of high-quality multi-center datasets. Moving forward, it is imperative to advance real-world validation and develop explainable AI techniques to expedite the clinical translation of these predictive models.

This paper summarizes Professor LU Fang's clinical experience in treating systemic lupus erythematosus （SLE） based on the differentiation and treatment of heat-toxin and blood-stasis in the collaterals. SLE is generally characterized by deficiency in origin with excess in manifestation. The core pathogenesis is heat-toxin obstructing the collaterals. During the acute active stage， the predominant pattern is blazing heat-toxin causing blood stasis， while in the chronic remitting stage， the main pattern is toxic stasis blocking the collaterals with qi and yin deficiency. Clinical treatment follows the basic principle that treat with salty-cold herbs， when heat invades internally and that assist with acrid-dispersing herbs when stasis obstructs the collaterals. The self-formulated Yimian Decoction （抑免汤） serves as the base formula and is applied in stages. During the acute active stage， it is often combined with herbs for clearing heat and detoxifying， cooling blood and resolving stasis， and unblocking the collaterals. In the chronic remitting stage， it is often combined with herbs for activating blood circulation and unblocking the collaterals， as well as tonifying qi and nourishing yin.

ObjectiveTo evaluate the antitumor activity of Periplaneta americana extract（PAE） against human-derived lung adenocarcinoma organoids（LUAD-PDOs） and to elucidate its potential mechanism based on transcriptomics. MethodsFresh tumor and adjacent normal tissues from patients with LUAD were collected to construct LUAD-PDOs and normal lung organoid（Nor-PDOs） models using 3D organoid culture technology. The effective intervention concentration of PAE was determined using the cell counting kit-8（CCK-8） assay. Experimental groups included the model group（LUAD-PDOs）， normal group， model administration group（LUAD-PDOs+PAE）， and normal administration group（Nor-PDOs+PAE）. Hematoxylin-eosin（HE） staining was used to observe the pathological structures of PDOs， immunohistochemistry（IHC） was performed to detect the expressions of the proliferation marker Ki-67 and lung adenocarcinoma differentiation markers cytokeratin-7（CK-7） and Napsin A， TUNEL staining was applied to detect cell apoptosis. RNA sequencing（RNA-Seq） was conducted to identify differentially expressed genes（DEGs）， followed by Gene Ontology（GO）， Kyoto Encyclopedia of Genes and Genomes（KEGG）， and Gene Set Enrichment Analysis（GSEA）， alongside protein-protein interaction（PPI） network analysis to screen core mechanisms. Finally， key targets were validated by integrating external database analysis with immunofluorescence（IF）. ResultsNor-PDOs and LUAD-PDOs that highly recapitulated the pathological characteristics of the primary tissues were successfully established. The CCK-8 assay determined that the effective intervention concentration of PAE was 16 g·L^-1. Morphological observation showed that Nor-PDOs exhibited lumen-forming structures， whereas LUAD-PDOs displayed dense， solid structures. CCK-8 and TUNEL assays revealed that， compared with the model group， PAE intervention inhibited the proliferation of LUAD-PDOs and promoted apoptosis in LUAD cells， while showing no significant effect on the viability of Nor-PDOs. Transcriptomic analysis identified 719 DEGs that were significantly reversed after PAE intervention（347 up-regulated and 372 down-regulated）（P<0.05）. GO enrichment analysis indicated that DEGs in the model administration group were significantly enriched in biological processes related to cell cycle regulation compared to the model group. KEGG pathway analysis revealed that PAE affected pathways related to proliferation and metabolism， including pathways in cancer and the p53 signaling pathway. GSEA further confirmed that PAE significantly enhanced the activity of the p53 signaling pathway（P<0.05）. PPI network analysis indicated that breast cancer type 1 susceptibility protein（BRCA1） and checkpoint kinase 1（CHEK1） were the core down-regulated targets in the p53 pathway. IF verified the high expression of BRCA1 and CHEK1 in LUAD-PDOs and their significant downregulation after PAE intervention（P<0.05）. Furthermore， survival analysis based on The Cancer Genome Atlas（TCGA） database indicated that low expression of BRCA1 and CHEK1 was significantly associated with prolonged overall survival in patients with LUAD（P<0.05）. ConclusionPAE effectively inhibits proliferation of LUAD-PDOs and promotes their apoptosis， its anti-tumor mechanism is potentially associated with the activation of the p53 signaling pathway， with BRCA1 and CHEK1 genes likely serving as key downstream targets for the effects of PAE.

