Sepsis Research Pushes Toward Predictive AI, Molecular Subtypes and Hydrogen Therapy

Sepsis remains one of the most formidable challenges in critical care, and a new editorial in Anesthesiology and Perioperative Science argues that the field is now advancing on several fronts at once, from interpretable artificial intelligence and molecular subtyping to emerging drug targets and hydrogen-based therapy. The editorial, published on 28 September 2026, defines sepsis as life-threatening organ dysfunction caused by a dysregulated host response to infection under the Sepsis-3 criteria, and frames its pathophysiology as a biphasic immune course: an initial hyperinflammatory “cytokine storm” phase followed by a state of immunosuppression that can culminate in persistent inflammation, immunosuppression and catabolism syndrome, known as PICS. Immunothrombosis and microcirculatory dysfunction are described as key drivers of multi-organ failure.

The burden behind those mechanisms is stark. Acute lung injury is reported as the most frequent complication of sepsis, occurring in 68.2 percent of septic patients and carrying a 90-day mortality of 35.5 percent. Globally, sepsis accounts for an estimated 48.9 million cases and 11.0 million deaths each year according to the global burden of disease study, and it remains a leading cause of death worldwide and a large proportion of intensive care unit admissions. Against that backdrop, the editorial reviews recent articles in the journal and sets out an integrated, human-centred agenda for critical care that stretches from before ICU admission to long after discharge.

The first pillar is predictive analytics. Early risk stratification is essential, yet conventional tools such as the Sequential Organ Failure Assessment score depend on subjective components. To address this, Huang and colleagues built a machine-learning model using only seven routine clinical features, including platelets, leukocytes, blood urea nitrogen, anion gap and ICU length of stay, with the XGBoost algorithm. Trained on the MIMIC-IV database and validated on the eICU Collaborative Research Database, the model achieved area under the curve values of 0.920, 0.900 and 0.839. Shapley Additive Explanations analysis revealed the dynamic contribution of each feature before and after antibiotic administration, providing an interpretable, treatment-responsive risk assessment tool.

The editorial argues that such models must prioritise clinical practicality and interpretability over technical sophistication. It points to broader reviews of artificial intelligence in perioperative and critical care that highlight the value of robust multi-centre datasets and standardised data formats, and to reviews of AI in anaesthesiology underscoring the potential of decision support systems to reduce cognitive load and improve clinical judgement. Beyond static prediction, the authors call for early warning systems that continuously synthesize bedside monitoring data, laboratory trends and clinical notes, capable of detecting subtle deterioration hours before overt shock or organ failure. The stated goal is human-centred AI that eases clinicians’ workload and augments, rather than replaces, clinical expertise, especially when linked to automated warning and response protocols.

The second pillar moves from prediction to precision medicine. A letter by Zou and colleagues draws attention to plasmacytoid dendritic cells as possible contributors to sepsis-associated acute lung injury. These cells are best known for producing antiviral interferon-alpha, but emerging evidence implicates them in non-viral inflammation as well. Their precise role in sepsis-associated lung injury remains unclear; aberrant function may drive inflammatory mediators that disrupt microcirculation, alter metabolic reprogramming and promote alveolar epithelial apoptosis. If confirmed, targeting these cells could open new therapeutic avenues. Separately, the capsaicin receptor TRPV1 has been shown to protect against sepsis-associated brain injury in mice by inhibiting pyroptosis: TRPV1 deficiency worsened sepsis-associated encephalopathy symptoms through activation of the NF-kappa-B and pyroptosis pathways, whereas capsaicin pretreatment improved outcomes in a TRPV1-dependent manner.

Immunometabolism has yielded further regulatory nodes. Phosphoglycerate dehydrogenase, or PHGDH, has been identified as a negative regulator of TLR4-mediated NF-kappa-B and MAPK signalling through direct binding and inhibition of TAK1 via a non-metabolic function. PHGDH is down-regulated in septic patients; myeloid-specific deletion exacerbates organ damage, while overexpression alleviates symptoms. In addition, high expression of interferon regulatory factor 4 driven by N6-methyladenosine modification has been shown to promote immunosuppression in late-phase sepsis, pointing to another intervention point. The editorial also notes recent precision medicine frameworks, including consensus immune dysregulation signatures and blood transcriptomic subtypes labelled CTS1, associated with inflammation and endothelial activation; CTS2, with heme and platelet function; and CTS3, with interferon response and anticoagulant features, which the authors describe as a molecular foundation for tailoring therapy.

