MDY Analytics

See sepsis deterioration coming, while there is still time to act.

Every sepsis instrument in use reads the present. PARIS AI Sepsis AEP reads forward, giving care teams advance warning of deterioration and evidence of stability, and it is built to clinical evidence standards.

In clinical validation with NHS and university partners
Request the evidence pack The need

Sepsis kills 21.4 million people a year, close to one third of all global deaths. Management typically begins two to six hours after the criteria are met, and each hour of delay adds mortality. GBD 2021, Lancet Global Health 2025; Surviving Sepsis Campaign literature

Advance warning chart Serum lactate observed over the past day rises slowly. Its projected course from now reaches the septic shock threshold. The interval between now and that point is marked as the time to act. -24 h -12 h +12 h +24 h lactate low time to act 4 mmol/L, septic shock now
Illustration: observed serum lactate, the projected course, and the window to act before the septic shock threshold is crossed.

Sepsis is behind one death in three worldwide, and its burden does not end at discharge.

The Global Burden of Disease 2021 study puts sepsis at 166 million cases and 21.4 million deaths a year, far above the figures most sepsis literature still quotes. Survival is decided in hours, and for those who survive, readmission and late mortality follow for years.

  • 21.4M

    Deaths a year, close to one third of all global deaths

    GBD 2021, Lancet Global Health 2025
  • 166M

    Sepsis cases a year worldwide; adult incidence up 230% since 1990

    GBD 2021
  • 2 to 6 h

    Typical interval between sepsis criteria being met and management beginning

    Large cohort studies
  • >40%

    Mortality once serum lactate exceeds 4 mmol/L in contemporary cohorts

    Sepsis-3 and septic shock literature
  • $52.1bn

    United States hospital costs for sepsis stays in 2021, the costliest condition in US hospitals

    AHRQ

Survivorship: the burden that begins at discharge

Readmission within thirty days affects one in five survivors and within a year almost half. Three quarters leave with a new medical, psychological or cognitive diagnosis, and cognitive impairment persists for years.

More survivors are dead at five years than were readmitted in the first year.

Mortality and readmission after sepsis discharge Readmission rises from 20 per cent at one month to 33 per cent at three months and 44 per cent at one year. Cumulative mortality rises from 15 per cent at one month to 42.3 per cent at five years. Logarithmic time axis; points are the reported observations. 0% 20% 40% 1 month 3 months 1 year 5 years 20% 33% 44% 15% 42.3% Readmission Cumulative mortality
Mortality and readmission after discharge. Logarithmic time axis. Points are the reported observations; no values are interpolated between them. Sources: 76-study systematic review, 2025; pooled reviews.

Every instrument in use reads the present. None reads forward.

The gap is not a shortage of monitoring data but a shortage of foresight in time. Each established sepsis instrument classifies a patient's current severity or current risk. None warns before that patient's vital signs breach the thresholds that define septic shock or multi-organ dysfunction.

InstrumentWhat it doesWhat it cannot do
Sepsis-3 definition and SOFAClassifies organ dysfunction that has already occurredIdentify which patients on the trajectory toward the definition will reach it, or when
Septic shock criteriaClassifies shock once vasopressors are required and lactate remains raisedWarn before the lactate or pressure thresholds are crossed
qSOFA, NEWS2, MEWSTrusted track-and-trigger charts that fire when a vital crosses a danger lineFire before the line is crossed; they are reactive by design
Machine-learning sepsis classifiersDetect patterns earlier than charts, from a model fitted on a populationGive advance warning with a clinical lead rather than a binary label; performance varies across institutions and alert volumes are high

What instruments in use answer

How severe is this patient now?

What the field has asked for

Will this patient cross the line, and is there time to act?

The need is named, not inferred

The scale of the problem and the shape of the gap define what sepsis care is missing: foresight on the signals already watched, so that a crisis is seen coming rather than reported once it has arrived.

The interval that decides the outcome A timeline from the moment sepsis criteria are met. Management typically begins two to six hours later. Each hour of delay adds mortality of the order of several per cent. The signals that could have warned earlier were already moving. the signals move first: heart rate creeping, pressure softening, lactate rising, each unremarkable alone each hour of delay adds several per cent to mortality management typically begins here two to six hours after the criteria are met criteria met 2 h 6 h serum lactate above 4 mmol/L: mortality upward of 40 per cent
The interval that decides the outcome. Management typically begins two to six hours after criteria are met, and the signals that could have warned earlier were already moving.
  • Outcomes are governed by timeManagement typically begins two to six hours after the criteria are met, and each hour of delay adds mortality of the order of several per cent.
  • The instruments in use cannot look forwardDefinitions, track-and-trigger charts and learned classifiers all read the present state. None warns before a defining threshold is crossed.
  • The gap is named by the fieldThe 2026 Surviving Sepsis Campaign guidelines and United Kingdom national guidance identify real-time prediction and trajectory monitoring as open research priorities.
  • The value is bidirectionalThe same foresight warns the few who are getting worse and confirms stability for the many who are safe, releasing clinical attention, reducing alarm fatigue and supporting earlier intervention and better-timed discharge.

