← back to loginAnesRecord++ V2 · planning document · no patient data

Monitor integration — how others do it, and what we can add now

2026-09-24 · AnesRecord++ V2 sandbox · research digest + wish list

1. How other facilities do it (what the research and vendors show)

The pattern everywhere is the same as ours: a box beside the bed translates the monitor, the record system never talks to the monitor directly.

Commercial hospitals (Epic, Cerner, Philips): the monitor plugs into a small device-integration hub at the bedside (Philips Capsule "Neuron": plug-and-play cable sets, a library of ~1,000 supported devices). The hub turns each vendor's private stream into HL7/FHIR and feeds the EHR. In the OR that data lands in Epic Anesthesia / OpTime. Our "OR laptop" is exactly that hub, plus the record itself.
Values are not "the record" until a clinician validates them. In Epic, device data sits in a temporary buffer (7 days) and must be reviewed/validated to become part of the permanent chart. Our confirm-first decision matches standard practice.
Artifacts are real and machines can't reliably remove them. In 88 procedures / 5,711 auto-captured BP points, 2–6 % were artifacts (flush, blood sampling, patient movement, cuff inflating on the same arm). Three machine-learning detectors reached at best moderate agreement (kappa 0.65). Conclusion of the field: keep a human in the loop, define "artifact" explicitly, keep the raw data next to the confirmed values.
Automated records help, without hurting vigilance. A randomized simulation study found equal vigilance with lower mental workload under automated record-keeping; automated records are more complete, and hand-charted values tend to be "nicer" (biased) than machine values.
Standards: today's plumbing is HL7 v2 ORU (IHE PCD "Device Enterprise Communication"); the future is IEEE 11073 SDC (Dräger, OR.NET) for device-to-device and FHIR Observation for the EHR. Vendor streams like Edan's sit upstream of all of these — the hub does the translation. When we later add an HL7/FHIR export, we speak the same language as the big systems.
Low-cost / research route: SNUH's Vital Recorder (free; Windows, Linux, Raspberry Pi; 98.5 % automatic-record success in the paper) and VSCapture (open source, 14 device protocol families incl. GE S/5 and Philips MIB) already contain the drivers for most other monitor brands and anesthesia machines. OpenICE/MDPnP shows the next step: smart alarms and apps on top of integrated device data.
Wards/PACU: the same hubs feed early-warning systems (e.g. Hamilton Health Sciences connecting bedside data to the EHR and its early-warning score).

2. What we already have (verified today)

Edan iM8 → direct cable → our box → decoded every second: HR, RR, SpO2, PR, NIBP (SYS/DIA/MAP), T1/T2/ΔT, alarm text, ECG/PLETH/RESP waveforms, patient demographics (name, patient no., doctor, DOB, admit date, height, weight, bed). Live Monitor Mirror tab. Everything marked provisional.

3. Wish list — what we can integrate NOW (no new hardware)

Ordered by value ÷ effort. ✅ = data already in hand.

#	Feature	Why (from the research)	Effort
A	Confirm-first charting — one tap puts the current HR/SBP/DBP/MAP/RR/Temp/SpO2 into the current time column, marked as monitor-sourced, with an audit entry	Matches Epic's validate-before-record; keeps the human in the loop	small ✅
B	NIBP result queue — every new cuff result appears as a pending chip ("133/77 at 19:41 — accept?"); one tap charts it	NIBP is discrete and the most-charted value; least artifact-prone	small ✅
C	Demographics prefill + identity check — offer the monitor's name/DOB/sex/height/weight/bed into the record header; warn loudly if the record's patient ≠ the monitor's patient	Vendors call patient-identity matching the #1 integration safety issue	small ✅
D	Alarm log → events row — monitor alarm texts (SENSOR OFF, LEAD OFF, T1 LOW, apnea…) with start/stop times, one tap to keep as an intra-op event	Complete record; explains gaps	small ✅
E	1-second trend archive per case — store the raw decoded stream with the record (compressed), viewable later	The "immutable registry" idea; needed for artifact review, QA and research	medium ✅
F	Mark-as-artifact on any charted point, keeping raw and confirmed side by side, with an explicit artifact definition in the UI	2–6 % of BP points are artifacts; rule-based flags help, humans decide	small
G	Waveform strip capture — button saves the last 10 s of ECG/PLETH as an image into the record's photos	Documents arrhythmias/ischemia at the moment they happen	small ✅
H	No-data / disconnect banner on the record page + last-good timestamp	Downtime awareness; standard in every AIMS	tiny ✅
I	Auto case timeline hints — first sustained SpO2 = "patient on monitor"; NIBP cadence; suggests anesthesia start/end marks (never places them)	Completeness without automation risk	small ✅
J	Rule-based quality flags before confirm — value outside physiologic range, jump > x % from previous, cuff-inflation window on the same limb, probe-off	Literature's simple filters catch the obvious artifacts	small
K	PACU board — one PACU box, 10 monitors (each with its own LOCAL NET NO), per-bed tiles, tap a bed to open its record	Same hub pattern as wards; needs per-bed connection keying in `monitor_edan.py`	medium
L	Central trend forwarding — the OR/PACU box posts the 1-s trend to the VPS; QA dashboard shows live rooms	Already have the sync path and a QA dashboard	medium
M	HL7 ORU / FHIR Observation export of confirmed values	Speaks the hospital's language when they eventually want the data	medium

