AyGLOO – Prescriptive Decision AI for Healthcare

The following use cases are an illustrative example per sector. Each snapshot shows what the agent evaluates before acting, what it executes autonomously when it can do so with guarantees, and what the analyst receives when it cannot.

Early sepsis warning

Early-warning models score patients at risk of sepsis. The problem is that 70–80% of alerts are ignored because clinicians receive a number with no context: they do not know which signals are driving risk, how fast it is evolving, or how much time they have to act. The result is alert fatigue — and in sepsis, every hour of delayed intervention increases mortality by 7%. AyGLOO adds an agentic layer on top of that model: the agent evaluates alert reliability for that specific patient and acts differently depending on the detected pattern. When deterioration is clear and reliability is total, it activates a critical alert with the protocol already prepared and auto-escalates if there is no response within the defined time. When the signal is ambiguous, it intensifies monitoring without consuming clinical capacity. The clinician always validates and executes: the agent removes all work before that decision.

The agent triggers a critical alert with a calibrated protocol when reliability is total — and auto-escalates if there is no response within the defined time. The clinician does not receive a number: they receive a pre-structured decision ready to validate.→ Alert response rate: from 20–30% to 80%+ when the alert includes context

ISA turns abstract probability into a concrete clinical reference: of the last 8 patients with this profile, 6 went into shock when intervention was delayed more than 2 hours.→ 30-day mortality: 34% without intervention in <2h for this segment

The agent has three operating modes — critical alert, intensive monitoring, active surveillance — depending on the deterioration pattern. It does not consume clinical capacity when the signal is weak, nor under-alert when risk is high.→ Escalated false positives reduced without losing sensitivity on critical signals

Simulations quantify the impact of each intervention before acting and the plan B if the first does not work — the clinician decides in seconds under pressure, not from scratch.→ Time to antibiotics: −45 min on average when the protocol arrives pre-structured

Today: what the clinician receives

Patient: Bed 4B · Post-op · 67 years Sepsis risk: 0.89 · HIGH Alert generated 14:23 · No additional context

With AyGLOO. Same alert, fully enriched

Bed 4B · 67 years · Day 2 post-laparotomy Sepsis risk: 0.89 · Window: <90 min The agent triggers a critical alert

XAI · Twin Alert rules with dominant signals ISA Clinical segment context CF · What-if Intervention & follow-up simulations Impact Clinical & economic impact by action Action Agent decision PDF Clinical traceability

1. Alert rules with dominant signals. Twin model (depth 7 · fidelity 95.9%)

100%IF lactate > 2.5 AND upward_trend_4h = true AND days_since_surgery ≤ 3 → Imminent sepsis: act within <90 minutes. Lactate 1.4 → 2.1 → 3.1 mmol/L in 4h (importance: 0.38), HR >100 bpm sustained 3h (0.29), WBC decreasing (0.22). The agent triggers a critical alert. The clinician validates and executes.

82%IF HR_100bpm_sustained_3h = true AND MAP < 70 AND WBC_downward_trend = true → Active haemodynamic deterioration. The agent triggers intensive monitoring with a lower re-alert threshold.

61%IF lactate > 2.0 with no other factors → Weak signal in isolation. The agent keeps active surveillance and re-evaluates in 2h without escalating.

Lactate trend is the dominant signal: a progressive rise in 4h in a day-2 post-op patient is the pattern most consistent with early sepsis. It is not an isolated high value; it is a direction. The agent responds differently depending on the pattern — not all alerts are equal.

XAITwin

2. Clinical segment context (ISA)

High-risk micro-segment: abdominal surgery, age >65, lactate >2.5 mmol/L · 30-day mortality if no action in <2h: 34% · 8 similar profiles in the last 18 months: 6 developed septic shock when intervention was delayed more than 2h · High reliability: act with confidence.

ISA is the agent’s double-check before triggering a critical alert: it confirms the model performs well in this specific segment. If the light were amber — unstable pattern or high false-positive rate for this profile — the agent would not trigger a critical alert even if the score is high. It would trigger intensive monitoring and route the clinician with full context.

ISA

3. Simulations. Impact of each intervention and follow-up by response

IV fluids (30 mL/kg) in 30 min: if lactate responds (↓ ≥10%), risk drops 0.89 → 0.52 · Blood cultures + antibiotics within 60 min: reduces 30-day mortality by 28% in this segment · If lactate does not decrease after fluids: risk remains 0.89 — escalate to vasopressors immediately, plan B already prepared · If HR drops below 100 bpm in 2h: risk drops to 0.61 · Lactate response to fluids is the most determinant follow-up signal

Simulations do two things: CF quantifies the impact of each intervention before acting, including plan B if lactate does not respond. What-if guides real-time follow-up during the first hours — the clinician knows exactly which signal to watch and what threshold triggers the next action.

CFWhat-if

4. Clinical impact by time-to-intervention

Act within <90 min ✓

6–9%
antibiotics + fluids in window

3–5 days
resolved without shock

Delay 2–4h

18–24%
+7% mortality per hour

8–12 days
septic shock risk

No intervention

34%
abdominal surgery profile >65

>14 days or death
shock in 6/8 similar cases

→ Each hour of delay: +7% mortality · €3,000–€5,000 additional ICU cost per day · The 90-min window is the KPI managed by the agent

Unlike other sectors, here economic impact is a consequence of clinical impact, not the other way around. The table does not optimise cost: it optimises time-to-intervention, the KPI that determines outcome. ICU cost connects clinical outcome to hospital management.

Impact

5. Agent decision

The agent triggers a critical alert — clinician validates and executes

Actively notifies the responsible clinician with the protocol already prepared · If there is no response in 15 minutes, auto-escalates to the on-call supervisor · Step-by-step protocol: [1] Blood cultures ×2 + procalcitonin + CRP (now) [2] Fluids 500 mL crystalloids <30 min, re-check lactate [3] Broad-spectrum empiric antibiotics <60 min [4] Notify ICU if lactate does not respond or MAP <65 after fluids · Clinical traceability exportable for MDR and EU AI Act

or, if reliability is medium or ISA detects high false-positive rate in this profile (82%)

→ The agent triggers intensive monitoring: increases vital-sign review frequency · defines the exact threshold that would trigger the critical alert · prepares the protocol on standby · the clinician receives a watch notice, not a maximum-urgency alert

or, if the signal is weak and isolated (61%)

→ The agent keeps active surveillance: monitors without escalating · re-evaluates in 2h · if the pattern consolidates, it triggers the appropriate alert automatically · it does not consume clinical capacity while the signal does not justify it

The agent does not have a single alert mode: it has three calibrated responses based on deterioration pattern. That solves alert fatigue without losing sensitivity where it matters. The agent does not replace clinical decision-making — prescription remains the clinician’s responsibility — but it removes all the work before that decision under pressure.

⚠ Antibiotic selection is the clinician’s responsibility, subject to local antibiograms, patient allergies, and formulary. This output does not constitute a prescribing recommendation. SaMD classification under MDR Annex VIII Rule 11.

ActionPDF

Illustrative example. Each deployment is adapted to each institution’s models, data, and operating procedures.

Estimated impact · wards with active early-warning

−7%

Mortality per hour of earlier detection

Surviving Sepsis Campaign reference · direct impact of shortening time to antibiotics

80%+

Alert response rate with context

Vs 20–30% with score-only alerts · less fatigue without losing sensitivity

−3 days

Average ICU stay per case detected in-window

€3,000–€5,000 avoided cost per day · direct impact on capacity and unit efficiency