Fall detection system and method

Abstract

The present disclosure provides a fall detection system comprising a wearable sensor device configured to capture movement data of a user. The wearable sensor device may include a housing with an inertial measurement unit (IMU) configured to generate movement data and a wireless communication module configured to transmit the movement data. A fastener may be configured to releasably attach the housing to the user. The system may include one or more processors configured to receive the movement data, process the received movement data using a first neural network model and a second neural network model operating in parallel, determine outputs from the first and second neural network models, and determine whether a fall event has occurred based at least on the outputs from the first and second neural network models.

Description

FALL DETECTION SYSTEM AND METHODRELATED APPLICATIONS[1] This application claims priority to U.S. Provisional Patent Application 63 / 612,608, the entire contents of which is incorporated by reference herein in its entirety.FIELD OF THE DISCLOSURE[2] The present disclosure relates to the field of fall detection and human movement assessment using wearable devices.BACKGROUND[3] Falls among the elderly and individuals with certain medical conditions are a significant health concern, often leading to serious injuries, loss of independence, and decreased quality of life. As the proportion of elderly individuals in the global population increases, the need for effective fall detection and prevention systems has become increasingly important. Traditional fall detection methods have relied on manual activation of alert systems or continuous monitoring by caregivers, which can be impractical and may not provide timely assistance in the event of a fall.[4] Existing fall detection solutions also present various challenges that affect their practical implementation and adoption. Some existing solutions may raise privacy concerns, particularly in sensitive areas such as bedrooms and bathrooms where falls frequently occur. Other existing solutions may experience accuracy limitations due to their single-point motion detection and may generate false positives during normal daily activities. Other limitations with existing solutions include limited battery life requiring frequent recharging and restricted detection range. The development of effective fall detection systems may involve addressing and balancing these various considerations to provide reliable and practical solutions for users and caregivers.[5] As the global population ages, there is also a significant economic incentive to address the problem of undetected falls among the elderly and individuals with certainmedical conditions. The average cost associated with an undetected fall may be very high, encompassing medical expenses, rehabilitation costs, and potential long-term care needs. In some cases, these costs may extend beyond immediate medical treatment to include extended hospital stays, specialized equipment, and ongoing support services. Additionally, undetected falls may lead to complications that further increase healthcare costs and reduce quality of life for affected individuals. By implementing effective fall detection systems, healthcare providers and caregivers may reduce these costs, improve patient outcomes, and alleviate the economic burden on healthcare systems and families.[6] The integration of artificial intelligence and machine learning techniques into fall detection systems has the potential to improve accuracy and reduce false positives. These advanced algorithms can analyze complex movement patterns and contextual information to better differentiate between fall events and normal activities.[7] Energy efficiency and user comfort are additional considerations in the design of wearable fall detection devices. Extended battery life is crucial for ensuring continuous monitoring, particularly for elderly users who may have difficulty regularly charging devices and for caretakers who may have limited time. Furthermore, the wearable nature of these devices necessitates a compact and unobtrusive design that does not interfere with the user's daily activities or cause discomfort.[8] As the field of fall detection technology continues to evolve, there is an ongoing need for improved systems that can provide accurate, reliable, and timely detection of fall events while maintaining user comfort and device longevity. Advancements in this area have the potential to significantly enhance the safety and independence of vulnerable individuals, as well as provide peace of mind to their caregivers and loved ones.SUMMARY[9] The following description presents a simplified summary in order to provide a basic understanding of some aspects described herein. This summary is not an extensiveoverview of the claimed subject matter. It is intended to neither identify key or critical elements of the claimed subject matter nor delineate the scope thereof.

[0010] According to aspects of the present disclosure, a fall detection system is provided. The fall detection system includes a wearable sensor device configured to capture movement data of a user. The wearable sensor device includes a housing including an inertial measurement unit (I MU) configured to generate movement data and a wireless communication module configured to transmit the movement data. The wearable sensor device also includes a fastener configured to releasably attach the housing to the user. The fall detection system further includes one or more processors configured to receive the movement data, process the received movement data using a first neural network model and a second neural network model operating in parallel, determine an output from the first neural network model and an output from the second neural network model, and determine whether a fall event has occurred based at least on the outputs from the first and second neural network models. The first neural network model is adapted to capture spatial and temporal features of the movement data. The second neural network model is adapted for image-based analysis of the movement data.

[0011] According to aspects of the present disclosure, the fall detection system may include one or more of the following features. The one or more processors may be further configured to determine whether the fall event has occurred based at least on a weighted combination of the outputs from the first and second neural network models. The outputs from the first and second neural network models may be weighted differently in determining whether the fall event has occurred. The one or more processors may process the received movement data using a third neural network model that operates in parallel to the first and second neural network models. The one more processors may determine an output from the third neural network model and determine whether the fall event has occurred based at least on a weighted combination of outputs from the first,second, and third neural network models. The fall detection system may further include a third neural network model, wherein when the first neural network model and the second neural network model disagree on a classification of the movement data, the third neural network model may be configured to side with either the first neural network model or the second neural network model to determine whether the fall event has occurred. The one or more processors may be further configured to receive time and location data, determine a level of isolation of the user based on the time and location data, and determine whether the fall event has occurred based at least on the outputs from the first and second neural network models and on the determined level of isolation. The I MU may comprise a 3-axis accelerometer and a 3-axis gyroscope, and the movement data may be 6-axis movement data. The wearable sensor device may further comprise a proximity sensor configured to detect whether the device is being worn by the user, wherein the wearable sensor device may be configured to automatically adjust an operation mode based on input from the proximity sensor, a manual activation button configured to allow the user to manually trigger an alert, and a barometric pressure sensor configured to detect changes in atmospheric pressure. The one or more processors may be further configured to provide notification of a determined fall event. The one or more processors may be further configured to analyze the historical movement data of the user and train the first and second neural network models based on the analyzed historical movement data. The fall detection system may further include a gateway device configured to receive the movement data from the wireless communication module of the wearable sensor device and transmit both the movement data and location data relating to a location of the gateway device to the one or more processors, and a server configured to communicate with the gateway device, wherein the server may include the one or more processors, wherein the system is further configured to emit a verbal message to the user after determining a fall event has occurred and prevent transmission of a fall notification if theuser responds to the verbal message that assistance is not needed. The fall detection system may further include a database configured to store one or more patient-specific parameters, wherein the one or more processors may be further configured to determine whether the fall event has occurred based at least on the outputs from the first and second neural network models and on the one or more patient-specific parameters stored in the database.

[0012] According to aspects of the present disclosure, a method for fall detection is provided. The method includes capturing movement data of a user using a wearable sensor device. The wearable sensor device includes a housing including an IMU configured to generate movement data and a wireless communication module configured to transmit the movement data. The wearable sensor device also includes a fastener configured to releasably attach the housing to the user. The method further includes receiving the movement data, processing the received movement data using a first neural network model and a second neural network model operating in parallel, determining an output from the first neural network model and an output from the second neural network model, and determining whether a fall event has occurred based at least on the outputs from the first and second neural network models. The first neural network model is adapted to capture spatial and temporal features of the movement data. The second neural network model is adapted for image-based analysis of the movement data.

[0013] According to aspects of the present disclosure, the method may include one or more of the following features. Determining whether the fall event has occurred may be based at least on a weighted combination of the outputs from the first and second neural network models. The outputs from the first and second neural network models may be weighted differently in determining whether the fall event has occurred. The method may further include processing the received movement data using a third neural network model operating in parallel to the first and second neural network models, determining an outputfrom the third neural network model, and determining whether the fall event has occurred based at least on a weighted combination of the outputs from the first neural network model, the second neural network model, and the third neural network model. The method may further include, when the first neural network model and the second neural network model disagree on a classification of the movement data, using a third neural network model to side with either the first neural network model or the second neural network model to determine whether the fall event has occurred. Determining whether the fall event has occurred may be based at least on the outputs from the first and second neural network models and on location and time data. The IMU may comprise a 3-axis accelerometer and a 3-axis gyroscope, and the movement data may be 6-axis movement data. The wearable sensor device may further comprise a proximity sensor configured to detect whether the device is being worn by the user, and the method may further comprise automatically adjusting an operation mode based on input from the proximity sensor. The wearable sensor device may further comprise a manual activation button configured to allow the user to manually trigger an alert. The wearable sensor device may further comprise a barometric pressure sensor configured to detect changes in atmospheric pressure.

