Causal Effects of Natural Language Processing-Enhanced Clinical Decision Support on Early Cognitive Impairment Detection: A Propensity Score Analysis Using Inverse Probability of Treatment Weighting

Background: Natural language processing (NLP) systems integrated into clinical workflows show promise for detecting early cognitive impairment, yet causal evidence from real-world implementation remains limited. Observational studies comparing outcomes between hospitals with and without NLP-enhanced clinical decision support (CDS) systems face significant confounding from systematic differences in patient populations and institutional characteristics. Objective: To estimate the causal effect of NLP-enhanced CDS implementation on time-to-diagnosis of mild cognitive impairment (MCI) and early dementia using propensity score analysis with inverse probability of treatment weighting (IPTW). Methods: We conducted a retrospective cohort study across 12 hospitals in Singapore (n=8,247 patients aged [≥]60 years presenting with memory complaints, 2022-2025). Six hospitals implemented NLP-based cognitive screening systems analyzing clinical notes; six used standard care protocols. Propensity scores were estimated using gradient-boosted models incorporating 42 patient-level and 18 institutional covariates. IPTW-adjusted Cox proportional hazards models estimated treatment effects on time-to-diagnosis. Sensitivity analyses included trimming, augmented IPTW (AIPTW), and E-value calculations. Results: After IPTW adjustment, the standardized mean differences for all covariates were <0.10, indicating adequate balance. NLP-enhanced CDS was associated with significantly earlier cognitive impairment diagnosis (IPTW-adjusted HR=1.58; 95% CI: 1.41-1.77; p<0.001), corresponding to a median reduction of 4.2 months in diagnostic delay. The average treatment effect on the treated (ATT) was 3.8 months earlier diagnosis (95% CI: 2.9-4.7 months). Results were robust across sensitivity analyses with E-values of 2.47 for point estimate and 1.98 for confidence interval bound. Conclusions: Using rigorous causal inference methods, we demonstrate that NLP-enhanced clinical decision support systems significantly accelerate cognitive impairment detection in routine clinical practice. These findings support broader implementation of AI-enhanced screening tools to facilitate earlier therapeutic intervention.