ObjectiveTo investigate the therapeutic effect of sitravatinib on carbon tetrachloride （CCl₄）-induced liver fibrosis in mice. MethodsA total of 30 male C57BL/6J mice， aged 8 weeks， were randomly divided into control group， CCl₄ model group， and low- （5 mg/kg）， middle- （10 mg/kg）， and high-dose （20 mg/kg） sitravatinib groups. All mice except those in the control group were given intraperitoneal injection of CCl₄ for 4 consecutive weeks to induce liver fibrosis， and since the first day of modeling， the mice in the low-， middle-， and high-dose sitravatinib groups were given sitravatinib at the corresponding dose by gavage every day. The serum levels of total cholesterol （TC）， triglyceride （TG）， and alanine aminotransferase （ALT） were measured for the mice in each group； hepatic hydroxyproline content was measured； HE staining， Masson staining， and Sirius Red staining were used to observe liver histopathological changes； quantitative real-time PCR and Western blot were used to measure the mRNA and protein expression levels of α-smooth muscle actin （α-SMA） and collagen type I alpha 1 （Col1a1） in liver tissue. The therapeutic effect of sitravatinib was assessed based on the above results. A one-way analysis of variance was used for comparison of continuous data between multiple groups， and the least significant difference t-test was used for further comparison between two groups. ResultsCompared with the control group， the model group had significant increases in the levels of TC， TG， and ALT （all P<0.05）， and there were no significant differences in the levels of TC， TG， and ALT between the model group and the low-， middle-， and high-dose sitravatinib groups （all P>0.05）. Hepatic hydroxyproline content decreased after sitravatinib intervention， with a significant difference between the middle-/high-dose sitravatinib groups and the CCl₄ model group （both P<0.05）. Histopathological staining showed that the sitravatinib treatment groups had a reduction in collagen deposition， along with thinning and fragmentation of fibrous septa， and in the high-dose sitravatinib group， 4 mice had a fibrosis stage of S0—S1 and 2 mice had a fibrosis stage of S2—S3， suggesting a certain degree of alleviation of liver fibrosis degree compared with the CCl₄ model group （mainly S3—S4）. The measurement of related molecules showed that sitravatinib downregulated the mRNA and protein expression levels of α-SMA and Col1a1 （all P<0.05）. ConclusionSitravatinib can effectively alleviate CCl₄-induced liver fibrosis in mice， possibly by inhibiting hepatic stellate cell activation and collagen synthesis.

Objective To analyze the main causes and risk factors of unplanned reoperation after liver transplantation. Methods The clinical data of 242 liver transplant recipients in the First Affiliated Hospital of Anhui Medical University from January 2015 to December 2024 were retrospectively analyzed. According to whether unplanned reoperation was performed during the same hospitalization after surgery, the recipients were divided into the reoperation group (n=36) and the non-reoperation group (n=206). The preoperative, intraoperative and postoperative data of the two groups, as well as donor and graft-related data, were compared to analyze the risk factors of unplanned reoperation after liver transplantation and the survival status of the two groups. Results Among the 242 liver transplant recipients, 36 underwent unplanned reoperations, with a total of 54 procedures including various laparotomies, endoscopic and interventional surgeries, among which there were 20 laparotomies, 18 endoscopic surgeries and 16 interventional surgeries. The most common cause of unplanned reoperation was biliary complications (20 times), followed by vascular complications (17 times). Compared with the non-reoperation group, the reoperation group had longer graft cold ischemia time, higher postoperative fatality rate of recipients, longer length of stay in the intensive care unit and postoperative hospital stay, and higher total hospitalization costs (all P<0.05). The incidence of unplanned reoperation was higher in recipients who underwent split liver transplantation (P<0.05). Multivariate analysis showed that intraoperative blood loss ≥1 000 mL, positive culture of graft perfusate and split liver transplantation were independent risk factors for unplanned reoperation (all P<0.05). The postoperative 7-day, 1-month, 3-month and 6-month survival rates of recipients in the reoperation group and the non-reoperation group were 100% vs. 98.1%, 88.9% vs. 94.2%, 69.4% vs. 90.8% and 66.7% vs. 90.8%, respectively, and the postoperative survival rate of recipients in the reoperation group was lower than that in the non-reoperation group (P<0.05). Conclusions The main causes of unplanned reoperation after liver transplantation are biliary complications, vascular complications, abdominal incision infection and intra-abdominal hemorrhage. Intraoperative massive blood loss, positive culture of graft perfusate and split liver transplantation are the risk factors associated with unplanned reoperation after liver transplantation.

