Introduction: Acute respiratory failure has been traditionally classified based on pathophysiologic parameters. We aimed to examine clinical notes and apply unsupervised machine learning methods for a more pragmatic approach to identify phenotypes.Methods: A cohort of 46,520 adult patients admitted to an intensive care unit at a tertiary academic center was examined. Inclusion criteria were a discharge diagnosis for acute respiratory failure and oxygen required upon ICU admission. Natural language processing tools were used to examine the spans of Unified Medical Language System (UMLS) for named entity mentions with mapping to a concept unique identifier (CUI). All CUIs were normalized using term frequency-inverse document frequency (TF-IDF), and the elbow method was used to estimate the number of clusters. A K-means algorithm was used to assign each patient encounter to a cluster. A multinomial logistic regression model included an elastic net penalty to predict cluster membership from CUIs. The model achieved an accuracy of 91% in 10-fold cross-validation. The top 50 CUIs from each cluster as ranked by beta coefficient were extracted for review to identify themes and describe clusters.Results: A total of 5,166 patients met inclusion criteria and 19,402 CUIs were extracted. Six clusters were identified and 2,987 CUI features remained after TF-IDF normalization. The greatest proportion of patients was in Cluster 6 (n = 1711, 33%) which comprised of patients from outside hospital transfers with prolonged respiratory failure with CUIs for tracheostomy and parenteral nutrition as top features. These patients had the shortest 28-day free ICU days (6.7 days, p<0.01) and highest mortality rate (77.2%, p<0.01). Cluster 1 (15%) had CUIs with a theme for malignancies and complications from metastases. Cluster 2 (16%) had CUIs with a theme for medical patients with sepsis. Cluster 3 (9%) had CUIs with a theme for obstructive airway disease. Cluster 4 (10%) had CUIs with a theme for neurologic disease and head trauma. Cluster 5 (17%) had CUIs with a theme for cardiac disease requiring both medical and surgical intervention. Table 1 represents patient characteristics for each cluster.Conclusion: Six distinct phenotypes were identified from the notes in patients with acute respiratory failure. Most phenotypes included specific organ injuries or diseases; however, the largest cluster was non-specific and comprised of outside hospital transfers with prolonged respiratory failure, longest lengths of stay, and the highest mortality. These phenotypes may represent an opportunity for better resource allocation after external validation of these findings.