Fall detection system and method using multiple sensors including air pressure sensor

Abstract

The wearable device includes a machine learning fall detection processor that issues an alert when a minimum number of sensors among a total number of sensors, the plurality of motion sensors, at least one height sensor, and a force sensor, indicate the occurrence of a fall. The minimum number of sensors is automatically set by the fall detection processor based in part on an expected activity level of the wearer. The fall detection processor updates the expected activity level of the wearer stored in the database to the actual activity level of the wearer.

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002]

[0001] This application claims priority to U.S. Provisional Application No. 63 / 271,194, filed October 24, 2021, the contents of which are incorporated herein by reference.

[0003] Aspects of the present disclosure generally relate to devices, systems, and methods for reducing false positives in fall detection using multiple sensors.

[0004]

[0003] According to the World Health Organization (WHO), a fall is defined as a person inadvertently coming to rest on the ground, floor, or at a level lower than the starting height or orientation (from vertical / standing to horizontal / on the ground). Several factors can indicate the occurrence of a fall, including a change in height, a change in direction, a change in speed of movement, and the detection of an impact.

[0005]

[0004] Of the 35 million Americans age 65 or older, about one in three falls each year, and 50% of people who fall need help getting up. Older adults are five times more likely to be hospitalized for injuries caused by falls than for injuries from any other cause. Falls are the leading cause of accidental death among older adults.

[0006]

[0005] Studies have shown that rapid assistance increases the chances of surviving such emergencies. If rapid assistance can be called for, the elderly are more likely to continue living independently. On the contrary, the longer a person spends in a helpless situation (e.g., falls but cannot return to an upright position), the more likely they are to move to supportive care. For elderly people living alone, becoming disabled and unable to get help is a common problem and usually marks the end of their independent life. Even if a fall or other emergency occurs, 90% of those who receive help within an hour are able to continue living independently, but if help does not arrive for 12 hours, only 10% of them will be able to continue living independently at home.

[0007]

[0006] Current fall detectors can be divided into two categories: impact or shock detection and orientation detection. In the first category, the device must experience a measurable impact (e.g., by impact) for a fall to be identified. Devices in the second category typically use a tilt switch (mercury switch) and must be equipped with an object oriented sensor so that the tilt switch is activated when the user is in a horizontal orientation. The first and second types can also include gyros and accelerometers. Summary of the Invention

[0008]

[0007] The wearable device includes a machine learning fall detection processor, a non-transitory computer readable medium, a plurality of motion sensors, at least one height sensor, and a force sensor. The non-transitory computer readable medium has a database configured to store an expected activity level of the wearer. The plurality of motion sensors each have an individual fall motion threshold, and each motion sensor sends a fall motion signal to the fall detection processor when the individual fall motion threshold is exceeded. The at least one height sensor has a fall height threshold. The at least one height sensor sends a fall height signal to the fall detection processor when the fall height threshold is exceeded. The force sensor has a force threshold. The force sensor sends a force signal to the fall detection processor when the force threshold is exceeded. The fall detection processor issues an alert if a minimum number of sensors among the total number of the plurality of motion sensors, the at least one height sensor, and the force sensor indicate the occurrence of a fall. The minimum number of sensors is automatically set by the fall detection processor based in part on the expected activity level of the wearer. The fall detection processor updates the wearer's predicted activity level stored in the database with the wearer's actual activity level.

[0009]

[0008] The wearable device performs a fall detection method. The non-transitory computer readable medium stores an expected activity level of the wearer in a database. The fall detection processor receives fall motion signals from a plurality of motion sensors, each having a fall motion threshold, when an individual fall motion threshold is exceeded. The fall detection processor receives a fall height signal from at least one height sensor having a fall height threshold when the fall height threshold is exceeded. The fall detection processor receives a force signal from a force sensor having a force threshold when the force threshold is exceeded. The fall detection processor issues an alert to the fall detection processor if a minimum number of sensors from the total number of sensors, the plurality of motion sensors, the at least one height sensor, and the force sensor, indicate the occurrence of a fall. The minimum number of sensors is automatically set by the fall detection processor based in part on the expected activity level of the wearer. The fall detection processor updates the expected activity level of the wearer stored in the database to the actual activity level of the wearer.

[0010]

[0009] A non-transitory computer readable storage medium is encoded with data and instructions that, when executed by a fall detection processor, cause the wearable device to perform a fall detection method. The non-transitory computer readable medium stores an expected activity level of the wearer in a database. The fall detection processor receives fall motion signals from a plurality of motion sensors, each having a fall motion threshold, when an individual fall motion threshold is exceeded. The fall detection processor receives a fall height signal from at least one height sensor having a fall height threshold when a fall height threshold is exceeded. The fall detection processor receives a force signal from a force sensor having a force threshold when a force threshold is exceeded. The fall detection processor issues an alert to the fall detection processor if a minimum number of sensors from the total number of sensors, the plurality of motion sensors, the at least one height sensor, and the force sensor, indicate the occurrence of a fall. The minimum number of sensors is automatically set by the fall detection processor based in part on the expected activity level of the wearer. The fall detection processor updates the expected activity level of the wearer stored in the database to the actual activity level of the wearer. [Brief description of the drawings]

[0011]

[0010] To better understand the nature and advantages of the present disclosure, please refer to the following description and the accompanying drawings. However, it should be understood that each figure is provided for illustrative purposes only and is not intended as a definition of the limits of the scope of the present disclosure. Moreover, as a general rule, and where elements in different figures use the same reference numbers, those elements are generally identical or at least similar in function or purpose, unless otherwise clear from the description. [Figure 1] FIG. 1 is a block diagram of a personal emergency response system (PERS) that detects when a person has fallen, as determined by measurements from multiple sensors. [Diagram 2] 1 shows a single neuron in an artificial neural network. [Diagram 3]1 is an embodiment of a more accurate neural network for fall detection. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Aspects of the present disclosure include a personal emergency response system (PERS) that detects when a person has fallen, as determined by measurements from a plurality of sensors. In such cases, the personal emergency response system is integrated into a pendant, watch, glasses, or other wearable device. The PERS may include a pressure sensor to track the height of the sensor above a reference. The one or more sensors may be capable of operating in a low power environment to conserve battery power.

[0013] Another aspect of the disclosure is the recognition that a PERS wearer may experience a shock or impact simply by bumping into a counter or table while walking through their home. Also, when a critical event occurs, such as sudden cardiac arrest, the person may not fall violently to the ground. Instead, for example, the person may fall slowly to the ground and grab a nearby object to cushion the fall. Such a slow fall may not trigger an impact or shock sensor.

[0014]

[0014] Also, take into consideration that whenever the wearer wants to lie down (eg, to take a nap or go to bed in the evening), the tilt sensor needs to be disabled.

[0015]

[0015] In addition to accurately detecting falls and signaling emergency assistance, false positive fall detection (i.e., the PERS detects and indicates a fall, but the wearer does not fall) is a major issue with current fall detection PERS. If a PERS frequently signals a fall event when no fall occurred, it may undermine user trust and discourage users from using or carrying a PERS. It is estimated / rumored that 30-50% of PERS fall detection features are disabled after device deployment due to such false positive alerts.