[0014] Further variations encompassed within the systems and methods are described in the detailed description of the invention below.BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings, which are incorporated in and constitute a part of this specification, show certain aspects of the subject matter disclosed herein and, together with the descriptions, help explain some of the principles associated with the disclosed implementations.

[0016] FIG. 1A depicts a first perspective view of a wearable sensor device according to aspects of the present disclosure.

[0017] FIG. 1 B depicts a second perspective view of a wearable sensor device according to aspects of the present disclosure.

[0018] FIG. 1 C depicts a third perspective view of a wearable sensor device according to aspects of the present disclosure.

[0019] FIG. 2 depicts a block diagram of a fall detection system according to aspects of the present disclosure.

[0020] FIG. 3 depicts a flow diagram of a process for fall detection according to aspects of the present disclosure.DETAILED DESCRIPTION

[0021] While aspects of the subject matter of the present disclosure may be embodied in a variety of forms, the following description and accompanying drawings are merely intended to disclose some of these forms as specific examples of the subject matter. Accordingly, the subject matter of this disclosure is not intended to be limited to the forms or embodiments so described and illustrated.

[0022] Referring to Figs. 1A-1C, a wearable device 100 may be a wearable sensor device configured to capture movement data of a user. The wearable device 100 may include a housing, which may be composed of a cover 101 and a base 102. The cover 101 and base 102 may be rectangular-shaped components with rounded corners, although in some cases, they may have other shapes. The cover 101 may be a removable or attachable part of the wearable device 100, allowing access to internal components of the wearable device 100. The base 102 may serve as the main body of the wearable device 100, housing various components of the wearable sensor device.

[0023] The wearable device 100 may also include a fastener 103, which may be configured to releasably attach the housing to the user. In some aspects, the fastener 103 may be a clip that allows the device to be securely attached to a user's clothing article, such as a belt or waistband of the user’s pants or other clothing article. In some aspects,the clip may be designed with a spring-loaded mechanism to provide a firm grip while allowing easy attachment and removal. The clip may be adjustable to accommodate different thicknesses of fabric or materials. In some cases, alternatives to clips may be used as the fastener 103. These alternatives may include magnetic attachments, which allow for quick and easy placement and removal of the device. Hook-and-loop fasteners may also be utilized, providing a customizable and secure attachment method. In some implementations, the fastener 103 may be a lanyard or necklace-style attachment, allowing the device to be worn around the neck. For users with specific needs or preferences, the fastener 103 may be interchangeable, allowing different attachment mechanisms to be used with the same wearable device. This may provide flexibility in how and where the device can be worn, accommodating various clothing styles and user activities.

[0024] The base 102 may also include one or more connection points or openings to ensure a tight and secure fit between the base 102 and the cover 101 when the wearable device 100 is assembled. In some aspects, the openings may be screw ports 104. The screw ports 104 may allow for the insertion of screws or other fastening mechanisms to firmly connect the cover 101 to the base 102, enhancing the overall structural integrity of the wearable device 100. The screw ports 104 may also facilitate easy access to the interior of the wearable device 100 for maintenance, repairs, or battery replacement when necessary. By removing the screws from the screw ports 104, the cover 101 may be easily detached from the base 102, allowing authorized personnel to access the internal components of the wearable sensor device. This design feature may enhance the serviceability and longevity of the wearable device 100, as it allows for routine maintenance and component replacement without compromising the device's structural integrity during normal use. In some cases, the number and placement of the screw ports 104 may be optimized to provide a balance between secure attachment and ease ofaccess. The screw ports 104 may be strategically positioned around the perimeter of the base 102 to distribute the securing force evenly and minimize the risk of water or dust ingress. Additionally, the screw ports 104 may be designed to accommodate different types of screws or fasteners, allowing for flexibility in the assembly and maintenance processes.

[0025] In some aspects, the base 102 may include one or more sensor mounts. The sensor mount may include a coupling point using, for example, adhesives or screws to provide secure attachment of sensors to the wearable device 100. This secure attachment may help maintain the integrity of the sensor casing and protect the sensitive electronic components housed within it, such as the IMU, proximity sensor, and wireless communication module(s) (as described in reference to Fig. 2). In some aspects, the coupling point may be designed to allow for easy replacement or adjustment of sensors while still providing a stable mounting surface. The housing and / or the sensor mount may be constructed from a durable material, such as acrylonitrile butadiene styrene (ABS), which may provide excellent impact resistance, structural strength, and dimensional stability. In some aspects, other durable materials may be used for the housing and / or sensor mount, such as polycarbonate, nylon, or reinforced thermoplastics, depending on specific requirements for strength, weight, and manufacturability. The use of durable materials may help protect the wearable device and its internal components from physical damage and ensure consistent performance over time.

[0026] In some aspects, the wearable device 100 may incorporate a Personal Emergency Response System (PERS) or manual activation button 118. This button may be positioned on the exterior of the device housing in a location that is easily accessible to the user. The PERS button 118 may allow the user to manually trigger an alert or call for assistance in situations where they feel unsafe or require immediate help, even if a fall has not been detected by the device's sensors.

[0027] The inclusion of the PERS button 118 may provide an additional layer of safety and peace of mind for users, allowing them to quickly and easily request help in various emergency situations beyond fall events. This feature may be particularly beneficial for users with medical conditions that may require urgent attention, orforthose who live alone and want an easily accessible way to call for help.

[0028] In some aspects, the PERS button 118 may also be used to confirm falls after a fall has been detected by the system. This feature may allow users to provide immediate feedback and verification of a detected fall event, improving the accuracy of the fall detection system and reducing false alarms.

[0029] In some aspects, the PERS button 118 may be designed with tactile feedback or a raised surface to make it easy for users to locate and activate by touch alone, which may be beneficial in emergency situations where the user's vision or motor control may be impaired. The button may also be designed to require a specific action, such as a long press or double-click, to prevent accidental activation.

[0030] Referring to Fig. 2, a fall detection system is depicted. In some aspects, the system may include a wearable sensor device 100, a gateway device 130, and a server 140, with the server 140 hosting one or more processors 142. The server 140 may be communicatively coupled to a database 160, a client device 137, and the gateway device 130 via a network 150, where the network 150 may be the Internet, a local area network (LAN), a wide area network (WAN), or another suitable wired or wireless communication network.

[0031] The wearable device 100 may be configured to capture the movement data of a user. In some aspects, the wearable device 100 may correspond to the wearable device shown in and described with respect to Figs. 1A-1C. The wearable device 100 may include an inertial measurement unit (IMU) 112 configured to generate movement data. In some aspects, the IMU 112 may comprise a 3-axis accelerometer and a 3-axisgyroscope, and the movement data may be 6-axis movement data. In some aspects, the IMU 112 may be implemented as a 6-axis sensorthat outputs 6-axis sensor data. In some aspects, the IMU 112 may have other configurations and the movement data may have other forms.

[0032] In some aspects, the wearable device 100 may further comprise a proximity sensor 114 configured to detect whether the device is being worn by the user. The proximity sensor 114 may use various sensing technologies, several examples of which are discussed below. In some aspects, the wearable device 100 may be configured to automatically adjust its operation mode based on input from the proximity sensor 114. The proximity sensor 114 may detect whether the wearable device 100 is being worn by the user. When the proximity sensor 114 detects that the device is not being worn, the wearable device 100 may enter a low-power mode to reduce battery usage while still maintaining some monitoring capabilities. In this low-power mode, the device may shut off data-gathering capabilities and Bluetooth Low Energy (BLE) sending, while keeping the proximity sensor 114 active. This may allow the device to detect when it is worn again and reactivate other processes. When the proximity sensor 114 detects that the device is being worn, the wearable device 100 may exit the low power mode and resume full functionality. This power management approach may help extend battery life and improve overall energy efficiency of the wearable device 100, and may also help minimize false positives.

