Aging
Survival Disparities among Alzheimer’s Patients in Hawaii: Exploring Racial/Ethnic Disparities and Machine Learning for Survival Prediction in the Hawaii Medicare Population Chathura Siriwardhana* Chathura Siriwardhana
Background: The survival outcomes following an Alzheimer’s disease (AD) diagnosis hold significant importance for health management, caregivers, patients, and their families. Hawaii is known as the most diverse ethnic population in the US and there exist racial health disparities. This study investigates racial/ethnic disparities in survival among AD patients in Hawaii and develops Machine Learning models for survival prediction, utilizing Hawaii Medicare data.
Methods: We analyzed nine years of Hawaii Medicare data to collect information on individuals developing AD after the age of 65, tracking them until all-cause survival or censoring. We assessed the impact of race/ethnicity, in conjunction with socioeconomic status (SE), on the risk of mortality. SE status was determined using the surrogate marker: Medicare/Medicaid dual eligibility. Cox regression analysis was performed on overall survival, accounting for age at AD onset, gender, and various comorbidities. Subsequently, a Survival Random Forest was employed to predict survival within a machine learning framework, incorporating K years of longitudinal health profiles, including demographics, chronic disease profiles, observed acute conditions, and hospitalization history.
Results: The study included n = 9,393 AD subjects. Our analysis revealed that American Asians (AA) had a later age at AD diagnosis (p<.001), with an average age of 85.9, compared to 82.7 and 83.3 years for Whites (WH) and Native Hawaiians and Pacific Islanders (NHPI), respectively. SE exhibited a marginal increase in hazard (Hazard Ratio [HR]=1.36, p<.001). In comparison to AA with high SE (AA + high SE), increased hazards were found for AA + low SE (1.29, p<.001), WH + high SE (1.19, p<.001), WH + low SE (1.52, p<.001), and NHPI + low SE (1.39, p<.001). The Random Forest model with K=2 setting demonstrated a Concordance (C) Index of 0.806 via five-fold cross-validation, exhibiting robust survival predictability of AD subjects. A permutation-based study identified factors influencing subject survival.
Conclusion: The onset of AD development and survival are influenced by race/ethnicity and SE status. Machine learning, when combined with longitudinal health data, demonstrates reasonable predictability of survival.