ClinLoop AI | Closed-Loop Follow-up of Clinical Results

김철수 (Kim, Chul-soo)

🚨 골든타임 임박 (24일 남음)

남성 52세 • 환자번호: PT-0004 • 내원: 응급의학과 외상진료 (갈비뼈 골절 치료 후 퇴원)

추적 관리 의무 기한

⏳ 24일 남음

안전 마감일자: 2026-10-15

1. 무엇이 발견되었나요?

흉부 CT 촬영 결과, 우상엽(RUL)에 7.8mm 크기의 침상형(Spiculated) 폐 결절이 우연히 발견되었습니다. 악성 종양(조기 폐암) 가능성이 높은 고위험 소견입니다.

2. 왜 즉시 조치가 필요한가요?

응급실에서 갈비뼈 골절 치료만 완료되고, 폐 결절에 대한 외래 추적검사나 호흡기내과 진료 예약이 누락된 채 66일 동안 병원 컴퓨터 안에 방치되어 있었습니다.

3. 권고되는 임상 조치는 무엇인가요?

Fleischner Society 2017 가이드라인에 따라 6–8mm 고형 결절은 6–12개월 내 흉부 CT로 추적해야 합니다. 침상형(Spiculated) 등 고위험 형태는 더 이른 추적을 고려합니다.

합성 시나리오 · 예측값 없음

1년 이내 전이암(4기) 진행 위험

추정하지 않음

5년 기대 생존율

추정하지 않음

환자별 위험 추정치는 제공되지 않습니다.

가상 업무 흐름 상태 · 치료 효과 아님

조기 치료 완치율

추정하지 않음

5년 기대 생존율

추정하지 않음

QALY, 생존율 또는 비용 절감 효과는 입증되지 않았습니다.

📈 Hypothetical trajectory (drawing only, no data)

Hand-drawn illustration, not a model output and not clinically validated

SC-0004 • RUL 7.8mm Nodule

❌ EMR Neglect Path

✨ ClinLoop Closed-Loop

Guideline Deadline

⚡ 지금 바로 할 수 있는 두 가지 조치

버튼 하나로 환자에게 안심 연락문을 발송하거나 담당 주치의에게 추적 처방 오더를 전달하여 진료 루프를 종결합니다.

Dynamic Temporal Hypergraph (DTH)

Loading clinical trajectory...

Event Node Open Obligation Follow-up recorded (example)

Fast-Forward Clinical Time:

35 Days

Cervical Intraepithelial Neoplasia (HSIL)

Hypothetical scenario (illustration only)

A hand-written illustration. No survival or progression model runs.

5-year survival differenceNot estimated
No outcome model exists

Quality-Adjusted Life Years

Not established

No simulation runs

Legal liability

Not estimated

No data

Guideline Authority

ASCCP 2020

Citation shown for context; no evidence grade assigned

Illustrative counterfactual — Intervention: do(FollowUp=True)

Finding → rule-based review → clinician review → follow-up record

🔴 Standard Care (Neglected Obligation Path)

Hypothetical delayed pathway

🟢 ClinLoop AI (Early Closed-Loop Intervention)

Hypothetical timely follow-up pathway

📖

Peer-Reviewed Clinical Rationale: ASCCP 2020 Guidelines show prompt colposcopic excision of HSIL prevents invasive cervical cancer progression in >95% of cases (PMID: 32243307).

Global Health Informatics Standards Grounding

Cross-institutional interoperability via OHDSI OMOP Common Data Model v5.4 and HL7 FHIR R4 JSON resources.

Example OMOP mapping FHIR-shaped sample Example timing rule

● OHDSI OMOP-CDM Standard Mapping

Concept ID: 4180497

Standard Concept Name High grade squamous intraepithelial lesion of cervix on cytology

Domain ID Measurement

Vocabulary & Concept Code SNOMED (416952002)

LOINC Observation Code 10524-7

Standardized Observation Raw Record:

● HL7 FHIR R4 Interoperability Resource

Resource Type: DiagnosticReport Example JSON · active

🛡️ Metric Temporal Logic (MTL) Safety Verification

Robustness Margin: ρ(Φ, t) ≥ 0

Φ = □ ( LabFinding(HSIL) ⟹ ◇[0, 30d] ColposcopyReferral )

ρ(Φ, t) = T_crit - Δt = 30d - 35d = -5d (Violated without ClinLoop)

Deterministic rules: Each obligation is checked against an explicit rule and deadline, so the same input always gives the same result and the reason is shown. This makes errors auditable; it does not guarantee that every finding is read correctly (see the accuracy evaluation).

⚡

Privacy review example

No live de-identification or external AI transmission is configured.