[0033] In some aspects, the proximity sensor 114 may be a micro switch or a miniature snap-action switch. In some aspects, the micro switch may include a pre-tensioned spring connected to a pivoting arm that snaps between two stable positions when a force threshold is reached. The micro switch may include terminals such as Common (COM) and Normally Open (NO) to detect whether the wearable device is being worn. When the wearable device is attached to the user's clothing using the fastener, the physical pressuremay actuate the micro switch, changing its state to indicate the device is being worn. The micro switch may be constructed with conductive metal contacts, such as gold, silver or copper alloys, which may provide low resistance, corrosion resistance, and durability under repeated mechanical stress. The housing of the micro switch may be made from thermoplastics or metal materials. In some aspects, the micro switch may consume minimal power since it operates mechanically by redirecting current flow rather than drawing power directly. When the force on the micro switch is released, such as when the device is removed, the spring mechanism may return the switch to its original position, indicating the device is no longer being worn. This mechanical detection method may provide reliable wear status monitoring while maintaining energy efficiency.

[0034] In some aspects, the proximity sensor 114 may be a magnetic Hall sensor. The magnetic Hall sensor may detect the presence and strength of magnetic fields and may operate based on the Hall effect. A decrease in magnetic field strength may indicate contact with the user, enabling activity tracking, navigation, and device monitoring. In some aspects, the sensor may be positioned to face the device's fastener (e.g., a clip-on attachment). A magnet may be installed within the hook-shaped component, aligning directly with the Hall sensor. When the wearable device 100 is not in use, the sensor may remain unobstructed, allowing it to detect the magnetic field generated by the magnet. Once the wearable device 100 is clipped onto the user’s clothing, the material may partially obstruct the sensor, resulting in a reduction of the magnetic field strength. This decrease in magnetic field strength may correspond to a lower voltage output from the Hall sensor, providing a signal indicating that the wearable device 100 is being worn by the user. Calibration, shielding, and alignment techniques may also be applied to maintain consistent results and reduce potential disturbances (e.g., magnetic interference from nearby objects or electronic devices).

[0035] The sensitivity of the Hall sensor may be selected or designed to enable the detection of subtle position shifts and even slight changes in magnetic field proximity, which may be useful for determining device contact and motion. The sensitivity may refer to the smallest detectable change in magnetic field strength, which may range for example from 14 - 100 mV / mT. High sensitivity may be beneficial in wearable applications. The Hall sensor may be low-power (e.g., consuming 5 to 15 mA), which may be beneficial for wearable devices that require minimal battery drain. Temperature variations may affect both the magnetic properties of nearby materials and the sensor’s electronic components, so this feature ensures consistent readings across diverse conditions. The Hall sensor may integrate temperature compensation circuits to improve accuracy and consistency even as ambient temperatures fluctuate.

[0036] In some aspects, the proximity sensor 114 may be a pressure sensor such as a force sensitive resistor. In some aspects, the force sensitive resistor may include a thin, flexible sensor that changes electrical resistance based on applied pressure or force. The force sensitive resistor may comprise a conductive polymer material layer that becomes more conductive as pressure increases, and a spacer layer that may help ensure consistent readings. When pressure is applied, such as from clothing when the device is clipped to a user's waistband, more of the conductive material may make contact, which may decrease the electrical resistance. In some aspects, the force sensitive resistor may be incorporated into a voltage divider circuit, where the output voltage corresponds to the applied pressure. This voltage may be read and processed by a microcontroller. The force sensitive resistor may be compact enough to fit within the device housing while still providing sufficient sensitivity to detect subtle pressure changes from clothing contact.

[0037] In some aspects, the force sensitive resistor may be mounted on an underside portion of the fastener to measure upward force exerted by clothing when the device is clipped to a waist or similar location. When the fastener is removed from clothing, thepressure may no longer be detected, providing an indication that the device is not being worn. In some aspects, the force sensitive resistor may be sampled at intervals rather than continuously monitored to help optimize power consumption. In some aspects, the sensitivity of the force sensitive resistor may be calibrated to detect appropriate pressure thresholds for determining device wear status. The force sensitive resistor may be connected to an analog input pin of a microcontroller for pressure data reading and processing.

[0038] In some aspects, alternative or additional types of proximity sensors may be used. In some aspects, an infrared sensor may be used. In some aspects, the infrared sensor may include an infrared emitter and detector pair positioned within the housing of the wearable device. The infrared emitter may emit infrared light, while the detector may measure the reflected infrared light. When the device is worn, the clothing material may reflect the infrared light back to the detector, indicating the presence of a surface near the sensor. The intensity of the reflected infrared light may vary depending on the proximity and material properties of the reflecting surface. In some aspects, the infrared sensor may be configured to operate at specific sampling intervals to conserve power while maintaining reliable wear detection. The sensor may include filtering mechanisms to reduce interference from ambient light and may incorporate automatic gain control to adjust for different clothing materials and environmental conditions. In some aspects, the infrared sensor's detection threshold may be adjustable to accommodate different wearing configurations and user preferences.

[0039] In some aspects, the wearable device 100 may include additional sensors. Additional sensors may include one or more pressure sensors for monitoring pressure exerted on the device and / or a temperature sensor for monitoring the user's medical condition and device usage. In some aspects, an acoustic sensor with a microphone may be incorporated in the wearable device 100 and may be used to enhance the accuracy offall detection. The distinct noise generated during a fall may be captured by the acoustic sensor, providing additional data points for the system to analyze and improve the overall fall detection performance. The microphone may also enable user communication when a voice assistant or other communication module is not present.

[0040] In some aspects, the wearable device 100 may include a barometric pressure sensor 116. The barometric pressure sensor 116 may be configured to measure atmospheric pressure, which can be used to detect minor changes in altitude or elevation. This sensor may provide insightful data that can be integrated into the fall detection system to enhance its accuracy and functionality. For example, the barometric pressure sensor 116 may help distinguish between different types of movements, such as sitting down abrupty, walking up or down stairs, riding in an elevator, stumbling, or experiencing a sudden change in elevation due to a fall. For example, the barometric pressure sensor 116 may monitor vertical displacement data, which may help distinguish between non-fall activities (e.g., dropping onto a sofa) and actual falls. The data from the barometric pressure sensor 116 may be combined with the movement data from the IMU 112 to provide a more comprehensive understanding of the user's activities and potential fall events. The inclusion of the barometric pressure sensor 116 in the wearable device 100 may further improve the system's ability to accurately detect falls and reduce false positives, enhancing the overall effectiveness of the fall detection system. The barometric pressure 116 sensor may also help monitor the post-fall position by tracking any further changes in altitude, which may be useful for determining if the person has remained on the ground after a fall or if they have gotten up.

[0041] In some aspects, the barometric pressure sensor 116 may incorporate MEMS technology (micro-electro-mechanical-systems) to integrate micro-mechanical and electronics elements onto a single silicon chip, which may make it beneficial for wearable and portable devices. The barometric pressure sensor 116 may measure absolutepressure along with temperature to correct temperature-related fluctuations in pressure readings which may provide more accurate estimations. In some aspects, the barometric pressure sensor 116 may include an integrated temperature sensor that monitors the sensor’s internal temperature, and may apply compensation algorithms to correct pressure readings for temperature-induced errors. This compensation may help ensure stable pressure readings across various environmental conditions. The barometric pressure sensor 116 may support multiple oversampling settings for different levels of resolution and power consumption. Higher oversampling may yield more accurate data but may require more processing power, which may be suitable for applications where accuracy is prioritized over response time. This may be beneficial in systems that include an IMU that provides quick response time. In some aspects, the barometric pressure sensor 116 may also use a filter to smooth out short-term fluctuations in pressure. This may be beneficial in environments with frequent pressure changes, such as moving between different altitudes (e.g., traveling between floors on an elevator) or in weather- affected regions or scenarios (e.g., opening a window).

[0042] In some aspects, the barometric pressure sensor may be processed using the same (or similar) neural network models used for processing the IMU movement data (as will be discussed in greater detail below), as both types of data represent time series measurements.

