(1470) Clustering Analysis as a Method of Patient Phenotyping in Critical Care

Introduction: It is crucial to understand disease patterns, prognosis, severity, and treatment to improve outcomes in heterogenous diseases, as it can aid in identifying and treating subtypes in other critical disease areas, leading to more tailored care. The CURE Drug Repurposing Collaboratory developed a collection of open source projects into an OHDSI stack to harmonize EHR data and facilitate observational research across multiple healthcare systems [1], and this study builds on the previous analysis of a single institution [2] in evaluating the data's utility by examining treatment patterns across a broad spectrum of clinical characteristics in hospitalized COVID-19 patients.

Methods: This study analyzed 124,684 hospitalized acute COVID-19 patients from eight US healthcare institutions between March 2020 to March 2024. Inclusion criteria were inpatients admitted for COVID-19, at least 18 years old, with complete demographic information. Laboratory measures included markers of liver injury, kidney function, blood measures, heart rate, and respiratory rate with greater than 90% data completeness. Using LASSO regression for feature selection, Factorial Analysis of Mixed Data (FAMD), and hierarchical clustering, distinct COVID-19 subtypes were identified with prognostic differences examined through hospitalization length, mortality, and Kaplan-Meier curves. Classification models (SVM, Random Forest, and XGBoost) were then constructed to classify these COVID-19 subtypes, and their performance was evaluated using precision, recall, and F1-score.

Results: Using LASSO regression to reduce input features before FAMD and hierarchical clustering, patients were divided into three clusters: Cluster 1 (mild disease, oldest patients), Cluster 2 (moderate severity, high prevalence of multiple treatments), and Cluster 3 (severe disease, high prevalence of ventilated patients). The XGBoost model achieved the highest performance with an F1-score of 0.88.

Conclusions: Understanding disease patterns, prognosis, and severity is crucial for improving patient outcomes. This study looked to evaluate the utility of the OHDSI stack data through discovering COVID-19 clusters with distinct characteristics, prognoses, and treatment patterns, and developed classification models to enhance clinical practice in critical care settings.