SMART-SEPSIS: Sequential MAchine learning for individualized Response to early Treatment in lung SEPSIS

TITLE OF PROPOSED STUDY: SMART-SEPSIS: Sequential MAchine learning for individualized Response to early Treatment in lung SEPSIS Title Sequential MAchine learning for individualized Response to early Treatment in lung SEPSIS Abstract Lung infection is the most frequent cause of sepsis and accounts for more than 40% of cases of acute respiratory failure. Among patients with pneumonia and sepsis, any delay in initiating the treatment increases the risk of respiratory failure and death. The hour-1 sepsis bundle recommends obtaining blood cultures and lactate measurement, initiating antibiotics, fluid resuscitation, and potentially vasopressors within the first hour. Hospital compliance with these guidelines remains moderate at 60%. The response to each element of the sepsis bundle is heterogeneous. While early treatment is desirable, indiscriminately exposing patients with suspected lung sepsis to antibiotics, fluids and/or vasopressors carries some risk. At the bedside, clinicians’ acumen, albeit augmented by early warning systems (EWSs), and machine-learned (ML) sepsis prediction algorithms is not sufficient to accurately identify lung sepsis and discriminate which patient should receive the hour-1 bundle. Thus, there is an unmet need to precisely select lung sepsis patients who will benefit from immediate application of the hour-1 bundle to reduce their risk of organ dysfunction, including respiratory failure and/or death. Our central hypothesis is that electronic health record (EHR) data and artificial intelligence/machine learning (AI/ML) can be used to model the individual response to sepsis treatment and promote an early but reasoned application of the hour-1 bundle in patients with suspected community-onset lung sepsis (COLS). We will use existing EHR data from 3 integrated health systems (UCSF, UPenn and UPMC) to train (Aim 1a), externally validate (Aim 1b), prospectively and silently test (Aim 2a) a real-time, interpretable clinical decision support system (CDSS) to guide early treatment in patients with suspected COLS. Our CDSS will derive from the estimated individual treatment effect (ITE) of antibiotics, fluid resuscitation and vasopressors. To optimize the acceptability by the clinician and the feasibility of a future clinical trial, we will design an EHR-embedded user interface that will provide a treatment recommendation based on the estimated ITE with a degree of confidence. (Aim 2b) We will iteratively improve its design using structured feedback from the clinicians. This research will allow us to identify the patients who should receive each individual element of the hour-1 bundle. Based on the previously reported impact of early treatment on organ dysfunction, including respiratory failure and mortality in patients with sepsis, this research has the potential to decrease mortality, morbidity and reduce inappropriate interventions among patients with COLS. Project Number: 1R01HL178952-01 | Fiscal Year: 2025 | NIH Institute/Center: National Heart Lung and Blood Institute (NHLBI) | Principal Investigator: Romain Pirracchio (+2 co-PIs) | Institution: UNIVERSITY OF CALIFORNIA, SAN FRANCISCO, SAN FRANCISCO, CA | Award Amount: $750,561 | Activity Code: R01 | Study Section: Clinical Informatics and Digital Health Study Section[CIDH] View on NIH RePORTER: https://reporter.nih.gov/project-details/1R01HL17895201