● EXAMPLE

Input (stays in this browser)

PHI 항목: 0개 탐지됨

🛡️

→

Stripped: 0

Pseudonymized: 0

Masked output (demo, in this browser only)

[De-identified output will appear here in real-time as you type above]

Optional · hospital server only

🔒

Local language model (optional)

Optional: a local language model (for example EXAONE 3.5 or Llama 3, via Ollama) can run on the hospital's own server, inside its firewall, on a CPU or a GPU. Patient text then never leaves the hospital. This public page has no model connected.

MedLlama2 7B ★ Llama 3.1 Medical DeepSeek-R1 Ollama Air-Gapped

Checking...

⚡ Runs on the hospital's server · CPU or GPU

Not built

🌐

Federated Learning

Idea for multi-hospital research: share model updates instead of records. Not built; ClinLoop's rules need no training data.

🏥 Seoul

🏥 Busan

🏥 Daegu

⚙️ Gradients Only

NVIDIA FLARE PySyft Multi-hospital

📊 Design goal: patient data stays on-premise · not implemented

Not built

📊

Differential Privacy (ε-DP)

Idea for published statistics: add calibrated noise so no individual can be singled out. Not built; the slider below only illustrates the idea.

Mechanism: X̃ = X + Lap(Δf/ε) where ε=0.1, Δf=sensitivity

Exact value:42.1%(re-identifies individual)

ε-DP output:42.1% ± 2.3%(population-level only)

Privacy Budget ε: 0.1

ε = 0.1 Laplace Noise Apple / Google standard

🔢 Illustration only · no privacy guarantee

Not built

💻

Confidential Computing (TEE)

Hardware-isolated processing in a cloud (Intel SGX / AMD SEV). Not built and not needed: ClinLoop runs on the hospital's own server.

Intel SGX Azure Confidential AMD SEV

🔐 Roadmap only · not built

Not built

🧬

Synthetic data for research

The public site uses hand-written synthetic patients only. Generating realistic synthetic copies of real records is not built.

Encrypted EMR Sync CTGAN MDClone standard

🧬 Not built · no compliance certification

Demo · in this browser

🔥

Identifier screening (demo)

Try the box above: it finds identifier patterns (phone, resident number, e-mail, dates) and masks them, in your browser only. Pattern matching misses identifiers it has no pattern for; do not enter real patient information.

18 PHI Categories Korean PIPA phi_deidentifier.py

🔥 Rule-based identifier removal · not measured on real records

Not built

🔮

Homomorphic Encryption (HE)

Computing on encrypted data. A research field, far too slow for this use today; not planned.

Microsoft SEAL IBM HElib Zama.ai

🔮 Not planned

Not in use

🏦

Pseudonymised identifiers

A pattern-based de-identification module exists in the code (phi_deidentifier.py) but nothing uses it, because no patient data is sent outside the hospital. Its output would need review before being called de-identified.

SAMPLE-01→PT-99F74B58

780312-1234567→[National ID removed]

1978-03-12→1978

SHA-256 Pseudonym On-Premise Vault PCI-DSS model

🏦 Module present · not used

📋

Example activity log

Example log of actions in this browser. On a server, ClinLoop keeps an append-only audit log of every detection and clinician action.

EXAMPLE Example log started (this browser only)

Summary

In use: no patient data to external AI; role-based access; audit trail

Optional: local language model · Demo: identifier screening

Not built: 6 of the ideas above

Regulatory status

Assessment pending No regulatory submission made

Report reading (synthetic, blind)

Clinical accuracy unestablished See Validation evidence for task-specific synthetic results

Clinical outcome benefit

Not studied No patient outcome data exist yet

Hospital integration

HL7 FHIR R4 (read) Tested against test servers only; no hospital connected

Deadline status (example)

Research score · not clinically validated

Rule R003

Severity CRITICAL

How the urgency score is calculated ▾

R(t) = S · [1 / (1 + e^(-8.0 · Δt))] Illustrative curve

Mandatory Action Missing

Loading obligation...

김철수 (Kim, Chul-soo)

1. 무엇이 발견되었나요?

2. 왜 즉시 조치가 필요한가요?

3. 권고되는 임상 조치는 무엇인가요?

⚡ 지금 바로 할 수 있는 두 가지 조치

Dynamic Temporal Hypergraph (DTH)

Hypothetical scenario (illustration only)

Global Health Informatics Standards Grounding

🛡️ Metric Temporal Logic (MTL) Safety Verification

Data protection: what is built and what is not

Local language model (optional)

Federated Learning

Differential Privacy (ε-DP)

Confidential Computing (TEE)

Synthetic data for research

Identifier screening (demo)

Homomorphic Encryption (HE)

Pseudonymised identifiers

Research Action Preview

Workflow review — design only

Questions for the existing workflow

What ClinLoop demonstrates

Evidence and guidelines

Loading guideline...

Patient message preview (example)

Patient messaging is not enabled

Ethics & safety charter

Principles the software follows today

What has not been shown yet

Before any clinical use

Compute: what runs where

Hospital AI & Biomedical API Credentials Zero-Leakage Security