RF Sensing for Video-less Healthcare Monitoring — Deep Research

Synthesized from 7 research tracks + 3 adversarial verifications, 2026-08-06. Where research and verification disagreed, verification findings are used and the correction is flagged. One research track (academic SOTA) returned empty; its ground is covered by the ESP32-salvage and through-wall verification tracks.

1. Executive verdict

(a) Can the ESP32/RuView rig be made to work? Not for the stated goal. The 5-node CSI mesh cannot deliver position tracking under any software fix — one antenna and 20/40 MHz of 802.11n bandwidth gives 7.5 m (20 MHz) range resolution and zero angular information, and commodity per-packet phase is corrupted by carrier/sampling offsets that only hardware clock distribution fixes (ESPARGOS paper, arXiv:2408.16377; arXiv:2410.11400). A 2026 systematic study on ESP32 specifically found multi-person separation plateaus at 39–56% accuracy with 97–99% feature overlap and concluded the limit is hardware, not algorithms (arXiv:2601.02177). What the boards can be salvaged into: (i) honest per-site-calibrated binary presence via ESPectre or Espressif's own esp-csi; (ii) — the recommended path — the compute/WiFi half of five per-room nodes with a ~$12 HLK-LD2450 24 GHz radar each, which delivers genuine 3-target x/y tracks at 10 Hz into Home Assistant the same afternoon (verified against manufacturer protocol docs; see §7).

(b) Has anyone made WiFi sensing work? Yes — for motion/presence only, never position. Labs: through-wall activity classification at 87–92% (per-site-trained, directional antenna — TU Wien ICVS 2023), single-person breathing at short range (PulseFi, arXiv:2510.24744), occupancy counting 0–3 people (~86–96%, IEEE 2024). Commercial: two licensors (Cognitive Systems, Origin AI) ship motion detection through ISP gateways — Xfinity WiFi Motion, Verizon Fios Home Awareness, Rogers, Plume — and Comcast's own docs state the system "will not tell the exact location of the movement" (Xfinity). The commercial ceiling is Origin's Aug-2025 room-name "Zone Detection" (PRNewswire). Verified: no commercial WiFi-sensing product on earth ships x/y position tracking. The tell: nami.ai, the most aggressive WiFi-sensing caretech player, is bolting on mmWave radar and a wearable for falls (nami, MWC 2026 release).

(c) What actually delivers? Two evidenced paths to video-less position tracking. In-room: 60 GHz FMCW mmWave radar — 10–30 cm accuracy, up to 20 tracks per sensor (TI TIDEP-01000), 97.9% multi-person fall detection in peer review (Sci. Reports 2026), commercially deployed in eldercare (Vayyar Care, Essex County Council 316+ sensors — case study). Important correction (verification overrides research): 60 GHz does NOT see through real interior walls — NIST measured 11.8–31.6 dB one-way through framed plasterboard walls, doubled for radar round-trip (NIST 60 GHz penetration campaign); the oft-quoted ~1 dB figure is a bare single gypsum sheet at normal incidence. Architecture = one sensor per room, full stop. Genuinely through-wall: sub-10 GHz UWB/FMCW radar — MIT WiTrack (10–13 cm through sheetrock, NSDI 2014), RF-Pose (CVPR 2018), Emerald Innovations clinically (pharma trials, Science Translational Medicine 50-patient Parkinson's study), and FDA-cleared vitals on UWB (Xandar Kardian XK300, 510(k) K202464) — but Emerald is un-buyable, and FCC 47 CFR 15.510 restricts through-wall imaging systems to law enforcement/fire/rescue (eCFR), a named regulatory landmine for any US healthcare product marketed as through-wall person tracking.

