Sepsis: Diagnosis, Phenotypes, and Treatment

• Sepsis is a life-threatening condition defined by organ dysfunction due to a dysregulated host response to infection, often measured by SOFA/qSOFA scores.
• It manifests as both hyperinflammatory and immunosuppressive states, with transcriptomic and computational methods enhancing early detection.
• Emerging diagnostic and AI-driven monitoring tools improve real-time recognition and personalized treatment, addressing delays that worsen outcomes.

Sepsis is a life-threatening condition characterized, in the Sepsis-3 consensus formulation, as “life-threatening organ dysfunction caused by a dysregulated host response to infection,” with organ dysfunction operationalized as an acute increase in SOFA score of at least 2 points (Gary et al., 2016). Across the supplied literature, sepsis appears simultaneously as a syndrome of rapidly evolving organ failure, a time-critical diagnostic problem, an immunopathologic state combining innate hyperactivation with adaptive suppression, and a major target for computational monitoring, phenotyping, and treatment optimization (Hu, 2013). The same corpus also shows that controversy over definitions and screening tools remains clinically consequential, because delays of even a few hours in recognition and treatment are repeatedly linked to worse outcomes and to the need for earlier, more reliable identification in emergency and intensive-care settings (Ivanov et al., 2022).

1. Definition, staging, and conceptual evolution

The modern definition of sepsis emerged through three major consensus stages. Sepsis-1, from the 1991 SCCM/ACCP conference, framed sepsis through the systemic inflammatory response syndrome, or SIRS, requiring at least two of four criteria: temperature abnormality, tachycardia, tachypnea or low $PaCO_2$, and abnormal white blood cell count (Gary et al., 2016). Severe sepsis was defined as sepsis with organ dysfunction or tissue hypoperfusion, and septic shock as sepsis-induced hypotension persisting despite adequate fluid resuscitation. Sepsis-2 retained this framework but expanded the diagnostic criteria from 4 to 21 bedside and laboratory variables and introduced PIRO staging—Predisposition, Infection, Response, Organ dysfunction—to capture heterogeneity more finely (Gary et al., 2016).

Sepsis-3, introduced in 2016, shifted emphasis away from inflammation alone and toward organ dysfunction caused by dysregulated host response. In this framework, the SOFA score is central: $$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$ with subscores for respiratory, coagulation, liver, cardiovascular, central nervous system, and renal systems (Gary et al., 2016). A total SOFA increase of at least 2 points from baseline indicates sepsis in the context of suspected infection. Septic shock is retained as a subset with profound circulatory and metabolic abnormalities (Gary et al., 2016).

The quick SOFA, or qSOFA, was introduced as a bedside screening tool. It assigns one point each for respiratory rate $\ge 22$ breaths/min, Glasgow Coma Scale $<15$, and systolic blood pressure $\le 100$ mmHg; a score of at least 2 suggests high risk of poor outcome and prompts fuller evaluation (Gary et al., 2016). However, the same review emphasizes the debate around universal adoption of Sepsis-3 and qSOFA, including concerns about sensitivity, exclusion of pediatric and neonatal populations, and potential loss of continuity with earlier quality metrics and studies (Gary et al., 2016).

A recurring theme across the corpus is that definition is not merely terminological. Label choice determines onset time, eligibility for alerts, and the apparent utility of interventions. DeepAISE explicitly compared Sepsis-3 and a CDC-based criterion through offline policy evaluation and found an expected approximately $8.2\%$ absolute survival benefit if antibiotics were given 6 h before $t_{\rm sepsis\mbox{-}3}$, versus only approximately $0.8\%$ for the CDC-based label, illustrating that clinically actionable labels need not coincide with administratively convenient ones (Shashikumar et al., 2019).