BACKGROUND:The Guilu Erxian Glue consists of Testudinis Plastrum,Cornu Cervi,Lycii Fructus,and Ginseng Radix.In earlier clinical observations,it is discovered that using Guilu Erxian Glue to treat patients with liver-kidney deficiency type knee osteoarthritis effectively alleviated knee pain,increased the range of motion,and improved walking ability.However,the exact mechanism by which oral administration of Guilu Erxian Glue can produce local therapeutic effects on the knee joint is still unclear.OBJECTIVE:To investigate the effects of Guilu Erxian Glue on gut microbiota in rats with knee osteoarthritis and to evaluate its mechanism using 16S rDNA sequencing and machine learning analysis.METHODS:Totally 18 female SD rats were randomly divided into three groups:blank group,model group,and Guilu Erxian Glue group,with 6 rats in each group.A knee osteoarthritis model was prepared using the destabilization of the medial meniscus surgical method.After successful modeling,the Guilu Erxian Glue group was given a decoction of Guilu Erxian Glue by gavage,while the blank and model groups were given an equal amount of distilled water.After 28 days of continuous intervention,high performance liquid chromatography was used to detect the active ingredients of Guilu Erxian Glue.MRI imaging was used to observe the condition of rat knee articular cartilage.Fecal samples were collected;DNA was extracted using a kit,amplified and purified by PCR,and an Illumina sequencing library was constructed.The Illumina MiSeq platform was used for high-throughput sequencing to generate raw sequence data.After obtaining the raw data,QIIME2 software was used to process the data.Linear Discriminant Analysis Effect Size analysis and random forest algorithm were used to screen for differential species in microbial data.KEGG and MetaCyc functional pathway analyses were used to explore the association between key microbial communities and experimental groups.Linear discriminant analysis effect values and random forest algorithm were used to screen for differential species.Association networks were used to analyze the interactions between microbial communities,and machine learning methods were used to analyze the composition and changes of gut microbiota.RESULTS AND CONCLUSION:(1)LC-MS component identification was conducted on the traditional Chinese medicine formula of Guilu Erxian Glue,and a total of 7 effective ingredients were identified.(2)MRI imaging showed that synovitis scope of high-density shadows in rats of the Guilu Erxian Glue group was reduced,and the degeneration of medial femoral condyle cartilage was less than that in the model group.(3)16S rDNA sequencing showed that the model group rats exhibited significant microbial imbalance,with a significant decrease in the abundance of Firmicutes and Bacteroidetes at the phylum level,while the proportion of Proteobacteria increased significantly(P＜0.05).The gut microbiota structure of rats in the Guilu Erxian Glue group was significantly improved,and the proportion of Firmicutes and Bacteroidetes increased,restoring a more diverse microbiota composition,approaching that of the blank group(P＜0.05).(4)KEGG and MetaCyc functional pathway analysis showed that the Guilu Erxian Glue group significantly activated multiple metabolic pathways,including amino acid metabolism,lipid metabolism,and biotin synthesis pathways(P＜0.05).(5)The results indicate that Guilu Erxian Glue contains seven active ingredients,and the changes in gut microbiota of knee osteoarthritis rats were analyzed using 16S rDNA sequencing.Guilu Erxian Glue can significantly improve the imbalance of gut microbiota,restore the abundance of beneficial bacteria,and have a significant impact on the composition of gut microbiota,providing scientific basis for the efficacy and mechanism of Guilu Erxian Glue.

