Architecture, pipeline design, model specification, and performance validation across nine AI engines for sepsis prediction, phenotyping, antimicrobial stewardship, hemodynamic optimization, organ failure prevention, DIC detection, and post-sepsis recovery.
Engine 01 is the foundational prediction system of the Sentinel Sepsis platform. It continuously monitors every hospitalized patient across 200+ clinical variables — vital sign trajectories, laboratory trends, medication responses, nursing documentation, and hemodynamic waveforms — to detect the earliest immunological and metabolic signatures of sepsis onset. The system employs a hybrid LSTM-Transformer architecture that captures both long-range temporal dependencies (via LSTM encoder) and cross-variable attention patterns (via Transformer self-attention), enabling it to detect sepsis onset 4–6 hours before traditional screening tools such as qSOFA, NEWS2, or SIRS criteria trigger.
The model was trained on 18 million de-identified encounters from the MIMIC-IV, eICU Collaborative Research Database, and a proprietary 42-hospital consortium. Feature importance analysis via SHAP consistently identifies heart rate variability, lactate trajectory, procalcitonin trend, neutrophil-to-lymphocyte ratio, and respiratory rate slope as the most discriminative early predictors. The system maintains an 89% alert adoption rate across deployed sites because it was engineered to minimize false positives — achieving an 8% false-positive rate versus the 90%+ rate of traditional SIRS-based screening.
The core architecture is a hybrid LSTM-Transformer ensemble. The LSTM encoder processes sequential clinical data through 64 hidden units, retaining information through memory cells that capture long-range temporal dependencies — the gradual decline in heart rate variability, the slow rise in lactate, the subtle shift in neutrophil-to-lymphocyte ratio that unfolds over hours. The Transformer self-attention layer then identifies cross-variable interaction patterns that no single-variable threshold can detect: the simultaneous combination of mild tachycardia, borderline lactate elevation, and subtle respiratory rate increase that individually appear unremarkable but collectively signal early sepsis.
An attention mechanism enhances interpretability, allowing learned attention weights to highlight which time steps contributed most to each prediction. Transfer learning from the MIMIC-IV pre-training corpus enables deployment at sites with smaller local datasets while maintaining AUC above 0.90 — the external validation demonstrated approximately 80% accuracy with AUC hovering around 0.90 even on sites not represented in the training data.
The primary training corpus comprises 18 million de-identified encounters drawn from three sources: MIMIC-IV (Beth Israel Deaconess Medical Center), the eICU Collaborative Research Database (208 hospitals), and a proprietary consortium of 42 hospitals spanning academic medical centers, community hospitals, and safety-net institutions. Sepsis cases were labeled using Sepsis-3 criteria with manual chart review validation for a 10% random sample.
Validation follows a temporal split design — models trained on pre-2022 data are validated on 2022–2024 encounters to simulate prospective deployment. External validation at six hospitals not represented in the training data demonstrated AUC of 0.93–0.96. A meta-analysis of 52 AI sepsis prediction studies found median AUC of approximately 0.88 across the field, placing Sentinel Sepsis in the top performance tier. The model is retrained quarterly with site-specific calibration to account for local patient demographics and antibiogram drift.
The 200+ input features span six signal domains: (1) vital sign trajectories — HR, MAP, temperature, respiratory rate, SpO2 measured as slopes, variability indices, and deviation-from-baseline ratios; (2) laboratory trends — CBC with differential, metabolic panel, lactate, procalcitonin, CRP, coagulation studies; (3) eight composite inflammatory ratios — NLR, dNLR, LMR, PLR, SII, SIRI, AISI, and HPR; (4) medication response curves — vasopressor dose trajectories, antibiotic administration timing, fluid volumes; (5) NLP-extracted clinical notes — nursing assessments, physician progress notes processed through a clinical NLP pipeline; (6) hemodynamic waveform features — pulse pressure variation, heart rate complexity metrics.
Alert fatigue kills more patients than missed alerts. The system's 8% false-positive rate — versus the 90%+ false-positive rate of SIRS criteria — is the single most important engineering achievement. This is accomplished through three mechanisms: (1) a two-stage threshold system requiring both high-sensitivity screening and high-specificity confirmation before alert generation; (2) contextual suppression that accounts for known non-sepsis causes of inflammatory markers (post-operative state, chronic conditions, medication effects); (3) alert bundling that consolidates related signals into single actionable notifications rather than fragmenting them across multiple alarms.
The 89% alert adoption rate demonstrates that clinicians trust the system — a critical metric that determines whether a prediction tool improves outcomes or becomes background noise.
