Hospitals are adopting "micro-segmentation": the network is split into many small zones, and a policy engine can cut off any device that starts to look compromised. The most principled engines keep a running trust score for each device and isolate it when the score falls too low. In an intensive care unit this rule can kill someone. Cutting off a compromised smart plug costs nothing. Cutting off the arterial-line monitor or vasopressor pump of a patient in septic shock, because an intrusion alarm fired falsely, may cost more than the attack it was meant to stop. We propose a policy engine that weighs two quantities before isolating a bedside device: the risk that the device is spreading an attack, and the clinical harm of losing it for this particular patient right now. The harm estimate comes from the same kind of early-warning model that ICUs already run on streaming vital signs.
The second idea concerns evaluation, and it comes from an unexpected place: the way experimental sepsis research induces disease. Sepsis researchers do not test a therapy against a single insult. They grade the severity of the insult (needle gauge in cecal ligation and puncture), compare a single toxin bolus with a continuously leaking infection source, run sham controls, cut candidate transmission routes (vagotomy) to learn which route carries damage between organs, and ask whether a treatment still works when given after onset rather than before. Micro-segmentation papers mostly report containment in one scenario with a policy that is already in place. We apply the sepsis researchers' protocol to network defense and report dose-response curves, channel-ablation results, and rescue-window curves, with harm measured in clinical terms.
Core Claim
For bedside devices in critical care, the best isolation threshold is not fixed. It should fall when the patient's dependence on the device is low and rise when dependence is high, as in a septic patient deteriorating toward shock. We claim that an acuity-conditioned policy engine keeps most of the containment of an aggressive trust-score engine while avoiding most of its self-inflicted outages, and that this advantage appears only under false-positive rates as high as those real detectors produce. We also claim that rescue-window curves, which plot containment against how late the isolation is applied, give a more honest account of real-world deployability than containment under pre-emptive policies. These curves are the network analogue of the open pre-treatment versus rescue question in sepsis pharmacology.
The decision rule is simple to state. Device $d$ is isolated at time $t$ when
$$p_d(t)\,C_d > \big(1-p_d(t)\big)\,H_d(t),$$
where $p_d(t)$ is the posterior probability that $d$ is compromised (from a Bayesian trust score), $C_d$ is the expected downstream damage if a compromised $d$ keeps its access, and $H_d(t)$ is the expected clinical harm of losing $d$'s function for the isolation period, given the patient's current state. The new ingredient is $H_d(t)$. We estimate it from a sepsis and deterioration early-warning model and a device-to-function map: which vital signs or therapies go dark if $d$ is isolated, and how much the patient's predicted trajectory depends on them.
Relation to Prior Work
Three lines of work motivate the gap. Zero Trust for Cyber Resilience models trust as a Bayesian score updated from intrusion alerts and solves an attacker-defender Markov game. It shows that micro-segmentation confines a lateral attacker to its entry segment on the Tennessee Eastman chemical plant and on 5G authentication graphs. It also warns that false-positive alerts drive benign trust scores down and cause self-inflicted denial of service. We reuse its trust score and its $(T,D)$-resilience metric, which asks whether the system recovers within time $T$ with performance loss at most $D$. Our changes are to make the isolation cost patient-dependent and to define $D$ in clinical units rather than process-control units. We do not attempt the full equilibrium computation, which that paper itself flags as intractable at scale.
Zero-Trust Foundation Models proposes learned behavioral trust scores that drive edge micro-segmentation. It reports more than 90% precision for LSTM anomaly detectors and 1.5-second isolation of rogue nodes. It treats every node as equally expendable, which is the assumption we remove. Enhancing Patient-Centric Healthcare through Secure and Resilient Intelligent Cyber-Physical Systems argues for zero trust and micro-segmentation around connected ventilators and infusion pumps, and it notes that disconnection can itself cause harm. It is a position paper with no experiments. We build the experiment it gestures at.