4. Needs hardware or a vendor step (next phase)

Anesthesia machine (Dräger Atlan A300): MEDIBUS.X over RS‑232 → Vital Recorder on the same box → EtCO2, FiO2, agents/MAC, TV, RR, Ppeak, PEEP, MV. Needs USB‑serial adapter + Dräger cable.
Other monitor brands in the hospital (Mindray ePM: native HL7 page; Philips/GE: Vital Recorder/VSCapture drivers).
ESP32 Wi‑Fi bridges for PACU beds (avoid cabling a switch) — custom firmware, needs review.
Push demographics TO the monitor (ADT) — Edan protocol is two-way but we only know the read side; write access to a clinical device needs its own safety review (Phase 7).
SDC / IEEE 11073 — when devices in the hospital support it (Dräger is shipping it); not before.

5. Deliberately NOT doing

Auto-charting straight into the legal record without confirmation.
Putting the monitor on hospital Wi‑Fi or the hospital LAN (keep the direct cable).
Machine-only artifact removal.

6. Where AI fits — and how VitalDB helps

What is in real clinical use today (all decision support, none of it charts on its own):

Hypotension prediction — Edwards Hypotension Prediction Index (HPI): AI on the arterial-pressure waveform, warns ~2–5 min (up to 15) before a hypotensive episode; sensitivity/specificity ≈ 86 % at 5 min; CE-marked, FDA-cleared, the most-studied model in anesthesia. Needs an arterial line (or their finger-cuff sensor).
Hypoxemia prediction — Prescience (UW): minute-by-minute record data from 50,000 cases; with it anesthesiologists anticipated 30 % of hypoxemia events vs 15 % without, with an explanation of why (which features drive the risk).
Deterioration / early-warning scores — Epic Deterioration Index (0–100, recomputed every 15 min from charted vitals + labs) on wards and PACU-type settings.
Depth-of-anesthesia and closed-loop drug delivery — BIS-guided propofol/remifentanil loops (still mostly research/regulated devices).
Artifact detection — machine learning reaches only moderate agreement with clinicians (kappa ≤ 0.65); used to flag, not to delete.

How VitalDB assists (this is its main purpose):

It is the open training and validation set behind most academic models above: 6,388 cases with high-resolution waveforms + outcomes. STEP-OP (5-minute hypotension warning), Lee et al. BJA 2021 (deep learning hypotension prediction), and Jeong et al. 2024 (hypotension prediction from non-invasive monitors only — NIBP/pleth/ECG) were all built on it. That last one matters for us: our Edan iM8 has no arterial line, and those models use exactly the signals we now capture.
Vital Recorder's filter engine runs algorithms in real time on the stream (its built-in JavaScript "filters": pulse-transit time, etc.) — a ready pattern for plugging a model into a live feed.
Replay for development: real VitalDB cases already drive our sandbox (load_vitaldb.py); the same files let us test any model against thousands of cases before it touches a live patient.