A solution must be early, like a learned model, and transparent and auditable, like a threshold chart. It must serve both directions of the clinical decision, be evidenced on what was knowable at the time, hold under every sepsis definition, and fit within existing pathways rather than displace them.

  • Surviving Sepsis Campaign, 2026 guidelines: real-time prediction and trajectory monitoring named as priority research directions.
  • NICE suspected-sepsis suite: real-time deterioration prediction identified as a research gap.
  • ADQI 28 consensus: early identification and sub-phenotyping of sepsis-associated kidney injury as central unmet needs.
  • Global Sepsis Alliance: calls for integrated approaches to recognition and antibiotic timing.
PARIS AI Sepsis AEP: Acute Events Prevention

One prediction, read in both directions

Where a patient is heading for a clinical boundary, the system warns ahead of it. Where a patient is not, it says so. One system, two clinical purposes, both validated by the same evidence.

One prediction, two clinical purposes Two courses from now. One rises and crosses the clinical boundary; the interval before it does so is the time to act, and the reading is a deterioration warning. The other stays in range, and the reading is a stability confirmation. now clinical boundary time to act Deterioration warning a course heading for the boundary Stability confirmation a course that stays in range time through the patient's stay
One prediction, two clinical purposes. The same forecast produces both readings: a course heading for the boundary is the warning; one that stays in range is the confirmation.

Deterioration warning

Ahead of the threshold

Where a patient is moving toward a threshold, the system warns before it is crossed. That warning is what allows action while there is still time to act.

  • Order investigations
  • Adjust therapy
  • Alert specialist teams
  • Prepare the treatment bundle before the threshold, not after

Stability confirmation

When no threshold is in view

Where a patient is not moving toward a threshold, the system says so. That is evidence, not the absence of an alarm, and it reaches the far larger number of patients who are safe.

  • Reduce observation frequency
  • Wean and stop support
  • Step down from intensive care
  • Confirm discharge readiness

PARIS AI Sepsis AEP is built to the requirements the need sets out: early, transparent and auditable, warning and confirmation from one system, evidenced on what was knowable at the time, and portable across institutions.

The need, in full

Eight clinical decisions, one prediction

Each clinical application module addresses one decision in the patient's stay and inherits its prediction from the core system. Warning modules act early in the stay; confirmation modules support recovery.

Admission

Septic shock: lactate trajectoryLead time to the thresholds that define septic shock.Warning
Sepsis bundle timingPreparation of the time-critical treatment bundle ahead of its triggers.Warning

Acute phase

Multi-organ dysfunctionOrgan dysfunction forecast against published scoring and guideline triggers.Warning
Sepsis-associated acute kidney injuryEarly identification and sub-phenotyping against the international staging criteria.Warning

Ongoing care

Infection surveillanceDeveloping hospital-acquired infection from routine observations. Prompts assessment and sampling; does not diagnose.Warning
Ventilator-associated pneumonia and eventsVentilator-associated events against the objective surveillance definition.Warning

Recovery

Antibiotic stewardshipBiomarker-guided de-escalation of antimicrobials. Supports stopping only.Confirmation
Vasopressor de-escalation and step-downStability evidence through weaning, discontinuation and step-down from intensive care.Confirmation

Sepsis is the beachhead, not the boundary

The same system generalises to other acute-deterioration problems and to settings beyond intensive care: emergency, ward, perioperative and longer-horizon monitoring. Auditability and standards alignment are what make it a candidate for the regulated clinical-software pathway rather than a research instrument.

PARIS AI AEP and its fourteen clinical domain applications

Six stages, one direction of travel

The development journey Six stages in sequence: data sources, cohort definition, wrangling, the engine, validation and reporting. Raw tables enter; an auditable warning and the trail to reproduce it leave. 1 Data sources Records registered through a catalogue 2 Cohort definition Definitions carried in parallel 3 Wrangling Checked and standardised 4 The engine Foresight from routine observations 5 Validation Pre-specified and reported separately 6 Reporting Auditable evidence package what enters: tables on different clocks, in different units, sampled irregularly what leaves: an auditable warning, and the trail to reproduce it
The development journey. Every downstream result can be traced back through the stages that produced it; no stage is a black box to the one after it.

Join the validation programme

Further NHS trusts, intensive care units and research institutions are welcome. The evidence pack, validation reports and commercial model are available under NDA.

Talk to us