[0043] In some aspects, the wearable device 100 may include one or more wireless communication modules. In some aspects, the wearable device 100 may include a Bluetooth low energy (BLE) wireless controller 111 configured to transmit captured multiaxis movement data to a gateway device 130 using BLE and SPI communication protocols. In some aspects, the wearable device 100 may utilize alternative wireless communication technologies and protocols for transmitting the captured multi-axis movement data to the gateway device 130. These alternatives may include Wi-Fi, Zigbee,Z-\Nave, or other low-power wireless communication standards. The choice of wireless technology may depend on factors such as power consumption, range requirements, and compatibility with existing infrastructure. In some implementations, the wearable device 100 may support multiple wireless communication protocols, allowing for flexibility in different usage scenarios or environments. The communication interface may also be designed to be upgradable or modular, enabling the device to adapt to future wireless communication standards or technologies.

[0044] In some aspects, the wearable device 100 may incorporate energy harvesting techniques, such as kinetic energy harvesting using piezoelectric materials and solar energy harvesting via, for example, high-efficiency photovoltaic cells 113, to extend its operational life. A power management system may include a coin cell battery 115 and may be configured to optimize energy consumption based on data from the temperature, proximity, pressure, and / or environment sensors. The device 100 may be powered by the coin cell battery 115, which may have high energy density and long shelf life. In some aspects, the wearable device 100 may utilize alternative power sources or energy storage technologies in addition to or instead of the coin cell battery 115. These may include rechargeable lithium-ion batteries, supercapacitors, thin-film batteries, or wireless power transfer systems, which may provide different combinations of energy density, charging speed, and operational lifetime to suit various use cases and user preferences.

[0045] In some aspects, the wearable sensor device 100 may further include a manual activation button 118 that allows the user to request assistance for non-fall-related emergencies. In some implementations, the wearable device 100 may be connected to or integrated into the user's clothing using a fastener 103, which as described above may be a clip-on mechanism and may provide a discreet and comfortable design.

[0046] In some aspects, the wearable device 100 may include an edge processor 117 with lightweight machine learning algorithms, which may enable rapid data processing and decision-making at the edge of the network and reduce latency.

[0047] In some aspects, the fall detection system may include a gateway device 130. The gateway device 130 may be configured to receive the movement data from the BLE controller 111. The gateway device 130 may also be configured to transmit both the movement data and location data relating to a location of the gateway device 130 to the one or more processors 142. In some aspects, the gateway device 130 may include a BLE wireless controller 131 to facilitate communication with the wearable device 100 using BLE. The gateway device 130 may further include a WiFi controller 132 for transmitting the received data to the server 140 and a 4G / 5G cellular module 134 to provide backup connectivity when WiFi is unavailable. The gateway device 130 may utilize WiFi and NB-loT / LTE protocols for robust and reliable data transmission directly to the Internet or a central server. The gateway device 130 may further include a companion processor 133 coupled to the controllers 131 , 132 and the cellular module 134 to assist in data processing and communication tasks, which may ensure efficient operation and enhance the gateway's capability to manage multiple sensors and data streams. In some aspects, the gateway device 130 may incorporate alternative communication protocols and controllers to enhance its versatility and adaptability in various environments.

[0048] In some aspects, the gateway device 130 may be associated with a specific location based on its placement within a user's environment. This association may allow the system to determine the approximate location of the user based on which gateway device 130 the wearable device 100 is connected to at any given time. The system may utilize Received Signal Strength Indicator (RSSI) to determine which gateway device 130 the wearable device 100 is most strongly connected to, thereby inferring the user's location.

[0049] For example, in a residential setting, multiple gateway devices 130 may be strategically placed throughout the home, each labeled with a specific location identifier such as "living room," "bedroom," or "kitchen." As the user moves through different areas of the home, the wearable device 100 may automatically connect to the nearest gateway device 130 based on the strongest RSSI. This connection information may be used to update the user's location in real-time.

[0050] The movement data captured by the wearable device 100 may be combined with this location information to provide a more comprehensive understanding of the user's activities and potential fall events. For instance, if a fall is detected while the user is connected to the "bathroom" gateway, the system may infer that a fall has occurred in the bathroom and the appropriate authority may be notified for immediate assistance. This combined movement and location data may enhance the accuracy of fall detection and provide valuable context for emergency responders or caregivers.

[0051] In some implementations, the system may also use the changing connections between the wearable device 100 and different gateway devices 130 to track the user's movement patterns throughout their environment. This information may be used to establish baseline activity levels, detect unusual patterns that might indicate health issues, or provide insights into the user's daily routines.

[0052] The location data may also be used to trigger location-specific alerts or actions. In some aspects, the location data can be used for exit detection or wander management in retirement home settings. For example, the location data may be used for exit detection to identify and alert staff when a resident leaves the retirement home, and for wander management to notify staff when a resident leaves their room at inappropriate times. For example, the system may detect when a resident's location changes during times they should be in their room, such as during sleeping hours, allowing staff to be alerted and assist the resident in returning to their room.

[0053] In some aspects, the fall detection system may include a server 140 configured to communicate with the gateway device 130. The server 140 may include one or more processors 142. The one or more processors 142 may be configured to receive the movement data and location data from the gateway device 130.

[0054] In some aspects, the fall detection system may operate through a multi-step process involving data collection, transmission, and analysis. The wearable device 100 may continuously collect movement data using its IMU 112. At predetermined intervals (e.g., every minute), the wearable device 100 may compile this movement data and transmit it to the nearest gateway device 130 using the BLE controller 111 or other wireless communication module.

[0055] The gateway device 130 may receive the movement data from the wearable device 100 and associate it with location information based on the gateway's specific placement within the user's environment. As described above, each gateway device 130 may be assigned a unique identifier or name that corresponds to its location. The gateway device 130 may then package the movement data along with this location identifier and a time indicator (e.g., a timestamp) indicating the time of day when the movement occurred into a file or data packet.

[0056] Upon receiving the movement data, associated location information, and timestamp, the gateway device 130 may transmit this combined data to the server 140 via the network 150. The server 140 may receive these data packets from one or more gateway devices 130 throughout the user's environment. This combination of movement data, location information, and time of day may allow the system to provide a more comprehensive analysis of the user's current activities.

[0057] Once the server 140 receives the data, its processors 142 may initiate the analysis process. The server 140 may utilize one or more neural network models, such as the first neural net 143, second neural net 144, and third neural net 145, to processthe movement data, as described in greater detail below. In some aspects, the server 140 may process the data in real-time as it is received from the gateway device 130. In other cases, the server 140 may batch-process the data at regular intervals or when a certain amount of data has accumulated. The results of this analysis may then be used to determine whether a fall event has occurred, assess the user's activity patterns, or generate other insights about the user's movements and behavior.

[0058] In some aspects, the one or more processors 142 may be configured to receive the movement, location and, time data, including via the gateway device 130 and the network 150. The processors 142 may be configured to process the received movement data using multiple neural network models operating in parallel. In some aspects, two or more neural network models may operate in parallel. In some aspects, three or more neural networks may operate in parallel. The parallel operation of the multiple neural network models may allow the system to analyze the movement data in different ways simultaneously, improving the accuracy and reliability of the fall detection. In some aspects, each neural network model may provide a unique perspective on the movement data.

[0059] A first neural network model, which may be the first neural net 143, may be adapted to capture spatial and temporal features of the movement data. The spatial features may include the position, velocity, or acceleration of the user in three-dimensional space, while the temporal features may include the changes in these spatial features over time. The first neural net 143 may use various machine learning algorithms to capture these features, such as convolutional neural networks, recurrent neural networks, or long short-term memory networks. As a non-limiting example, the first neural net 143 may combine convolutional neural networks (CNN) for spatial feature extraction with long short-term memory (LSTM) networks for temporal sequence modeling, thus forming a CNN-LSTM hybrid model. This architecture may allow the first neural net 143 toeffectively capture both the spatial and temporal aspects of the movement data, which may improve its ability to detect fall events.