(d) Why RuView got popular despite not working. A hype-loop artifact, forensically documented: legitimate CMU "DensePose From WiFi" branding (arXiv:2301.00250) + AI-generated production-grade appearance (~35 commits at 8,365 stars = ~239 stars/commit per independent audit) + suspected star inflation (1.3k→3k overnight with zero commits in 6 months — HN) + GitHub-trending feedback loops (#1 trending, +4,557 stars/day, Feb 2026 — byteiota) + README-repeating aggregator citogenesis + deletion of critical issues (issue #12 documenting np.random.rand() "parsers" was deleted by the repo). Scott's zero-discrimination A/B result is independently replicated at least four ways in the repo's own tracker: #1440 (presence AUC ≈ 0.58 = chance once time-of-day drift controlled), #1521 (shipped presence head mathematically cannot output "absent" — bias 8.188 > ‖w‖ 3.668), #1448 (the literal inverted-variance bug Scott measured), #1525 (confidence-0 pose stub). The only verified-working layer — raw CSI capture — is exactly the layer Scott already has.

2. What each technology ACTUALLY delivers

3. The academic truth about WiFi-CSI

What labs genuinely achieved, and under what conditions:

• Through-wall activity classification (not tracking): ESP32-S3 + directional biquad antenna spanning five rooms, CNN on spectrograms, 92.0±3.5% NLOS accuracy — with per-site training data and a purpose-built antenna (TU Wien, ICVS 2023; dataset Wallhack1.8k).
• One-sided through-wall occupancy (presence + walking direction, both radios outside the wall): the Hernandez/Bulut VCU line on the canonical ESP32-CSI-Tool (ICC 2021).
• Breathing: single static subject, controlled environment — MAD 1.38–2.6 brpm (2024 sleep study), MAE 0.09 brpm with ML (PulseFi).
• Counting: ~98% presence, ~86% counting up to 3 people, per-site ML (IEEE 2024).
• The headline pose/position papers people cite are not WiFi-CSI-on-commodity: RF-Pose/WiTrack used custom FMCW radios sweeping ~5.4–7.2 GHz with GHz-scale bandwidth plus vertical + horizontal antenna arrays — roughly two orders of magnitude more bandwidth and 8–16× more aperture than an ESP32 (RF-Pose CVPR 2018; WiTrack). CMU's DensePose-From-WiFi was camera-supervised, two commercial routers, lab-only, never productized (arXiv:2301.00250).

Why homes fail: models are environment-specific and collapse on layout change; CSI extraction depends on fragmented per-chipset hacks (the discontinued Intel 5300 era); multi-person signals superpose irrecoverably on single-antenna hardware (failure-mode analysis, Apr 2026; arXiv:2601.02177). Espressif itself concedes its detection algorithm "is still being optimized" (esp-csi README). The ESPARGOS project's framing is the cleanest indictment: commodity devices "lack sufficient spatial diversity or phase synchronization" for sensing — their fix was a custom 8-radio phase-coherent array with hardware clock distribution, not yet purchasable (espargos.net; arXiv:2408.16377).

Standards tailwind, honestly sized: IEEE 802.11bf-2025 is published (WG-approved Mar 2025 — IEEE TGbf; IEEE SA). It standardizes sensing measurement sessions — not detection or localization algorithms, and it adds no antennas or bandwidth to existing silicon (NIST overview). First explicitly-sensing silicon: Infineon AIROC ACW741x dev kits, Jan 2026; Wi-Fi CERTIFIED Sensing expected ~2026 (wi-fi.org). It fixes CSI-extraction fragility, not the physics.

The overlooked academic path that fits Scott's hardware class: RTI. Variance-based radio tomography tracked a person through exterior brick walls at ~0.9 m error — but with ~30 nodes surrounding the house and surveyed geometry, solving a per-link inverse problem (arXiv:0909.5417); modern CNN variants reach ~0.23 m (Sci. Reports 2025). RuView's global fused-variance threshold discards exactly the spatial structure that makes RTI work — Scott's 132-vs-141 non-result is consistent with the literature, not an anomaly. Five nodes is well below working deployments (12–34), and RTI is still static-person-blind.

