Pipeline / 04 · Infrastructure

GEDM-ICU

A research-grade, multimodal, trajectory-based intelligence layer — deployed first at the University of Antwerp's ferroptosis-oriented critical care program.

HOW GEDM-ICU REASONS Six stages, one governed loop. 01 Perceive Multimodal intake, normalization, vectorization 02 Harmonize Baseline, deviation, velocity, acceleration 03 Remember Patient-specific state + trajectory history 04 Reason Monte Carlo, Bayesian posterior 05 Validate Stability, calibration, confidence thresholds 06 Explain Clinician-facing rationale, audit trail Closed loop: clinician feedback + observed outcomes continuously refine the engine.
The six-stage reasoning loop — perceive, harmonize, remember, reason, validate, explain. Every output traces through this sequence; the order itself is the scientific guarantee. Closed loop: clinician feedback and observed outcomes refine the engine.
Premise

The problem we are solving.

The intensive care unit is where decision pressure is highest and information density is most extreme. Clinicians integrate dozens of data streams — vitals, labs, imaging, pharmacology, intuition — under acute time pressure. Most clinical AI today gives a binary answer; ICU medicine demands a continuous one.

Ferroptosis — a regulated form of iron-dependent cell death — is a particularly hard problem. It cannot be diagnosed from a single biomarker. The recent GPX4 fin-loop paper (Lorenz et al., Cell 2026) showed that ferroptosis vulnerability can arise even when GPX4 expression and catalytic activity look preserved. The real defect is loss of membrane anchoring. One lab value can mislead.

GEDM-ICU is built precisely for this kind of reasoning — inferring a systems state from converging multimodal signals over time, not from a single snapshot. The first deployment is a research collaboration with Prof. Dr. Tom Vanden Berghe at the University of Antwerp's Cell Death Signaling Lab.

Approach

How we tackle it.

GEDM-ICU is a research-grade reasoning layer, not a therapy recommender. Clinical decisions stay with clinicians. The system reads ICU data streams, builds a patient-specific evolving state, and reasons over candidate actions using constrained optimization, stability theory, probabilistic forecasting, and Bayesian belief updating.

Every output traces back through six stages — perceive, harmonize, remember, reason, validate, explain — and every recommendation links to its evidence, rationale, confidence, and stability check. The order itself is the scientific guarantee. Not a black box. Not an autonomous actor. A governed partner.

Three findings from the Antwerp program create an almost perfect brief for this approach: cfDNA tissue-of-origin tracking (Vanden Berghe et al.), the GPX4 fin-loop ferroptosis vulnerability paper, and longitudinal response phenotyping in neuroblastoma ferroptosis (Koeken et al., CDD 2025). All three are problems where static classification fails and trajectory-based reasoning adds measurable value.

SCIENTIFIC FIT Three Antwerp findings → one reasoning architecture. cfDNA tissue-of-origin Vanden Berghe et al. FINDING A single plasma cfDNA assay reports tissue-of-origin injury and pathogen presence — total cfDNA correlates with SOFA, hepatocyte cfDNA with ALT, cardiac with troponin, lung with PaO₂/FiO₂. GEDM FIT GEDM reasons over multi-stream integration over time. Infection vs sterile injury separation becomes a first-class reasoning output, not a post-hoc interpretation. GPX4 fin-loop Lorenz et al., Cell 2026 FINDING Ferroptosis vulnerability can arise even when GPX4 expression and catalytic activity look preserved — the real defect is loss of membrane anchoring. One lab value can mislead. GEDM FIT Ferroptosis must be inferred as a systems state from converging signals. GEDM's trajectory + cross-biomarker attention is designed precisely for this. Neuroblastoma ferroptosis Koeken et al., CDD 2025 FINDING Baseline lipidome and transcriptome did not reliably predict ferroptosis sensitivity. Response is heterogeneous and evolves over time. GEDM FIT Static classification fails here. Longitudinal response phenotyping — direction and speed of multi-signal change — is where GEDM adds measurable value.
Three findings from Prof. Vanden Berghe's Antwerp program — cfDNA tissue-of-origin, GPX4 fin-loop ferroptosis vulnerability, longitudinal response phenotyping — each map onto a specific reasoning capability of GEDM-ICU.
i.
Perceive
Multimodal intake — vitals, labs, SOFA, cfDNA tissue-of-origin, pathogen cfDNA, lipid peroxidation markers — normalized and vectorized.
ii.
Harmonize
Baseline, deviation, velocity, acceleration. Direction and speed of change, not single-point snapshots.
iii.
Remember
Patient-specific state plus trajectory history. The reasoning is over evolution, not instants.
iv.
Reason
Candidate evaluation, Monte Carlo rollout, Bayesian posterior with conformal-calibrated uncertainty bounds.
v.
Validate
Stability check, calibration, confidence thresholds. Outputs that fail the validation step are flagged, not delivered.
vi.
Explain
Clinician-facing rationale, audit trail, feedback loop. Every output linked to the signals and formulas that produced it.
Capabilities

What makes this real.

01
Trajectory-based reasoning
Tracks direction and speed of multi-signal change — not snapshots. Velocity, acceleration, and deviation from patient-specific baseline.
02
Multimodal integration
Vitals, labs, cfDNA, pathogen signals, lipid peroxidation markers, intervention metadata reasoned over together as a single evolving state.
03
Conformal uncertainty
Monte Carlo rollout with calibrated confidence bounds. Every recommendation arrives with its uncertainty, not after it.
04
Fully auditable
Every output traceable to inputs, formulas, and decisions. Designed for shadow-mode pilots and ethics-board scrutiny — not deployment-by-fiat.
Where It Connects

Part of an integrated platform.

01 · Platforms
GEDM-3DQ Decision Engine
Gradient Equilibrium Decision Modeling in 3-Dimensional Quadrant — the AI spine of every MACRO HRD program.
04 · Infrastructure
SA Trial Hospital
A fast-track clinical trial hospital platform in South Africa — accelerating first-in-human studies for therapeutics targeting high-burden African diseases.
04 · Infrastructure
Ambient Cell Logistics
Room-temperature cell logistics — a bioinspired sporulation strategy that ends cold-chain dependency for cellular therapies.
02 · Infectious Disease
HIV Functional Cure Program
An integrated therapeutic approach combining tri-specific antibodies, CAR-like NK effectors, AH-D peptide activation, and GEDM-3DQ-guided coordination.
⸻ Continue the platform

“ICU intelligence cannot be a verdict. It must be a conversation — auditable, governed, and earned over time.”

See all ventures → Engage with MACRO HRD