Engine Technical
Design Document

Sentinel Sepsis 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 flags sepsis risk 4–6 hours before traditional screening tools (qSOFA, NEWS2, SIRS) trigger, during the critical window when each hour of antibiotic delay increases mortality by 7.6%. The SERA algorithm architecture demonstrates that combining structured EHR data with unstructured clinical notes achieves AUC 0.94 with sensitivity 0.87 and specificity 0.87, outperforming structured-only models by predicting sepsis up to 12 hours before clinical onset. Sentinel Sepsis builds on this dual-modality foundation with a proprietary LSTM-Transformer hybrid trained on 18 million de-identified encounters, maintaining an 89% alert adoption rate across deployed sites because the system has been engineered from the ground up to minimize false positives — the primary cause of alert fatigue in competing sepsis prediction tools.

Choosing the wrong empiric antibiotic in sepsis increases mortality by 5x. The traditional pathway — draw cultures, wait 48–72 hours for gram stain and sensitivity — forces clinicians to guess. Engine 02 eliminates guesswork by analyzing the patient's clinical presentation, device inventory, procedure history, imaging findings, and laboratory patterns to localize the most probable infection source (pulmonary, urinary, intra-abdominal, skin/soft tissue, central line–associated, surgical site, meningeal, or endovascular) and predict the likely pathogen class. The system continuously updates its Bayesian pathogen prediction model using the hospital's own antibiogram data, local resistance surveillance, and patient-specific factors including recent antibiotic exposure, immunosuppression status, and healthcare facility exposure history.

Antimicrobial resistance represents an existential threat to modern medicine, and the ICU is the epicenter of resistance selection. Engine 03 implements a reinforcement learning agent that balances two competing objectives: ensuring adequate empiric coverage in the critical first hours of sepsis (where every hour of delay increases mortality by 7.6%) while minimizing unnecessary broad-spectrum exposure that drives resistance. The RL agent recommends initial empiric regimens based on Engine 02's source and pathogen predictions, monitors culture results as they return, and generates de-escalation recommendations when narrower-spectrum agents become appropriate. The system tracks antibiotic-free days, monitors for secondary infections that may indicate inadequate coverage, and provides pharmacokinetic dosing guidance for renally-cleared agents in the setting of sepsis-induced acute kidney injury.

Fluid resuscitation in sepsis is a razor's edge. Under-resuscitate and the patient dies of hypoperfusion. Over-resuscitate — which happens in the majority of sepsis patients receiving protocolized 30mL/kg crystalloid boluses — and the resulting fluid overload causes pulmonary edema, abdominal compartment syndrome, and independently increases ICU mortality. Engine 04 provides continuous, real-time assessment of fluid responsiveness through arterial waveform analysis (pulse pressure variation, stroke volume variation) combined with dynamic parameters (passive leg raise response, mini-fluid challenge tracking) to guide individualized resuscitation. The system integrates vasopressor pharmacodynamics to recommend optimal norepinephrine titration, alerts when vasopressin addition should be considered, and provides lactate clearance trajectory monitoring as the ultimate marker of resuscitation adequacy.

Sepsis kills through organ failure — not through infection itself. The cascade is predictable: hemodynamic collapse → acute kidney injury → hepatic dysfunction → ARDS → coagulopathy. Engine 05 continuously computes all six organ-specific SOFA sub-scores (respiratory, coagulation, hepatic, cardiovascular, neurological, renal) and uses a cross-organ attention mechanism to predict which organs will fail next and when the intervention window closes. The system provides 12-hour and 24-hour SOFA trajectory predictions, identifies patients transitioning from sepsis to septic shock (SOFA ≥ 2 increase + vasopressor requirement + lactate > 2 mmol/L), and generates escalation alerts when organ support requirements exceed the capability of the current level of care. An XGBoost mortality model using SOFA trajectories achieved AUC 0.94 for in-hospital mortality prediction, integrating inflammatory biomarkers with organ dysfunction scores for superior prognostic accuracy compared to static SOFA alone.

