QIS Protocol · Critical Care Intelligence Series
A 67-year-old arrives in your ICU. Suspected sepsis. Lactate 4.2 mmol/L. Heart rate 118. Blood pressure 88/60. History of recent community-acquired pneumonia. You have the Sepsis-3 criteria (Singer et al., JAMA, 2016). You have your hospital's bundle protocol. You have your attending's clinical experience.
What you do not have: the outcomes from the 400 other ICUs in the United States that have seen a clinically identical presentation in the last 30 days. Not their patient records. Not their charts. Just the answer to the question you actually need answered: for a patient like this, what worked?
That question has an answer. It's distributed across hundreds of hospitals, synthesized by nobody, available to you at the moment of decision: never.
The Numbers Are Clear and the Cause Is Not What You Think
Sepsis is the leading cause of death in U.S. intensive care units. The CDC estimates 1.7 million Americans develop sepsis each year. 270,000 die — more than prostate cancer and breast cancer combined (Rhee et al., JAMA, 2017).
Here is the number that should end the conversation about whether this is a funding problem:
Sepsis mortality varies 15 to 30 percent between hospitals with comparable patient populations and comparable resources (Seymour et al., NEJM, 2017). Hospitals with identical case mixes, similar staffing ratios, and access to the same published protocols routinely differ by 15 percentage points in 28-day survival.
That gap is not explained by equipment. It is not explained by physician training. It is not explained by resource allocation.
It is explained by whether or not a hospital has internalized the specific protocols, timing adjustments, and antibiotic selection patterns that have been empirically validated at other hospitals — and how quickly that empirical knowledge propagates.
The propagation mechanism today is: published papers. Clinical guidelines updated every three to five years. Conferences. Word of mouth between physicians. Peer-reviewed journals with 18-month publication lags.
Meanwhile: every hour of delay in antibiotic administration in septic shock increases mortality by approximately 7 percent (Kumar et al., Critical Care Medicine, 2006). The protocol gap is measured in percentage points. The time sensitivity is measured in hours.
The mismatch is architectural.
What the Data Looks Like and Where It Stays
Every ICU in the United States is generating outcome data continuously. For every sepsis case:
- Presenting labs (lactate, procalcitonin, WBC, creatinine)
- Antibiotic selection and timing
- Fluid resuscitation approach and volume
- Vasopressor initiation timing and agent selection
- Ventilator initiation criteria if applicable
- 28-day survival
- ICU length of stay
- Readmission at 30 days
This is structured, timestamped, domain-tagged outcome data. Millions of data points generated annually across thousands of ICUs.
None of it synthesizes across institutions in real time.
Each hospital analyzes its own outcomes in quarterly or annual quality improvement reviews. Some large academic medical centers have internal analytics teams. Some participate in multi-site registries like STOP-IT or the SCCM's ARDS Network.
But participation in registries is voluntary and slow. Data sharing is legally constrained. Raw patient data cannot cross institutional boundaries without IRB protocols, data use agreements, HIPAA authorizations, and months of procurement. The constraint is permanent — sovereignty over patient data is not an obstacle to route around, it is a structural feature of healthcare.
And federated learning doesn't cleanly solve this either. Federated learning requires enough local data per site to generate meaningful gradients — a rural ICU admitting 12 sepsis patients a month cannot contribute meaningfully. Training rounds operate on schedules, not in real time. And federated learning still requires a central aggregation server that becomes a bottleneck when every site is updating simultaneously during a surge.
The result: 5,000 ICUs, each learning in isolation, each generating outcomes that are invisible to every other ICU seeing the same patient profile tomorrow.
The Architecture That Closes the Loop
Christopher Thomas Trevethan discovered a protocol that routes pre-distilled outcome intelligence without moving raw data — Quadratic Intelligence Swarm (QIS), covered by 39 provisional patents. The core of it is simple enough to state in two sentences:
Rather than centralizing raw data for synthesis, each node distills its local outcomes into a small, compressed outcome packet (~512 bytes) and deposits it to an address determined by the semantic fingerprint of the problem it solved. Any node facing a semantically similar problem can query that address and pull outcome packets from every node that has deposited there — then synthesize locally, without the packets ever touching each other.
Applied to sepsis:
Step 1 — Fingerprint the presentation. A 67-year-old with lactate 4.2, suspected pneumonia-source sepsis, and no recent antibiotic exposure generates a semantic fingerprint. The fingerprint is defined by domain experts — in this case, intensivists and infectious disease specialists who specify which clinical features actually predict outcome similarity. This is what the QIS architecture calls the First Election: not a vote, but the act of getting the right domain expert to define similarity for your network. An ICU physician defines it for sepsis. An oncologist defines it for oncology. Get the best person for each domain.
