Exploratory Construction of a Random Forest Prediction Model for Mild Cognitive Impairment Through Combined Detection of Multiple Blood Biomarkers and Machine Learning

Objective: This study aims to screen for indicators that significantly differ between Mild Cognitive Impairment (MCI) and Control Group (HC) through combined detection of multiple blood biomarkers, and to explore and construct a Random Forest model using these indicators as feature parameters to attempt to predict the occurrence of MCI. Methods: This study involved 83 elderly participants. All participants met the inclusion criteria and signed informed consent. Blood samples were collected from the subjects via fasting venipuncture between 7:00 and 9:00 am, then immediately centrifuged for analysis or stored at -80°C. Subsequently, cognitive status was assessed using neuropsychological scales, and blood biomarkers were analyzed. Information such as age, gender, height, and weight of the subjects was recorded. Results: A comparison of basic subject information and blood biomarker differences between the MCI and HC groups revealed significant differences in age (P=0.027) and white blood cell count (WBC) (P=0.017). Therefore, Propensity Score Matching (PSM) was used to eliminate age differences, leaving 56 subjects. The results showed significant differences in TAT (P=0.017), TG (P=0.035), WBC (P=0.003), and P-Tau181 (P=0.042). Based on the post-PSM differential data, TAT, TG, WBC, and Tau181 were used as feature parameters to construct a Random Forest model for predicting MCI. The model demonstrated excellent performance in 10-fold cross-validation, achieving an accuracy of 87.5%, sensitivity of 85.7%, and specificity of 89.3%. Additionally, the model’s Area Under the Curve-Receiver Operating Characteristic (AUC-ROC) value was 0.92. Conclusion: The Random Forest model constructed through blood multi-biomarker detection can effectively predict the occurrence of Mild Cognitive Impairment (MCI), indicating that the combination of blood biomarkers and machine learning methods has significant potential in the early screening of MCI, providing theoretical and practical support for the development of non-invasive and efficient MCI prediction tools in the future.