Longitudinal Impact of NLP-Augmented Clinical Decision Support on Cognitive Decline Detection in German Geriatric Primary Care: A Dynamic Panel Data Analysis Using System GMM Estimation

Background: Cognitive decline and dementia represent major public health challenges in aging populations. Natural language processing (NLP)-augmented clinical decision support systems (CDSS) offer promising tools for early detection, yet causal evidence on their longitudinal impact at the health system level remains sparse. This study examines whether the phased adoption of NLP-augmented CDSS across German geriatric primary care practices causally improved cognitive decline detection rates over a six-year period. Methods: We employed a System Generalized Method of Moments (System GMM) estimator, specifically the Arellano-Bover/Blundell-Bond approach, to analyze a dynamic panel dataset comprising 847 geriatric primary care practices across 14 German federal states from 2019 to 2024. The dependent variable was the annual cognitive decline detection rate per 1,000 registered patients aged 65 and older. The key exposure was the intensity of NLP-CDSS utilization, measured as the proportion of clinical encounters with NLP-assisted documentation review. We controlled for time-varying confounders including practice size, physician-to-patient ratio, urbanization, regional deprivation indices, and state-level health expenditure. Instrument validity was assessed using Hansens J-test for overidentifying restrictions and the Arellano-Bond test for serial correlation. Results: System GMM estimation revealed that a one-standard-deviation increase in NLP-CDSS utilization intensity was associated with a 4.73-point increase in the cognitive decline detection rate per 1,000 patients (95% CI: 2.41 to 7.05, p < 0.001), controlling for persistence in detection rates ( = 0.614, SE = 0.047). The Hansen J-statistic (p = 0.372) confirmed instrument validity, and the AR(2) test (p = 0.489) indicated no second-order serial correlation. Heterogeneity analyses revealed stronger effects in rural practices ({beta} = 6.12, p < 0.001) compared to urban practices ({beta} = 3.28, p = 0.004). The lagged dependent variable coefficient indicated substantial persistence, suggesting that practices with historically higher detection rates continued to outperform, independent of NLP-CDSS adoption. Conclusions: NLP-augmented clinical decision support significantly and causally improved cognitive decline detection rates in German geriatric primary care, with particularly pronounced effects in rural and underserved settings. Dynamic panel data methods accounting for endogeneity and state dependence provide stronger causal evidence than cross-sectional analyses. These findings support the scaled implementation of NLP-based screening tools as a cost-effective strategy for addressing the growing dementia burden in aging populations.