[0016]

[0016] It is well known that determining whether a monitored subject has fallen is difficult. According to one embodiment of the present invention, multiple sensors are employed to improve the probability of accurate fall detection and limit the number of false positives. It is important to determine whether a monitored subject is about to fall or has fallen. The more sensor inputs available, the clearer the answer becomes. Certain sensors can more accurately indicate a fall, while other measurements make the fall detection system less reliable. Thus, the most accurate PERS use several sensors, taking into account (including weighting) each measurement as necessary, e.g., as a set of data value pairs, arrays, or other data formats, to improve their accuracy in fall detection. Some such systems are described below.

[0017]

[0017] One promising fall detection technology uses an AI / ML (artificial intelligence / machine learning) decision-making process and data fed from multiple sensors, each providing data used in the AI / ML decision-making process. This data can form a rich corpus for the ML / AI methodology and subsequent decision-making process.

[0018]

[0018] Use of AI / ML techniques may include, for example, identification of data representative of changes in state of sensors providing such data and the probability that such data may be indicative of a fall, either partially or for some time. ML / AI may detect feature sets representative of such state changes and may be precursors to those state changes.

[0019]

[0019] Additionally, as described further below, PERS uses comparative data and data patterns from the community of PERS users as another input to the monitored person's fall detection process. The monitored person's past movements, activities, and data patterns, as well as data patterns from the community at large, are also used as determinants of falls. Applying ML / AI techniques to the corpus of community data supports the identification of the monitored person's repetitive activities and the contextualization of those activities relative to the community.

[0020] For example, a PERS AI / ML-based algorithm may look at data from others who are similar in age, sex, health, time, day of the week, etc. to determine whether similar patterns lead to similar potential fall circumstances or whether they lead to falls. This creates a sufficient corpus of data for the ML / AI to apply techniques such as similarity assessment, matching and comparison, prediction, Bayesian, fuzzy logic, recursive, and deep learning to provide an adaptive decision-making processing system. If there are similarities in the measured data patterns, the aggregate group can be expanded and / or extended to include people of different sex, health, age, weight, etc. For example, the system can review cohorts that are not similar in age, health, sex, etc., but present similar data that led to falls. Identification of such similarities can employ machine learning feature detection techniques.

[0021]

[0021] Sensors typically found on a PERS can easily detect changes in direction (gyroscope), changes in speed (accelerometer), impacts (accelerometer detects sudden stop (speed suddenly drops to zero)), and tilt or tendency (tilt switch). However, determining the starting direction and / or height of the device prior to a fall can be problematic, especially since the detection device may be worn or carried in any of several locations, each likely to be a different distance from the ground or floor (reference level). For example, the monitored person may wear the PERS on their wrist or a strap, clip it to clothing, or wear it in some other manner. Additionally, the height of the PERS wearer must be taken into account insofar as this may contribute to the establishment of at least one height criterion, and such height criterion may be used as part of the fall detection assessment.

[0022]

[0022] However, each of these locations is at a different distance from the reference floor height. Variations in these parameters, as found in many situations, may be evaluated and used, in part or in whole, as data for an adaptive decision-making system. For example, variations in the rate of change in the environment may be measured to establish differential values ​​essential for accurate detection of a fall with low false alarm rates, which may make such accurate detection difficult. Although the reference height cannot be determined from common PERS components such as accelerometers, gyroscopes, tilt switches, etc., without this reference height information (i.e., the distance to the ground or floor on which the wearer is standing), it is difficult to accurately determine whether a person has experienced a fall.

[0023]

[0023] One embodiment of the present invention uses an atmospheric pressure sensor that periodically determines and stores pressure data in a FIFO (first in, first out) memory buffer, so that the system can evaluate the data prior to the occurrence of, for example, a possible fall, as indicated by at least one other sensor. The data is also sent to a cloud-based backup memory. The time the pressure was measured is also recorded so that it can be compared to the timing of other events (such as an impact event, acceleration, data provided by other sensors, etc.). Height information can be easily determined from the atmospheric pressure. Height information obtained from the pressure measurement can be used to determine a reference level or reference floor height. After the reference level is determined, if subsequent height values ​​are determined to be within a few inches of the reference height, this is one indication that the monitored person may have fallen.

[0024] In one embodiment, the pressure sensor comprises a low power pressure sensor to conserve battery power.

[0025]

[0025] Cloud backup of measurements and data can occur when the PERS is at full power and during "sleep" (low power) periods. The timing of data backup to the cloud is also variable and can depend on the monitored subject's activity level and other relevant factors, just like the data measurements.

[0026]

[0026] The rate at which pressure data (and other sensor data) is determined may be variable and dependent on the monitored person's activity level, concurrent measurements, and / or previous measurements. For example, measuring pressures and other fall indicators once an hour may be sufficient during periods of low or light activity, while a more active monitored person may need to measure pressures and other fall indicators more frequently.

[0027] The sensed pressure may be a function of temperature, and therefore, in certain embodiments, it may be necessary to correct the pressure measurement based on the ambient temperature.

[0028]

[0028] The speed of data collection may also be slowed down when the wearer (monitored subject) is within a predefined geofence area (e.g., lying in bed).

[0029]

[0029] Additionally, the measurement time (i.e., the interval that the sensor remains on to collect data) can be variable and dependent on a variety of factors that contribute to the fall detection process, including the configuration of the sensor and the system that receives the data and provides the monitoring function. The rate at which data is collected also varies with various configurations and other factors.

[0030]

[0030] The monitoring period may also be divided into time slices, with a particular predefined sensor being activated during each time slice, or different sensors may be activated in different time slices. Such an approach can ensure that a set of sensors can be operated to provide sufficient data to determine the conditions for assessment of a fall condition, including any false positives.

[0031]

[0031] The order in which the sensors are activated may also be variable and wearer dependent. For example, it may be advantageous to check the pressure sensor first. If there is no change from the previous reading, there may be no need to measure acceleration or other fall detection parameters.

[0032]

[0032] The use of low power sensors in one embodiment conserves battery power and allows the main processor (which analyzes pressure and other sensor data) to remain in a "sleep" or low power state while data is collected. When one of the low power sensors in the PERS (which detects acceleration, impact, rotation, motion, etc.) detects a condition that may be indicative of a fall, the processor is awakened to further analyze the measurement data.

[0033]

[0033] In another embodiment, the battery capacity in the PERS is sufficient for all sensors to be active or active at all times. Furthermore, if the battery capacity is sufficient, the processor may also maintain an active state at all times, except perhaps when the monitored subject is asleep or engaged in other low-motion activities. The state of the sensors may be changed in light of the battery capacity without degrading the sensor coverage of potential fall conditions. The time may be ideal to charge the PERS battery.

[0034]

[0034] Each sensor measurement (acceleration, impact, rotation, angular velocity, etc.) is analyzed by the processor in relation to predefined thresholds (obtained from a look-up table of fixed values, or created using AI or machine learning techniques, or obtained based on previous sensing / measurements) and the time the measurement was made. For example, this may be embodied as a graph database, a schema, or a schema on a read database. These thresholds generally indicate a fall, so when one or more are exceeded, the processor accesses the air sensor measurements from a FIFO memory buffer (and / or a cloud-based memory where the PERS can download the required data and operating procedures) and compares the air data collected before, during, and immediately after the detected event. Again, thresholds (static or dynamic values) may be employed. A sudden change in air pressure (height) is an indication of a fall, especially if supported by other sensor data analyzed by the processor.

