Identification of Cognitive Subtypes and Risk Networks in Chinese Older Adults: Based on Latent Profile Analysis and Explainable Machine Learning

Background Traditional cognitive screening often relies on total scores, which can easily mask asynchronous impairment patterns across different cognitive domains. In particular, calculation ability, a sensitive indicator of executive function, is frequently misattributed to normal aging and overlooked clinically. This study aims to utilize data-driven methods to identify specific concealed cognitive phenotypes in the older population and deconstruct their underlying sociodemographic and clinical driving mechanisms. Methods This study included 7,159 older participants from the China Health and Retirement Longitudinal Study (CHARLS). Latent Profile Analysis (LPA) was employed to identify latent cognitive subtypes based on three dimensions: orientation, calculation, and memory. Prediction models were constructed using Boruta feature selection and the Random Forest algorithm. Furthermore, SHAP (SHapley Additive exPlanations) and Gaussian Graphical Models (GGM) were utilized to quantify the nonlinear contributions and network interactions of risk factors. Results LPA identified a unique cognitive impairment subtype termed Calculation-Dominant Impairment, accounting for 24.6% of the sample population. Although the majority of respondents were classified as cognitively intact (55.0%), this specific quarter of the impaired population exhibited a disproportionately severe decline in calculation function (V-shaped deficit), while episodic memory was relatively preserved. The Random Forest model demonstrated superior performance in identifying this subtype (AUC = 0.781), significantly outperforming traditional linear models, and achieved high specificity (specificity = 0.960). Interpretability analysis revealed a pattern dominated by significant social determinants: education level and internet use were the primary predictors, with importance weights significantly exceeding those of somatic comorbidities such as hypertension and diabetes. Network analysis further confirmed that internet use acts as a critical "bridge node" connecting age and education level, exerting a potential cognitive reserve effect in mitigating calculation decline. Conclusion This study defines a cognitive impairment subtype primarily driven by education poverty and digital exclusion, which is prone to be missed in routine memory-centric screenings. The findings suggest that biomedical interventions alone may be insufficient to prevent such decline; promoting digital inclusion among older adults should be regarded as a critical public health strategy to compensate for early-life educational deficits and prevent the decline of calculation and executive functions.