BACKGROUND:The Guilu Erxian Glue consists of Testudinis Plastrum,Cornu Cervi,Lycii Fructus,and Ginseng Radix.In earlier clinical observations,it is discovered that using Guilu Erxian Glue to treat patients with liver-kidney deficiency type knee osteoarthritis effectively alleviated knee pain,increased the range of motion,and improved walking ability.However,the exact mechanism by which oral administration of Guilu Erxian Glue can produce local therapeutic effects on the knee joint is still unclear.OBJECTIVE:To investigate the effects of Guilu Erxian Glue on gut microbiota in rats with knee osteoarthritis and to evaluate its mechanism using 16S rDNA sequencing and machine learning analysis.METHODS:Totally 18 female SD rats were randomly divided into three groups:blank group,model group,and Guilu Erxian Glue group,with 6 rats in each group.A knee osteoarthritis model was prepared using the destabilization of the medial meniscus surgical method.After successful modeling,the Guilu Erxian Glue group was given a decoction of Guilu Erxian Glue by gavage,while the blank and model groups were given an equal amount of distilled water.After 28 days of continuous intervention,high performance liquid chromatography was used to detect the active ingredients of Guilu Erxian Glue.MRI imaging was used to observe the condition of rat knee articular cartilage.Fecal samples were collected;DNA was extracted using a kit,amplified and purified by PCR,and an Illumina sequencing library was constructed.The Illumina MiSeq platform was used for high-throughput sequencing to generate raw sequence data.After obtaining the raw data,QIIME2 software was used to process the data.Linear Discriminant Analysis Effect Size analysis and random forest algorithm were used to screen for differential species in microbial data.KEGG and MetaCyc functional pathway analyses were used to explore the association between key microbial communities and experimental groups.Linear discriminant analysis effect values and random forest algorithm were used to screen for differential species.Association networks were used to analyze the interactions between microbial communities,and machine learning methods were used to analyze the composition and changes of gut microbiota.RESULTS AND CONCLUSION:(1)LC-MS component identification was conducted on the traditional Chinese medicine formula of Guilu Erxian Glue,and a total of 7 effective ingredients were identified.(2)MRI imaging showed that synovitis scope of high-density shadows in rats of the Guilu Erxian Glue group was reduced,and the degeneration of medial femoral condyle cartilage was less than that in the model group.(3)16S rDNA sequencing showed that the model group rats exhibited significant microbial imbalance,with a significant decrease in the abundance of Firmicutes and Bacteroidetes at the phylum level,while the proportion of Proteobacteria increased significantly(P＜0.05).The gut microbiota structure of rats in the Guilu Erxian Glue group was significantly improved,and the proportion of Firmicutes and Bacteroidetes increased,restoring a more diverse microbiota composition,approaching that of the blank group(P＜0.05).(4)KEGG and MetaCyc functional pathway analysis showed that the Guilu Erxian Glue group significantly activated multiple metabolic pathways,including amino acid metabolism,lipid metabolism,and biotin synthesis pathways(P＜0.05).(5)The results indicate that Guilu Erxian Glue contains seven active ingredients,and the changes in gut microbiota of knee osteoarthritis rats were analyzed using 16S rDNA sequencing.Guilu Erxian Glue can significantly improve the imbalance of gut microbiota,restore the abundance of beneficial bacteria,and have a significant impact on the composition of gut microbiota,providing scientific basis for the efficacy and mechanism of Guilu Erxian Glue.

