Quantifying clinical indicators for identifying and prognosticating aspiration pneumonia in Chinese geriatric patients: a retrospective cohort study integrating multivariable regression and interpretable machine learning

Objective: Aspiration pneumonia (AP) substantially increases mortality risk among elderly patients, necessitating prompt diagnostic recognition and prognostic stratification. However, standardized diagnostic criteria and prognostic markers remain elusive. This investigation seeks to quantify high-contribution clinical parameters for the diagnosis and evaluation of AP prognosis, elucidate nonlinear risk interactions, and establish a comprehensive risk stratification framework tailored for Chinese geriatric populations. Methods: This retrospective cohort analysis (2017-2018) enrolled 295 patients with pneumonia (aged 65-98 years) from Guangzhou First People's Hospital. The exposure variable was the diagnosis of AP, with the primary endpoints being 3-month mortality and recurrence rates. The adjustment variables encompassed demographic characteristics, comorbidities, biomarkers, and functional status. A dual-modeling approach was implemented: (1) multivariable multinomial logistic regression; (2) interpretable random forest algorithms incorporating SHAP/PDP analyses. We identified AP as an independent predictor using multinomial logistic regression of mortality (OR = 4.082, 95%CI: 1.041-15.997) and recurrence (OR = 3.329, 95%CI: 1.370-8.090). Regarding AP prognosis, consciousness impairment, muscle strength and swallowing function emerged as pivotal assessment indicators. Machine learning insights (AUC = 0.92) identified critical quantitative thresholds associated with AP classification, including in-hospital antibiotic duration >15 days and Barthel Index ≤2, along with nonlinear interaction effects between antibiotic duration and Barthel score; antibiotic duration was interpreted as a clinical-course marker rather than a causal factor. Additionally, prognostic quantitative markers for AP were established: albumin <30 g/L, with interactive effects between albumin levels and nutritional risk. Conclusion: This dual-framework methodology amalgamates epidemiological rigor with artificial intelligence interpretability, allowing AP identification and prognostic evaluation through routine clinical parameters. This approach establishes a foundation for dynamic risk stratification in AP patients and contributes to improved clinical outcomes.