The Clinical and Regulatory Case for Contactless Fall Detection
One in three adults over the age of 65 falls every year. Half of those who fall and lie unattended for more than an hour die within six months. The clinical literature on long-lie outcomes is consistent and grim. Reducing the time between a fall and a caregiver response is the single most important intervention in elderly-care safety, and it has remained an unsolved problem for decades.
Existing solutions exist but adoption is poor. Pull-cords assume the resident can reach the cord. Pendant buttons assume the resident is wearing it and is conscious. Camera-based monitoring is rejected by residents, families, and data-protection regulators in many jurisdictions. The result is that most falls in private dwellings and many in care homes go undetected until a caregiver finds the resident hours later.
Contactless 60 GHz radar fall detection changes the equation. The resident wears nothing. The radar sees nothing identifiable. A fall triggers an alert within seconds. Adoption by residents and by data-protection officers becomes feasible. The technology is ready; the engineering and regulatory work is what stands between a research demonstration and a sellable Class IIa medical device.
Why Wearables and Cameras Fall Short
- Pendant alarms have measured adherence rates below 30 percent in independent-living settings and below 50 percent in supervised settings. The pendant is taken off at night, when most serious falls occur. It is forgotten when changing clothes. It is removed during showering, also a common fall scenario.
- Smartwatch fall detection works algorithmically but inherits the wearability problem. Elderly residents stop wearing smartwatches consistently within months of issue. Battery life and visible-technology stigma both contribute.
- Camera-based fall detection works technically and has been demonstrated in laboratories for over a decade. Real-world deployment is blocked by privacy regulations in bedrooms, bathrooms, and any private space, which are precisely the rooms where falls most often happen.
- Pressure mats detect a fall after the fact when somebody steps on them, useful only for known walking paths. They cannot detect a person who falls and remains on the bed or in the chair.
The mmWave Approach: Detection, Classification, Confidence
A modern radar fall detector is a small piece of engineering with three layered behaviours.
Detection. The radar tracks targets in three dimensions, including elevation. A sudden downward velocity exceeding a threshold raises a candidate fall event. The threshold is tuned so that all genuine falls trigger, accepting some false alarms at this stage.
Pose verification. Once a candidate fall is raised, the system holds a 15 to 30 second post-event window. During this window it checks whether the person is in a prone or supine pose at the bottom of the trajectory. A controlled lying-down on a bed or sofa has a different pose profile from a true fall.
Classification with confidence. A small on-device machine-learning model, trained on micro-Doppler and pose features, outputs a fall-probability score. Scores above an upper threshold raise an immediate alert. Scores in an intermediate band raise a check-on-resident notification, which the caregiver platform handles with a lower-priority response.
Two confidence thresholds, not one, are what makes the system clinically useful. A single threshold either floods caregivers with false alarms or misses subtle falls. Layered confidence matches the way human-judgement responses already work in care settings.
Bonus Capability: Continuous Vital Signs from the Same Sensor
The same radar that detects falls can also extract respiration rate and heart rate from the chest reflections of a stationary person, at no additional hardware cost.
Respiration is the easier signal. The chest moves by 4 to 12 millimetres at 60 GHz wavelengths, well above the radar's micro-Doppler floor. Respiration rate is extracted reliably at distances up to 4 to 6 metres. The signal is useful as both a clinical indicator and as a presence-confirmation check in fall detection (if a person is breathing, they are alive).
Heart rate is harder. The chest displacement from heartbeat is sub-millimetre, near the noise floor. With careful signal processing (typically harmonics-aware filtering on the micro-Doppler spectrum), heart rate is extractable at distances up to 2 metres in a quiet pose. This is sufficient for sleep monitoring and for confirming consciousness after a fall. It is not a substitute for ECG-grade measurement.
The clinical literature on the ballistocardiogram (BCG) and radar-based vital-sign monitoring continues to grow. For products in the AAL and remote-patient-monitoring space, presenting both fall detection and vital-sign monitoring on a single Class IIa device is a strong commercial and clinical proposition.
Performance Targets: Recall, Precision, False-Alarm Rate
Engineering target numbers from radar fall-detection products we have delivered to clinical pilot stages:
- Fall recall above 95 percent on a balanced dataset including scripted falls, real-world simulated falls, and provoked-fall datasets from clinical-research partners.
- False-alarm rate below 1.5 alarms per occupied room per week. The room population matters: a bedroom is quieter than a living room with active grandchildren.
- Detection latency below 8 seconds from impact to alert dispatched, including the post-event verification window.
- Coverage range 1 to 6 metres from sensor, covering a typical bedroom or living-room footprint with one device.
- Vital-sign respiration accuracy within plus or minus 2 breaths per minute against capnograph reference, at 2 metres distance.
- Vital-sign heart-rate accuracy within plus or minus 5 beats per minute against ECG reference, at 1 metre distance, on a calm stationary subject.
These numbers improve every year as ML model training data and radar SoC compute both expand. The 95 percent recall figure today was 88 percent five years ago.
MDR Class IIa Regulatory Pathway
A contactless fall-detection product intended for use in elderly-care settings and marketed with health-outcome claims falls under the EU Medical Device Regulation (MDR 2017/745). Classification is most commonly Class IIa for a device that alerts on a clinically relevant event without itself making a therapeutic decision.
The technical file required by a notified body for Class IIa covers:
- Quality Management System compliant with ISO 13485:2016. Required from the legal manufacturer.
- Risk management per ISO 14971:2019, including risk control measures traceable to the design.
- Electrical safety per IEC 60601-1, particularly the 4th edition supplemented by IEC 60601-1-11 for the home healthcare environment.
- Software lifecycle per IEC 62304, typically Class B for a fall-detection device whose failure could result in a missed emergency response.