Two works constrain the design. Protected or Porous audited eight commercial IoT safeguards on 79 real devices. The best detected about 35% of threats, only three could identify even one device without manual help, alert latency ranged from 30 seconds to 3 minutes, and some added more than 60% traffic overhead. Any realistic evaluation must therefore assume noisy, slow, and incomplete detection. We sweep false-positive rate and detection latency over ranges that include these values instead of assuming a clean detector. Yet Another ICU Benchmark supplies harmonized sepsis-onset and mortality tasks across MIMIC-III/IV, eICU, HiRID and AUMCdb, with GRU, Transformer and gradient-boosting models. It shows that cohort and preprocessing choices matter more than architecture, and that sepsis prevalence is about 1% of hourly slices. We use it as the source of both the harm estimator and the downstream clinical metric. Unlike that benchmark, we do not assume the incoming data arrive intact.
The sepsis induction literature supplies the evaluation method, not the biology. Graded cecal puncture and the contrast between a single lipopolysaccharide bolus and a continuous leak, as in the CLP work of Ferrostatin-1 improves acute sepsis-induced cardiomyopathy, become graded attacker capability and burst-versus-persistent compromise. The route-cutting logic of The vagus nerve-dependent lung-brain axis becomes channel ablation. That study cut the cervical or subdiaphragmatic vagus and found that brain injury depended on one route even though local lung injury was unchanged. The Ferrostatin-1 study rescued about 70% 48-hour CLP mortality only with pre-treatment. That result becomes our rescue-window question: does isolation still help if it arrives minutes after compromise? We do not claim that networks and organs share dynamics. We borrow the discipline of the experimental design, and no animal work is proposed.
Research Program
We will build an emulated ICU network in a software-defined network emulator. Each simulated bed has a monitor, an infusion pump and a ventilator endpoint. These feed a central gateway, an early-warning server running the sepsis model, and an electronic health record (EHR) store, with realistic allowed flows between them. Patient streams are replayed from YAIB cohorts, so every bed carries a real stay's vitals, labs and therapy timeline, including vasopressor and ventilation periods.
The harm estimator $H_d(t)$ combines three things. The first is a YAIB-trained deterioration and sepsis model. The second is a sensitivity analysis: how much the model's prediction and uncertainty change when the channels served by $d$ are removed. The third is a therapy-dependence term that is large when the patient is actively receiving what $d$ delivers. $C_d$ is computed from reachability in the segment graph.
Attacks come in two insult types. Burst attacks inject spoofed vitals or pump commands once. Persistent attacks come from a compromised node that probes and spreads laterally at low rate. Each type is graded by attacker capability: the number of exploitable services and the credential reach. The attack categories follow the threat classes audited in Protected or Porous. Sham runs replay identical traffic with no attack, so that false-positive lockouts can be measured on their own. Detector quality is a controlled variable, with false-positive rate and alert latency swept over ranges that include the audited commercial values and the 1.5-second edge figure.
Baselines
The naive reference is a flat network with no dynamic isolation. It sets the ceiling on attack spread and the floor on self-inflicted outage.
The strongest practical competitor is static micro-segmentation: medical devices sit in their own VLAN or role-based zones, as in Securing Public Cloud Networks with Efficient Role-based Micro-Segmentation, with allowlisted flows and no automatic isolation. It is strong because it never locks out a benign pump.
The strongest dynamic competitor is the Bayesian trust-score engine of Zero Trust for Cyber Resilience, which uses a single isolation threshold for all devices. We give it a moving-horizon policy, as its authors suggest for scalability. For this engine, and for the network side generally, no reported result was found on Lacuna for containment or isolation latency in healthcare telemetry. That paper reports no numeric containment figure for the Tennessee Eastman or 5G experiments. The only related numbers come from other papers: more than 90% anomaly precision and 1.5-second isolation reported in Zero-Trust Foundation Models, and 35% detection with 30-second to 3-minute latency in Protected or Porous. On the clinical side, YAIB reports mortality AUROC of 0.84 to 0.87 across datasets. No reported result was found on Lacuna for YAIB's sepsis-onset score, so we will reproduce YAIB's sepsis baseline on clean replay and treat it as the reference that every defense tries to preserve.
All dynamic engines receive the same detector outputs, the same attack and sham seeds, the same replayed stays, and the same budget for threshold tuning on a held-out development split. Thresholds for the uniform engine are tuned to the same false-lockout rate on sham runs before containment is compared, so that no method wins simply by being more trigger-happy.
