(B-197) A Fair Individualized Risk Score for Identifying Risks of Falls among Patients with Alzheimer Disease and Related Dementia

Background: Alzheimer's Disease and Related Dementias (ADRD) is an insidious issue associated with aging, which is also linked to higher risks of falls. Emerging research has increasingly recognized the role of social determinants of health (SDOH) in influencing ADRD outcomes, however, few studies have explored how SDOH contributed to falls.

Objectives: We aim to identify the key SDOH associated with falls in patients with ADRD.

Methods: With a retrospective cohort study design, we identified real-world patients with ADRD from the University of Florida Health Integrated Data Repository from Jan 1, 2012, to Jan 31, 2022. The index date was defined as the first date of diagnosis with mild cognitive impairment or ADRD. The outcome was falls within 6 months after the index date, defined as a composite of falls, syncope, fractures, and traumatic brain injury. Baseline covariates included patient demographics, clinical comorbidities, and individual- and contextual-level SDOH. The data were split into training, validation, and testing with a ratio of 7:1:2 and the hyperparameters were tuned with 5-fold cross-validation. To account for class imbalance, we used random over-sampling technique. This SDOH-based model, namely individualized polysocial risk score (iPsRS), was developed to predict 6-month falls using logistic regression with regularization and Extreme Gradient Boosting (XGBoost). The model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), F-1 score, precision, recall, and specificity; and the model fairness of the best-performing model was assessed through predictive parity. We then used SHapley Additive exPlanations (SHAP) framework to interpret the best-performing model and identify the key SDOH predictors.

Results: The cohort consists of 10,993 patients with 47.8% females, 69% non-Hispanic Whites, and a mean age of 72. 2,376 (21.6%) were identified with the outcome. The best predictive model was LR, achieving an AUROC of 0.698, F-1 score of 0.440, precision of 0.343, recall of 0.611, and specificity of 0.68, which also achieved predictive parity (similar positive and negative predictive values across race and sex groups) across race and sex groups. Emergency department visits in the baseline period was the strongest predictor for 6-month falls followed by inpatient visits in baseline period and patients’ age.

Conclusions: Our machine learning pipeline iPsRS can fairly and accurately screen for patients who have increased social risk leading to hospitalization among patients with ADRD. Emergency department and inpatient visits in the baseline period were identified as strong predictors for falls.