2. Immunopathophysiology and organ dysfunction

The immunologic literature supplied here presents sepsis as neither purely hyperinflammatory nor purely immunosuppressive. Whole-blood microarray analysis comparing 21 septic patients with 21 healthy controls reported significant up-regulation of innate-immunity genes including CD14, TLR1, TLR2, TLR4, TLR5, TLR8, HSP70, CEBP proteins, AP1 family members, TGF-$\beta$, IL-6 pathway components, S100 proteins, complement machinery, neutrophil enzymes, caspases, and Fc receptors (Hu, 2013). In the same study, many adaptive-immunity genes were down-regulated, including MHC-related genes, TCR genes, granzymes, perforin, CD40, CD8, CD3, TCR signaling components, BCR signaling components, and TH17 helper-specific transcription factors such as STAT3, RORA, and REL (Hu, 2013). Treg-related genes, including $TGF\beta$, IL-15, STAT5B, SMAD2/4, CD36, and thrombospondin, were up-regulated (Hu, 2013).

That study therefore interpreted sepsis as a syndrome with hyperactivity of TH17-like innate immunity and hypoactivity of adaptive immunity, reconciling the long-standing “hyperimmune” and “hypoimmune” theories rather than treating them as mutually exclusive (Hu, 2013). A plausible implication is that sepsis progression may involve simultaneous tissue-damaging inflammatory amplification and impaired pathogen-clearing adaptive control.

Mathematical modeling work extends this view into dynamical systems language. An improved sepsis model was formulated as 20 coupled ordinary differential equations representing pathogen dynamics, Kupffer cells, neutrophils, monocytes/macrophages, TNF-$$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$0, HMGB-1, IL-10, CD4$$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$1 and CD8$$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$2 T cells, B cells, and antibodies (Chen et al., 2022). Numerical bifurcation analysis identified saddle-node bifurcations in key subsystems, and under some parameter and initial-value settings the model exhibited persistent inflammation, sustained oscillations, or limit cycles in pathogen and immune-cell variables in the absence of control (Chen et al., 2022). In that formulation, lower pathogen replication or sufficiently high pathogen-killing rates allowed settling to an infection-cleared equilibrium, whereas threshold crossing led to coexistence of unstable infection-cleared states and stable high-inflammation or high-pathogen states (Chen et al., 2022).

Organ dysfunction is the clinical expression of these dysregulated processes. The supplied state-analysis study on 16,546 distinct adult sepsis patients in MIMIC-III identified six states, including a mild state A3, a moderate state A1, an inflammatory state A2, and three MODS subtypes A4–A6, each with distinct patterns in liver, kidney, coagulation, respiratory, cardiovascular, and neurologic variables (Fang et al., 2020). This state structure supports the view that organ dysfunction in sepsis is not a monolithic endpoint but a family of pathophysiologic trajectories.

3. Diagnostic modalities, biomarkers, and screening logic

Conventional diagnosis in the supplied literature includes clinical scores, microbiology, and host biomarkers, each with well-defined limitations. Blood culture remains standard but, in one transcriptomics study, conventional microbiological culture methods were described as requiring 24 h for positive results, being positive in only approximately $$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$3 of clinically diagnosed cases, and being vulnerable to both false negatives from prior antibiotic use and false positives from contamination (Yang et al., 2020). Host biomarkers such as procalcitonin and C-reactive protein were reported with pooled sensitivity $$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$4–$$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$5 and specificity $$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$6–$$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$7, which that study described as suboptimal for reliable early screening (Yang et al., 2020).

Transcriptomic diagnostics were proposed as an alternative route to early recognition. Recurrent Logistic Regression was used to derive a five-gene immune-related signature, LIFTS—LRRN3, IL2RB, FCER1A, TLR5, and S100A12—from blood transcriptome data (Yang et al., 2020). The resulting logistic model included the coefficients

$$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$8

and across nine validation cohorts spanning three microarray platforms, LIFTS achieved an average AUROC of $$\mathrm{SOFA}=\sum_{k=1}^{6}\mathrm{score}_k,\qquad \mathrm{score}_k\in\{0,1,2,3,4\},$$9 with $\ge 22$0 CI $\ge 22$1 (Yang et al., 2020). The paper further interpreted these genes as hubs across innate and adaptive immune pathways, linking the diagnostic signature to the immunopathology described above (Yang et al., 2020).

