Abstract Body: Background: The post-acute journey for stroke patients is highly variable. We evaluated patient and health-system factors that contribute to the variability in 1-year home time (HT) by employing unsupervised learning (cluster analyses) among Acute Ischemic Stroke (AIS) patients.

Methods: Texas statewide medical claims data on post-stroke care transitions were extracted from a CMS Qualified Entity housing data for ≥80% of the population (including 100% of Medicare Fee-for-Service). Clinical information on a matched sub-population was linked from the EMR of a 7-hospital certified stroke health system. K-means clustering was performed on 1) the statewide dataset on post-discharge care utilization, and 2) the hospital dataset which further included clinical and acute care features. The optimal number (k) of cluster centers was compared algorithmically using direct (average silhouette) and gap statistic methods. Cluster performance, with respect to distinguishing high risk patients more likely to have low HT post-discharge (i.e., total 1-year home time of <30 days), was evaluated by precision (positive predictive value), recall (sensitivity), and F1 score.

Results: Between 2016 and 2020, 75,478 statewide AIS admissions were captured, while 2,889 AIS admissions were analyzed for the matched hospital system subgroup. Cluster models with k = 2 centers (vs. 3) provided distinctive risk profiles (Fig. 1A / 1B) and yielded better performance in classifying patients at higher risk of low HT (patients depicted in red cluster in Fig. 1C / 1D). The 2-cluster model demonstrated a high recall and correctly identified 93.1% of all high risk (low HT) patients (Fig. 1E). Precision was modest, with 57.2% of those patients flagged as high risk representing true predictions. Demographic and clinical patient characteristic differences were observed across HT strata, in both the statewide dataset and the hospital sub-group. These included advanced age (both cohorts), higher stroke severity (hospital cohort), and higher comorbidity burden (hospital cohort). Post acute care at an inpatient rehabilitation facility minimized the risk of low HT (vs. high HT: ≥ 300 days) in both the statewide (16.2% vs. 28.3%) and the hospital cohort (15.1% vs. 28.5%) (Fig. 2).

Conclusion: Cluster analyses can elucidate high-risk patient profiles and guide in identifying intervenable targets and prioritizing provision of appropriate care, including rehabilitation, to optimize long-term patient outcomes.