Step 2 — Query outcome packets. The fingerprint maps to a deterministic address. The physician (or clinical decision support system) queries that address. Back come compressed outcome packets from every ICU in the network that deposited results for a semantically similar patient presentation — in the last 7 days, last 30 days, whatever time window the network specifies. No raw data. No patient records. Just: what happened, and what worked.
Step 3 — Synthesize locally. The synthesis runs on the clinical workstation. In milliseconds. The result: for a patient like yours, across 340 ICUs that contributed outcome data for similar presentations last month, the bundle using piperacillin-tazobactam plus early vasopressor initiation at MAP < 65 despite adequate fluid resuscitation showed a 28-day survival rate of 89%. The bundle delaying vasopressors pending additional fluid showed 74%. The difference was driven by 23 ICUs in cold-climate regions where gram-negative pneumonia bacteremia had higher prevalence in December–February.
No ICU's data left any server. No PHI crossed any institutional boundary. The synthesis reflects 340 ICUs' worth of outcomes — and it happened before the attending walked back into the patient's room.
The Math Behind Why This Is Not Incremental
The discovery at the core of QIS is that when you route pre-distilled outcome packets by semantic similarity, the number of unique synthesis opportunities scales as N(N-1)/2, where N is the number of participating nodes.
- 10 ICUs: 45 synthesis pairs
- 100 ICUs: 4,950 synthesis pairs
- 1,000 ICUs: 499,500 synthesis pairs
- 5,000 ICUs (the U.S. network): approximately 12.5 million synthesis pairs
Each synthesis pair is a potential learning relationship — a channel through which one ICU's hard-won experience can inform another's decision in real time. Today, the number of active synthesis pairs in U.S. intensive care is effectively zero. Every ICU is an island.
The quadratic scaling comes from the complete loop — the architecture of depositing distilled outcomes to semantic addresses and querying across them — not from any single transport mechanism. The same loop works over DHT-based routing (O(log N) cost, fully decentralized), vector similarity databases (O(1) lookup with approximate nearest neighbor), REST APIs, or message queues. The routing mechanism is a choice; the architecture is the discovery.
A Runnable Implementation
import hashlib
import json
from datetime import datetime
# Semantic fingerprint for a sepsis presentation
def fingerprint_sepsis_case(lactate, suspected_source, prior_antibiotics_30d, age_decade):
"""
Domain experts (intensivists) define which features predict outcome similarity.
Fingerprint is deterministic — same presentation always routes to same address.
"""
feature_string = f"sepsis|lactate:{round(lactate, 1)}|source:{suspected_source}|abx_naive:{not prior_antibiotics_30d}|age_decade:{age_decade}"
return hashlib.sha256(feature_string.encode()).hexdigest()[:32]
# Local ICU distills outcome into a packet — no raw data
def create_outcome_packet(case_id_hash, antibiotic_bundle, vasopressor_timing_hrs,
fluid_volume_ml, survival_28d, icu_los_days, icu_region):
return {
"packet_id": hashlib.sha256(f"{case_id_hash}{datetime.utcnow().isoformat()}".encode()).hexdigest()[:16],
"timestamp": datetime.utcnow().isoformat(),
"bundle": antibiotic_bundle,
"vasopressor_hrs": vasopressor_timing_hrs,
"fluid_ml": fluid_volume_ml,
"survival_28d": survival_28d,
"icu_los": icu_los_days,
"region": icu_region
# No patient ID, no demographics, no raw labs
}
# Synthesize received packets — runs locally on the clinical workstation
def synthesize_outcomes(packets):
if not packets:
return {"status": "insufficient_data", "count": 0}
by_bundle = {}
for p in packets:
bundle = p["bundle"]
if bundle not in by_bundle:
by_bundle[bundle] = {"survival": [], "los": [], "vasopressor_hrs": []}
by_bundle[bundle]["survival"].append(p["survival_28d"])
by_bundle[bundle]["los"].append(p["icu_los"])
by_bundle[bundle]["vasopressor_hrs"].append(p["vasopressor_hrs"])
results = []
for bundle, data in by_bundle.items():
n = len(data["survival"])
survival_rate = sum(data["survival"]) / n
avg_los = sum(data["los"]) / n
results.append({
"bundle": bundle,
"n": n,
"survival_28d": f"{survival_rate:.1%}",
"avg_icu_los_days": round(avg_los, 1)
})
results.sort(key=lambda x: float(x["survival_28d"].strip('%')), reverse=True)
return {"count": len(packets), "by_bundle": results}
# --- Demonstration ---
# Patient arrives: 67yo, lactate 4.2, suspected pneumonia-source sepsis, antibiotic-naive
address = fingerprint_sepsis_case(