4. Commercial landscape

WiFi sensing (shipping): Cognitive Systems licenses "WiFi Motion" into Xfinity (launched ~Aug 2025, motion-only, max 3 coverage zones, explicitly no location — Xfinity), Rogers, Plume; eldercare via Caregiver Aware/Electronic Caregiver — activity patterns via ~4 smart plugs, no fall detection, no position (PRNewswire 2024). Claims 19–20M homes; independent corroboration ~1–9M (Communitech). Its 2026 site's "localization / vitals" language is roadmap, not product. Origin AI powers Verizon Fios Home Awareness (2022, motion-only), Hex Home, TruShield + Zone Detection (room names, Aug 2025), Aloe Care (whose shipping fall SKUs still lean on a wearable — Aloe Care). Linksys Aware, the first consumer product (2019), was discontinued mid-2024 (Kaspersky). Both leaders are B2B license-only — no self-serve SDK exists. Signify ships WiZ SpaceSense (WiFi) and Hue MotionAware (Zigbee, Oct 2025) — per-room motion triggers only.

Healthcare RF monitoring:

• Vayyar Care (60 GHz 4D radar): the closest commercial analog. In-room point-cloud position, posture, bed-exit, bathroom analytics, fall detection through steam; ~16 m²/device, ~3 devices/home at $250 + ~$20/mo historical B2C (Calcalist); Essex County Council 316+ sensors with claimed elimination of fall-admission "long lies" (Vayyar). Verification corrections: zero FDA records of any kind (openFDA 510(k)/PMA/registration queried 2026-08-05); consumer channel is dead (Alexa Together discontinued May 21, 2025); company downsized ~2024, narrowed to smart-home + automotive (CTech); its Jan-2026 "clinical" credential is UK DCB0129 — a process standard, not a clearance; no peer-reviewed sensitivity/specificity exists. Position comes from one sensor per room — no through-wall magic.
• Emerald Innovations (MIT/Katabi, custom FMCW-class RF, not commodity WiFi): the scientific high-water mark — Sci Transl Med 2022 Parkinson's gait (50 patients, 1 yr, 200k measurements — paper), Nature Medicine 2022 (PD from nocturnal breathing, n=7,671). Verification corrections: zero FDA clearances (openFDA confirmed); not buyable — B2B pharma-trial instrumentation only (Verge Genomics ALS, Aspen Neuroscience, Rett — BusinessWire); public news trail effectively stops in 2023.
• Xandar Kardian XK300: the regulatory exemplar — FDA 510(k) Class II cleared (K202464) for contactless resting HR/RR/motion/presence, on 6.5–8 GHz impulse UWB, not mmWave (the band choice is the tell about penetration); 50 US facility deployments by Aug 2023 (Fierce Biotech; PR).
• The rest: SafelyYou (camera-based contrast case, ~$102M raised, published ER-visit reductions — AJMC); Butlr thermal ($38M B, senior living); Sensi.AI audio ($98M+); Nobi smart-lamp ($37M B, ~$240–525/mo/room — McKnight's); EchoCare sub-10 GHz through-wall apartment monitor (~40 m²/unit — ISRAEL21c; regulatory status unverified — diligence item); EarlySense under-mattress (multiple 510(k)s). Market: fall-detection ~$517M (2025) → $940M (2033) (Grand View). The consumer channel is a graveyard (Alexa Together dead, Walabot HOME a $150 commodity); every funded player sells B2B into senior living/home care.

5. The RuView post-mortem

Who: Reuven "rUv" Cohen — real, non-anonymous serial evangelist (founded Enomaly 2004, Wikipedia; ex-Forbes contributor; cloud → NFT → agentic-AI hype cycles). ~199 public repos plus ~751 registry artifacts, openly "vibe-coded" (AY Automate breakdown). Not a pure facade: his claude-flow/Ruflo has 67k stars and ~80k npm downloads/month with genuine users. But HN's verdict on the portfolio is brutal: "AI generated lists of AI generated projects… Time crystal detection? 100% Accuracy? Give me a break" (HN); "the whole thing is a scam" (HN).