[0035]

[0035] The atmospheric pressure data and acceleration, impact, rotation, velocity, etc. data are analyzed relative to respective thresholds and the relative timing of each measurement to determine whether to classify the event as a fall. Such a combination of sensor data can effectively detect falls and false positives where one or more sensor data have changed from a threshold but such change is not confirmed or is inconsistent with other sensor data. For example, comparison of atmospheric pressure data before, during, and after a fall can more accurately determine whether a fall has occurred, especially when evaluated in light of data related to direction of movement (i.e., movement only along the horizontal axis or both horizontal and vertical, as can be determined from pressure / height data), velocity of movement (immediate change or slowly crawling into bed) and impact.

[0036]

[0036] The thresholds and timing of measurements may also be based on profiles created for users classified into different classes. For example, different profiles, and therefore different thresholds and timing of measurements, may be created for adults, seniors, teenagers, infants, etc. And the parameters sensed or measured may also depend on the class to which the wearer belongs.

[0037]

[0037] Such personal information can be easily collected before the first use of the PERS and collected continuously thereafter. Various data thresholds, data patterns, timing of data collection, sensor measurements, etc. can then be automatically implemented within the PERS, allowing for more accurate fall detection.

[0038] In addition to the personal information, a profile of the monitored subject is created. The profile identifies characteristics of the monitored subject relevant to fall detection, such as daily activities, normal gait, normal heart rate, etc. Such a profile can be used, in part or in whole, to configure at least one threshold, pattern, timing data and / or sensor and system parameters, such alignment of the profile with the personal information supports identification of false positives that might otherwise be overlooked, triggering undesirable events.

[0039]

[0039] It is also suggested that different fall profiles may require different information to reliably determine that a fall has occurred. For example, the following fall profiles: tip over or forward, travel several blocks then fall, walk and then fall, fall up to the knees, fall onto the wearer's buttocks, fall onto the wearer's side suggest the need for different measurement and analysis techniques.

[0040]

[0040] The collected data that is linked to falls in the context of machine learning / artificial intelligence embodiments (see further details herein) is stored in a database (on at least one device and / or in the cloud). A fall is then suggested if the current data is highly correlated with previous data. This use of fall data, including data of actual falls, may be used to detect false positives using ML / AI and may incorporate data sources such as profiles that include patterns.

[0041]

[0041] PERS also includes a "learning" feature where the software "remembers" if, for example, the wearer exercises vigorously between 10 and 11 a.m. every day. Such information may be incorporated into the wearer's profile. Thus, the PERS processor will discount (or assign a lower weight to) certain measurements taken during that time (see discussion of neural network embodiments herein). In addition to PERS learning the monitored subject's times of vigorous activity as they recur at approximately the same times each day, the monitored subject can manually set the PERS based on anticipated vigorous activity for future days / times. Finally, just prior to engaging in vigorous activity, the monitored subject can manually adjust the PERS (e.g., sensor thresholds) so that certain sensor measurements are ignored or the impact of the measurements is discounted relative to other sensed values ​​from other sensors. This control function may change (e.g., increase) thresholds for certain measurement parameters that are directly affected by vigorous activity. This may be presented to the wearer in the form of an application, including a wellness dashboard, where user profile settings can be changed. Such changes may be stored on the device and / or in the cloud and can be recalled and / or used by machine learning techniques to further expand and / or enhance the profile.

[0042]

[0042] Thus, the system learns in the sense that past behavior (especially past repetitive behavior) influences the current thresholds and analysis. If the wearer slumps low during her yoga routine at 10am for a few days, and the fall notification issued for each such "fall" is cancelled by the wearer (she did not fall), the PERS will retain that information. Any subsequent "falls" around 10am will be discounted and other measurement data will be relied upon to indicate falls. This can significantly reduce the number of false positives detected and enhance the overall operation of the system for the benefit of the wearer and all systems such as emergency services, medical or care teams.

[0043]

[0043] In addition to using time as a parameter in determining that a fall has (or has not) occurred, the location of the wearer may also be considered. A PERS includes a location determining device (e.g., using a GPS receiver or signal triangulation techniques). If the determined location is a gymnasium or yoga studio, the measured slope value or other measurements indicative of vigorous activity are discounted. When the wearer leaves the gym or yoga studio (as determined by the location determining device), the measured slope value is given more weight in analyzing subsequent fall events.

[0044]

[0044] The fall determination algorithm within the PERS can be modified (either manually by the monitored subject or automatically based on previous fall or no-fall determinations) to be more sensitive to certain sensors under certain conditions and less sensitive to certain sensors under other conditions, resulting in more accurate fall determinations and fewer false positive alerts. This adaptive approach to reducing false positives provides significant benefits in the performance of the system and the inherent reliability and subsequent trust placed in the events triggered by those systems, the conditions that govern its sensitivity are numerous, some of which are described herein, including, for example, the time of day of the monitored subject, activity level, location of the monitored subject, expected future activity level, or location based on previous activity level, or location at the same time of day.

[0045]

[0045] The conditions and measurement data that triggered the fall alarm are then neutralized and stored for later comparison with the measurement parameters. A high correlation between the past data that generated the false alarm and the current data is a good indicator that a fall has not occurred, but also suggests that additional sensor data is needed for more accurate fall detection to reduce the number of false positives.

[0046] In another embodiment, when the processor wakes up, it analyzes measurements obtained from all sensors, including the pressure sensor. By analyzing all sensed data, this embodiment is distinguished from the embodiment above, in which pressure data is analyzed only if certain other measurements lead to a tentative conclusion that a fall has occurred.

[0047]

[0047] Upon detecting a fall, the PERS issues an alert or notification reporting the fall, including a request for emergency assistance. The notification may be issued to parties identified by the wearer and / or emergency personnel.

[0048]

[0048] In addition to notifying the wearer that they have fallen, the alert may include additional information related to the fall, such as the location, time of the fall, the direction in which the wearer fell, the strength of the impact of the fall, current heart rate, current respiration rate, and other data generated by sensors in any placement.

[0049] In addition to the low-power sensors as described herein, in one embodiment, the PERS may also include a low-power transmitter. Typically, such transmitters have limited signal coverage range and can only communicate with nearby devices. In such an embodiment, the low-power transmitter issues an alert (e.g., the monitored subject has fallen) to nearby devices that operate at a higher power level and therefore have a larger signal coverage area. The latter high-power device issues a fall notification to parties designated by the monitored subject, such as family members or emergency responders. For example, Bluetooth or WiFi.

[0050]

[0050] If the PERS has sufficient power, it can replace the low power transmitter with a high power transmitter and issue fall notifications directly to emergency personnel, family members, etc. via a nearby WiFi router that transmits a notification signal to identified receiving parties via cellular service or the Internet (e.g., email and text notifications).

