Construction and Validation of an Artificial Intelligence-Assisted Diagnostic Model for Glioma Based on Laboratory Indicators: A Single-Center Retrospective Cohort Study

Background The diagnosis of brain glioma relies on tissue biopsy and imaging examinations, which have invasiveness, sampling errors, and limitations in MRI differentiation. Biomarker research is mostly limited to single-index analysis, while artificial intelligence (AI) shows significant advantages in multi-dimensional data modeling. This study aims to construct an AI-assisted diagnostic model based on routine laboratory indicators to achieve non-invasive and accurate diagnosis and promote clinical transformation. Methods A retrospective analysis was performed on 71 laboratory indicators of 502 intracranial lesion patients (251 glioma cases and 251 control cases) from January 2006 to January 2024. Logistic regression, Softmax, and three-layer multi-layer perceptron (MLP) neural network were used for modeling, with model optimization through Min-Max normalization and SHAP value analysis. Results The MLP model showed the best performance, with a test set accuracy of 0.88, AUC of 0.933, sensitivity of 0.89, and specificity of 0.86. Key indicators were white blood cell count (SHAP 0.18), total bilirubin (0.15), triglycerides (0.13), and urine specific gravity (0.12), which were associated with tumor inflammation, liver metabolic reprogramming, lipid metabolism abnormalities, and water-electrolyte metabolism disorder, respectively. The model reduced the missed diagnosis rate from 23.7% to 5.8% in primary care hospitals and shortened the emergency diagnosis time to 2.5 hours. Conclusion This study first constructs a diagnostic model by integrating multi-dimensional laboratory indicators through AI, providing a new path for non-invasive screening of glioma. Multi-center studies are needed to verify its generalizability.