Hype mechanics: created 2025-06-07, most code in a single day; rode CMU DensePose-From-WiFi's credibility (HN 176 pts, Jan 2023); 1.3k→3k stars overnight during 6 months of commit dormancy (HN: "I suspect these stars being bought"); #1 GitHub trending Feb 2026 (+4,557 stars/day) during the OpenClaw trending mania; 88,647 stars by Aug 2026 (GitHub API) — while every HN submission of the repo itself maxed at 2–32 points, one titled "(Scam Alert)" (Algolia). Amplification came from README-repeating SEO aggregators, not users. Context: the StarScout study documents 4.5M+ suspected fake GitHub stars ecosystem-wide (arXiv:2412.13459); no direct proof RuView's were bought — the evidence is circumstantial but heavy.

What other users report — Scott's failure mode, replicated: independent audit found np.random.rand() in the "parser," hardcoded CSI strings, placeholder Doppler, untrained networks (audit fork); critical issue #12 deleted; Cohen's rebuttal (#37) defends the real capture layer as proof of the undemonstrated detection claims. Controlled reproductions Jul–Aug 2026: #1440 — pipeline "genuinely works end to end" but presence AUC ≈ 0.58 (chance) once confounds controlled; #1521 — presence head can only ever say "present" (trained on one 51-min recording with zero empty-room negatives); #1448 — the exact inverted-variance bug; #1525/#1526 — confidence-0 pose canvas and "LIVE — ESP32 Hardware Connected" shown with no hardware. Of 443 visible issues, 90 are self-filed by ruvnet plus 343 self-PRs — the repo is substantially agent-operated. What's real: the firmware calls Espressif's actual esp_wifi_set_csi API and the UDP CSI stream is genuine (hand-decoded in #1440) — which is exactly what Scott measured: real radio, fictional detection. The honest comparable, ESPectre (8.9k stars, actively maintained), claims motion/presence only — roughly where commodity CSI truth sits.

6. Scott's healthcare play

Where the open gaps are (validated against buyer behavior, not tech novelty):

1. Nobody commercially ships whole-home multi-person position — Vayyar/Xandar are per-room, Origin/Cognitive are motion events, Emerald is locked in pharma. But note the honest framing: the industry converged on per-room ambient sensing because that's what the physics and FCC allow. The winning architecture everywhere (Essence, EarlySense, Butlr, Vayyar) is: ambient per-room sensor (radar/thermal) + contact-adjacent vitals (bed piezo/UWB) + door/PIR ADL rhythm + longitudinal baseline model — not a through-wall tracker (Essence Care@Home).
2. Integration, not sensing, is the #1 buyer complaint: only 15% of senior-living operators report full EHR/tool integration; 50% little or none (McKnight's); nurse-call integration is "often the hardest part of any fall detection deployment" (B&B Integrations). A nurse-call/EHR-native data layer beats a better sensor.
3. False alarms are the operational killer — Origin markets "lowest false-alarm rate" because buyers are burned; bed-alarm RCTs show no fall reduction and alarm fatigue (PMC2744312).
4. The evidence bar is outcomes, not detection accuracy — SafelyYou wins with published ER-visit reduction (AJMC); almost no ambient vendor besides SafelyYou/Emerald has peer-reviewed outcomes. A validation-first entrant (Scott's DiD instincts) has an open lane.
5. Reimbursement arbitrage: RPM requires an FDA-defined device transmitting physiologic data (ACP) — wellness-tier ambient products can't bill it, but a 510(k)-cleared radar RR/HR device (Xandar precedent) can, and 2026's new 99445 code (2–15 days data, ~$52) makes passive capture newly monetizable (Nixon Law Group on the 2026 final rule).