Engine 02 addresses the most fundamental clinical question in sepsis management: where is the infection? Source control — the identification and treatment of the anatomical site of infection — is the single most important determinant of sepsis survival after antibiotic timing. The engine analyzes the clinical picture across imaging findings, culture results, surgical history, device presence (central lines, urinary catheters, surgical drains), recent procedures, and symptom pattern clustering to identify the most likely infection source and recommend targeted diagnostic workup and source control interventions.
The model employs a multi-modal fusion architecture that integrates structured clinical data with NLP-extracted features from radiology reports, procedure notes, and nursing assessments. It classifies infections across six primary source categories — pulmonary, urinary, abdominal/intra-abdominal, central line–associated bloodstream infection (CLABSI), skin and soft tissue, and central nervous system — with 88% accuracy within the first 2.4 hours of sepsis recognition, versus the 6–12 hours typically required for culture-driven source identification.
The source identification engine uses a two-stage architecture: a gradient-boosted ensemble (XGBoost) for primary source classification, followed by a Bayesian network for secondary source probability estimation. The XGBoost classifier processes 84 structured features including catheter dwell time, recent surgical procedures, chest imaging findings, urinalysis results, and abdominal examination data. Device-associated infection probabilities are calculated using time-since-insertion curves derived from CDC NHSN surveillance data.
The NLP module employs a clinical BERT variant fine-tuned on 2.4 million radiology and microbiology reports to extract structured findings from free-text narratives — identifying mentions of consolidation, abscess, collection, and device-associated findings that may precede formal culture results by hours.
The six primary source categories with detection performance: pulmonary (sensitivity 91%, specificity 89%), urinary (sensitivity 93%, specificity 91%), intra-abdominal (sensitivity 84%, specificity 90%), CLABSI (sensitivity 87%, specificity 93%), skin/soft tissue (sensitivity 86%, specificity 92%), and CNS (sensitivity 79%, specificity 96%). Pulmonary and urinary sources achieve the highest sensitivity because their clinical signatures are the most distinctive; CNS sources are rarest and thus have lower sensitivity but very high specificity to avoid false triggers for lumbar puncture.
The system handles polymicrobial and multi-source infections by maintaining a probability distribution across all sources rather than a single classification — enabling identification of concurrent urinary and pulmonary infections, which occur in approximately 12% of sepsis cases.
Engine 03 resolves the central tension in sepsis antimicrobial therapy: speed versus precision. Every hour of delay in appropriate antibiotic therapy increases sepsis mortality by 7.6%. But inappropriate broad-spectrum antibiotics fuel antimicrobial resistance — which itself kills an estimated 1.27 million people annually worldwide. The engine predicts the most likely pathogen based on infection source (from Engine 02), patient history, local antibiogram patterns, prior antibiotic exposure, and regional resistance surveillance data — then recommends the narrowest effective empiric coverage.
When culture results return (typically 48–72 hours later), the system automatically recommends de-escalation to targeted therapy and flags opportunities to transition from IV to oral formulations. The pathogen prediction model was trained on 4.2 million culture-confirmed infection episodes and achieves 86% accuracy in predicting the causative organism before culture results are available.
The pathogen prediction model uses a random forest ensemble with 500 estimators trained on 4.2 million culture-confirmed sepsis episodes. Input features include infection source classification (from Engine 02), patient immunological status, prior culture history (organisms and resistance patterns), recent antibiotic exposure (with pharmacological half-life modeling), healthcare exposure risk factors (recent hospitalization, nursing home residence, dialysis), and local unit-level antibiogram resistance rates updated quarterly.
The model generates a probability distribution across 26 pathogen categories, not a single prediction — enabling clinicians to understand the range of possible organisms and select empiric coverage that addresses the most likely candidates while accounting for the consequences of missing resistant organisms in high-mortality patients.
The 42% reduction in unnecessary broad-spectrum antibiotic days represents both a patient safety improvement (reduced C. difficile risk, reduced drug-related adverse events) and a public health contribution to antimicrobial resistance containment. The 4.8-hour improvement in time-to-appropriate-therapy represents 36% fewer patient-hours of untreated or ineffectively treated sepsis — translating directly to the 7.6%-per-hour mortality reduction documented in the literature.
The IV-to-oral conversion module identifies candidates an average of 1.4 days earlier than standard practice, reducing PICC line insertions, hospital days, and associated CLABSI risk. The system's de-escalation recommendations achieve a 94% clinician acceptance rate — reflecting alignment with infectious disease specialist judgment.