Clinical scoring systems remain deeply embedded in practice. SIRS, qSOFA, SOFA, NEWS, and MEWS recur across the studies as baselines or comparator tools (Gary et al., 2016). Their strengths are interpretability and clinical familiarity; their limitations are incompleteness, delayed applicability, or low specificity or sensitivity depending on the setting. The 2024 SepsisCalc framework makes this tension explicit by arguing that AI sepsis prediction models often generate only a single risk score without incorporating clinician-trusted clinical calculators for organ dysfunction, and by attempting to integrate those calculators directly into a dynamic temporal graph representation of the electronic health record (Yin et al., 2024).

A distinct diagnostic axis is rapid pathogen identification from blood smears. A 2025 image-analysis study used 16,637 Gram-stained microscopic images from positive blood-culture smears of septic patients, Cellpose 3 for segmentation, and Attention-based Deep Multiple Instance Learning for classification of 14 bacterial species and 3 yeast-like fungi (Sroka-Oleksiak et al., 17 Mar 2025). The model achieved overall accuracy $\ge 22$2 for bacteria and $\ge 22$3 for fungi, with ROC AUC $\ge 22$4 and $\ge 22$5, respectively; best individual values reached $\ge 22$6 for selected species, while morphologically similar organisms such as Staphylococcus hominis and Staphylococcus haemolyticus remained difficult to distinguish (Sroka-Oleksiak et al., 17 Mar 2025). This approach addresses etiologic identification rather than syndrome detection, but it directly targets the diagnostic-time bottleneck central to sepsis care.

4. Early recognition and machine-learning surveillance

A large portion of the supplied research treats sepsis as an early-warning problem on irregular, incomplete, and heterogeneous clinical time series. In the emergency department triage setting, KATE Sepsis was trained on patient encounters from 16 U.S. hospitals using only triage data, including age, six vital signs, pain, Glasgow Coma Scale, point-of-care glucose, categorical triage metadata, and NLP-derived UMLS concepts from free text (Ivanov et al., 2022). Evaluated retrospectively on 512,949 adult encounters with 9,257 clinician-assigned sepsis diagnoses made within 24 hours of arrival, KATE Sepsis achieved AUC $\ge 22$7, sensitivity $\ge 22$8, and specificity $\ge 22$9, compared with AUC $<15$0, sensitivity $<15$1, and specificity $<15$2 for a SIRS-based standard screening protocol (Ivanov et al., 2022). Sensitivity for severe sepsis and septic shock was $<15$3 and $<15$4, respectively, versus $<15$5 and $<15$6 for the standard protocol (Ivanov et al., 2022).

In ICU time-series modeling, different architectures were proposed for different failure modes of earlier systems. Moor et al. derived an hourly Sepsis-3 label from MIMIC-III and framed early detection as supervised time-series classification, using either a Multi-task Gaussian Process Adapter plus Temporal Convolutional Network or a Dynamic Time Warping ensemble (Moor et al., 2019). Seven hours before sepsis onset, the previous GP-RNN baseline reached AUPRC approximately $<15$7, whereas MGP-TCN reached approximately $<15$8 and DTW-KNN approximately $<15$9 (Moor et al., 2019). DeepAISE instead cast the task as recurrent neural survival modeling with a Weibull baseline hazard and GRU-based feature extractor, producing hourly risk scores beginning at ICU +4 h (Shashikumar et al., 2019). At 4 h lead time and sensitivity fixed at $\le 100$0, it achieved AUC $\le 100$1 with false alarm rate $\le 100$2 on the internal Emory cohort and AUC $\le 100$3 with false alarm rate $\le 100$4 on external MIMIC-III data (Shashikumar et al., 2019).

Generalizability across hospitals and countries remains a central concern. A harmonized multi-national study assembled 156,309 ICU admissions from MIMIC-III, eICU, HiRID, AUMC, and Emory, with 26,734 septic stays labeled using hourly Sepsis-3 annotations (Moor et al., 2021). A deep self-attention model achieved AUROC $\le 100$5 in internal out-of-sample validation and $\le 100$6 in external validation. For harmonized prevalence of $\le 100$7, at $\le 100$8 recall the model reached $\le 100$9 precision with median lead time $8.2\%$0 h internally, and $8.2\%$1 precision with median lead time $8.2\%$2 h externally (Moor et al., 2021). These values make explicit the performance drop under cross-site deployment.

