Healthcare / Critical Care
AI-powered oxygen therapy: Intelligent FiO2 titration system for critical care
Development of a closed-loop oxygen therapy system, with design documentation structured to IEC 62304 and ISO 13485 expectations, that adjusts FiO2 automatically from real-time SpO2 monitoring and predictive algorithms. Design targets: 35% less oxygen consumption than manual titration and 98.5% time in the prescribed SpO2 range, with less hyperoxia and nursing workload.
What we did: We designed the bedside controller electronics and safety-critical PCB, integrated the SpO2 module and proportional air/O2 blender valves, developed the IEC 62304 Class B closed-loop control firmware (with an independent hardware safety supervisor enforcing FiO2 and low-SpO2 limits as the external risk control) and prediction models, built the hospital monitoring dashboard, and prepared the CE technical file for the client's notified-body review.
Medical equipment manufacturer, India16 months
Published Last reviewed:
Illustrative imageThe challenge
Manual oxygen titration in hospitals is labor-intensive, often delayed, and frequently results in hyperoxia or hypoxia. A medical equipment manufacturer needed an intelligent system to automatically maintain optimal oxygen levels while reducing oxygen waste and nursing workload.
Manual titration delays
Nurses check SpO2 and adjust oxygen flow every 1-2 hours. Between checks, patients may experience dangerous oxygen fluctuations that go undetected.
Oxygen overconsumption
Conservative clinical practice leads to unnecessarily high FiO2 settings. Hospitals waste 30-40% of supplemental oxygen, increasing costs significantly.
Hyperoxia complications
Prolonged high oxygen levels cause lung damage, ROP in premature infants, and increased mortality in some patient populations. Prevention requires precise control.
Nursing workload
ICU nurses spend significant time on manual oxygen adjustments. Automated titration would free them for more critical patient care activities.
Impact figures describe the situation before the project, as the client reported it to us; we did not measure them.
Our solution
We developed an AI-powered closed-loop oxygen therapy system that continuously monitors SpO2, predicts oxygenation trends, and automatically adjusts FiO2 to maintain target saturation levels, designed for 98.5% time-in-target and 35% lower oxygen consumption.
System architecture
Complete automated oxygen therapy ecosystem with bedside controller, central monitoring, and AI analytics.
Bedside controller
- High-precision SpO2 monitoring (Masimo SET)
- Proportional air/O2 blender valves (21–100% FiO2)
- Flow rate sensing (0.1 LPM resolution)
- FiO2 analyzer for delivery verification
- Touch display for manual override
- WiFi + Ethernet connectivity
AI engine
- Real-time SpO2 trend analysis
- Predictive oxygenation modeling
- PID control with adaptive tuning
- Patient-specific learning
- Anomaly detection and alerts
- Edge inference (< 100ms latency)
Hospital platform
- Multi-patient monitoring dashboard
- Oxygen consumption analytics
- Clinical decision support
- EMR integration (HL7 FHIR)
- Regulatory audit logging
- Remote specialist consultation
Custom hardware design
- MCU
- STM32H7 with independent hardware safety supervisor
- SpO2 Module
- Masimo SET (motion-tolerant)
- Air/O2 Blender
- Proportional air and O2 valves (21–100% FiO2, up to 15 LPM total flow)
- FiO2 Sensor
- Paramagnetic (21-100%)
- Flow Sensor
- Mass flow (0.1 LPM resolution)
- Display
- 7" Medical-grade touchscreen
- Power
- Medical PSU + 2hr UPS backup
- Enclosure
- IP54; designed to IEC 60601-1
Safety-critical firmware (IEC 62304 Class B)
- Independent hardware safety supervisor enforcing FiO2 and low-SpO2 limits (external risk control for the Class B classification)
- Real-time SpO2 processing at 100 Hz
- Closed-loop control with 200ms cycle time
- Watchdog monitoring with safe-state fallback
- Comprehensive event logging (30-day storage)
- Secure OTA updates with rollback
- Continuous self-test routines
- Audit trail designed to FDA 21 CFR Part 11
Implementation timeline
Phase 1: Clinical requirements & AI research
10 weeks- Clinical workflow analysis in ICUs
- Oxygen therapy protocols review
- AI/ML algorithm research and selection
- Regulatory strategy for AI-based SaMD
Phase 2: Hardware development
14 weeks- Proportional valve selection and testing
- SpO2 module integration (Masimo SET)
- Safety-critical PCB design
- Enclosure design per IEC 60601
Phase 3: AI model development
16 weeks- Training and validation dataset preparation
- SpO2 prediction model development
- Adaptive PID controller optimization
- Edge deployment and optimization
Phase 4: Firmware development
14 weeks- IEC 62304 Class B development process
- Closed-loop control implementation
- Safety monitoring and alarm system
- EMR integration protocols
Phase 5: Verification & validation
12 weeks- IEC 60601-1 pre-compliance testing (formal testing by the client at an accredited lab)
- Clinical accuracy study support
- AI performance validation
- Usability testing with ICU staff
Phase 6: Submission package preparation
8 weeks- CE technical file compilation
- AI/ML documentation (GMLP)
- Clinical evaluation report
- Manufacturing validation
Results & impact
The system was designed to reduce oxygen waste and nursing workload through closed-loop oxygenation control; the figures below are design targets checked in bench and simulation testing. The device is pre-submission and has no clinical outcome data. CDSCO licensing under the Medical Devices Rules, 2017 (in the risk class CDSCO assigns) is the client's step, as are the CE and FDA filings.
How to read the labels: Design target = an engineering target, not a field measurement; Client-reported baseline = the client's own figure for the situation before the project; we did not measure it. Figures without a label are specifications or project scope.
Oxygen savings
Design target35%
Reduced consumption vs manual titration
Time in target
Design target98.5%
SpO2 within prescribed range
Response time
Design target< 30 sec
To correct desaturation
Nursing time
Design target15%
Freed for other care
Technologies used
- STM32H7
- Masimo SET
- Proportional Valves
- TensorFlow Lite
- LSTM
- React
- Node.js
- AWS IoT Core
- PostgreSQL
- HL7 FHIR
- MQTT/TLS
Related case studies
Healthcare / Neonatal Care
Automated nitric oxide therapy system: Connected iNO delivery for NICU patients
Healthcare / Medical Devices