Regulatory notes: Fall/activity/wellness alerts need no FDA clearance under the General Wellness policy (RegDesk summary) — that's how Vayyar/Butlr/Origin/Sensi ship. HR/RR claims → Class II 510(k) (Xandar path). HIPAA: vendor operates as a business associate; BAA before PHI flows; "anonymous by physics" (no images) is a genuine selling point (AccountableHQ). FCC 15.510: do not market anything as "through-wall person tracking" in the US — that phrase describes a device class restricted to law enforcement/fire/rescue (eCFR 15.503/15.510); shipping products describe "whole apartment" coverage instead.

A credible MVP stack (per-room-cell architecture, all COTS, no cloud, no camera): per monitored room — LD2450 (x/y tracks + zones, $12) + LD2410 (stationary presence backstop, $3) on an ESP32, MR60FDA2 60 GHz fall module on the bathroom ceiling ($25–37, ESPHome-native — Seeed), MR60BHA2 pointed at the bed for respiration ($25 — Seeed) — ~$110–130/room vs Vayyar's $250/room + subscription — feeding Home Assistant/MQTT into a baseline-deviation model with tiered escalation, positioned as wellness-tier (no clearance needed), with the Xandar 510(k) route as the upgrade path if vitals claims ever matter. Research-grade tier: TI IWR6843ISK (~$150, Octopart) running the prebuilt TIDEP-01000 20-track people-counting firmware — the platform the fall-detection literature builds on. Given SGA context: the nearest in-house application isn't elder care but practice ops — chair/operatory occupancy, patient-flow, and wait-time analytics across 260 practices is the same sensor stack with zero PHI and no FDA exposure.

7. Salvaging the 5 boards

Worth doing on esp-csi (1–2 boards, R&D tier only):

• Flash esp-radar/wifi_sensing_demo — on-site training, Web Serial diagnostics (jitter_value, enter/exit thresholds). This is what RuView's detection layer should have been; expect calibrated binary presence, environment-dependent, false positives from fans/microwaves/neighboring APs.
• Or ESPectre (ESPHome/HA-native, automatic subcarrier selection, F1>96% after per-site tuning, README honestly states environment dependence) as a whole-house "something moved" tripwire arbitrating the radar zones.
• Optional science: replicate the VCU one-sided through-wall occupancy setup, or contribute captures to Wallhack1.8k-lineage datasets.

The LD2450 upgrade path (the main event, verified against primary sources):

• Buy 5× HLK-LD2450 ($9–15: Hi-Link official, AliExpress, Amazon) + a few LD2410 ($1–3, AliExpress). Wire to hardware UART @256000 baud; use the first-party ESPHome ld2450 component (module FW ≥ V2.02.23090617, updatable via HLKRadarTool app) + ld2410.
• Verified capability envelope: 3 simultaneous targets, int16 x/y in mm + speed at 10 Hz (30-byte frame, header AA FF 03 00 — protocol v1.03 PDF); 6 m practical (ignore the 8 m marketing number); ±0.15 m radial but 2–20° angular error → ±2 m lateral at FOV edges (mount for center-of-room coverage, 1.5–2 m high); moving-target tracker — static sleepers drop out (pair with LD2410); ghost targets off mirrors/metal (zone exclusion + Kalman, see PeterkoCZ91/HLK-LD2450-security); track IDs not persistent; clean 5 V ≥200 mA supply mandatory (Screek warning); 2D only — no height axis, so no intrinsic fall detection.
• Proven builds on this exact stack: Everything Presence Lite ($35 commercial kit, same architecture), NAMO on-device floorplan tracking, multi-LD2450 fused indoor positioning, esp32.co.uk tutorial, TillFleisch polygon zones. Room-to-room "tracking" emerges from zone handoff between nodes — honest architecture: a $12 radar on each side of the wall, not through it.
• Add-ons: fall = MR60FDA2 ceiling-mount (ESPHome seeed_mr60fda2); identity = BLE via ESPresense or ESPHome BT proxy + Bermuda (wearable-dependent, adjunct only). The S3's 8MB PSRAM runs ESPHome + UART radar + BLE proxy simultaneously.