[0060] A second neural network model, which may be the second neural net 144, may be adapted for image-based analysis of the movement data. The image-based analysis may involve converting the movement data into a form that can be processed as an image, such as a spectrogram or a time-frequency representation. In some aspects, the second neural net 144 may process the movement data as a graph-like image representation. This representation may include two Y-axes with fixed scales, one for gyroscope data and one for acceleration data, while the X-axis may represent the time elapsed during the data interval packet. The gyroscope and acceleration data from the I MU may be plotted on their respective Y-axes, creating a visual representation of the movement data over time. This graph-like image may allow the second neural net 144 to leverage image processing techniques to analyze patterns and features in the movement data that may be indicative of fall events. By representing the movement data in this visual format, the second neural net 144 may be able to identify subtle patterns or relationships between the gyroscope and acceleration data that may not be as easily detectable through other analysis methods. This approach may complement the spatial and temporal analysis performed by the first neural net 143, enhancing the overall accuracy and reliability of the fall detection system.

[0061] The second neural net 144 may use image processing algorithms, such as convolutional neural networks, to analyze the converted movement data. In some cases, as described above, the second neural net 144 may be able to detect patterns or features in the movement data that are not easily captured by the first neural net 143. As a nonlimiting example, the second neural net 144 may utilize a VGG16-based CNN model to process the movement data represented as images. The VGG16 architecture, which maybe pre-trained on large image datasets, may be adapted for the specific task of fall detection.

[0062] In some aspects, the fall detection system may include a third neural network model, which may be the third neural net 145. The third neural net 145 may operate in parallel to the first and second neural nets 143 and 144 and be configured to act as a tiebreaker or decision-maker when the first neural net 143 and the second neural net 144 disagree on the classification of the movement data. In such cases, the third neural net 145 may analyze the movement data and side with either the first neural net 143 or the second neural net 144 to determine whether a fall event has occurred. This approach may help to resolve conflicting outputs from the other neural networks and improve the overall accuracy of fall detection. As a non-limiting example, the third neural net 145 may be implemented as a DenseNet-based time-series classification model using TensorFlow and Keras.

[0063] In some aspects, the third neural net 145 may be adapted to process sequential movement data through a densely connected convolutional architecture. The model may implement a DenseNet architecture modified for one-dimensional time series data, where each layer may be directly connected to every other layer in a feed-forward fashion. This dense connectivity pattern may allow for improved feature reuse, stronger gradient flow, and more efficient learning. The architecture may include multiple dense blocks, where each block may contain a series of layers performing batch normalization, ReLU activation, and 1 D convolution operations. In some aspects, each layer may receive feature maps from all preceding layers within its dense block, utilizing a specified growth rate (e.g., 16) to control the number of new features added per layer. This design may help the network learn diverse and complementary features from the movement data.

[0064] In some aspects, transition blocks between dense blocks may perform feature map pooling and channel reduction, which may help control the model's complexity andcomputational requirements. In some aspects, the network may employ multiple (e.g., three) dense blocks of increasing depth (e.g., 3, 6, and 12 layers respectively), separated by transition layers that may reduce the feature maps by a specified amount (e.g., 50%). This progressive structure may allow the network to build increasingly abstract representations of the movement patterns while maintaining computational efficiency. The final layers of the network may include global average pooling to reduce spatial dimensions, followed by a dense layer with regularization and dropout. This architecture may be effective at learning hierarchical features from time series movement data while maintaining robustness through its various regularization mechanisms.

[0065] In some aspects, the processors 142 may be configured to determine whether a fall event has occurred based at least on the outputs from the first neural net 143 and the second neural net 144. In some aspects, the determination may involve a weighted combination of the outputs from the first and second neural nets 143 and 144. The determination may involve comparing the outputs to a threshold, combining the outputs in a certain way, or using other decision-making algorithms. In some cases, the outputs from the first and second neural nets 143 and 144 may be weighted differently in determining whether the fall event has occurred.

[0066] In some aspects, the processors 142 may be configured to determine whether a fall event has occurred based at least on the outputs from the first neural net 143, the second neural net 144, and the third neural net 145. In some aspects, the determination may involve a weighted combination of the outputs from the first, second, and third neural nets 143, 144, and 145. The determination may involve comparing the outputs to a threshold, combining the outputs in a certain way, or using other decision-making algorithms. In some cases, the outputs from the first, second, and third neural nets 143, 144, and 145 may be weighted differently in determining whether the fall event has occurred.

[0067] The server 140 may further be configured to determine whether a fall event has occurred based on a weighted combination of outputs from the neural network models 143, 144, 145 and patient-specific parameters stored in the database 160 (further detailed below).

[0068] In some aspects, data to train the three neural network models may be gathered from various sources, such as retirement homes or assisted living facilities. This raw data may then be optimized based on the specific requirements of each neural network model. For example, in the case of the VGG16-based neural network, the raw dataset may be converted into images through a standard format to facilitate image-based analysis. This data preparation process may help ensure that each neural network model receives input data in a format that is optimized for its specific architecture and analysis approach, improving the overall performance and accuracy of the fall detection system.

[0069] In some aspects, the barometric pressure data from the barometric pressure sensor 116 may be processed using the same neural network models used for processing the IMU movement data, as both types of data represent time series measurements. In some aspects, the first neural network model may capture spatial and temporal features of the barometric pressure data, while the second neural network model may analyze image-based representations of the pressure readings. In some aspects, the barometric pressure data may be processed in parallel with the IMU data, allowing the system to simultaneously analyze both data streams. The neural network models may be trained on datasets that include both IMU and barometric pressure data, which may enable the system to identify correlations between changes in movement and atmospheric pressure. This integrated approach to data processing may enhance the system's ability to distinguish between different types of movements and activities, improving the accuracy of fall detection. The third neural network model may also process the barometric pressuredata when acting as a tiebreaker, providing additional context for resolving disagreements between the first and second neural network models.

[0070] In some aspects, the processors 142 may be configured to provide notification of a determined fall event. The notification may be sent to various recipients, such as the user, a caregiver, ora medical professional. In some aspects, after a fall event is detected, the system may initiate an automated emergency response protocol. This protocol may involve sending a message (e.g., such as through a mobile app, an email, a text message) and / or placing a call to emergency staff or services or another appropriate authority, providing critical information about the incident. The system may relay a message indicating that a fall has been detected for a specific resident, along with the precise location of the fall based on the gateway device's identifier. This automated communication may help expedite the response time and ensure that emergency personnel have accurate information about the situation and location, improving the outcome for the fallen individual. The system may also be configured to provide additional relevant details, such as the resident's medical history or specific care instructions, to further assist emergency responders in providing appropriate care.

[0071] In some aspects, the system may include voice assistant integration (e.g., for in-room fall verification) and / or an activity recognition algorithm (ARA) 154 (e.g., for verifying falls outside the room). The notification provided by the system can include the user's location indicated by the gateway 130 for 24 / 7 location monitoring. A notification engine may be configured to send immediate alerts to a dashboard displayed on a humanmachine interface 180 when a fall is detected and record it in the database 160. A calling engine may be configured to contact the appropriate nursing staff or medical services. The human-machine interface (HMI) 180 may serve as a dashboard for families and staff to monitor and analyze the sensor data. In some aspects, nursing staff may access realtime information about each wearable device through the human-machine interface 180,such as the current location of each user or the remaining battery life of their wearable sensor device. This information may help staff monitor resident safety and ensure devices are charged and functioning properly. In some embodiments, an API engine 182 may facilitate the creation, management, and handling of APIs for integrating various components of the fall detection system. In some aspects, fall verification may be conducted using the voice assistant and / or ARA 154, depending on the sensor location.

[0072] In some aspects, the fall detection system may incorporate a voice assistant as a separate device used forfall verification. This voice assistant may be placed in the user's room and may serve as an additional measure to prevent false positives in fall detection. When a potential fall is detected and the user's location is determined to be within their room, the voice assistant may be activated to engage with the user. The voice assistant may ask the user if they have fallen, providing an opportunity for direct user feedback. In some aspects, the gateway device 130 may function as the voice assistant, which may eliminate the need for a separate device. For example, in settings such as a retirement home setting where one gateway device 130 may be placed in each room, the gateway device 130 may include a speaker and microphone to enable voice interaction capabilities. When integrated with voice assistant functionality, the gateway device 130 may communicate directly with users to verify potential fall events. For example, if the system detects a possible fall and determines that the user is in their room, the gateway device 130 may activate its voice interaction features to ask the user if they have fallen and need assistance.