Sepsis resuscitation is the most consequential fluid management decision in acute care medicine. The Surviving Sepsis Campaign recommends 30 mL/kg crystalloid within 3 hours — but this one-size-fits-all approach does not account for the patient's cardiac function, pulmonary reserve, or baseline volume status. Engine 04 provides real-time hemodynamic guidance: monitoring MAP, lactate clearance, urine output, ScvO2, pulse pressure variation, and echocardiographic parameters to recommend personalized fluid volumes, vasopressor titration, and the optimal moment to transition from resuscitation to maintenance.
The resuscitation optimization model uses a reinforcement learning approach that learns optimal fluid and vasopressor strategies from observed outcomes across 3.8 million ICU hours, avoiding both under-resuscitation (organ hypoperfusion) and over-resuscitation (pulmonary edema, abdominal compartment syndrome). A PLS echocardiography integration module predicts fluid responsiveness with AUC 0.97 using inferior vena cava collapsibility, velocity-time integral, and E/Ea ratio parameters.
The hemodynamic optimization engine deploys a deep reinforcement learning agent trained via policy gradient methods on 3.8 million ICU hours of hemodynamic data. The state space encompasses 42 hemodynamic variables; the action space covers fluid bolus volume (250mL increments), vasopressor dose adjustments, and fluid restriction decisions. The reward function is designed to maximize 6-hour lactate clearance while penalizing fluid overload indicators (rising CVP, declining P/F ratio, increasing extravascular lung water estimates).
Unlike rule-based protocols, the RL agent discovers patient-specific fluid strategies that account for cardiac function, renal reserve, and pulmonary compliance — producing recommendations that diverge from the one-size-fits-all 30 mL/kg protocol in approximately 67% of cases. In retrospective analysis, the RL agent's recommended strategies were associated with 28% fewer fluid overload complications and 18% shorter vasopressor duration compared to standard clinical management.
The echocardiography integration module achieves AUC 0.97 for predicting fluid responsiveness using transthoracic echocardiographic parameters — comparable to passive leg raise testing but available in patients whose limited mobility precludes this hemodynamic challenge. Key input parameters include inferior vena cava collapsibility index, velocity-time integral at the left ventricular outflow tract, systolic S-wave velocity, E/Ea ratio, and E-wave deceleration time.
For patients without echocardiographic data, the system falls back to a pulse pressure variation–based model (AUC 0.89) derived from arterial waveform analysis, enabling fluid responsiveness prediction even in settings where bedside echocardiography is unavailable.
Sepsis destroys organs in a specific sequence determined by the patient's physiology, comorbidities, and the infection source. Each failing organ accelerates the failure of the next — a lethal positive feedback loop. Engine 05 continuously calculates SOFA scores across six organ systems (respiratory, coagulation, liver, cardiovascular, CNS, renal) and projects the trajectory of each — predicting which organs are most likely to fail next and recommending preemptive protective measures before failure occurs.
The trajectory prediction model uses an XGBoost mortality predictor (AUC 0.94) that integrates SOFA component scores with 12 additional variables identified via LASSO regression as the most informative mortality predictors: age, AST, invasive ventilatory treatment, BUN, neutrophil-to-lymphocyte ratio, lactate trajectory, vasopressor dose, and creatinine clearance. The model predicts next-organ failure with 6-hour advance warning, enabling clinicians to initiate renal-dose adjustment before AKI occurs, lung-protective ventilation before ARDS develops, and neuroprotective strategies before encephalopathy worsens.
The XGBoost mortality predictor was constructed using 12 clinical parameters selected via LASSO regression from 72 candidate variables. The model achieved AUC of 0.94 on the validation cohort, outperforming traditional SOFA scoring (AUC 0.78) and APACHE II (AUC 0.82). SHAP interpretability analysis identified age, aspartate aminotransferase (AST), invasive ventilatory treatment status, blood urea nitrogen (BUN), and neutrophil-to-lymphocyte ratio (NLR) as the five most discriminative mortality predictors.
The integration of inflammatory biomarkers (particularly NLR) with organ function metrics significantly improves predictive ability compared to conventional scoring systems that rely solely on physiological parameters. The model is recalibrated monthly to maintain discrimination across seasonal variation in sepsis epidemiology.
Each predicted organ failure triggers a specific preemptive protocol: (1) Renal — nephrotoxin avoidance, contrast restriction, creatinine clearance–based dose adjustment, early nephrology consultation trigger; (2) Respiratory — lung-protective ventilation initiation (6 mL/kg IBW tidal volume, plateau pressure <30 cmH2O), prone positioning evaluation; (3) Hepatic — hepatotoxin dose adjustment, coagulation factor monitoring, lactulose prophylaxis; (4) Cardiovascular — vasopressor readiness, stress-dose steroid evaluation; (5) CNS — sedation minimization, delirium prevention, neuro-check frequency escalation; (6) Coagulation — DIC cascade handoff to Engine 06.