- Usability engineering per IEC 62366-1, with summative usability validation involving genuine end-users (residents and caregivers).
- Clinical evaluation per MDR Annex XIV, including a clinical investigation if equivalent device data is insufficient.
- Cybersecurity per MDCG 2019-16 and IEC 81001-5-1.
- Radio compliance per CE RED with ETSI EN 305 550, parallel to the medical CE marking.
Our work on Class IIa programs typically runs alongside our functional safety engineering team for the IEC 62304 software lifecycle, and integrates with the client's regulatory affairs function from the first scoping workshop.
Integration With Care Platforms and Nurse Call Systems
A fall-detection sensor in isolation has no clinical value. It needs to alert a person who can respond. Integration patterns we have engineered:
- Nurse call systems (Hill-Rom NaviCare, Ascom, RAULAND-BORG). Most use either an analogue dry-contact closure or an SIP-based notification protocol. Direct integration via local gateway is standard.
- Modern care platforms (Caretower, KISI, Tunstall Connected Care). These typically expose REST and webhook APIs. The radar publishes events to the cloud platform, which routes alerts based on configurable rules.
- Family notification apps. Lighter-weight Bluetooth-LE plus phone-app stacks for independent-living deployments where the responder is a family member rather than a professional caregiver.
- Smart-home hubs for IoT-tier consumer deployments. Matter is the emerging open standard; HomeKit and Google Home are available where licensed.
The integration layer is part of the Class IIa scope. Failures of the alert pathway directly affect clinical outcome, so the connectivity stack is subject to the same risk-management discipline as the radar firmware.
Privacy by Physics: Why Radar Is GDPR-Friendly for Bedrooms
Camera-based monitoring is rejected by elderly residents and by data-protection regulators in private spaces. Radar at 60 GHz is fundamentally a different proposition. The data leaving the sensor consists of point clouds, motion classifications, and aggregate health metrics. There is no image. There is no identifiable feature. A range-Doppler-derived event cannot identify the individual it concerns.
Under the EU GDPR, the absence of identifiable data simplifies the controller-processor relationships and reduces the regulatory exposure of the care operator. Under HIPAA in the United States, the same logic applies to Protected Health Information. The privacy story is what unlocks installation in bedrooms and bathrooms, which is exactly where falls happen most.
Engineering this correctly requires that the radar product never expose raw data outside the device, never store individual-resident point clouds beyond the post-event verification window, and offer a clearly documented data flow to the data-protection officer. Our mmWave occupancy sensing work uses the same principles, and many of the design patterns transfer cleanly.
Engineering for Long-Term Reliability in 24/7 Operation
Fall-detection products operate continuously for many years, often in residential settings without IT support. Reliability engineering is therefore central to the product cost-of-ownership.
- MTBF target above 50,000 hours. Component selection (no electrolytic capacitors in the signal path, conservative thermal design) supports this.
- Watchdog supervision at three levels: hardware, firmware, and connectivity. A radar that hangs without raising its own watchdog alert is worse than a radar that fails closed.
- Configurable self-test schedule. A nightly automated self-test (when residents are asleep and minimal motion is expected) verifies RF chain, signal-processing, classifier, and connectivity. Failures escalate to the operator before they become missed-fall events.
- Field upgradeability via signed over-the-air updates. Classifier improvements arrive regularly; deploying them without site visits is essential to clinical-quality maintenance.
- End-of-life predictability documented in the technical file. The legal manufacturer is liable for safety throughout the device lifetime. Designing for a known, planned end-of-life is part of MDR compliance, not optional.
Frequently Asked Questions
How accurate is radar fall detection?
A well-engineered radar fall detector achieves above 95 percent fall recall with under 1.5 false alarms per occupied room per week, under typical bedroom and living-room conditions. Accuracy depends heavily on the training data used for the classifier, the mounting geometry, and the post-event verification logic that filters lying-down events from genuine falls.
Can mmWave radar measure heart rate?
Yes, at short range. A 60 GHz radar can extract heart rate from the phase modulation of returns from the chest at distances up to 2 metres. Respiration is easier and works at up to 4 to 6 metres. Both signals are extracted as a by-product of the same radar that performs fall detection, with no additional hardware.
Is contactless fall detection MDR-certifiable?
Yes. Under the EU Medical Device Regulation (MDR 2017/745), a contactless fall detection product intended to support emergency response in elderly-care settings is typically classified as Class IIa. The technical file is built around ISO 14971 risk management, IEC 60601-1 electrical safety, IEC 62304 software lifecycle, and clinical evaluation under MDR Annex XIV.
What is the false-alarm rate of radar fall detection?
Below 1.5 false alarms per occupied room per week is achievable with proper engineering. Most false alarms come from intentional lying-down on a bed or sofa with rapid descent. A confidence threshold combined with a 15 to 30 second post-event verification window suppresses most of these. Caregiver feedback during pilot installations is the most important input to tuning the classifier.
How does radar tell a fall from someone lying down to rest?
By the vertical velocity profile and the post-event behaviour. A fall has a steep downward velocity transient followed by a static prone or supine pose. Lying down on a bed has a slower, controlled descent and usually ends in a specific known location (the bed). The classifier combines vertical velocity, pose, dwell time, and location context to distinguish the two events.
Discuss your fall-detection product
Thirty minutes with our principal engineer. We will scope the MDR pathway, technical architecture, and clinical-pilot design for your product.
Lukas Hofstätter
Founder and Managing Director, HALready GmbH
Lukas leads HALready's embedded engineering practice, including medical device projects under the EU MDR. His radar fall-detection work spans technical architecture, IEC 62304 software lifecycle, ISO 14971 risk management, and notified-body coordination for Class IIa devices.