[0073] In some aspects, the system may employ Natural Language Processing (NLP) techniques to interpret the user's response accurately, allowing for nuanced understanding of various affirmative or negative responses. If the user confirms that they have fallen, or if no response is received within a predetermined time frame (which may suggest the user is unconscious or unable to respond), the system may automaticallyinitiate an emergency call or alert. Conversely, if the user indicates that they have not fallen, the system may refrain from sending out an alert, reducing false alarms and unnecessary interventions. This integration of voice assistant technology may enhance the overall accuracy and reliability of the fall detection system, providing an additional layer of verification before emergency protocols are activated.

[0074] In some aspects, the gateway device 130 and / or the voice assistant may provide additional functionalities beyond fall detection and verification, which may help enhance resident well-being and quality of life. In some aspects, the gateway device 130 may serve as an entertainment and communication hub, offering features such as music playback, audiobook streaming, or news updates. These entertainment options may help reduce feelings of isolation and depression among residents, particularly those who may have limited social interactions. The voice assistant capabilities (whether provided by separate device or by the gateway device 130) may also help residents maintain their daily schedules by providing reminders for meals, medications, or activities. In some aspects, the voice assistant capabilities may facilitate communication between residents and their families or caregivers through voice messaging or calls. The voice assistant capabilities may also allow the system to respond to various non-fall-related queries, such as providing weather updates, setting alarms, or answering general questions, which may help residents maintain independence and engagement in their daily activities. These additional features may complement the fall detection capabilities while providing broader support for resident well-being and social connection.

[0075] In some aspects, the processors 142 may be configured to analyze historical movement data of the user. The historical movement data may include past movement data captured by the wearable device 100, and may provide information about the user's typical movements, activity levels, or patterns of behavior. The analysis of the historical movement data may involve various data analysis techniques, such as statistical analysis,time series analysis, or machine learning algorithms. The results of the analysis may be used for various purposes, such as identifying trends or anomalies in the user's movements, predicting future movements, or personalizing the operation of the fall detection system.

[0076] In some aspects, the processors 142 may be configured to train the first neural net 143, the second neural net 144, and / or the third neural net 145 based on the analyzed historical movement data. The training may involve adjusting the parameters of the neural nets 143, 144, and 145 to minimize the difference between the outputs of the neural nets and the actual outcomes, as indicated by the historical movement data. The training may use various machine learning techniques, such as gradient descent, backpropagation, or genetic algorithms. The trained neural nets 143, 144, and 145 may be more accurate and reliable in analyzing the movement data and detecting fall events. In some cases, the training may be performed periodically or continuously to adapt to changes in the user's movements or conditions.

[0077] In some aspects, the system 100 may include a database 160 coupled to the server 140, wherein the database 160 is configured to store various types of data generated and required by the system 100. The stored data may include raw multi-axis movement data from the wearable sensor device 100, processed data from the neural network models 143, 144, 145, user-specific patterns and risk factors, location data, and / or fall event records. The database 160 may also store user profiles, personalized fall prevention plans, and data related to the system's performance and operation. By maintaining a comprehensive database 160, the system 100 may be enabled for longterm data analysis, machine learning model improvements, and the generation of valuable insights for enhancing fall detection.

[0078] In some aspects, to analyze movement data overtime and identify user-specific patterns and risk factors, the server 140 may employ machine learning algorithms thatprocess the historical multi-axis movement data stored in the database 160. These algorithms, which may comprise clustering, anomaly detection, and pattern recognition techniques, may be configured to detect recurring patterns, deviations from normal behavior, and specific characteristics in the user's movement data.

[0079] In some aspects, the machine learning algorithms may be trained on datasets that include separate classifications for different activities of daily living (ADLs), such as walking, standing, sitting, and lying down. By continuously monitoring and analyzing this data in combination with the user's location information, the server 140 can establish a baseline for the user's typical movement patterns and identify the most likely ADL at any given point in time. This may enable the detection of any significant changes or anomalies that may indicate an increased risk of falling. The analysis may take into account various factors, comprising the user's location, environmental conditions, and the specific ADL being performed, and these factors may be stored in the database 160.

[0080] Referring to Fig. 3, a method for fall detection using a wearable sensor device is depicted. The method may include several steps that work together to capture, process, and analyze movement data to detect fall events.

[0081] In step 302, movement data of a user may be captured using a wearable sensor device. In some aspects, this wearable sensor device may correspond to any of the wearable devices described herein. The movement data may be generated by the IMU 112, which may include a 3-axis accelerometer and a 3-axis gyroscope, providing 6-axis movement data.

[0082] In step 304, the movement data may be transmitted from the wearable sensor device to a gateway device. This transmission may occur via one or more wireless communication modules of the wearable device, which may utilize Bluetooth Low Energy (BLE) or other wireless communication protocols. In some aspects, this transmission may occur at predetermined intervals. For example, the wearable sensor device may compilemovement data over a set period, such as every 30 seconds, 1 minute, or 5 minutes, and then transmit the compiled data to the gateway device. The interval length may be adjustable based on various factors, including battery life considerations, the user's activity level, or the perceived risk of falls.

[0083] In some aspects, the wearable sensor device may capture movement data at a specific sample rate, which may be the number of I MU readings per second. For example, the device may capture 20 samples per second. The sample rate may be adjustable based on the level of detail required for the movement data and battery life considerations, as a higher sample rate may provide more detailed information but may also consume more power, reducing the device's battery life. The combination of sample rate and transmission interval may allow for a balance between data granularity and energy efficiency in the fall detection system.

[0084] In step 306, the gateway device may transmit the movement data along with location and time data to a server (e.g., in JSON format). The location data may be associated with the specific placement of the gateway device within the user's environment, while the time data may include a timestamp indicating when the movement occurred. In some aspects, each gateway may labeled as a specific location based on its location. The wearable device may connect to gateways based on signal strength (e.g., RSS I). The gateway the resident is connected to may indicate their location.

[0085] In step 308, the received movement data may be processed using neural network models operating in parallel. This step may involve separate data processing and preparation for each neural net. In some aspects, a sliding window technique may be applied to create time windows of the 6-axis data.

[0086] This step may utilize multiple neural network models, such as the first neural net 143, second neural net 144, and third neural net 145 described earlier. Each neuralnetwork model may analyze the movement data from a different perspective, improving the overall accuracy of fall detection.

[0087] In step 310, an output may be determined from each of the neural network models. These outputs may represent the individual assessments of each neural network regarding the likelihood of a fall event based on the analyzed movement data.

[0088] In some aspects, a Convl D-LSTM model may be used for the first neural net. Feature extraction may be performed on the prepared data to obtain the input for the model. Data augmentation techniques may be applied to enhance the training and realtime prediction capabilities of the model. The Convl D-LSTM model may process the augmented data and generate an output (e.g., a confidence score).

[0089] In some aspects, a VGG16-based CNN model may be used for the second neural net. The data may pass through 10 layers of CNN, each of which may be followed by a Rectified Linear Unit (ReLU) activation function and batch normalization. The output from the CNN layers may be flattened using a Flatten Layer. The flattened data may then be processed by a Classification Dense Layer with ReLU activation and normalization. The data may pass through a Dense Layer with a Softmax activation function, which may generate an output (e.g., a confidence score).

[0090] In some aspects, a DenseNet-based time-series classification model may be used for the third neural net. The third net may output a prediction and be configured to act as a tie-breaker or decision-maker when the first neural net 143 and the second neural net 144 disagree on the classification of the movement data. As with the first and second models, the third model may generate an output (e.g., a confidence score).

[0091] In some aspects, all three neural network models may process the same dataset concurrently. The first neural net 143, second neural net 144, and third neural net 145 may thus operate in parallel, analyzing the same movement data simultaneously. This concurrent processing may allow for real-time analysis and may improve the overallaccuracy and reliability of fall detection. By processing the same dataset concurrently, the system may leverage the strengths of each model simultaneously. This approach may enable the system to capture different aspects of the movement data, leading to more comprehensive and accurate fall detection. The parallel processing may also contribute to faster decision-making, as the system does not need to wait for sequential processing of the data through each model.