[0051]

[0051] Thus, by recording / storing, buffering, and analyzing the sudden change in atmospheric pressure, PERS can confirm a fall by comparing the height (atmospheric change) before and after a detected fall event. This change in air pressure, along with the initial detection of a fall using other sensors, can confirm if a person has experienced a fall. Being able to compare air pressure before, during, and after the fall can accurately confirm the direction, speed, and impact if it occurred on the horizontal axis only, or if the vertical axis also included movement and distance.

[0052]

[0052] Sensor placement

[0053]

[0053] The sensors and / or processing components of the various described embodiments may be encapsulated in a wearable, encapsulated in a pendant, or affixed to clothing. In general, wearables include electronic devices that may be worn as an accessory, embedded in clothing, attached to clothing, embedded in the wearer's body, and / or tattooed on the skin in any configuration. For example, such devices may incorporate sensors that are worn on the skin and may be further connected to sensors and / or devices that are carried as a pendant.

[0054]

[0054] Such electronic devices may also be encapsulated within, for example, clothing, shoes, hats, headbands, eyeglasses, hearing aids (hearing aids are particularly attractive because they contain multiple electronic components, some of which can be used to function as a PERS), chokers, bracelets, necklaces, rings, watches (i.e., any jewelry item), belts, tie clasps, earphones, GPS tracking devices, etc. Any device that is typically carried by a user (such as a cell phone or keys), while not strictly considered a wearable, may also be adapted to perform the functions of a PERS.

[0055] Description of the embodiment of FIG.

[0056]

[0056] Figure 1 shows another fall detection algorithm. The PERS is worn by the person being monitored and an accelerometer / gyroscope 100 senses / measures acceleration and acceleration direction. In decision block 102, the measurements are analyzed to determine if the wearer has moved or not. If not, processing returns to measurement block 100. Thus, blocks 100 and 102 determine if there has been any movement of the person being monitored; if there has been no movement, there is no need to process the measurements further or measure values ​​that may actually indicate a fall, although such values ​​may be recorded as part of the person's profile or pattern.

[0057]

[0057] However, if it is detected that the subject has moved, processing transitions from decision block 102 to dual paths 110, 112. Path 110 measures or determines the magnitude of a signal vector which essentially estimates the degree of intensity of the movement in each one of three axes or directions, X, Y, and Z. Here, the magnitude of the signal vector is the square root of the sum of the squares of each of these three component values. Thus, movement or motion along each axis determines the magnitude of the signal vector and contributes equally to determining whether there has been an impact based on that value.

[0058] As can be seen, the magnitude of the signal vector determined in each of these three directions is processed through an infinite impulse response filter 116 and the magnitude of the signal vector is determined in block 118. The result from block 118 is buffered in a buffer 120. In decision block 122 the calculated signal vector magnitude is compared to a signal vector magnitude threshold. If the result from that decision block is negative, processing returns to blocks 100 and 102 to again determine movement or actions by the monitored subject.

[0059] If the outcome from that decision block is positive, processing continues to block 136, as described below.

[0060]

[0060] Pressure is measured along path 112 at block 130, filtered through infinite impulse response filter 132, and buffered in buffer memory 134. The final pressure value determined at block 136 is compared to a threshold pressure value in decision block 138. An affirmative response from decision block 138 indicates that there has been at least a 2 foot change in elevation at the location of the PERS.

[0061] If the determination at decision block 138a is negative, processing returns to the initial blocks 100,102.

[0062]

[0062] It should be noted that the results of the shock analysis (decision block 122) also lead to a pressure measurement block 136.

[0063]

[0063] If the determination from decision block 138 is positive, processing moves to block 140 where a tilt value is determined. In decision block 142, the determined tilt value is compared to a tilt threshold. A negative response from decision block 142 indicates that the pendant or PERS has not tilted since the time the impact was determined. A positive response from decision block 142 indicates that there has been tilt of the PERS and may indicate that the wearer has fallen, resulting in a tilted position of the PERS wearer. Processing then proceeds to decision block 146.

[0064]

[0064] In decision block 146, the signal amplitude area (SMA) parameter is compared to a threshold value. SMA is another parameter for determining the wearer's activity intensity and can therefore be used to distinguish between rest and movement of the wearer.

[0065]

[0065] Note that both SVM and SMA values ​​can be used to determine movement. SVM is a real-time value based on current (or buffered) acceleration values, so it can determine quick changes or movements (or impacts). SMA value is the area under the curve that helps determine the intensity of the movement. This is done over a short period of time (such as 1 second) to determine the amount of movement that has occurred in that period. Therefore, it is used to determine if the user is moving after a fall.

[0066]

[0066] If the result from decision block 146 is positive, processing moves to block 148 where tilt is again measured. This is followed by decision block 150 where the measured tilt parameter is again compared to the tilt threshold. The objective here is to determine if the wearer has stopped moving, and therefore the tilt value must be less than the tilt threshold to generate a positive response at decision block 150. A fall is indicated at block 152 following a positive response from decision block 150.

[0067] Note that negative responses from decision blocks 138, 142, 146, and 150 result in processing being returned to blocks 100 and 102.

[0068]

[0068] The data generated through these blocks can be stored in at least one repository to form a corpus for at least one machine learning technique to operate on, for example, to detect feature sets indicative of a fall or no fall. These feature sets can then be compared to the data generated in real time to predict and / or identify the probability of a fall, thereby increasing the accuracy of fall detection and its false positives.

[0069] Accelerometer and Gyroscope Embodiments

[0070]

[0070] The pendant or PERS may include an accelerometer (i.e., a three-axis accelerometer) that detects either speed or direction changes. The second sensor is a gyroscope that detects changes in the wearer's orientation, more specifically, changes in the orientation of the pendant that surrounds the gyroscope. For example, if the gyroscope is attached to a belt or strap of the wearer and the belt or strap changes direction, this is another indication that a fall may occur. Using the accelerometer sensor measurement and the gyro sensor measurement allows for more accurate fall detection.

[0071] In one embodiment, another sensor is activated when either the accelerometer or gyroscope measurements exceed a threshold, or when both the accelerometer and gyroscope measurements exceed their respective thresholds. Using two or more detection values ​​to determine if a fall has occurred is an improved approach over using a single sensor.

[0072]

[0072] If the change in orientation or direction detected by the gyroscope occurs immediately after a change in acceleration determined by the accelerometer, this is a more conclusive indication that the wearer has fallen.

[0073]

[0073] If both the accelerometer and gyroscope generate signals indicative of a possible fall, then a third sensor may collect additional information to make the determination of a fall more reliable. In such a case, the third sensor may be a height sensor that uses the value of atmospheric pressure to determine height above sea level or a reference altitude. When the wearer is upright, the pressure sensor / height detector measures a first pressure value. If the wearer falls, the height pressure sensor measures a second pressure value that is lower than the first measured pressure value. Although a simple change in pressure value may indicate that the wearer has fallen (or simply lay down), using the accelerometer value, the gyroscope value, and the height pressure value provides a more accurate fall detection process.

[0074]

[0074] Also, if the wearer moves suddenly, this movement will be detected by the accelerometer and the readings alone may indicate a false fall detection, however, using the accelerometer values ​​together with the gyroscope and pressure sensor values ​​reduces the chance of a false positive.

