Identification of Sepsis Subphenotypes Based on Bi-directional Long Short-Term Memory Auto-encoder Using Real-Time Laboratory Data Collected from Intensive Care Units

Sepsis is a heterogeneous syndrome characterized by a dysregulated immunological response to infection that results in organ dysfunction and often death. Identification of clinically relevant subphenotypes, which could potentially lead to personalized sepsis management. A clustering framework was proposed to identify sepsis subphenotypes using the bidirectional long short-term memory Auto-Encoder (BiLSTM-AE) together with the k-means algorithm. The subphnotypes were determined according to the proposed algorithm. After that, each subphnotype was evaluated according to their inpatient mortality rate. In addition, sensitivity was performed to evaluate the robust of the predictiveness of the model-learned representation and clinical significance. A total of 971 patients with 4905 records meeting the clustering criteria after data processing were included in the study. There were 748 patients (77.0%) with an overall in-hospital mortality rate of 2.9% in subphenotype 1, 120 (12.4%) with 10.0% mortality rate in subphenotype 2, and 103 (10.6%) with 78.6% mortality rate in subphenotype 3. Some laboratory biomarkers showed significant differences between different sepsis subphenotypes, such as AST (subphenotype 1: 61.0 vs. subphenotype 2: 75.5 vs. subphenotype 3: 298.6, units/L; P < 0.001) and ALT (subphenotype 1: 49.8 vs. subphenotype 2: 59.3 vs. subphenotype 3: 200.2, units/L; P < 0.001). The area under the receiver operating characteristic curve (AUROC) in this study was 0.89, and the area under precision recall curve (AUPRC) was 0.65. In this study, sepsis subphenotypes were identified using clustering framework based on BiLSTM-AE and k-means. This research may drive the personalization of sepsis management and provide new insights and guidance for subphenotype research.