Ten microliters of unstained whole blood, imaged in 12 minutes: a three-class model flags sepsis at 100% sensitivity and 85.3% specificity

In critically ill adults, separating sepsis from non-infectious critical illness and from healthy controls can be done from rapid imaging of unstained whole blood alone, without cultures, molecular testing, or a single biomarker. Existing sepsis diagnostics either rely on blood cultures and molecular pathogen detection that take hours to days, or on host markers such as C-reactive protein and procalcitonin with limited specificity, or on transcriptional and cellular assays that need specialized sample processing and centralized testing. In a prospective study, multi-physician adjudication under Sepsis-3 criteria classified 29 critically ill adults as sepsis (13), non-infectious critical illness (12), or indeterminate (4), with 22 healthy controls; indeterminate cases were excluded from training and performance analyses.

The model treats each patient as a bag of images and finds disease-associated morphology on its own, rather than being told which cell types or shape features to measure. Conventional practice requires investigators to predefine the informative cell types or morphologic features and then measure them. Remo-ID uses a ResNet-50 encoder and gated attention pooling to aggregate image embeddings into a bag-level representation and then predict a patient-level class; evaluation used leave-one-patient-out cross-validation with a fixed training epoch across folds, averaging technical replicates per patient.

Tuning the decision threshold toward ruling sepsis out keeps sensitivity high while accepting a defined false-positive fraction, matching the critical-care need to exclude infection first. At Youden's J, sepsis versus critically ill controls gave 76.9% sensitivity and 100.0% specificity; tuned for maximum sensitivity it gave 100.0% sensitivity and 75.0% specificity. For sepsis versus all non-sepsis, 100.0% sensitivity came at 85.3% specificity, with 3 of 12 ICU controls and 2 of 22 healthy participants misclassified; 43 of 47 participants were classified correctly, and errors concentrated between sepsis and ICU controls.

What is demonstrated is three-class separation of sepsis, non-infectious critical illness, and health in critically ill adults from a single-center prospective cohort, using 10 microliters of unstained whole blood and about 12 minutes of acquisition yielding 20,000 images; the intended readers are researchers and engineers working on rapid critical-care infection diagnostics and point-of-care testing. The source proposes independent external validation and follow-up work linking model outputs to blinded hematopathology review, transcriptional profiling, and clinical laboratory values.

The source lists a limited sample size, a single-site design, lack of an independent external validation cohort, a young healthy control group, and limited representation of viral sepsis. Cross-validation prevented data leakage during training, but the source notes it may not have eliminated all instrument effects, sampling errors, and outlier vulnerability; cells experience moderate shear and planar confinement on the instrument, so observed morphologic features may not be directly comparable to other methods, and how cellular rigidity influences shear distortion rates is not yet understood.