Objective To prepare the recombinant Echinococcus granulosus Polo-like kinase 2 (rEgPLK2) protein and evaluate its immunoprotective efficacy against cystic echinococcosis, so as to provide insights into research and development of novel vaccines against echinococcosis. Methods The Polo-like kinase (PLK) protein sequences were retrieved from 12 species in the NCBI protein database, including E. granulosus and E. multilocularis. Multiple sequence alignment was performed using the Clustal Omega program, and structural visualization and homology analysis were conducted using the ESPript 3.2 program. The recombinant plasmid pET-30a-EgPLK2 was transformed into BL21(DE3) competent cells. Protein expression was induced with isopropyl-β-D-thiogalactoside (IPTG), and sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) was performed to characterize the expression and molecular weight of the rEgPLK2 protein. The purified rEgPLK2 protein was thoroughly emulsified with Freund’s complete adjuvant at a 1 : 1 volume ratio. Two New Zealand white rabbits were immunized with multipoint subcutaneous injection on the back at a dose of 300 μg per rabbit for primary immunization. For booster immunizations, the protein was emulsified with Freund’s incomplete adjuvant at a 1 : 1 volume ratio and administered on days 14, 28, and 42 after the primary immunization at a dose of 150 μg per rabbit. Serum was sampled from the rabbit ear vein on day 7 after the final immunization to yield anti-rEgPLK2 polyclonal antibodies. Antibody titer was determined by indirect enzyme-linked immunosorbent assay (ELISA), and antibody specificity was verified by Western blotting. The tissue localization of the EgPLK2 protein was detected in E. granulosus protoscoleces and adult worms using immunofluorescence assay (IFA). Eighteen 6- to 8-week-old female SPF-grade BALB/c mice were randomly divided into three groups, including the blank control group, rEgPLK2-ISA immunization group, and PBS-ISA adjuvant control group, of 6 mice each group. Mice in the rEgPLK2-ISA immunization group and PBSISA group received three primary immunizations via intramuscular injection, and animals in the rEgPLK2-ISA immunization group was inoculated with immunogens prepared by emulsifying rEgPLK2 protein with ISA 201 adjuvant at a 1 : 1 volume ratio (6 μg per mouse), while mice in the PBS-ISA adjuvant control group received an equal volume of PBS emulsified with ISA adjuvant at a 1 : 1 volume ratio. A fourth booster immunization was administered via intraperitoneal injection. Mice in the rEgPLK2-ISA immunization group received a booster immunization with 8 μg of rEgPLK2 protein per mouse, and animals in the PBS-ISA group received an equal volume of PBS, with immunizations given at 2-week intervals. Mice in the blank control group were given no treatment, and housed under standard conditions. Tail vein blood was collected from all mice 7 days after the final immunization, and levels of specific anti-rEgPLK2 IgG antibody and its subclasses (IgG1, IgG2a, IgG2b, IgG3) were measured by indirect ELISA. E. granulosus infection was modelled in mice through injection with 1 000 E. granulosus protoscoleces via intrahepatic portal vein in the rEgPLK2-ISA immunization group and PBS-ISA adjuvant control group 2 weeks after the last immunization. All mice were sacrificed and dissected. The number of cysts was counted in mouse livers, and the cyst reduction rate was calculated. Liver tissues were processed for paraffin sectioning and stained with hematoxylin and eosin (HE), and histopathological changes were examined under a light microscope. Results Sequence analysis revealed that EgPLK2 shared a high amino acid sequence homology with