Skip: any further CSI positioning/pose work on this hardware (physics-closed); CSI multi-person anything (39–56% ceiling); a 5-node RTI attempt (working deployments used 12–34 surrounding nodes; still static-blind); ultrasonic and device-free BLE-RSSI (dominated); waiting for 802.11bf to rescue CSI (it standardizes measurements, not miracles); anything from the RuView repo above the capture layer.

8. Recommended next moves (ranked)

1. This weekend, ~$60–90, low effort: order 5× LD2450 + 3× LD2410; convert the existing ESP32-S3 nodes to per-room radar nodes on ESPHome/HA. Deliverable: live 3-target x/y floor-plan tracking with zone occupancy — the exact thing RuView could never do — by Sunday. (ESPHome ld2450)
2. +$50–75, low effort: add MR60FDA2 (bathroom ceiling — where ~80% of falls occur per Vayyar) and MR60BHA2 (bed-pointed respiration). Full elder-care cell ~$110–130/room, no camera, no cloud, no wearable.
3. ~$150, medium effort: one TI IWR6843ISK/AOPEVM running the prebuilt 3D people-tracking firmware, parsed with pymmw — the research-grade benchmark to quantify what the $12 modules miss, and the published platform for a fall classifier if this ever becomes a product.
4. Zero cost, medium effort: re-flash 1–2 boards with ESPectre or wifi_sensing_demo as a whole-house motion tripwire — and re-run the A/B protocol against it for closure on what honest commodity CSI actually delivers.
5. Business exploration, no hardware: if pursuing healthcare seriously, scope the wellness-tier per-room stack against SGA-adjacent use cases (practice flow/operatory occupancy first — zero PHI, zero FDA; senior-care second), with the Xandar 510(k) + 2026 RPM 99445 route as the reimbursement upgrade path. Diligence items before any buy/build: pull FCC grant class for EchoCare/any UWB candidate (15.510 exposure), and never use the phrase "through-wall tracking" in US marketing.
6. Skip entirely: further RuView engagement; CSI position/pose research on ESP32; Vayyar-as-component (no FDA clearance, consumer channel dead, B2B quote-only); waiting on 802.11bf silicon to change the physics.

One-line verdict: the radios were never the problem — RuView's detection layer was fiction; position through walls is either restricted-band UWB or a dense tomography mesh Scott doesn't have; and the fastest honest path to "presence + position, video-less" is a $12 radar per room on the ESP32s he already owns.

LD2450 Radar Mesh — Order Spec & Build Plan

Converting the 5× ESP32-S3-DevKitC-1 (RuView survivors) into a real video-less presence-and-position mesh: per-room x/y tracking of up to 3 people, zone analytics, and (optional add-ons) fall + breathing detection. Everything below is verified against datasheets, first-party ESPHome docs, and published builds — sources in RESEARCH-rf-sensing-healthcare.md §7.

1. Order list

Core (the position mesh) — ~$60–75

Healthcare add-ons (the elder-care room cell) — +$55–65

Total: ~$120–140 for the full kit; ~$60–75 for the position mesh alone. (Optional benchmark tier, later: TI IWR6843ISK ~$150 with the prebuilt TIDEP-01000 20-track people-counting firmware — research-grade comparison if this heads toward a product.)

2. Wiring (per node, 10 minutes)

LD2450 → ESP32-S3:

LD2450 5V  → ESP32 5V (VBUS)     LD2450 GND → ESP32 GND
LD2450 TX  → ESP32 GPIO16 (RX)   LD2450 RX  → ESP32 GPIO17 (TX)
UART: 256000 baud (hardware UART, not software)

LD2410 (where fitted) → second UART (e.g. GPIO18/19) @256000. Check module firmware ≥ V2.02.23090617 via the HLKRadarTool phone app (BLE) — older FW predates the ESPHome protocol.