The key ablation replaces $H_d(t)$ with a device-class constant: pumps and ventilators get high cost, everything else low cost. This isolates the value of patient-specific, time-varying acuity from the value of merely knowing device types.
The proposal beats the uniform Bayesian engine if, at matched containment, it has significantly fewer acuity-weighted lost monitoring hours and a smaller drop in sepsis-alarm sensitivity. It beats static segmentation if it contains attacks within the medical zone that static segmentation cannot.
Why not just…?
Why not just use static VLAN segmentation of medical devices and never isolate anything automatically? This is a real threat, because it avoids false lockouts by construction. Its weakness is that it cannot act on an attack that travels over allowed flows. A compromised monitor that sends plausible spoofed vitals to the early-warning server uses exactly the flow the allowlist permits. The Tennessee Eastman results in Zero Trust for Cyber Resilience needed dynamic segmentation to protect core assets once the attacker was inside. Our persistent-insult experiments measure how often static zoning lets corrupted data reach the sepsis model. If that rate is negligible at realistic attacker capability, we will report that dynamic isolation is not worth its risk in this setting.
Why not just exempt life-critical device classes from isolation? This is the device-class ablation above, and it may come close to the full method. Our reason to expect a gap is that dependence changes within a stay. A pump that is idle for a stable patient can be cut off safely, and exempting it leaves a durable foothold. Meanwhile the same class of device becomes untouchable once vasopressors start. If the ablation matches the full method on both harm and containment, we will conclude that acuity conditioning adds nothing and recommend the simpler rule.
Why not just make the sepsis model robust to corrupted inputs instead of changing the network? Input plausibility checks and robust training are complementary, and we include them as an additional layer on every configuration. They cannot stop pump-command tampering or lateral spread to the EHR store. They also miss spoofing that stays physiologically plausible, which is exactly what a capable attacker who replays another patient's stream produces. The burst-spoof experiments measure how much of the clinical-metric damage input checks alone recover.
Proposed Experiments
We run the replay testbed on MIMIC-IV for development and hold out eICU and HiRID for testing. Holding out whole databases forces the harm estimator to transfer across hospitals, which YAIB showed is where single-center models degrade.
The first decisive check is the false-positive sweep. We expect the acuity-conditioned engine and the uniform engine to be indistinguishable when detectors are clean. The proposal's advantage should appear, and grow, as the false-positive rate moves toward the levels implied by the commercial audit. If no such crossover appears, the core claim fails.
The second check is the dose-response curve. We plot containment and acuity-weighted harm against graded attacker capability for burst and persistent insults, and we expect static segmentation to hold up well against bursts and degrade against persistent lateral movement.
The third check is channel ablation. We cut candidate inter-segment flows one at a time: monitor to early-warning server, gateway to EHR, pump management. This identifies which channel carries clinical harm, as distinct from which carries data exfiltration. By analogy with the vagotomy result, we expect these to differ.
The fourth check is the rescue-window curve. We delay isolation by fixed amounts after the true compromise time, from zero through the 1.5-second edge figure to the multi-minute commercial latencies. We report where containment collapses, giving each policy an operational point of no return.
Across all runs, the outcome metrics are:
- the fraction of runs in which the attacker reaches the early-warning server, pump control or EHR store;
- $(T,D)$-resilience, with $D$ defined as the drop in sepsis-onset AUPRC on replayed stays together with missed-alarm hours weighted by acuity;
- self-inflicted outage minutes on sham runs.
References
- Zero Trust for Cyber Resilience
- Zero-Trust Foundation Models: A New Paradigm for Secure and Collaborative Artificial Intelligence for Internet of Things
- Enhancing Patient-Centric Healthcare through Secure and Resilient Intelligent Cyber-Physical Systems and the Internet of Things
- Protected or Porous: A Comparative Analysis of Threat Detection Capability of IoT Safeguards
- Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML
- Ferrostatin-1 improves acute sepsis-induced cardiomyopathy via inhibiting neutrophil infiltration through impaired chemokine axis
- The vagus nerve-dependent lung-brain axis mediates brain demyelination following acute lung injury
- Securing Public Cloud Networks with Efficient Role-based Micro-Segmentation