lactate=4.2, suspected_source="pneumonia", prior_antibiotics_30d=False, age_decade=6
)
print(f"Routing address: {address}")
# → always the same for this presentation, deterministic
# Simulating packets pulled from network (these came from 4 ICUs, no raw data)
received_packets = [
create_outcome_packet("anon1", "pip-tazo+vanco+early-vaso", 1.5, 2800, True, 4, "northeast"),
create_outcome_packet("anon2", "pip-tazo+vanco+early-vaso", 2.0, 3200, True, 5, "midwest"),
create_outcome_packet("anon3", "pip-tazo+vanco+early-vaso", 1.0, 2500, True, 3, "south"),
create_outcome_packet("anon4", "pip-tazo+delayed-vaso", 4.5, 4100, False, 9, "midwest"),
create_outcome_packet("anon5", "pip-tazo+delayed-vaso", 6.0, 4800, False, 11, "south"),
]
synthesis = synthesize_outcomes(received_packets)
print(json.dumps(synthesis, indent=2))
Output:
{
"count": 5,
"by_bundle": [
{
"bundle": "pip-tazo+vanco+early-vaso",
"n": 3,
"survival_28d": "100.0%",
"avg_icu_los_days": 4.0
},
{
"bundle": "pip-tazo+delayed-vaso",
"n": 2,
"survival_28d": "0.0%",
"avg_icu_los_days": 10.0
}
]
}
This runs in milliseconds on any clinical workstation. The network contribution from each ICU is a packet smaller than a single lab result. Privacy is not a configuration option — it is an architectural property. There is no mechanism to extract patient data from an outcome packet because patient data is never put into one.
Why Rural ICUs Are Not Left Out
One of the structural failures of federated learning is the N=1 problem: a rural ICU admitting 12 sepsis patients per month cannot generate a gradient with statistical power. It cannot meaningfully participate in federated training rounds. Its local outcomes are too sparse.
QIS has no minimum site threshold. A rural ICU admitting its third sepsis patient this month can deposit an outcome packet. It cannot synthesize much locally — three packets are not statistically powerful. But it can receive synthesis from the 600 other ICUs that contributed. And its three packets contribute to the synthesis that a high-volume urban ICU will receive next week.
The math: N(N-1)/2 synthesis paths. Every additional site adds N new paths. A rural ICU with 12 cases/year adds hundreds of synthesis relationships to the network simply by participating. Its rarity makes it a more valuable contributor for the specific patient profiles it sees — the presentations that the high-volume urban centers see rarely.
That is not an edge case accommodation. It is the architecture working as designed.
What Happens When 5,000 U.S. ICUs Are on the Network
Current state: 0 real-time synthesis paths across U.S. intensive care.
QIS network at full penetration: N(N-1)/2 ≈ 12.5 million active synthesis paths.
The protocol moment is not a technology upgrade. It is a phase transition. The network at 5,000 nodes does not do what the network at 50 nodes does, only faster. It does something qualitatively different — because the semantic fingerprinting begins to resolve rare presentation profiles that appear in aggregate even when any individual site sees them once a year.
Sepsis is not one disease. It is thousands of presentations — organism, source, patient comorbidity, antibiotic history, timing — each requiring its own evidence base. A centralized database could store this. It cannot synthesize it in real time across sovereign institutions without violating every privacy constraint that governs healthcare data.
QIS does both.
The Same Architecture, Every Domain
Sepsis is one problem. The architecture is universal.
Christopher Thomas Trevethan's discovery is not a sepsis tool. It is a protocol for how intelligence can flow across any network of sovereign, privacy-constrained nodes — by routing pre-distilled outcome packets to semantic addresses rather than centralizing raw data. The 39 provisional patents cover the complete architecture: the loop, the semantic fingerprinting, the outcome packets, the local synthesis. Not any specific routing implementation.
The humanitarian licensing structure ensures that what works in the U.S. ICU network reaches the ICU in Lagos without a procurement department. Outcome packets are small enough for SMS. The synthesis runs on a phone. The protocol does not care about infrastructure.
270,000 deaths per year. 15–30% inter-hospital mortality variance. The gap has a name now.
QIS (Quadratic Intelligence Swarm) was discovered by Christopher Thomas Trevethan. 39 provisional patents filed. Protocol documentation: qisprotocol.com. Technical series: dev.to/roryqis.
Top comments (0)