[0092] In some aspects, the concurrent processing of the same dataset by all three models may facilitate more robust fall detection. If one model misses a potential fall event, the others may still detect it. Additionally, this approach may allow for dynamic weighting of the models' outputs based on their performance on specific types of movement data or in certain contexts.

[0093] In step 312, whether a fall event has occurred may be determined based on the outputs from the neural network models. This determination may involve a weighted combination of the outputs. The weighted combination may be compared to a predetermined threshold or dynamically adjusting threshold. In some aspects, other decision-making algorithms may be used. The system may also consider additional factors such as location and time data in making this determination.

[0094] In some aspects, a final fall prediction score, Xf, may be calculated by applying a weighted combination of the individual model outputs and scores: Xf = w1*X1 + w2*X2 + w3*X3, where w1 , w2, and w3 are the weights assigned to each model and X1 , X2, and X3 are the outputs or scores generated by each model. In some aspects, if Xf is greater than or equal to a predefined threshold, a fall is detected; otherwise, no fall is detected.

[0095] In some aspects, the outputs from the three neural network models may be weighted differently in the fall detection process. The system may assign different importance or significance to each neural network's output based on various factors. For example, the first neural net's output may be given a higher weight if it has shown betterperformance in capturing spatial and temporal features of fall-like movements. In some aspects, it may be advantageous to give a higher weight to the first neural net's output in scenarios where the first neural net is more accurate. For example, the first neural net may demonstrate superior performance in detecting falls that involve rapid changes in acceleration or orientation, which are typical of many fall events. In some aspects, the second neural net's output may be weighted more heavily in scenarios where imagebased analysis of movement data has proven more accurate.

[0096] The weighting of the neural network outputs may be dynamic and adaptable. In some cases, the system may adjust the weights based on the specific user's movement patterns, the time of day, or the location where the movement is detected. For instance, if historical data shows that one neural network model performs better for a particular user or in a specific environment, its output may be given more weight in those circumstances. In some aspects, the system may employ a machine learning algorithm to optimize the weighting of the neural network outputs over time. This algorithm may analyze the accuracy of fall detection results and adjust the weights accordingly to improve overall system performance. The weighting mechanism may also incorporate feedback from confirmed fall events or false alarms, allowing the system to refine its decision-making process continually.

[0097] In the event of a detected fall, the system may initiate fall verification based on the user's location. In some aspects, if the user is located in a room, a voice assistant may be used for verification. The voice assistant may prompt the user to confirm if they are safe. If the user confirms they are safe, no further action may be taken. However, if the user indicates they need help or there is no response for a predetermined amount of time (e.g., 15 seconds), a fall alert may be sent out (as described below). If the user is located outside the room, an ARA may be used for verification. The ARA may use the user's location and time to assess their level of isolation. It may increase the likelihood ofthe model detecting a fall if the location and time suggest that the user is in a vulnerable area.

[0098] For example, in some aspects, the ARA may utilize location and time data to provide context and adjust the fall detection sensitivity. For example, if the user is in the dining room at 12 PM (e.g., at or around lunchtime), the ARA may recognize that there are likely many people around who could provide assistance in case of a fall. In this scenario, the ARA may multiply all the scores associated with this location and time combination by a factor, slightly reducing the likelihood of fall detection. This adjustment may help to reduce false positives in situations where immediate assistance is more readily available and fall detection may therefore be less valuable. The ARA may apply similar contextual adjustments for various locations and times throughout the day, considering factors such as staff schedules, meal times, and typical activity patterns in different areas of the facility. By incorporating this contextual information, the system may improve its ability to distinguish between genuine fall events and other movements, reducing both false positives and false negatives in fall detection. In some aspects, a dictionary with location-time combinations and corresponding ARA impact factors may be used to enhance the contextual analysis of fall detection. This dictionary may contain predefined entries that associate specific locations and time periods with corresponding ARA impact factors. The system may utilize this dictionary to quickly look up and apply the appropriate ARA impact factor based on the current location and time of the user. This approach may allow for efficient and consistent application of contextual adjustments across various scenarios, improving the overall accuracy of fall detection. The dictionary may be periodically updated based on new insights, changing schedules, or feedback from caregivers to ensure its relevance and effectiveness in different care environments.

[0099] Finally, in step 314, if a fall event is determined to have occurred and an alert deemed necessary, a notification of the determined fall event may be provided. Asdescribed above, the decision to send an alert may depend on the value outputted by the ARA 154. The alert or notification may be sent to various recipients, such as the user, a caregiver, or emergency services, and may include relevant information about the fall event, such as its location and time. For example, the system may relay a message indicating that a fall has been detected for a specific resident, along with the precise location of the fall based on the gateway device's identifier.

[0100] The method may be implemented in a continuous or periodic manner, allowing for real-time or near-real-time fall detection. In some aspects, the method may also incorporate feedback mechanisms, allowing the system to learn and improve its fall detection accuracy over time based on confirmed fall events and false alarms.Examples

[0101] Exemplary embodiments of the systems and methods disclosed herein are described in the numbered paragraphs below.A1. A fall detection system comprising: a wearable sensor device configured to capture movement data of a user, wherein the wearable sensor device includes: a housing including: an inertial measurement unit (IMU) configured to generate movement data; and a wireless communication module configured to transmit the movement data; and a fastener configured to releasably attach the housing to the user; one or more processors configured to: receive the movement data; process the received movement data using a first neural network model and a second neural network model operating in parallel, wherein:the first neural network model is adapted to capture spatial and temporal features of the movement data; and the second neural network model is adapted for image-based analysis of the movement data; determine an output from the first neural network model and an output from the second neural network model; and determine whether a fall event has occurred based at least on the outputs from the first and second neural network models.A2. The fall detection system of A1 , wherein the one or more processors are further configured to determine whether the fall event has occurred based at least on a weighted combination of the outputs from the first and second neural network models.A3. The fall detection system of A2, wherein the outputs from the first and second neural network models are weighted differently in determining whether the fall event has occurred.A4. The fall detection system of any of A1-A3, wherein the one or more processors are further configured to: process the received movement data using a third neural network model operating in parallel to the first and second neural network models; determine an output from the third neural network model; and determine whether the fall event has occurred based at least on a weighted combination of the outputs from the first neural network model, the second neural network model, and the third neural network model.A5. The fall detection system of any of A1-A3, further comprising a third neural network model, wherein when the first neural network model and the second neural network model disagree on a classification of the movement data, the third neuralnetwork model is configured to side with either the first neural network model or the second neural network model to determine whether the fall event has occurred.A6. The fall detection system of any of A1-A5, wherein the one or more processors are further configured to: receive time and location data; determine a level of isolation of the user based on the time and location data; determine whether the fall event has occurred based at least on the outputs from the first and second neural network models and on the determined level of isolation.A7. The fall detection system of any of A1-A6, wherein the IMU comprises a 3- axis accelerometer and a 3-axis gyroscope, and wherein the movement data is 6-axis movement data.A8. The fall detection system of any of A1-A7, wherein the wearable sensor device further comprises: a proximity sensor configured to detect whether the device is being worn by the user, wherein the wearable sensor device is configured to automatically adjust an operation mode based on input from the proximity sensor; a manual activation button configured to allow the user to manually trigger an alert; and a barometric pressure sensor configured to detect changes in atmospheric pressure.A9. The fall detection system of any of A1-A8, wherein the one or more processors are further configured to provide notification of a determined fall event.A10. The fall detection system of any of A1 -A9, wherein the one or more processors are further configured to: analyze historical movement data of the user; andtrain the first and second neural network models based on the analyzed historical movement data.A11. The fall detection system of any of A1-A10, further comprising: a gateway device configured to receive the movement data from the wireless communication module of the wearable sensor device and transmit both the movement data and location data relating to a location of the gateway device to the one or more processors; and a server configured to communicate with the gateway device, wherein the server includes the one or more processors; wherein the system is further configured to: emit a verbal message to the user after determining a fall event has occurred; and prevent transmission of a fall notification if the user responds to the verbal message that assistance is not needed.A12. The fall detection system of any of A1-A11 , further comprising a database configured to store one or more patient-specific parameters, wherein the one or more processors are further configured to determine whether the fall event has occurred based at least on the outputs from the first and second neural network models and on the one or more patient-specific parameters stored in the database.A13. A method for fall detection comprising: capturing movement data of a user using a wearable sensor device, wherein the wearable sensor device includes: a housing including: an inertial measurement unit (IMU) configured to generate movement data; anda wireless communication module configured to transmit the movement data; and a fastener configured to releasably attach the housing to the user; receiving the movement data; processing the received movement data using a first neural network model and a second neural network model operating in parallel, wherein: the first neural network model is adapted to capture spatial and temporal features of the movement data; and the second neural network model is adapted for image-based analysis of the movement data; determining an output from the first neural network model and an output from the second neural network model; and determining whether a fall event has occurred based at least on the outputs from the first and second neural network models.A14. The method of A13, wherein determining whether the fall event has occurred is based at least on a weighted combination of the outputs from the first and second neural network models.A15. The method of A14, wherein the outputs from the first and second neural network models are weighted differently in determining whether the fall event has occurred.A16. The method of any of A13-A15, further comprising: processing the received movement data using a third neural network model operating in parallel to the first and second neural network models; determining an output from the third neural network model; anddetermining whether the fall event has occurred based at least on a weighted combination of the outputs from the first neural network model, the second neural network model, and the third neural network model.A17. The method of any of A13-A15, further comprising: when the first neural network model and the second neural network model disagree on a classification of the movement data, using a third neural network model to side with either the first neural network model or the second neural network model to determine whether the fall event has occurred.A18. The method of any of A13-A17, wherein determining whether the fall event has occurred is based at least on the outputs from the first and second neural network models and on location and time data.A19. The method of any of A13-A18, wherein the IMU comprises a 3-axis accelerometer and a 3-axis gyroscope, and wherein the movement data is 6-axis movement data.A20. The method of any of A13-A19, wherein the wearable sensor device further comprises: a proximity sensor configured to detect whether the device is being worn by the user, and wherein the method further comprises automatically adjusting an operation mode based on input from the proximity sensor; a manual activation button configured to allow the user to manually trigger an alert; and a barometric pressure sensor configured to detect changes in atmospheric pressure.

