Lesson 1Foundations of sepsis pathophysiology and clinical criteria (SIRS, qSOFA, SOFA, Sepsis-3)We review sepsis biology and body response, linking these to signs doctors see like low blood pressure, fast heart rate, and organ issues. We compare SIRS, qSOFA, SOFA, and Sepsis-3 criteria and how they're used in emergency triage.
Host response to infection and organ dysfunctionHemodynamic changes and microcirculatory failureSIRS, qSOFA, SOFA: components and thresholdsSepsis-3 definition and septic shock criteriaLimitations of clinical scores in the EDLesson 2Safety, failure modes, and mitigation: false positives/negatives, model drift, data quality issues, adversarial inputsWe identify safety risks like false alarms, missed cases, model drift, and dodgy data quality. We look at tricky or surprise inputs, strong monitoring, safety barriers, doctor oversight, and safe ways to update models.
False positives, false negatives, and harm modesData quality checks and anomaly detectionModel drift, recalibration, and retrainingAdversarial or unexpected input handlingHuman oversight, overrides, and governanceLesson 3Evaluation metrics and validation strategies for sepsis prediction: AUROC, AUPRC, calibration, lead time, decision curve analysisWe define key measures for sepsis prediction like AUROC, AUPRC, calibration, and early warning time. We explain internal/external checks, time-based validation, and decision curve analysis to assess real clinical value.
AUROC, AUPRC, and class imbalanceCalibration curves and risk stratificationLead time and horizon-specific performanceInternal, external, and temporal validationDecision curve analysis and net benefitLesson 4Feature engineering and temporal modeling: time-series preprocessing, sliding windows, trend extractionWe explain cleaning and aligning emergency time-series data for models. We cover resampling, irregular gaps, sliding windows, trend/variability features, and coding interventions and clinical context over time.
Time alignment, resampling, and interpolationSliding windows and prediction horizonsTrend, variability, and derivative featuresEncoding interventions and care escalationHandling irregular and sparse time-seriesLesson 5Machine learning models for acute risk prediction: logistic regression, gradient boosted trees, RNNs, temporal convolutional networks, transformer-based time-seriesWe compare approaches for sudden sepsis risk prediction, from logistic regression to boosted trees and deep sequence models. We highlight strengths, limits, explainability, and fit for emergency time pressures.
Logistic regression and regularization choicesGradient boosted trees and feature importanceRecurrent neural networks for sequencesTemporal convolutional networks for time-seriesTransformers for clinical time-series dataLesson 6Data modalities for real-time sepsis detection: vitals, labs, nursing notes, medication, waveformsWe describe key live data streams in emergencies, like vital signs, lab tests, meds, nurse notes, and body waveforms. We discuss sampling speeds, reliability, and how each signals worsening sepsis.
Vital signs and continuous monitoring feedsLaboratory panels, cultures, and turnaround timesMedication orders, fluids, and vasopressorsNursing notes, triage text, and flowsheetsWaveforms from monitors and bedside devicesLesson 7Integration with ED workflows and EHR systems: event streams, FHIR, HL7, SMART on FHIR apps, CDS HooksWe describe fitting AI sepsis models into emergency workflows and patient records. We review event streams, HL7, FHIR resources, SMART apps, and CDS Hooks, stressing ease, reliability, and minimal workflow disruption.
Event-driven architectures and data streamsHL7 and FHIR resources for sepsis signalsSMART on FHIR apps for bedside decision supportCDS Hooks for context-aware recommendationsWorkflow mapping and usability testingLesson 8Clinical alert design and human factors: thresholds, alarm fatigue mitigation, escalating workflows, who receives alertsWe cover alert design for emergency doctors, including threshold picks, tiered alerts, and routing to right roles. We tackle alert fatigue, timing, escalation paths, and showing explanations/context.
Choosing thresholds and alert tiersAlarm fatigue and suppression strategiesWho receives alerts and on which channelsEscalation workflows and handoff supportExplaining alerts and providing contextLesson 9Regulatory and evidence requirements for diagnostic AI: FDA/CMS considerations, clinical validation study design, prospective pilots, reporting standards (TRIPOD, CONSORT-AI)We outline rules and proof needs for diagnostic AI in sepsis, including FDA paths, CMS views, and clinical checks. We review forward pilots and standards like TRIPOD and CONSORT-AI.
FDA pathways for diagnostic support toolsCMS, reimbursement, and quality programsDesigning robust clinical validation studiesProspective pilots and phased rolloutsTRIPOD and CONSORT-AI reporting guidanceLesson 10Deployment frequency and latency considerations: near-real-time streaming vs batch scoring, handling missing and delayed dataWe discuss model rollout setups for sepsis, comparing live streaming vs batch checks. We address speed limits, missing/delayed data, backfills, and monitoring data flows in emergencies.
Near-real-time streaming vs batch scoringLatency budgets and SLA definitionsImputation for missing and delayed inputsBackfilling, replay, and late-arriving dataMonitoring pipelines and system resilience
