Identification of physiological clusters in acute hypoxemic respiratory failure patients undergoing non-invasive respiratory support using EIT-based t-SNE and spectral clustering

Purpose: Identifying physiological clusters in acute hypoxemic respiratory failure (AHRF) may help to personalize non-invasive respiratory support (NIRS). Electrical impedance tomography (EIT) provides real-time, regional information on tidal ventilation, but its value for clustering AHRF patients undergoing NIRS has not been established. Methods: We conducted a single-center observational study including adults with AHRF monitored with EIT during NIRS. Tidal ventilation images were pre-processed, normalized, and embedded into a 2-dimensional space using t-SNE. Spectral clustering was applied to identify distinct imaging patterns. Clinical, physiological and laboratory variables were compared across clusters. The association between cluster membership and intubation at 7 days was assessed using penalized Cox regression adjusted for age, BMI, PaCO₂ and ROX index. Results: Thirty-two patients were enrolled. Spectral clustering identified three distinct clusters of tidal images. Clusters differed in clinical severity and physiological profile: Cluster 1 was characterized by shorter stature and higher SAPS II; Cluster 2 showed the highest pendelluft; Cluster 3 exhibited symmetric ventilation with low pendelluft. These phenotypes also differed in hemodynamics, including heart rate and shock index. Cluster membership was independently associated with intubation at 7 days. Compared with Cluster 3, both Cluster 1 and Cluster 2 showed a significantly lower hazard of intubation (HR 0.115, p = 0.017 and 0.042, p = 0.002, respectively). Conclusions: Unsupervised clustering of EIT tidal images is feasible in AHRF and identifies distinct physiological clusters with different short-term outcomes. These findings support the potential role of EIT-based imaging patterns for early stratification of patients undergoing NIRS.