Multi-sensor streaming
Holoscan pipeline ingests CGM, blood pressure, SpO2, and wearable signals from multiple concurrent patients.
Vigil streams biometric signals from patient devices through Holoscan, detects deterioration via GPU inference on Jetson and cloud GPUs, and surfaces clinician-reviewed intervention recommendations grounded in the patient's Scribara encounter history.
Physical AI layer — extends the Scribara platform to the patient's home
Specialty patients with CHF, diabetes, psoriasis, and COPD deteriorate between visits. By the time a referral reaches the specialist, the patient is in the ED. Existing RPM platforms produce context-free alerts that nurses triage without specialist insight.
Holoscan ingests CGM, blood pressure, SpO2, and wearable data in real time — multi-sensor, time-aligned, sub-500 ms latency.
TensorRT-compiled anomaly models (on Jetson at the edge or L4 in the cloud) score every patient's signal against their personal baseline — not a population threshold.
Alerts are enriched with the patient's last Scribara encounter — last codes, medications, imaging findings — and a NeMo-drafted intervention recommendation. [ASPIRATIONAL]
Approved recommendations trigger Authra for medication auths and Codexa for CCM billing codes — closing the loop without leaving the workflow. [ASPIRATIONAL]
Holoscan pipeline ingests CGM, blood pressure, SpO2, and wearable signals from multiple concurrent patients.
Per-patient baseline models flag deviations; TensorRT-compiled for sub-500 ms inference.
Jetson Orin processes sensor streams on-device at the clinic — PHI stays local. [ASPIRATIONAL]
Every alert enriched with the patient's last Scribara encounter — notes, codes, medications, imaging.
NeMo models forecast CHF, diabetes, and COPD acute events 24–72 hours in advance. [ASPIRATIONAL]
CMS CPT 99453/99454/99457/99458 billing codes logged automatically — turns RPM into revenue.
Holoscan multi-sensor stream processing across dozens of concurrent patients requires GPU-parallel operations. TensorRT anomaly inference must return in under 500 ms for life-critical signals. NeMo exacerbation models run as continuous inference, not batch jobs. CPU cannot meet any of these requirements.
Vigil uses encounter history and feeds monitoring outcomes back to the platform.
CGM (Dexcom, Abbott Libre), blood pressure (Withings, Omron), SpO2, and common wearables via HL7 FHIR Observation resources. Device integration APIs are [UNKNOWN] — confirmed at implementation.
Yes — the Jetson edge node is optional. Holoscan and TensorRT inference run equally well on cloud GPU (L4/A100) for clinics without an on-prem device.
Vigil logs the required device-supply, monitoring, and care-management events and generates CPT 99453/99454/99457/99458 billing records automatically for export to your billing system.
See Vigil monitoring a chronic cardiology panel in a live demo.