[0075]

[0075] Additional sensor accuracy to improve fall detection function

[0076]

[0076] The three main sensors described herein include tilt sensors, accelerometers, and gyroscopes. Values ​​measured by additional sensors can increase fall detection accuracy and reduce false positives. Such additional sensors include audio sensors (e.g., microphones or any sound collection devices such as extended home speakers and television speakers), video sensors (e.g., cameras including cameras capable of capturing still images, moving images, and videos), and force sensors (pads for detecting the application of force to the pads).

[0077]

[0077] For example, a microphone may act as a sensor, where detection of noise may indicate a fall and absence of such noise may indicate that a fall has not occurred. The microphone may also detect breathing.

[0078]

[0078] For example, a camera could be used to track specific points, potentially at lower resolution, and identify rates of change, etc. to provide further data regarding the occurrence or non-occurrence of a fall.

[0079]

[0079] The sensors may measure various physiological and biometric parameters (heart rate, respiratory rate, muscle tension, blood pressure) and these measurements may also be taken into account along with the data collected from the various sensors described herein.

[0080]

[0080] A voice sensor can establish an audio link to the monitored person, thereby setting up an audio link. With a family member or first responder on one end of the link and the monitored person on the other, the latter can report if she has fallen. If the monitored person is unable or does not respond to vocalize her situation, the other party can infer that she needs assistance. This audio link can provide a definitive answer to the question of whether the monitored person has fallen.

[0081]

[0081] Microphones can be easily placed on any wearable device. In another embodiment, microphones can be distributed in the living area of ​​the monitored person. These microphones can be always on and activated by listening or voice as needed. In case of a fall or not being able to access a mobile phone to call for help, the monitored person can simply shout to activate the microphone and talk to the other side of the link. Voice comments can also be collected as data along with the sensor measurement data (e.g., from the wearable device) and sent to a cloud-based (or server-based) fall detection algorithm (in one embodiment, artificial intelligence / machine learning (AI / ML)).

[0082]

[0082] In addition to establishing an audio link via the microphone, the activated microphone can detect sounds directly caused by a loud crash fall, such as a loud thud. That audio signal can be analyzed by the PERS, perhaps by comparing it to stored thuds and loud crash sounds, thereby providing additional input data to more accurately determine if a person has fallen. If no thuds or loud crash sounds are heard from the open microphone, then the monitored person may not have fallen.

[0083]

[0083] Hearing aids have been identified as wearable PERS, and the devices can easily be modified to also function as microphones. In many cases, headphones (both wired and wireless, such as based on the Bluetooth standard) that contain both a microphone and a speaker also function as wearable PERS and microphones.

[0084]

[0084] Similarly, the video link provides video information that allows the subject's condition to be easily determined. For example, the video can provide data to determine the initial height and distance difference at the time the subject fell. The video can also provide information that allows the subject's location to be identified.

[0085]

[0085] Finally, properly placed force sensors can also provide useful information for detecting falls. For example, if a pad is placed on a person's back and the person falls to the floor with their back hitting the floor, the force sensor will generate a signal representative of the force applied to the pad. Of course, a person can fall and land in many different positions, and therefore it may be advantageous to use several such force sensors placed at various locations on the body of the person being monitored.

[0086]

[0086] If battery capacity and power conservation are concerns compared to PERS, a schedule can be determined for activating and deactivating certain sensors of the sensors. Algorithms can be created or adjusted manually or based on the activity level and location of the monitored individual. In an AI / ML implementation, the engine learns to create or adjust the battery life of the PERS unit based on activity level, location, recurrence, time of day, etc., so that less battery usage occurs during times of low activity, but battery capacity is available during times of high activity.

[0087] Multiple Sensors for Interim and Definitive Fall Detection

[0088]

[0088] Each sensor of the PERS can operate in combination with a different threshold. That is, a particular movement detected by the accelerometer may not generate a tentative fall indication if the accelerometer value does not exceed a first threshold. The threshold can be established by measuring sensor values ​​associated with normal wearer movements. That is, slow movements may indicate that the wearer is simply walking, while sudden accelerations or movements may indicate that the wearer has fallen. Normal or typical (non-fall) sensor values ​​can be determined during a test or pre-operational period, where multiple sensor values ​​are measured during the wearer's daily activities. Sensor thresholds are then established beyond those values ​​indicative of simple movements by the wearer. Obviously, sensor thresholds must be established empirically and personalized for each wearer.

[0089]

[0089] The sensor values ​​can be measured repeatedly, for example with a measurement rate determined by the time of day. At night, when the wearer is asleep, measurements can be taken less frequently, but during the day when she is active, all sensors can measure their respective parameters more frequently.

[0090]

[0090] In some embodiments, the fall detector described herein operates in a computationally intensive environment and requires more than a simple comparator to compare the detected value to a threshold. An algorithm controlling the operation of the fall detector can be written to employ different thresholds for different sensors at different times of the day and set the relationship between the thresholds of the sensors. For example, if a first sensor measures a value of X, a second sensor is activated and the threshold of the second sensor is Y. Thus, if the second sensor measures a value greater than Y, it indicates a fall. However, if the first sensor measures a value A, the threshold of the second sensor is set to D, a threshold greater than Y. According to this scenario, the second sensor needs to measure a value greater than D to indicate a fall.

[0091]

[0091] The thresholds of the PERS must also take into account the physical condition of the wearer. For example, is the wearer agile or slow? Is the wearer young and energetic or old and slow? Again, the thresholds can only be determined in an empirical pre-operational stage, where the thresholds are established personally for each individual wearer. Such values ​​of the thresholds may be held in the individual's profile and / or pattern and may be represented in the aggregate, for example by an average or arithmetic mean, in at least one template.

[0092]

[0092] Multiple sensor inputs

[0093]

[0093] Various embodiments are described herein in which measurements or sensing values ​​from one or a small number of sensors are used, either alone or as a group, to determine if the wearer has fallen. This concept can be extended to n sensors, relying on a subset of the n sensors to determine if a fall has occurred. For example, if n sensors are monitoring a wearer and measurements from w of the n sensors indicate that a fall has occurred, the PERS will issue a fall alert. The metric in this case is that an alert is issued if a minimum of w sensors out of the n sensors indicate that a fall has occurred.

[0094]

[0094] The numbers n and w are not uniquely determinable but depend on several factors such as the wearer's age and activity level. The numbers n and w can be fixed in the PERS or can be determined and reset by the wearer. Obviously, a smaller difference between n and w suggests the possibility of fewer false alarms (i.e., a fall being detected when the wearer has not fallen). On the other hand, a larger gap between n and w suggests the possibility of more false alarms.

[0095]

[0095] Additionally, the selection of w sensors among n sensors in a PERS can be customized to the wearer and may be manually set by the wearer as desired, depending in part on the wearer's expected activity level. Alternatively, the number of w sensors and which sensors are included in a group of w sensors can be automatically set by an AI / ML algorithm that receives and processes the measurement data. Of course, in an AI / ML-based embodiment, the monitored person need not consider such issues, since a trained algorithm can automatically select and configure the number and particular sensors.