Interpretability enters through multiple routes. One proof-of-concept study used Jensen–Shannon divergence between patient-specific kernel density estimates and a non-septic reference cohort to create a real-time anomaly score $8.2\%$3 and feature-wise divergence scores $8.2\%$4 (Smith et al., 2022). Another, SepsisCalc, constructed temporal heterogeneous graphs with clinical variable, organ, and calculator nodes, dynamically added estimated clinical calculators when confidence exceeded $8.2\%$5, and reported that the model outperformed state-of-the-art baselines by $8.2\%$6–$8.2\%$7 AUC while yielding earlier alerts by approximately 1 h on average (Yin et al., 2024). This suggests a broader methodological trend: sepsis surveillance is moving from opaque scalar risk scoring toward models that expose organ-specific mechanisms or clinician-trusted intermediate quantities.

5. Phenotypes, states, and temporal structure

The supplied literature repeatedly describes sepsis as heterogeneous in presentation, organ involvement, and recovery trajectory. One line of work uses unsupervised or semi-supervised representation learning to identify disease states. Archetypal analysis of MIMIC-III data identified six distinct sepsis states, with high clustering stability (NMI $8.2\%$8, ARI $8.2\%$9) and statistically significant pairwise differences across all state pairs by Hotelling $t_{\rm sepsis\mbox{-}3}$0 tests with $t_{\rm sepsis\mbox{-}3}$1 (Fang et al., 2020). A3 was labeled mild, A1 moderate, A2 inflammatory, and A4–A6 MODS subtypes with mortality approximately $t_{\rm sepsis\mbox{-}3}$2, $t_{\rm sepsis\mbox{-}3}$3, and $t_{\rm sepsis\mbox{-}3}$4, respectively (Fang et al., 2020). Third-order Markov analysis further showed highly persistent trajectories, including a nearly absorbing “111” state sequence with probability approximately $t_{\rm sepsis\mbox{-}3}$5 (Fang et al., 2020).

A second line of work used time-aware soft clustering with clinical guidance from organ-dysfunction labels. Jiang et al. modeled each patient as a $t_{\rm sepsis\mbox{-}3}$6 trajectory and minimized a fuzzy clustering objective augmented by semi-supervised pull-toward and push-away terms for clinically labeled organ dysfunction (Jiang et al., 2023). With $t_{\rm sepsis\mbox{-}3}$7 organ-guided clusters and a subsequent K-medoids step on membership vectors plus an ABM severity indicator, the study derived six hybrid sub-phenotypes H1–H6 (Jiang et al., 2023). In MIMIC-IV, H1 had nearly equal membership across the three organ clusters and the lowest ABM, corresponding to multi-organ failure and the worst survival; H4–H6 were milder, more organ-dominant profiles (Jiang et al., 2023). Their early-warning phenotype classifier, a one-vs-rest logistic regression model using feature summaries from the first 12, 24, 48, or 120 hours, performed best at 24 h with precision $t_{\rm sepsis\mbox{-}3}$8, recall $t_{\rm sepsis\mbox{-}3}$9, accuracy $0.8\%$0, and AUPRC $0.8\%$1 on MIMIC-IV (Jiang et al., 2023).

Earlier symbolic temporal mining approached the same problem without neural representation learning. Knowledge-based temporal abstraction converted 26 ICU concepts into interval-valued states and gradients, and KarmaLego mining then discovered frequent Time-Interval Relation Patterns in septic and non-septic patients (Sheetrit et al., 2017). In the last 12 h before onset, 26,968 frequent patterns were found, including 6,168 exclusive to septic and 6,416 exclusive to non-septic cohorts; in the 6 h window, 22,422 patterns were found (Sheetrit et al., 2017). Kolmogorov–Smirnov tests showed significant differences in pattern distributions between septic and non-septic populations in both 12 h and 6 h windows (Sheetrit et al., 2017). Although that paper did not report final classifier metrics, it argued that these temporal patterns are promising constructed features for early diagnosis (Sheetrit et al., 2017).