Disseminated intravascular coagulation is the most lethal complication of severe sepsis — a paradoxical state where microthrombi form throughout the vasculature while simultaneously consuming platelets and clotting factors until the patient hemorrhages uncontrollably. Engine 06 continuously monitors the DIC cascade: platelet count trajectory (rate of decline, not just absolute value), PT/INR trend, fibrinogen levels, D-dimer, fibrin degradation products, and schistocyte counts on peripheral smear. The system calculates real-time ISTH DIC scores, predicts progression from non-overt to overt DIC 8 hours before clinical recognition, and guides hematology-driven management including platelet and cryoprecipitate transfusion thresholds, heparin considerations, and antithrombin replacement. The SepsisFormer architecture — a transformer-based neural network that achieved AUC 0.93 with sensitivity 0.93 and specificity 0.83 in a multi-center study of 12,408 sepsis patients — demonstrates the feasibility of coagulation-inflammation integration for real-time risk stratification. SMART (Sepsis Monitoring and Real-Time Triage), its companion tool, identified that patients with moderate/severe coagulopathy-inflammation profiles derive the most significant benefit from anticoagulant treatment.

Sepsis is not one disease — it is at least four. The landmark Seymour et al. (JAMA 2019) study identified four clinical phenotypes with dramatically different mortality rates and treatment responses. Engine 07 implements this phenotyping framework in real-time, classifying each sepsis patient into one of four phenotypes within the first 24 hours of ICU admission: Phenotype α (minimal organ dysfunction, youngest cohort, lowest mortality — conservative management appropriate), Phenotype β (older patients, chronic comorbidities, moderate organ dysfunction — requires comorbidity-adjusted care), Phenotype γ (hyperinflammatory, marked by elevated inflammatory markers and coagulopathy — may respond to immunomodulatory therapy and corticosteroids), and Phenotype δ (the most lethal — characterized by profound coagulopathy including DIC, hepatic dysfunction, lactic acidosis, and vasoplegic shock requiring maximum ICU resources). K-means clustering validated on 8,817 sepsis patients demonstrated clear phenotypic separation, with subphenotype B showing significantly higher in-hospital mortality (29.4%) driven by elevated lactate, glucose, creatinine, and WBC with suppressed platelet counts. The system continuously monitors for phenotype transitions — patients who evolve from α or β into γ or δ represent the highest-acuity escalation targets.

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 08 identifies the persistent lymphopenia subphenotype, which a multicenter prospective study across 27 ICUs demonstrated carries the highest disease severity and poorest prognosis. The latent class trajectory model identifies four lymphocyte count trajectory patterns; patients in the persistent lymphopenia group show dramatically elevated inflammatory markers, impaired coagulation, and the highest ICU mortality. The system monitors for immune reconstitution markers and generates alerts when the patient enters the immunoparalysis window — enabling prophylactic antifungal consideration, isolation intensification, and potential immunostimulatory therapy (GM-CSF, IFN-γ) in appropriate candidates.

Surviving sepsis is only the beginning. Up to 40% of sepsis survivors die within one year of discharge, and 75% experience persistent cognitive impairment, PTSD, chronic fatigue, or functional disability collectively known as Post-Sepsis Syndrome. Despite this, the overwhelming majority of hospitals provide no structured post-sepsis follow-up — patients are discharged with a prescription list and told to follow up with their primary care provider, who may have no sepsis-specific expertise. Engine 09 implements a comprehensive post-discharge monitoring framework: 30-day and 90-day readmission risk prediction, cognitive assessment scheduling based on ICU delirium duration, screening for PTSD and depression, renal function trajectory monitoring (sepsis-induced AKI progresses to CKD in 30% of patients), and cardiovascular event risk prediction (sepsis survivors have 2–3x elevated myocardial infarction and stroke risk for years post-discharge). The system generates structured follow-up care plans, triggers primary care provider alerts when risk thresholds are crossed, and provides patient-facing recovery dashboards.