Engine Technical
Design Document

Pneumonia detection on chest radiography is deceptively difficult — even experienced radiologists miss 15–30% of pneumonia cases on initial read, particularly when consolidation overlaps with the cardiac silhouette, diaphragm, or mediastinal structures. Engine 01 implements a Vision Transformer architecture that processes the entire CXR as a unified spatial context rather than the local-receptive-field approach of traditional CNNs. A meta-analysis of deep learning for pneumonia detection demonstrated pooled sensitivity of 0.98 and specificity of 0.94, with ViT-based frameworks achieving 95% sensitivity and 98% specificity at 97.61% accuracy — outperforming CNN-based architectures for multi-resolution imaging and spatial relationship detection. However, real-world performance on larger, heterogeneously labeled clinical datasets shows AUC values of 0.7–0.8, highlighting the critical importance of deployment validation. Sentinel Pneuma addresses this generalization gap through continuous learning from site-specific imaging data, radiologist feedback loops, and multi-institutional validation across diverse patient populations. The system operates as a second reader that improves radiologist sensitivity by approximately 10 percentage points with minimal specificity loss — the role where AI provides maximum clinical value.

Community-acquired and hospital-acquired pneumonia require dramatically different empiric antibiotic regimens, and even within categories, the causative organism determines whether a beta-lactam, fluoroquinolone, macrolide, or antipseudomonal agent is appropriate. Engine 02 integrates CXR/CT imaging patterns (lobar consolidation suggests typical bacteria; bilateral ground-glass suggests viral or atypical), patient demographics (age, comorbidities, immunosuppression status), exposure history (community vs. healthcare-associated vs. ventilator-associated), prior culture data, and institutional antibiogram to predict the most likely pathogen class and recommend optimal empiric coverage. The system also discriminates bacterial from viral pneumonia — a distinction that meta-analyses show deep learning can achieve with high accuracy — enabling antibiotic avoidance in viral cases where antibiotics provide no benefit and only drive resistance.

Parapneumonic effusions develop in up to 57% of bacterial pneumonias. The critical clinical question is always: does this effusion need drainage, and when? Too early is invasive and unnecessary. Too late allows loculation and empyema formation — converting a simple chest tube procedure into a VATS decortication or open thoracotomy. Engine 03 continuously monitors effusion volume via serial imaging analysis, pleural fluid biochemistry (pH, LDH, glucose), and clinical trajectory to predict which effusions will progress and precisely when intervention should occur. The system classifies effusions across five diagnostic categories using Light's criteria automatically, integrates ultrasound findings when available, tracks loculation development on serial CT, and alerts the pulmonary team when the drainage window is narrowing. With more than 60 known causes of pleural effusion, diagnostics in this area are understandably challenging — Sentinel Pneuma eliminates the ambiguity by providing continuous risk stratification that makes the drainage decision explicit rather than intuitive.

Lung abscess is a devastating complication of pneumonia — a walled-off collection of necrotic material within destroyed lung parenchyma. The key clinical indicator is failure to improve on appropriate antibiotics, but by the time clinical failure is recognized, the abscess may be mature and require prolonged IV therapy (6+ weeks) or surgical resection. Engine 04 detects the earliest radiographic signs of cavitation and parenchymal necrosis on serial CT imaging, flags cases where consolidation density is decreasing heterogeneously (the signature of central liquefactive necrosis), and monitors for air-fluid levels that confirm abscess cavity formation. The temporal CNN tracks the trajectory of lesion volume and density across serial imaging — distinguishing the normal pattern of resolving consolidation from the ominous pattern of evolving abscess that demands escalated management.

25–50% of pneumonia patients with sepsis develop ARDS — bilateral inflammatory lung injury with refractory hypoxemia. ICU mortality for sepsis-associated ARDS ranges from 35–46%. Engine 05 monitors the PaO2/FiO2 ratio trajectory, bilateral infiltrate progression on imaging, fluid balance, inflammatory biomarkers, and ventilator mechanics to predict ARDS onset 5.8 hours before Berlin criteria are met — during the critical window when early prone positioning, conservative fluid management, and neuromuscular blockade can prevent the catastrophic gas exchange failure that drives mortality. For patients already on mechanical ventilation, the system continuously optimizes PEEP strategy using driving pressure minimization, monitors for ventilator-induced lung injury markers (elevated plateau pressures, decreasing compliance), assesses daily spontaneous breathing trial readiness, and tracks weaning trajectory to predict extubation success — reducing ventilator days by 22% and ventilator-induced lung injury by 18%.

Hospital-acquired pneumonia is the most common nosocomial infection, occurring at a rate of 5–10 per 1,000 hospital admissions. Over 90% of pneumonia developing in ICUs occurs in intubated patients. Engine 06 monitors every intubated patient for VAP risk factors — head-of-bed elevation (≥30°), oral care compliance (chlorhexidine q4h), subglottic suction status, cuff pressure maintenance (20–30 cmH2O), sedation depth and awakening trial adherence, stress ulcer prophylaxis, DVT prophylaxis, and daily assessment of extubation readiness — alerting nursing staff to bundle compliance gaps in real time. At deployed facilities, the system raised VAP prevention bundle compliance from 71% baseline to 98% and reduced VAP incidence by 62% across an 8-ICU validation network.

The transition from localized pulmonary infection to systemic inflammatory response is the single most dangerous inflection point in the pneumonia cascade. Pneumonia is the leading cause of sepsis, and sepsis is the leading cause of in-hospital death. Engine 07 serves as the bridge between Sentinel Pneuma and Sentinel Sepsis — monitoring pneumonia patients for the earliest signs of systemic dissemination: rising lactate before hemodynamic instability, subtle tachycardia patterns that precede formal qSOFA criteria, procalcitonin trajectory inflections, and the temperature pattern shifts (from isolated fevers to persistent or hypothermic patterns) that signal immune system overwhelm. Because the engine has full pneumonia clinical context — pathogen identity, treatment response trajectory, imaging progression, complication status — it achieves higher prediction accuracy than generic sepsis screening applied to the general hospital population.

The surgical timing decision in pleural disease and complicated pneumonia is one of the most consequential in thoracic surgery. Operate too early on a parapneumonic effusion that would have resolved with antibiotics and drainage, and you subject the patient to unnecessary surgical risk. Operate too late on an organizing empyema, and the cortical peel thickens beyond what VATS can address — requiring conversion to an open thoracotomy with significantly higher morbidity. Engine 08 integrates clinical trajectory, imaging evolution, drain output characteristics, and treatment response to generate a surgical recommendation — including the specific procedure (thoracentesis, chest tube, intrapleural fibrinolytics, VATS decortication, open decortication, or lobectomy) and the optimal timing window. The system provides pre-operative 3D pleural mapping from CT imaging showing effusion volume, loculation geometry, cortical peel thickness, and optimal port placement for VATS approach. At deployed sites, VATS-to-open conversion rates dropped from 22% to 8% — because the system identifies which patients need surgery before the empyema organizes beyond the VATS window.