Taken together, these studies indicate that “sepsis” in EHR data often decomposes into recurring temporal and organ-specific configurations rather than a single homogeneous syndrome. This suggests that detection, prognosis, and intervention may benefit from phenotype- or state-aware rather than population-average strategies.

6. Treatment optimization, simulation, and emerging platforms

Treatment-oriented studies in the supplied corpus concentrate on fluid management, control-theoretic intervention, and policy simulation. A human-in-the-loop AI framework trained on 1,122 MIMIC-III ICU sepsis stays modeled mortality as $0.8\%$2 and used constrained inverse classification to recommend only incremental changes to physician-prescribed IV fluid volumes (Gupta et al., 2020). On 224 held-out test cases, the baseline average $0.8\%$3 under physician dosing was $0.8\%$4; at maximal budget $0.8\%$5, the average probability fell to $0.8\%$6, a $0.8\%$7 relative reduction (Gupta et al., 2020). Even budgets $0.8\%$8–$0.8\%$9 yielded $\beta$0–$\beta$1 mortality reduction (Gupta et al., 2020). The recommendations systematically increased D5LR, D5HNS, and D5W while decreasing NS and LR relative to clinician choices (Gupta et al., 2020).

A more mechanistic control formulation came from the improved nonlinear sepsis model described above. That work defined antibiotic control $\beta$2 for the early high-pathogen phase and anti–TNF-$\beta$3 control $\beta$4 for the late high-inflammation phase, optimizing biomarker ratios such as $\beta$5, $\beta$6, and $\beta$7 (Chen et al., 2022). An RNN-BO algorithm combined Bayesian optimization with a recurrent neural network so that, after learning from historical optimal-control trajectories, it could predict a corresponding time-series optimal control for a new initial condition in approximately 2 s, compared with approximately 25–45 s for alternative BO-based procedures in the reported simulations (Chen et al., 2022).

Policy learning also appears through simulation. The “Sepsis World Model” used a variational auto-encoder and an MDN-RNN to build an OpenAI Gym simulator from MIMIC data, representing patient state in a 30-dimensional latent space and actions as 25 discrete vasopressor–fluid dosage combinations (Kiani et al., 2019). The authors evaluated fidelity by comparing simulated rollouts against real trajectories and found that the MDN component was critical for representing uncertainty and realistic variance (Kiani et al., 2019). This line of work does not provide a ground-truth optimal policy, but it creates a platform for testing reinforcement-learning strategies under learned trajectory dynamics.

Finally, sepsis monitoring is extending beyond the ICU and even beyond hospital infrastructure. SepAl used six digitally acquirable vital signs—heart rate, respiratory rate, systolic and diastolic blood pressure, SpO$\beta$8, and core body temperature—from low-power wearable sensors, processed by a fully quantized temporal convolutional network deployable on an ARM Cortex-M33 (Giordano et al., 2024). In a retrospective 4 h prediction-horizon task it achieved sensitivity $\beta$9 and specificity $TGF\beta$0; in a real-time sliding-window task, median predicted time to sepsis onset was 9.8 h before clinical onset with sensitivity $TGF\beta$1 and specificity $TGF\beta$2 over 5-fold cross-validation (Giordano et al., 2024). A separate heart-rate-only study on PhysioNet/CinC Challenge 2019 data optimized wearable-friendly models by genetic algorithm and reported, for the LSTM model, AUROC $TGF\beta$3 and AUPR $TGF\beta$4 at 1 h, and AUROC $TGF\beta$5 and AUPR $TGF\beta$6 at 4 h, all at a decision threshold yielding $TGF\beta$7 sensitivity (Rafiei et al., 30 Dec 2025).

These treatment and deployment studies collectively suggest that sepsis research is converging on a coupled agenda: earlier detection, more individualized control, explicit handling of temporal dynamics, and implementation on platforms ranging from bedside dashboards to low-power wearables.