Sepsis is a leading cause of death, and improved approaches for disease diagnosis and detection of etiologic pathogens are urgently needed. Here, we carried out integrated host and pathogen metagenomic next generation sequencing (mNGS) of whole blood (n=221) and plasma RNA and DNA (n=138) from critically ill patients following hospital admission. We assigned patients into sepsis groups based on clinical and microbiological criteria: 1) sepsis with bloodstream infection (SepsisBSI), 2) sepsis with peripheral site infection but not bloodstream infection (Sepsisnon-BSI), 3) suspected sepsis with negative clinical microbiological testing; 4) no evidence of infection (No-Sepsis), and 5) indeterminant sepsis status. From whole blood gene expression data, we first trained a bagged support vector machine (bSVM) classifier to distinguish SepsisBSI and Sepsisnon-BSI patients from No-Sepsis patients, using 75% of the cohort. This classifier performed with an area under the receiver operating characteristic curve (AUC) of 0.81 in the training set (75% of cohort) and an AUC of 0.82 in a held-out validation set (25% of cohort). Surprisingly, we found that plasma RNA also yielded a biologically relevant transcriptional signature of sepsis which included several genes previously reported as sepsis biomarkers (e.g., HLA-DRA, CD-177). A bSVM classifier for sepsis diagnosis trained on RNA gene expression data performed with an AUC of 0.97 in the training set and an AUC of 0.77 in a held-out validation set. We subsequently assessed the pathogen-detection performance of DNA and RNA mNGS by comparing against a practical reference standard of clinical bacterial culture and respiratory viral PCR. We found that sensitivity varied based on site of infection and pathogen, with an overall sensitivity of 83%, and a per-pathogen sensitivity of 100% for several key sepsis pathogens including S. aureus, E. coli, K. pneumoniae and P. aeruginosa. Pathogenic bacteria were also identified in 10/37 (27%) of patients in the No-Sepsis group. To improve detection of sepsis due to viral infections, we developed a secondary RNA host transcriptomic classifier which performed with an AUC of 0.94 in the training set and an AUC of 0.96 in the validation set. Finally, we combined host and microbial features to develop a proof-of-concept integrated sepsis diagnostic model that identified 72/73 (99%) of microbiologically confirmed sepsis cases, and predicted sepsis in 14/19 (74%) of suspected, and 8/9 (89%) of indeterminate sepsis cases. In summary, our findings suggest that integrating host transcriptional profiling and broad-range metagenomic pathogen detection from nucleic acid may hold promise as a tool for sepsis diagnosis.