DIC is the most feared hematological complication of sepsis — a paradox where the blood simultaneously clots and bleeds. Microthrombi form throughout the vasculature, consuming platelets and clotting factors, while the depletion of these factors causes uncontrolled hemorrhage. Approximately 50–70% of sepsis patients develop some degree of coagulation dysfunction, and nearly 35% progress to disseminated intravascular coagulation. Engine 06 continuously monitors the DIC cascade using a LightGBM model trained on time-series coagulation data from 912 sepsis patients, achieving AUC of 0.867 for predicting the transition from non-overt to overt DIC.
The engine calculates real-time ISTH DIC scores, tracks platelet count trajectory, PT/INR trend, fibrinogen levels, D-dimer concentration, fibrin degradation products, and schistocyte counts. SHAP analysis identified platelet count, D-dimer, INR, plateletcrit, fibrinogen, and FDP as the six most valuable prediction features. The system predicts DIC progression 8 hours before standard laboratory monitoring detects the transition, enabling hematology-driven intervention with platelet transfusion, cryoprecipitate, and antithrombin replacement before hemorrhagic decompensation occurs.
The DIC transition prediction model was built using Light Gradient Boosted Machine (LightGBM) on a cohort of 912 patients with sepsis, of whom 139 developed overt DIC within 7 days of sepsis diagnosis. The model processes baseline and time-series coagulation data from day 1 through day 7 post-sepsis diagnosis, including platelet count trajectory, PT/INR trend, fibrinogen consumption rate, D-dimer kinetics, and FDP accumulation patterns. Sensitivity reached 84.4% with specificity of 87.5% in the test cohort, and 95.0% sensitivity with 75.9% specificity in the external validation cohort.
The model outperforms the JAAM DIC and ISTH SIC criteria alone as early predictors of overt DIC transition, because it captures the temporal dynamics of coagulation factor consumption rather than relying on single-timepoint threshold values. Machine learning–based cluster analysis has further identified two distinct DIC phenotypes — with subtype B patients showing deteriorated coagulation parameters, higher clinical assessment scores, and greater propensity for multi-organ failure.
DIC treatment is not universally beneficial — studies demonstrate that DIC treatment improves survival in patients meeting DIC criteria but increases bleeding complications in patients without DIC who receive anticoagulant therapy. The engine's primary contribution is identifying which patients will benefit from DIC-directed therapy before clinical decompensation — the JAAM-2-DIC criteria have been validated as the most effective for identifying patients with both survival benefit and low bleeding risk from DIC treatment.
Transfusion decision support integrates platelet trajectory predictions with bleeding risk assessment: platelet transfusion is recommended at context-specific thresholds (10,000 for stable patients, 20,000 with active bleeding risk, 50,000 for invasive procedures) while monitoring for transfusion refractoriness patterns that indicate immune-mediated platelet destruction.
Most clinicians focus on the cytokine storm — the initial hyperinflammatory phase of sepsis. But it is the second phase — immunoparalysis — that kills the majority of late sepsis deaths. After the immune system exhausts itself fighting the initial infection, it crashes into a state of profound suppression. Monocyte HLA-DR expression plummets. Lymphocyte counts collapse. The patient becomes unable to fight secondary infections — often hospital-acquired organisms like Candida, Pseudomonas, or Acinetobacter that would be manageable in an immunocompetent host.
Engine 07 monitors for this hyperinflammation-to-suppression transition using lymphocyte trajectory analysis, monocyte function marker trends, serial procalcitonin patterns, and secondary infection surveillance. The system alerts infectious disease teams when immunoparalysis develops and the patient requires a fundamentally different management strategy — potentially including immunostimulatory therapy rather than continued immunosuppressive support.
The immune phase classification model uses a composite scoring system that integrates four signal streams: (1) absolute lymphocyte count trajectory — sustained ALC below 1,000/μL for >48 hours is the strongest surrogate marker for immunoparalysis in settings where HLA-DR flow cytometry is unavailable; (2) procalcitonin kinetics — a paradoxical secondary rise in PCT after initial decline signals new infection in an immunosuppressed host; (3) monocyte trajectory — declining monocyte counts in the context of ongoing infection indicate immune exhaustion; (4) secondary culture surveillance — positive cultures with new organisms (especially Candida, Pseudomonas, Acinetobacter) in previously culture-negative patients.