E. multilocularis PLK2 (EmPLK2) and contained the typical domains of the Polo-like kinase family, including the serine/threonine protein kinase catalytic domain (STKc) and Polo-box. The IPTG-induced rEgPLK2 protein was mainly expressed in the form of inclusion bodies, and the purified rEgPLK2 protein showed a relative molecular mass of approximately 70 kDa. The prepared rabbit anti-rEgPLK2 polyclonal antibody had a titer of 1 : 256 000, and Western blotting assay showed that this anti-body specifically recognized the rEgPLK2 protein with a relative molecular mass of approximately 70 kDa. Immunofluorescence assay showed that the EgPLK2 protein was localized in the excretory bladder and rostellum of E. granulosus protoscoleces, as well as the tegument, suckers, and inter-proglottid junctions of adult worms. Immunoprotective assay showed that the serum levels of specific anti-rEgPLK2 IgG, IgG1, IgG2a, and IgG2b antibodies were 2.92 ± 0.49, 0.33 ± 0.10, 0.31 (0.36), and 3.12 (1.73) in mice in the rEgPLK2-ISA immunization group, which were all significantly higher than those in the PBS-ISA adjuvant control group (0.14 ± 0.04, 0.07 ± 0.01, 0.12 ± 0.04, and 0.11 ± 0.04, respectively) (t = 19.28 and 8.46, Z = 3.75 and 4.15; all P values < 0.001); however, there was no significant difference in the serum anti-IgG3 antibody level between the rEgPLK2-ISA immunization group and the PBS-ISA adjuvant control group [0.07 (0.01) vs. 0.073 (0.07); Z = 0.69, P > 0.05)]. In the mouse model of E. granulosus infections, the area of hepatic lesions was reduced and the inflammatory infiltration was alleviated in the rEgPLK2-ISA immunization group than in the PBS-ISA adjuvant control group, and the number of hepatic cysts was higher in the PBS-ISA adjuvant control group than in the rEgPLK2-ISA immunization group [8.00 (2.00) vs. 1.00 (0.75); Z = −2.93, P < 0.01], with a cyst reduction rate of 80.40%. Indirect ELISA assay measured higher serum levels of specific anti-rEgPLK2 IgG (3.28 ± 0.48 vs. 0.11 ± 0.04; t = 15.86, P < 0.01), IgG1 (0.29 ± 0.02 vs. 0.09 ± 0.01; t = 15.67, P < 0.01), IgG2a [3.71 (1.09) vs. 0.08 (0.03); Z = 2.88, P < 0.01], and IgG2b antibodies [3.34 (1.01) vs. 0.08 (0.03); Z = 2.88, P < 0.01] in the rEgPLK2-ISA immunization group than in the PBS-ISA adjuvant control group, and there was no significant difference in the serum level of the specific anti-rEgPLK2 IgG3 antibody between the rEgPLK2-ISA immunization group and the PBS-ISA adjuvant control group (0.07 ± 0.01 vs. 0.07 ± 0.01; t = 1.29, P > 0.05). Conclusions The prokaryotic expression system has been successfully constructed for the EgPLK2 gene and the anti-rEgPLK2 polyclonal antibody has been obtained. The rEgPLK2 protein exhibits a high immunogenicity, and is effective to protect against E. granulosus infection, and inhibits cyst development, which is a promising candidate vaccine target against cystic echinococcosis.

Flow diverter (FD) can significantly improve the occlusion rate and reduce the recurrence rate of intracranial aneurysms, and has become an irreplaceable endovascular treatment option. Delayed aneurysm rupture (DAR) is one of the most serious complications after FD implantation, with a very high mortality rate. It is a major challenge to the safety of FD implantation. The mechanism of DAR is currently not fully understood and may be associated with the changes in hemodynamics after FD implantation, inflammation and mechanical stress within aneurysms, and changes in postoperative coagulation function. This article reviews the research progress on the occurrence mechanism of DAR and outlines future research directions, with the aim of reducing DAR occurrence and optimizing clinical decision-making.