[0102] While the subject matter of this disclosure has been described and shown in considerable detail with reference to certain illustrative embodiments, including various combinations and sub-combinations of features, those skilled in the art willreadily appreciate other embodiments and variations and modifications thereof as encompassed within the scope of the present disclosure. Moreover, the descriptions of such embodiments, combinations, and sub-combinations are not intended to convey that the claimed subject matter requires features or combinations of features other than those expressly recited in the claims. Accordingly, the scope of this disclosure is intended to include all modifications and variations encompassed within the spirit and scope of the following appended claims.

Claims

CLAIMS1 . A fall detection system comprising: a wearable sensor device configured to capture movement data of a user, wherein the wearable sensor device includes: a housing including: an inertial measurement unit (IMU) configured to generate movement data; and a wireless communication module configured to transmit the movement data; and a fastener configured to releasably attach the housing to the user; one or more processors configured to: receive the movement data; process the received movement data using a first neural network model and a second neural network model operating in parallel, wherein: the first neural network model is adapted to capture spatial and temporal features of the movement data; and the second neural network model is adapted for image-based analysis of the movement data; determine an output from the first neural network model and an output from the second neural network model; and determine whether a fall event has occurred based at least on the outputs from the first and second neural network models.

2. The fall detection system of claim 1 , wherein the one or more processors are further configured to determine whether the fall event has occurred based at least on a weighted combination of the outputs from the first and second neural network models.

3. The fall detection system of claim 2, wherein the outputs from the first and second neural network models are weighted differently in determining whether the fall event has occurred.

4. The fall detection system of any of claims 1-3, wherein the one or more processors are further configured to: process the received movement data using a third neural network model operating in parallel to the first and second neural network models; determine an output from the third neural network model; and determine whether the fall event has occurred based at least on a weighted combination of the outputs from the first neural network model, the second neural network model, and the third neural network model.

5. The fall detection system of any of claims 1-3, further comprising a third neural network model, wherein when the first neural network model and the second neural network model disagree on a classification of the movement data, the third neural network model is configured to side with either the first neural network model or the second neural network model to determine whether the fall event has occurred.

6. The fall detection system of any of claims 1-3, wherein the one or more processors are further configured to: receive time and location data; determine a level of isolation of the user based on the time and location data; determine whether the fall event has occurred based at least on the outputs from the first and second neural network models and on the determined level of isolation.

7. The fall detection system of any of claims 1-3, wherein the IMU comprises a 3- axis accelerometer and a 3-axis gyroscope, and wherein the movement data is 6-axis movement data.

8. The fall detection system of any of claims 1-3, wherein the wearable sensor device further comprises: a proximity sensor configured to detect whether the device is being worn by the user, wherein the wearable sensor device is configured to automatically adjust an operation mode based on input from the proximity sensor; a manual activation button configured to allow the user to manually trigger an alert; and a barometric pressure sensor configured to detect changes in atmospheric pressure.

9. The fall detection system of any of claims 1-3, wherein the one or more processors are further configured to provide notification of a determined fall event.

10. The fall detection system of any of claims 1 -3, wherein the one or more processors are further configured to: analyze historical movement data of the user; and train the first and second neural network models based on the analyzed historical movement data.11 . The fall detection system of any of claims 1-3, further comprising:a gateway device configured to receive the movement data from the wireless communication module of the wearable sensor device and transmit both the movement data and location data relating to a location of the gateway device to the one or more processors; and a server configured to communicate with the gateway device, wherein the server includes the one or more processors; wherein the system is further configured to: emit a verbal message to the user after determining a fall event has occurred; and prevent transmission of a fall notification if the user responds to the verbal message that assistance is not needed.

12. The fall detection system of any of claims 1 -3, further comprising a database configured to store one or more patient-specific parameters, wherein the one or more processors are further configured to determine whether the fall event has occurred based at least on the outputs from the first and second neural network models and on the one or more patient-specific parameters stored in the database.

13. A method for fall detection comprising: capturing movement data of a user using a wearable sensor device, wherein the wearable sensor device includes: a housing including: an inertial measurement unit (IMU) configured to generate movement data; and a wireless communication module configured to transmit the movement data; anda fastener configured to releasably attach the housing to the user; receiving the movement data; processing the received movement data using a first neural network model and a second neural network model operating in parallel, wherein: the first neural network model is adapted to capture spatial and temporal features of the movement data; and the second neural network model is adapted for image-based analysis of the movement data; determining an output from the first neural network model and an output from the second neural network model; and determining whether a fall event has occurred based at least on the outputs from the first and second neural network models.

14. The method of claim 13, wherein determining whether the fall event has occurred is based at least on a weighted combination of the outputs from the first and second neural network models.

15. The method of claim 14, wherein the outputs from the first and second neural network models are weighted differently in determining whether the fall event has occurred.

16. The method of any of claims 13-15, further comprising: processing the received movement data using a third neural network model operating in parallel to the first and second neural network models; determining an output from the third neural network model; anddetermining whether the fall event has occurred based at least on a weighted combination of the outputs from the first neural network model, the second neural network model, and the third neural network model.

17. The method of any of claims 13-15, further comprising: when the first neural network model and the second neural network model disagree on a classification of the movement data, using a third neural network model to side with either the first neural network model or the second neural network model to determine whether the fall event has occurred.

18. The method of any of claims 13-15, wherein determining whether the fall event has occurred is based at least on the outputs from the first and second neural network models and on location and time data.

19. The method of any of claims 13-15, wherein the IMU comprises a 3-axis accelerometer and a 3-axis gyroscope, and wherein the movement data is 6-axis movement data.

20. The method of any of claims 13-15, wherein the wearable sensor device further comprises: a proximity sensor configured to detect whether the device is being worn by the user, and wherein the method further comprises automatically adjusting an operation mode based on input from the proximity sensor; a manual activation button configured to allow the user to manually trigger an alert; anda barometric pressure sensor configured to detect changes in atmospheric pressure.

Images