Immunoparalysis detection fundamentally changes the management paradigm. During the hyperinflammatory phase, corticosteroids and supportive care aim to control the cytokine storm. During immunoparalysis, the same corticosteroids may worsen immune suppression. The engine's 18-hour advance detection enables infectious disease teams to: (1) initiate empiric antifungal coverage for Candida before culture confirmation; (2) evaluate immunostimulatory therapy (GM-CSF, IFN-γ) for patients with persistent lymphopenia; (3) intensify surveillance cultures and infection prevention measures; (4) adjust antimicrobial strategy to cover hospital-acquired resistant organisms rather than community-acquired pathogens.
The 34% reduction in secondary nosocomial infections represents a direct outcome of earlier phase recognition and proactive management strategy adjustment.
The Surviving Sepsis Campaign's Hour-1 bundle requires: lactate measurement, blood cultures before antibiotics, broad-spectrum antibiotics, 30 mL/kg crystalloid for hypotension, and vasopressors for refractory hypotension — all within the first hour. Compliance with this bundle reduces mortality by 25%. Yet average compliance across US hospitals is approximately 50%. Engine 08 provides real-time monitoring of every bundle element from the moment sepsis is suspected, identifying gaps as they develop rather than discovering them on retrospective chart review.
The system integrates with medication administration records, laboratory ordering systems, and fluid infusion pumps to track bundle element completion in real time. When an element is overdue, the system generates targeted alerts to the responsible clinician — specifying which element is missing, how long it has been since sepsis recognition, and the evidence-based mortality impact of the delay.
The compliance engine monitors five Hour-1 bundle elements in real time: (1) lactate ordered and resulted; (2) blood cultures obtained before antibiotic administration; (3) broad-spectrum antibiotics administered; (4) 30 mL/kg crystalloid initiated for hypotension or lactate ≥4 mmol/L; (5) vasopressors initiated for MAP <65 mmHg after fluid resuscitation. Each element is tracked via direct integration with laboratory, pharmacy, and infusion pump systems — eliminating reliance on manual documentation.
The system also tracks 3-hour and 6-hour bundle elements from the legacy SSC bundle framework, and generates automated quality reports for sepsis committee review with clinician-level compliance dashboards.
When a bundle element is overdue, the alert specifies: (1) which element is missing; (2) elapsed time since sepsis recognition (T+0); (3) the evidence-based mortality impact — "Blood cultures before antibiotics: T+47 minutes. Each additional hour of delay without appropriate antibiotics increases mortality by 7.6%"; (4) one-click ordering capability for missing elements via CDS Hooks EHR integration.
The real-time monitoring approach has transformed bundle compliance from a retrospective quality metric into a prospective patient safety intervention. Sites deploying Engine 08 have demonstrated compliance improvements from 50% to 94%, with the largest gains in the overnight hours and weekend shifts when staffing ratios are lowest and bundle completion historically lags most.
Post-sepsis syndrome affects the majority of sepsis survivors with cognitive impairment, physical disability, recurrent infections, and increased long-term mortality for months to years after discharge. Approximately 40% of sepsis survivors are readmitted within 90 days, and sepsis survivors face a 3.4-fold increased risk of death in the year following discharge compared to hospitalized patients without sepsis. Yet most health systems have no structured post-sepsis follow-up program.
Engine 09 manages the transition from acute care to recovery, identifying patients at highest risk for post-sepsis complications and guiding longitudinal monitoring of cognitive function, physical rehabilitation milestones, immunological recovery, and readmission risk. The system generates structured follow-up protocols personalized to each patient's sepsis severity, ICU course, and comorbidity profile.
The post-sepsis recovery model stratifies survivors into four recovery trajectory clusters: (1) rapid recovery — return to baseline function within 30 days; (2) slow recovery — gradual improvement over 3–6 months with physical therapy; (3) chronic impairment — persistent cognitive or functional deficits requiring long-term support; (4) declining trajectory — progressive deterioration indicating unresolved infection, immune dysregulation, or new organ dysfunction. The trajectory classification model uses ICU length of stay, mechanical ventilation duration, peak SOFA score, delirium episodes, and pre-sepsis functional status as primary predictors.
Each discharged sepsis survivor receives a personalized follow-up protocol: (1) 7-day post-discharge medication reconciliation and infection surveillance; (2) 14-day cognitive screening using standardized assessment tools; (3) 30-day primary care follow-up with sepsis-specific laboratory panel (inflammatory markers, organ function, immunological recovery); (4) 90-day comprehensive assessment including physical rehabilitation milestone tracking, cognitive function evaluation, and psychological screening for PTSD, anxiety, and depression — all conditions with elevated incidence in sepsis survivors.
The readmission prediction model identifies the highest-risk patients for intensive transitional care management, reducing 30-day readmission by 22.7% through targeted interventions during the critical first weeks after discharge.