[0096]

[0096] Certain sensors can be configured to include being deactivated or activated depending on the wearer's activity level. For example, if the wearer plays tennis at 3:00 PM on a weekday, the tilt sensor can be manually disabled before the match begins. Alternatively, the AI / ML algorithms may determine (based on recurrence time and other factors) to automatically adjust the fall detector settings to minimize false positives while the monitored person is engaged in vigorous activity such as tennis. Additionally, the wearer can set the PERS to automatically disable the tilt sensor at 3:00 PM on weekdays.

[0097]

[0097] Settings can be configured (manually or as determined by AI / ML algorithms) for all sensors in the PERS.

[0098]

[0098] Not all sensors or groups of sensors are necessarily active all day, every day. The wearer can customize the times each sensor is active or inactive. A strict pattern of activation and inactivation may not be desirable. Customization by the wearer is desirable as long as it does not compromise the effectiveness of the PERS.

[0099]

[0099] In some embodiments, at least one AI / ML algorithm within and / or accessible from the PERS customizes when each sensor is activated or deactivated, managing battery life while relieving the monitored person of the burden of manually reconfiguring sensors. The AI / ML algorithm learns about the monitored person's repetitive and overall activities and adjusts the sensors as needed.

[0100]

[0100] However, the choice of n and w also involves a trade-off: a small difference between n and w reduces the frequency of false alarms, but the more operational sensors there are, the greater the energy capacity (battery capacity) required.

[0101]

[0101] In addition to activating the w sensors and taking measurements, the order in which the w sensors are activated can be controlled, with activation of a particular subsequent sensor being conditional on measurements obtained from previously activated sensors. Again, the trained AI / ML algorithm is expected to make these adjustments as needed, refining the adjustments as it learns about the monitored subject's activities.

[0102]

[0102] Detection of specific patterns in sensor data may also be used to detect falls among multiple sensors. In particular, if the current data pattern matches a previous data pattern after a fall has occurred, observing the current pattern can be a strong indicator that the fall is significant. Again, trained AI / ML algorithms can provide this functionality.

[0103]

[0103] The AI / ML component of the PERS (whether cloud-based or not) will not only analyze patterns and symptoms for a particular individual, but will also consider all individuals in the community (or all individuals who have a PERS and send data to the AI / ML algorithms) looking for common precursors to notable fall events. For example, such precursors may be represented in the form of templates that may include all or part of profiles and / or patterns. These may include similar demographics, similar sensor data, and similar data patterns. For example, if a sequence of fall-related activities occurs for a monitored individual, that sequence will be recorded by the AI / ML algorithms of the NXTD / LogicMark system (the assignee of the present application) as representative of potential fall activities for that individual. However, that sequence may precede a false positive alert.

[0104]

[0104] However, if the data is correlated with data from a larger population in PERS (including individuals of a similar age, or with other similar characteristics, or similar demographics) (or even larger aggregate groups whose data are not included in the system), then the data is more likely to actually represent a fall rather than a false alarm (or vice versa), improving the accuracy of both fall detection and identification of false positives.

[0105]

[0105] For example, if a pattern emerges from Simmons' measurement data, but it is uncertain whether the data pattern indicates a potential fall, the data can be correlated with data from the collective (or larger) community to search for similar previous patterns that were indicative of a fall event. A high correlation indicates a potential fall. The wearer is contacted to determine if she has fallen. She may answer "no" to the question. In this case, the algorithm is modified to reflect that a data pattern does not necessarily indicate an expected fall event.

[0106]

[0106] The AL / ML algorithm notes (i.e., learns) that this type of pattern results in a "no fall" determination for Mr. Simmons at this time and in this location. The algorithm then validates this conclusion based on the data pattern to determine whether a repeat of this activity (i.e., similar measurement data) would lead to a subsequent situation where Mr. Simmons did not fall. The validation process may include several similar data sets to ensure that the algorithm's "no fall" determination is in fact a correct determination that Mr. Simmons did not fall.

[0107]

[0107] The AI / ML algorithm could also compare Simmons' data patterns with other data patterns in the community to determine whether a fall may or may not have occurred, revalidating the algorithm applied to Simmons. Thus, the algorithm may determine that similar future data patterns for Simmons would generate a "no fall" determination.

[0108]

[0108] Furthermore, because Mr. Simmons' data patterns are aggregated and shared anonymously across the community, if a similar data pattern is observed for another person of a similar age or other similar characteristics to Mr. Simmons (e.g., a similar-aged woman, Ms. Smith), the system will not indicate this data pattern as preceding a fall event. Mr. Smith will also be sent a verification question to ensure that he did not fall, and the cycle begins again.

[0109]

[0109] The fall detection system (mainly at least one AI / ML algorithm) searches for an individual's data patterns to determine that person's unique fall detection pattern. Additionally, the AI / ML algorithm is run against other individuals (of similar demographics) in the community to verify whether similar patterns have led to falls in others. Thus, the AI / ML algorithm (embodied as cloud-based software) is constantly learning about patterns, sensor measurements, sequences, etc., that indicate a fall is likely to occur. This learning process allows the system to note that, for example, when the sequence 1, 2, 3 occurs in the sensor data, blocks 4, 5 and 6 are likely to be triggered next in the sensor, which may result in a fall. Or, the sequence 1, 2, 3 (or other data), when considered with similar data of other PERS wearers in the community, means that a fall is more likely. Alternatively, the sequence 1, 2, 3 is considered along with the wearer's past history data to determine if that data precedes a fall event. Such analysis can also be applied to determine when a fall is initially indicated, but does not occur. For example, if the sequence 1, 2, and 3 occurs, the likelihood of a fall may be expected to be N%, but if the time between events 1, 2, and 3 is sufficiently different from the sequence when a fall occurs, this may indicate that the likelihood of a fall is low and a positive detection of a fall is avoided.

[0110]

[0110] Typically sensors output specific measurements, but the use of data ranges is also beneficial. Data values ​​within or outside of range represent conditions for activating (or not activating) other sensors. For example, if the measurement data from sensor 1 is within a range between X and Y, sensor 2 is activated. If the measurement data is less than X, no further measurements are taken. And if the measurement data is greater than Y, sensor 3 is activated. Using such conditional activation of others (e.g., more sensitive sensors or sensors that are better "fall" predictors) can measurably improve the accuracy of fall detection and reduce false positives.

[0111]

[0111] Sensor directionality

[0112]

[0112] Many sensor types are described herein and are commonly used in PERS devices, such as gyroscopes for detecting tilt or orientation relative to a reference, accelerometers for detecting velocity and acceleration, and pressure measurements for detecting height or height difference relative to a reference.

[0113]

[0113] The accuracy of fall detection can be further improved by adding a three-dimensional reference space (X-axis, Y-axis, Z-axis). For example, acceleration primarily in the X- or Y-direction (with minimal movement detected in the Z-direction) may indicate that the wearer is walking or jogging. However, a large change along the Z-axis indicates that the wearer has fallen. During such a fall, small changes along the X- and Y-axes are also expected, further suggesting a fall.