For precision sepsis care to work in practice, the editorial sets out four requirements. Molecular subtyping should integrate multi-omics data beyond transcriptomics and distinguish early hyperinflammatory from late immunosuppressive phenotypes. Immune status, organ function and treatment response should be monitored dynamically, because these features change during sepsis. Interventions should be matched to the identified subtype. And AI should assist clinical judgement rather than replace it. The authors acknowledge the translational limits of these frameworks: subtyping must move beyond transcriptomics to proteomic, metabolomic and single-cell technologies, monitoring must be dynamic rather than a single static classification, and the synergy between AI and clinicians must be carefully orchestrated. The utility of transcriptomic subtyping in guiding real-time, prospective therapeutic decisions has yet to be demonstrated.

The third pillar is therapeutics. Molecular hydrogen has emerged as a promising agent, with demonstrated protective effects across multiple organ systems including the lungs, brain, kidneys, heart and intestines in preclinical work. Mechanistically, hydrogen has been proposed to act as a selective antioxidant, scavenging harmful hydroxyl radicals while potentially preserving beneficial signalling reactive oxygen species. It inhibits NF-kappa-B signalling and the NLRP3 inflammasome, improves mitochondrial dynamics and biogenesis via PGC-1alpha and NRF2 activation, and promotes mitophagy through SIRT1 and Pink1/Parkin-dependent pathways that also suppress the cGAS-STING-IRF3 axis. The pleiotropic molecular networks of hydrogen have been mapped systemically using iTRAQ-based proteomic analyses, and these preclinical findings have prompted initial clinical exploration of hydrogen-oxygen mixtures in various settings. However, the editorial stresses that in vivo kinetics and the relative contributions of radical scavenging versus signalling modulation remain areas of active investigation and debate, and that widespread clinical adoption will require robust, large-scale randomised controlled trials with hard clinical endpoints.

Beyond hydrogen, emerging pharmacotherapies such as centhaquine and agents modulating parasympathetic tone illustrate ongoing refinement of haemodynamic and anti-inflammatory strategies. But the heterogeneity of sepsis means no single agent is likely to be universally effective; future progress will require combining subtype-specific targeted therapies with optimised supportive care, including careful oxygen management and judicious sedation, tailored to the patient’s evolving pathophysiology. The editorial also proposes viewing sepsis care as a continuum it calls peri-critical care medicine, analogous to perioperative medicine: pre-ICU recognition and early warning, in-ICU organ support and precision therapy addressing pulmonary, neurological, renal, hepatic and intestinal injury, and post-ICU rehabilitation and survivorship, requiring coordination among emergency services, intensivists, nurses, rehabilitation specialists, psychologists and primary care providers.

Within that continuum, the authors advocate accelerated sepsis recovery: early mobilisation and physical therapy to counter ICU-acquired weakness, tailored nutritional support with immunomodulatory nutrients guided by metabolic monitoring, psychological and cognitive interventions for post-traumatic stress, anxiety and cognitive dysfunction, family engagement and shared decision-making, and structured follow-up including tele-rehabilitation. Many sepsis survivors face new or worsening disabilities, recurrent infections and reduced quality of life, and post-discharge readmission rates remain alarmingly high, although recent randomised trials have shown that even comprehensive post-sepsis transition programs may not improve composite outcomes. Among the priorities for the next decade, the editorial lists multi-omics subtyping, dynamic real-time predictive models with built-in interpretability, validation of novel targets including non-metabolic functions of metabolic enzymes, large-scale multicentre trials of hydrogen therapy to define optimal dosing, timing and patient selection, refined animal models, rigorous evaluation of traditional medicine formulations, strengthened international data sharing, and prioritisation of long-term survivorship. The authors are candid about the field’s limitations: AI models rest mostly on retrospective datasets with unproven external validity and calibration, hydrogen therapy lacks large-scale randomised trials with hard endpoints, and subtyping frameworks await prospective clinical deployment. Acknowledging these gaps, they conclude, is a necessary step toward designing rigorous studies that can truly advance patient-centred sepsis care.

Subject of Research: Sepsis research and critical care advances, from predictive analytics to novel therapeutics

Article Title: Advancing sepsis research and critical care: from predictive analytics to novel therapeutics

Article References: Cui, Y., & Xie, K. (2026). Advancing sepsis research and critical care: from predictive analytics to novel therapeutics. Anesthesiology and Perioperative Science, 4(1), Article 50. https://doi.org/10.1007/s44254-026-00197-8

Image Credits: AI Generated

DOI: 10.1007/s44254-026-00197-8

Keywords: sepsis, critical care, artificial intelligence, machine learning, molecular subtyping, molecular hydrogen, immunometabolism, acute lung injury, peri-critical care, sepsis survivorship, randomised controlled trials, Advancing