What we can do now (ordered):

#	AI item	Data needed	Status
N1	Rule-based smart flags (range, jump, probe-off, cuff window) — see item J	our stream ✅	LIVE on the review screen; per-minute copy in shadow
N2	Trend-based early warnings — MAP falling x mmHg over y min, SpO2 slope, HR/PR mismatch, temp drift; shown as a "watch" chip, never an alarm	our 1-s trend ✅	BUILT — shadow mode (scores every minute, labelled at Stop, admin report /ai/shadow)
N3	LLM case summary / handoff — we already generate SBAR; extend to a draft narrative from the confirmed chart (drafts only, clinician edits)	record data ✅	BUILT — shadow mode (draft stored in ai_drafts at Stop; rules text until ANTHROPIC_API_KEY is set)
N4	Non-invasive hypotension early warning — run a VitalDB-trained model (Jeong 2024 class) on pleth + NIBP + HR, output "risk ↑ in next 5 min" as a provisional hint	pleth 256 Hz + NIBP ✅	BUILT — shadow mode (scores every minute, labelled at Stop, admin report /ai/shadow)
N5	Hypoxemia risk hint (Prescience-style features: SpO2 trend, FiO2, BMI, RR, procedure)	record + stream ✅ (FiO2 after the Atlan link)	BUILT — shadow mode (scores every minute, labelled at Stop, admin report /ai/shadow)
N6	Artifact suggestion — model flags likely artifacts for one-tap review (never auto-removal)	1-s archive (item E)	BUILT — shadow mode (scored per plotted value at Stop; labelled by the resident's amendments on save)
N7	PACU deterioration score (NEWS2-style, then a learned index) once the PACU board exists	item K	BUILT — shadow mode (scores every minute, labelled at Stop, admin report /ai/shadow)

Guard-rails we keep: every AI output is labelled hint / provisional, sits outside the legal record, never silences a monitor alarm, never doses anything; models are validated on our own archived cases (item E) before use; a clear "what this is not" note in the UI (not a regulated medical device).

Sources: HPI feasibility/RCTs https://pmc.ncbi.nlm.nih.gov/articles/PMC12482379/ · https://pmc.ncbi.nlm.nih.gov/articles/PMC12577059/ · hypotension prediction review https://pmc.ncbi.nlm.nih.gov/articles/PMC12140563/ · Prescience https://pubmed.ncbi.nlm.nih.gov/31001455/ · Epic DI https://sites.northwestern.edu/imqicommittee/2021/02/05/epic-deterioration-index-identifying-patient-acuity/ · STEP-OP (VitalDB) https://medinform.jmir.org/2021/9/e31311/ · non-invasive hypotension model on VitalDB https://link.springer.com/article/10.1007/s10877-024-01206-6 · cross-centre evaluation https://link.springer.com/article/10.1007/s10877-025-01357-0

Sources

Philips Capsule / Neuron device integration: https://capsuletech.com/blog/expand-medical-device-integration-to-drive-operational-clinical-and-safety-benefits · https://www.philips.co.in/healthcare/case-study/medical-device-information-platform-benefits
Epic Anesthesia device integration & validation: https://www.mindbowser.com/epic-anesthesia-guide/ · https://epicsupport.sites.uiowa.edu/epic-resources/philips-device-integration · https://digitalhealth.folio3.com/blog/criteria-to-evaluate-epic-compatible-device-integration/
Artifacts in auto-captured BP (88 procedures, 5,711 points; kappa ≤ 0.65): https://pmc.ncbi.nlm.nih.gov/articles/PMC7943498/ · machine artifact analysis: https://link.springer.com/article/10.1007/s10877-020-00589-6
Automated vs manual record keeping RCT (simulation): https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7327599/
IHE PCD framework / SDC: https://www.ihe.net/Technical_Framework/upload/IHE_PCD_TF_rev1-2.pdf · https://wiki.ihe.net/index.php/SDC@IHE · https://www.draeger.com/en_sa/Hospital/Interoperability-With-IEEE-11073-SDC
Vital Recorder paper: https://www.nature.com/articles/s41598-018-20062-4 · VSCapture: https://pubmed.ncbi.nlm.nih.gov/24106390/ · OpenICE: https://mdpnp.org/MD_PnP_Program___OpenICE.html
Bedside data → EHR + early warning (Hamilton Health Sciences): https://www.hamiltonhealthsciences.ca/share/connecting-bedside-data/
Immutable real-time vitals registry concept: https://www.medrxiv.org/content/10.1101/2021.12.18.21267876v1.full.pdf