[0114]

[0114] Machine Learning and Artificial Intelligence Embodiments

[0115]

[0115] In one embodiment, the present invention is embodied in at least one machine learning algorithm (i.e., an example of artificial intelligence analysis). These algorithms are dynamic models implemented in software that assimilate vast amounts of input data to generate an output, typically either a binary decision (e.g., yes / no, accept / reject) or a numerical prediction (e.g., 97%). Machine learning algorithms are distinguished from static computational algorithms by their ability to independently change (i.e., learn) the details of the algorithm to affect the outcome, with minimal human intervention.

[0116]

[0116] Machine learning algorithms are programmed to self-adjust to reduce error in subsequent predictions in order to more accurately predict actual outcomes. Over time and through multiple iterations, these algorithms learn to improve the accuracy of their predictions.

[0117]

[0117] Subcategories of machine learning algorithms include supervised learning and unsupervised learning, such as artificial neural networks, decision trees, random forests, support vector regression, etc. Many of these algorithms are best suited for specific classes of problems.

[0118]

[0118] Commercially available machine learning algorithms are available from: TensorFlow (available from Google, Inc., Mountain View, Calif.), Theano (developed by the University of Montreal), Sci-kitlearn (Machine Learning with Python, Pedregosa et al., JMLR12, pp. 2825-2830, 2011), or CNTK (available from Microsoft Corporation, Redmond, Wash.).

[0119]

[0119] The use of regression techniques may be employed to predict certain values, for example the data provided by the sensors may be evaluated by at least one regression function to create a prediction of the data set that the sensors are likely to provide at some point in the future. When sets of sensors as described herein each have a predictive data set, these sets may be compared to identify complementary or conflicting data from sensors enabling, for example, a multivariate approach to determining the state of a fall, thereby allowing for detection and remediation of false positives that may be exhibited by isolated sensors or a set of sensors measuring motion in the X, Y, Z planes, etc.

[0120]

[0120] For example, a decision tree may include a decision matrix that may be represented as a data set for evaluation by other systems and / or may form the basis of actions including events and / or triggers. One example of such an approach is a graph-based representation.

[0121] In some embodiments, the decision matrix may include thresholds and may be dynamic in that the thresholds may change in response to context. For example, data held in a profile as described herein may be used to vary the thresholds incorporated into the decision matrix.

[0122]

[0122] In some embodiments, a support vector machine (SVM) may be used for template matching, e.g., profiles and / or patterns are represented by templates, which may then be used to establish relative similarity between such profiles and / or patterns. For example, if the SVM includes a Gaussian kernel, the similarity of X in pattern A may be determined to X1 in pattern B, e.g., both patterns have a common template in whole or in part. SVMs may also be used to predict whether a class is positive or negative, e.g., to represent the state of a set of sensor data.

[0123]

[0123] One implementation of the invention involves the use of an artificial neural network, which may, for example, comprise multiple neural sub-networks in any arrangement. In some embodiments, the artificial neural network is a collection of "neurons" organized in one or more "layers."

[0124]

[0124] Figure 2 shows a single neuron in an artificial neural network. The single neuron 10 embodies a mathematical activation function 12, i.e. a simple static function that converts multiple weighted inputs from sensors into a single output value. Neuron 10 receives input data from sensors X1, X2, X3 to which weight values ​​W1, W2, W3 have been applied. The activation function 12 is applied to the sum of the weighted values ​​as shown. The output Y is the result.

[0125]

[0125] As applied to the present invention, the input data includes measurements from fall detection sensors, including, for example, accelerometers, gyroscopes, and pressure sensors. And the output Y indicates that a fall has or has not occurred, and may include a probability weighting factor, such as a 93% confidence index. However, this is an oversimplified description of a neural network applied to the current fall detection problem. Figure 3, described below, is a more accurate neural network embodiment for fall detection.

[0126]

[0126] The particular activation function used in the neural network depends on the type of output required. The activation function can be as simple or complex as necessary to adequately represent the combined effect of the lower layers of the neural network or the combination of input data. For example, to predict a binary classification such as whether a fall occurred or did not occur, a sigmoid activation function is suggested to be used, which outputs 0 or 1. Outputs above a threshold (usually 0.5) are predicted to be in one category (e.g., a fall occurred) and outputs below the threshold are classified in the other category (e.g., a fall did not occur). Such categories can be formed from data held in the profiles described herein.

[0127] Regardless of the activation function, the computed output (Y) can be passed as an input to another neuron or can represent the final output of the complete neural network. Those skilled in implementing artificial neural networks can apply the appropriate activation function, determined by the type of input and the desired output.

[0128]

[0128] Figure 2 is a schematic diagram of a simple artificial neural network for determining whether a fall has occurred. A number of sensor inputs are shown on the left side of Figure 2, with each input representing a value input to each node in column 21. The multiple inputs are represented by sensors described herein and other sensors that can provide information useful in determining whether a fall has occurred.

[0129]

[0129] Each input to a neuron in column 21 is the numerical output (sensed or measured value) of a particular sensor. Sensor values ​​can be measured continuously or periodically, at fixed times each day with fixed periods, at periods that vary throughout the day, or at variable periods determined by values ​​measured by one or more of the other sensors. Thus, the granularity of the sensor measurements can be determined by the wearer and / or the PERS designer.

[0130]

[0130] For example, several lines extending from node 21A represent values ​​measured by sensor 1. Each is multiplied by a weight from bubble 30 (the same or different weights applied to each sensor value), and each terminates at a node in column 22. The weight value thus modifies the output value from node 21A, and the weighted value is input to the corresponding node in column 22.

[0131]

[0131] This process is repeated with various weights multiplied by the sensor output values ​​from column 21. The products of the detection values ​​and weights are input to each node in column 22. This process is repeated with the outputs from column 22 multiplied by weights provided by bubble 32. The resulting products are input to each node in column 24. Finally, the output values ​​from column 24 are combined at node 30 and the output value from node 30 indicates whether a fall has or has not occurred.

[0132]

[0132] In general, the weight values ​​in Figure 3 indicate the influence that the sensor reading has on each node in the next column, and ultimately on the final decision that a fall did or did not occur. Note that the influence of each value from column 21 may be different for each node in the subsequent columns, as it is reflected in a different weight value. Furthermore, note that if an input feature does not influence the next node, the weight may be zero.

[0133]

[0133] These weights may be based on results from previous iterations through the neural network 20, thus establishing a starting point (or constant adjustment) for the input values ​​at each node. As the algorithm is trained (or later updated), the weight values ​​are changed to optimize the results from the neural network, improving the accuracy of fall detection and therefore improving the identification of false positives.

[0134]

[0134] As shown in Figure 3, artificial neural networks typically involve a complex fusion of individual neurons, diverse input data, and multiple activation functions (represented by nodes), with different weights applied to the output of each neuron, in order to accurately predict outcomes.

[0135]

[0135] This approach significantly improves the detection of false positives. For example, the loss function calculates the magnitude and direction of the prediction error. The prediction error is then backpropagated through the network, and the weights applied to the output from each neuron are updated to reduce the prediction error during subsequent iterations of the neural network. Thus, the neural network is not a static function, but a function that changes with each training episode.