Objectives:To evaluate the predictive value of Tabular Prior-data Fitted Network(TabPFN) for short-term outcome in patients with spontaneous intracerebral hemorrhage (sICH), and compared with the Extreme Gradient Boosting (XGboost) model and traditional logistic regression (LR) model. Methods:Patients with sICH admitted to the Department of Neurology, Hefei Second People's Hospital from January 2018 to March 2024 were included retrospectively. The demographic and baseline data were collected. At 3 months after onset, the modified Rankin Scale score was used to determine the outcome, 0-2 was defined as good outcome and >2 was defined as poor outcome. All enrolled patients were randomly divided into a training set and a testing set at a ratio of 7:3. Feature selection was performed using recursive feature elimination (RFE) method, and then the selected feature variables were included into TabPFN, XGboost, and LR models for training and testing. The area under the curve (AUC) of receiver operating characteristic (ROC) curve was used to evaluate the predictive ability of the models. Shapley additive explanations (SHAP) method was used for model interpretation.Results:A total of 547 patients with sICH were enrolled, including 367 males (67.1%), with a median age of 65 (interquartile range, 54-76) years. Two hundred twenty-six patients (41.3%) had poor outcome. Age, baseline blood pressure (systolic blood pressure, diastolic blood pressure), baseline laboratory tests (white blood cell count, red blood cell count, platelet count, neutrophil count, hemoglobin, fasting blood glucose, creatinine, uric acid, urea nitrogen, alanine aminotransferase, aspartate aminotransferase), hematoma rupture into the ventricle, island sign, baseline hematoma volume, and baseline National Institutes of Health Stroke Scale (NIHSS) score were selected as characteristic variables using RFE method. ROC curve analysis showed that the ROC AUC for TabPFN, Xgboost, and LR models predicting poor short-term outcome in the testing set were 0.918 (95% confidence interval [ CI] 0.870-0.966], 0.883 (95% CI 0.826-0.940), and 0.905 (95% CI 0.854-0.957), respectively. SHAP analysis showed that the top four important variables in the TabPFN model were baseline NIHSS score, baseline hematoma volume, baseline aspartate aminotransferase, and age. Conclusions:The TabPFN model is superior to the LR model and the XGBoost model in predicting poor outcome in patients with sICH. In the TabPFN model, baseline NIHSS score, baseline hematoma volume, aspartate aminotransferase, and age are the most important predictors of poor outcome in patients with sICH.Objectives To evaluate the predictive value of Tabular Prior-data Fitted Network(TabPFN) for short-term outcome in patients with spontaneous intracerebral hemorrhage (sICH), and compared with the Extreme Gradient Boosting (XGboost) model and traditional logistic regression (LR) model. Methods Patients with sICH admitted to the Department of Neurology, Hefei Second People's Hospital from January 2018 to March 2024 were included retrospectively. The demographic and baseline data were collected. At 3 months after onset, the modified Rankin Scale score was used to determine the outcome, 0-2 was defined as good outcome and >2 was defined as poor outcome. All enrolled patients were randomly divided into a training set and a testing set at a ratio of 7:3. Feature selection was performed using recursive feature elimination (RFE) method, and then the selected feature variables were included into TabPFN, XGboost, and LR models for training and testing. The area under the curve (AUC) of receiver operating characteristic (ROC) curve was used to evaluate the predictive ability of the models. Shapley additive explanations (SHAP) method was used for model interpretation. Results A total of 547 patients with sICH were enrolled, including 367 males (67.1%), with a median age of 65 (interquartile range, 54-76) years. Two hundred twenty-six patients (41.3%) had poor outcome. Age, baseline blood pressure (systolic blood pressure, diastolic blood pressure), baseline laboratory tests (white blood cell count, red blood cell count, platelet count, neutrophil count, hemoglobin, fasting blood glucose, creatinine, uric acid, urea nitrogen, alanine aminotransferase, aspartate aminotransferase), hematoma rupture into the ventricle, island sign, baseline hematoma volume, and baseline National Institutes of Health Stroke Scale (NIHSS) score were selected as characteristic variables using RFE method. ROC curve analysis showed that the ROC AUC for TabPFN, Xgboost, and LR models predicting poor short-term outcome in the testing set were 0.918 (95% confidence interval [ CI] 0.870-0.966], 0.883 (95% CI 0.826-0.940), and 0.905 (95% CI 0.854-0.957), respectively. SHAP analysis showed that the top four important variables in the TabPFN model were baseline NIHSS score, baseline hematoma volume, baseline aspartate aminotransferase, and age. Conclusions The TabPFN model is superior to the LR model and the XGBoost model in predicting poor outcome in patients with sICH. In the TabPFN model, baseline NIHSS score, baseline hematoma volume, aspartate aminotransferase, and age are the most important predictors of poor outcome in patients with sICH.