3. Software (per node, ~15 min once the first works)

ESPHome (runs fine alongside Home Assistant; the S3's 8 MB PSRAM is plenty):

1. esphome new project → board esp32-s3-devkitc-1.
2. Add the first-party components: ld2450 (targets → x, y, speed, zones) + ld2410 (still-presence). No custom code — these are maintained upstream.
3. Define 2–3 polygon zones per room (bed, chair, doorway) — TillFleisch/ESPHome-HLK-LD2450 for polygon support; exclusion zones kill mirror/metal ghosts.
4. Home Assistant auto-discovers each node → floor-plan card with live target dots (NAMO build, multi-LD2450 positioning thread).

(No Home Assistant yet? It runs as a Docker container on this PC in ~10 min — same pattern as the RuView sim. I'll set it up when the parts arrive.)

4. Placement (same 5 rooms as before, better rules)

• Center of a wall, 1.5–2 m high, facing the room — angular error is worst at FOV edges (2–20° → up to ±2 m lateral at 6 m), so aim the ±60° cone at where people actually are.
• One node per room. Room-to-room "tracking" = zone handoff at doorways (define a doorway zone in each adjacent room).
• Avoid facing mirrors/large metal (or add exclusion zones). Keep the 6 m practical range in mind (ignore the 8 m marketing figure).
• Fall module: bathroom ceiling, centered. Breathing module: aimed at the bed, ~1–1.5 m.

5. Acceptance test (the same rigor that caught RuView)

Repeat the empty-vs-walking A/B — this time it should pass decisively:

1. 2 min empty house → zero targets in every room (not "PRESENT_MOVING").
2. Walk a scripted route (bedroom → hall → kitchen) → the HA floor plan shows one dot making that exact journey with correct room handoffs.
3. Two people in one room → two distinct tracks.
4. Sit motionless 5 min → LD2410 still-presence holds while LD2450 track drops (expected). Log it like last time; keep the transcript as the validation record.

6. What this mesh will NOT do (so expectations stay honest)

• No identity (anonymous by physics — for healthcare that's a feature; add BLE/ESPresense later if "who" matters).
• Max 3 tracked targets per room.
• No through-wall sensing from a single sensor — and per FCC 15.510, never market anything as "through-wall tracking" in the US; the compliant framing is "whole-home, camera-free."
• No fall detection from the LD2450 itself (2D) — that's the MR60FDA2's job.

7. Where this goes if the healthcare idea has legs

• Wellness-tier product (no FDA clearance needed): zones + dwell + night-wandering + bathroom-duration + fall alert, per-room cell ~$110–130 vs Vayyar's $250+subscription. Differentiate on integration (EHR/nurse-call) + published outcomes — the two things buyers say are missing (report §6).
• Reimbursement path: 510(k) radar vitals (Xandar K202464 precedent) + 2026 RPM code 99445 makes passive capture billable.
• SGA-adjacent zero-risk pilot: the identical stack does operatory/chair occupancy + patient-flow analytics — no PHI, no FDA, 260 practices of addressable footprint, and the pilot hardware is already on your desk.

8. Scaling to a 3,000 sq ft practice (patient-flow deployment)

Layout assumption: waiting room, front desk/entry, ~8 operatories, halls/sterilization → ~13 nodes.

Node plan:

Cost: DIY ~$350–500 hardware + ~$150 mini-PC server (or existing) → under $600 all-in. Compare: TI-grade ~$2K; Vayyar-class $3.3K+ with subscription.

What it answers: door→checkout journey timelines, true wait times, chair utilization by hour/day, op turnover gaps — the physical-bottleneck data behind visits-per-provider-per-day that no PMS measures directly.

Caveats to validate in a 1-op pilot first: metal-heavy operatories (chairs/lights/cabinets) cause radar ghosts → placement + exclusion zones; waiting rooms >3 concurrent people need the 2-node split; sensor layer is identity-free (PHI-free by design) — any per-patient linkage happens later, in the data layer, against check-in timestamps.

Rollout sequence: home mesh (owned hardware, ~$60) → one-operatory pilot → 13-node practice build.

Next action: order the Core list. When parts land, say the word — first node takes ~1 hr (wiring + ESPHome + HA), remaining four ~20 min each, then the acceptance test.