[0136] In one scenario, initial weights may be randomly selected for each neuron. Training is then performed using historical data until the weights converge to values ​​that perform well enough to match the input data (sensor values) with the correct output data (whether a fall has occurred). Once the machine learning algorithm has "learned" by selecting appropriate weights to reduce errors in predicting falls or no falls, the processor executes the algorithm based on real-time sensor inputs. The training of the algorithm can continue even after it has been used by the optimization and backpropagation techniques described herein (commonly referred to as updating the algorithm).

[0137]

[0137] A neural network trained on many observations of historical data can achieve robust and accurate prediction capabilities for fall or no-fall judgments, improving the accuracy of fall detection and reducing false positives that may result in triggering events. Once the network has learned a stable predictive pattern, it can be deployed to make real-time predictions. Unlike the algorithm training process, employing an algorithm for fall prediction is very quick, allowing the algorithm to make decisions in near real-time. Furthermore, training (commonly referred to as updating the algorithm once it is deployed in operation) can continue even after the algorithm is operational given new data, and can be used in conjunction with reinforcement learning techniques to optimize learning and further improve it to increase the accuracy of fall or no-fall judgments.

[0138]

[0138] Many different sensor types and analysis processes are described herein for collecting and analyzing data indicative of a possibly anticipated fall. Generally, such sensors comprise physical sensors that output a signal using hardware and software elements. In one embodiment, such sensors are attached to or incorporated into a wearable device. In one embodiment, the measurement data is sent to a cloud-based AI / ML algorithm for processing and fall detection. A service provider (entity) can receive the measurement data from the monitored person and process the data through its algorithms to provide the algorithm. If the algorithm determines that a fall has occurred, the service provider issues a notification to parties identified by the monitored person.

[0139]

[0139] The previous description of the embodiments is provided to enable any person skilled in the art to practice the disclosure. Thus, the present disclosure is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and features disclosed herein.

Claims

1. A fall detection processor using machine learning; a non-transitory computer readable medium comprising a database configured to store an expected activity level of a wearer; a plurality of motion sensors, each having an individual fall motion threshold, the motion sensors configured to transmit a fall motion signal to the fall detection processor when the individual fall motion threshold is exceeded; at least one height sensor having a fall height threshold, the at least one height sensor being configured to transmit a fall height signal to the fall detection processor when the fall height threshold is exceeded; a force sensor configured to send a force signal to the fall detection processor when a force threshold is exceeded, the force sensor having a force threshold; Equipped with configured to issue an alert if a minimum number of sensors among the plurality of motion sensors, the at least one height sensor, and the force sensor indicate the occurrence of a fall; the minimum number of sensors is automatically set by the fall detection processor based in part on the expected activity level of the wearer; The wearable device, wherein the fall detection processor updates the predicted activity level of the wearer stored in the database to an actual activity level of the wearer.

2. 10. The wearable device of claim 1, wherein the fall detection processor is further configured to vary the minimum number of sensors based in part on the expected activity level of the wearer.

3. 3. The wearable device of claim 2, wherein the fall detection processor is further configured to adjust the individual's fall motion threshold, the fall height threshold, or the force threshold based in part on the expected activity level of the wearer.

4. The wearable device of claim 3 , wherein the fall detection processor is further configured to turn sensors on or off based in part on the expected activity level of the wearer.

5. The wearable device of claim 4 , further comprising a user interface for enabling the wearer to adjust the expected activity level of the wearer.

6. the user interface includes a microphone with an audio processor having an impact detection threshold, the audio processor further configured to transmit an impact signal to the fall detection processor when the impact detection threshold is exceeded; The wearable device of claim 5 , wherein the total number of sensors further includes the microphone.

7. the motion sensor is a gyro sensor having a tilt-detection threshold, the gyro sensor being further configured to transmit a tilt signal to the fall detection processor when the tilt-detection threshold is exceeded; The wearable device of claim 5 , wherein the total number of sensors further includes the gyroscope sensor.

8. The wearable device of claim 5 , wherein the fall detection processor adjusts the expected activity level of the wearer based on a time of day.

9. The wearable device of claim 5 , wherein the fall detection processor uses regression techniques, decision trees or decision matrices to help determine the expected activity level of the wearer.

10. The wearable device of claim 5 , wherein the fall detection processor repairs false positive fall signals by isolating sensors.

11. 6. The wearable device of claim 5, wherein the fall detection processor uses a support vector machine (SVM) for template matching to help determine the expected activity level of the wearer.

12. The wearable device of claim 5 , wherein the fall detection processor is a neural network.

13. storing the wearer's expected activity level in a database on a non-transitory computer readable medium; receiving, by a fall detection processor, fall motion signals from a plurality of motion sensors, each having a fall motion threshold, when a respective fall motion threshold is exceeded; receiving, by the fall detection processor, a fall height signal from at least one height sensor having a fall height threshold when the fall height threshold is exceeded; receiving, by the fall detection processor, a force signal from a force sensor having a force threshold if the force threshold is exceeded; issuing an alert to the fall detection processor if a minimum number of sensors among the plurality of motion sensors, the at least one height sensor, and the force sensor indicate the occurrence of a fall; the minimum number of sensors is automatically set by the fall detection processor based in part on the expected activity level of the wearer; updating, using the fall detection processor, the predicted activity level of the wearer stored in the database to an actual activity level of the wearer; A method for detecting a fall of a wearable device comprising:

14. 14. The method of claim 13, wherein the fall detection processor is further configured to vary the minimum number of sensors based in part on the expected activity level of the wearer.

15. 15. The fall detection method of claim 14, wherein the fall detection processor is further configured to adjust the individual fall motion thresholds, the fall height thresholds, or the force thresholds based in part on the expected activity level of the wearer.

16. 16. The method of claim 15, wherein the fall detection processor is further configured to turn sensors on or off based in part on the expected activity level of the wearer.

17. 17. The method of claim 16, further comprising adjusting the expected activity level of the wearer with input from a user interface.

18. the user interface includes a microphone with an audio processor having an impact detection threshold, the audio processor further configured to transmit an impact signal to the fall detection processor when the impact detection threshold is exceeded; The method of claim 17 , wherein the total number of sensors further includes the microphone.

19. the motion sensor is a gyro sensor having a tilt-detection threshold, the gyro sensor being further configured to transmit a tilt signal to the fall detection processor when the tilt-detection threshold is exceeded; 20. The method of claim 17, wherein the total number of sensors further includes the gyroscope sensor.

20. When executed by the fall detection processor, the wearable device is storing the wearer's predicted activity level in a database on a non-transitory computer readable medium; receiving, by the fall detection processor, fall motion signals from a plurality of motion sensors, each having a fall motion threshold, when an individual fall motion threshold is exceeded; receiving, by the fall detection processor, a fall height signal from at least one height sensor having the fall height threshold if the fall height threshold is exceeded; receiving, by the fall detection processor, a force signal from a force sensor having a force threshold when the force threshold is exceeded; causing the fall detection processor to issue an alert if a minimum number of sensors among the plurality of motion sensors, the at least one height sensor, and the force sensor indicate the occurrence of a fall; the minimum number of sensors is automatically set by the fall detection processor based in part on the expected activity level of the wearer; A non-transitory computer-readable storage medium encoded with data and instructions, wherein the fall detection processor updates the predicted activity level of the wearer stored in the database to an actual activity level of the wearer.

Images