Machine learning for fall prevention via pressure sensors

Abstract

An embodiment of the present disclosure relates to machine learning. Techniques for improved machine learning are provided. Pressure data corresponding to one or both feet of a user is received, and a predicted fall risk is generated by processing the pressure data using one or more trained machine learning models. In response to determining that the predicted fall risk meets one or more defined criteria, an intervention is selected for the user, and application of the intervention is initiated.

Description

[Technical field]

[0001] (CROSS REFERENCE TO RELATED APPLICATIONS) This application claims the benefit of co-pending U.S. Provisional Patent Application No. 63 / 264,138, filed November 16, 2021, and co-pending U.S. Provisional Patent Application No. 63 / 265,083, filed December 7, 2021. The aforementioned related patent applications are hereby incorporated by reference in their entireties.

[0002] Introduction FIELD OF THE DISCLOSURE Embodiments of the present disclosure relate to machine learning and, more particularly, to evaluating pressure sensor data using machine learning.

[0003] In a wide variety of medical (and non-medical) settings, falling (e.g., from a standing, sitting, or walking position) is one of the most common and most dangerous risks for users. This is especially true in people with reduced mobility, e.g., due to age, physical conditions, injuries, and the like. Traditionally, fall prevention involves general-purpose and most simple solutions, such as the use of canes, the provision of handrails, and potentially physical therapy to improve strength. However, these approaches do not consider a wide variety of practical issues and do not provide adequate or complete solutions. Moreover, they do not enable targeted prediction and specific intervention of fall risk.

[0004] Improved systems and techniques for predicting and preventing balance concerns and falls are needed. Summary of the Invention [Means for solving the problem]

[0005] According to one embodiment presented in the present disclosure, a method for training a machine learning model is provided, the method including receiving pressure data corresponding to one or both feet of a user, determining an indicator indicating whether the user has fallen within a defined time window after the pressure data is collected, and training one or more machine learning models to predict a risk of falling based on the pressure data and the indicator.

[0006] According to one embodiment presented in the present disclosure, a method for predicting an event using a machine learning model is provided, the method including receiving pressure data corresponding to one or both feet of a user, generating a predicted fall risk by processing the pressure data using one or more trained machine learning models, and selecting an intervention for the user in response to determining that the predicted fall risk meets one or more defined criteria, and initiating application of the intervention.

[0007] The following description and the annexed drawings set forth in detail certain illustrative features of the one or more embodiments. [Brief description of the drawings]

[0008] The accompanying drawings depict certain aspects of one or more embodiments and therefore should not be considered as limiting the scope of the present disclosure.

[0009] [Figure 1] FIG. 1 depicts an example environment for using machine learning to evaluate sensor data and predict fall events.

[0010] [Diagram 2] FIG. 2 depicts an example environment for using an intermediary communication device and machine learning to evaluate sensor data and predict fall events.

[0011] [Diagram 3]FIG. 3 depicts an exemplary workflow for generating labeled data, training a machine learning model, and predicting fall events.

[0012] [Figure 4] FIG. 4 depicts an example floor plan in which determined fall risk areas are shown.

[0013] [Diagram 5] FIG. 5 is a flow diagram depicting an exemplary method for generating labeled data, training a machine learning model, and predicting falls.

[0014] [Figure 6] FIG. 6 is a flow diagram depicting an example method for training a machine learning model to predict a fall event based on sensor data.

[0015] [Figure 7] FIG. 7 is a flow diagram depicting an exemplary method for training likelihood and severity models to predict fall events.

[0016] [Figure 8] FIG. 8 is a flow diagram depicting an example method for evaluating sensor data using a trained machine learning model.

[0017] [Figure 9] FIG. 9 is a flow diagram depicting an exemplary method for predicting fall risk and potential severity using a machine learning model.

[0018] [Figure 10] FIG. 10 is a flow diagram depicting an exemplary method for predicting fall events and severity using machine learning.

[0019] [Figure 11]FIG. 11 is a flow diagram depicting an example method for initiating preventative and / or corrective actions based on machine learning model output.

[0020] [Figure 12] FIG. 12 is a flow diagram depicting an exemplary method for evaluating and selecting a personalized intervention using machine learning.

[0021] [Figure 13] FIG. 13 is a flow diagram depicting an exemplary method for determining aggregate fall risk using machine learning.

[0022] [Figure 14] FIG. 14 is a flow diagram depicting an exemplary method for training one or more machine learning models to predict fall risk.

[0023] [Figure 15] FIG. 15 is a flow diagram depicting an exemplary method for predicting fall risk using machine learning.

[0024] [Figure 16] FIG. 16 depicts an exemplary computing device configured to implement various aspects of the disclosure.

[0025] To facilitate understanding, the same reference numbers have been used, where possible, to designate the same elements common to the figures. It is contemplated that elements and features of one embodiment may be beneficially incorporated into other embodiments without further recitation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026] Detailed Description Aspects of the present disclosure provide apparatus, methods, processing systems, and computer-readable media for improved machine learning models for sensor evaluation and initiation of corrective and / or preventative actions.

[0027] In some embodiments, pressure data indicative of pressure at one or more locations on a user's foot is evaluated using one or more trained machine learning models to predict the likelihood of a future fall and / or the severity of such a potential fall. As used herein, pressure data from a "user's foot" generally encompasses pressure on the user's actual foot, pressure on the user's leg or foot prosthesis, and the like. In some aspects, pressure data relating to the user's foot may also include pressure distribution on an assistive device used by the user, such as a cane, walker, and the like.

[0028] As used herein, fall risk (also referred to as risk of falling) may generally refer to the likelihood or probability of such a fall in the future, the severity of such a potential fall, or a combination of the likelihood and severity. Based on this analysis, certain preventative and / or corrective measures can be initiated.

[0029] In some embodiments, pressure data is collected via one or more insoles in the user's shoe, the insoles configured with one or more sensors (e.g., pressure sensors) to detect pressure exerted by the user's foot at one or more locations on the foot. Using this pressure data, the system is able to perform a balance assessment (e.g., using pressure on each portion of each insole). While pressure data is used in various examples discussed herein, in some embodiments, other sensor data (such as accelerometer data, inertial sensor data, and the like) can be collected for the user as well.

[0030] In some embodiments, pressure data can be monitored as a user walks and otherwise moves (e.g., in their normal environment, in shoes they normally wear, and the like). Generally, this pressure data can be used in various computing systems, including to predict and / or prevent falls, to inform or improve treatment options (e.g., to suggest improved physical therapy options), and the like.

[0031] In some embodiments, the system evaluates the pressure data at discrete instants in time. In other embodiments, the system evaluates the pressure data over a time window and analyzes the user's movement or walking patterns. For example, the system may evaluate the data and determine if the user typically walks on one foot while somewhat dragging the other (which may indicate an increased risk of falling).

[0032] In some embodiments, the data is evaluated using one or more machine learning (ML) and / or artificial intelligence (AI) models that may enable the system to determine which pressure data is normal and / or acceptable and which pressure data may trigger further concerns.

[0033] In some embodiments, in addition to or instead of evaluating the data and predicting the potential for a fall, the system uses a model to predict the fall severity of such a fall. For example, if a user falls forward, they may be more likely to suffer reduced harm compared to falling backwards. In some embodiments, certain pressure data may be used by the trained model to predict this severity. In some embodiments, the pressure data may thus be used to predict the direction in which the user will fall. In some embodiments, potential outcomes are also determined or generated (e.g., head injury from falling backwards, wrist fracture from falling forward, hip fracture from falling sideways, and the like).

[0034] In some embodiments, to train these models, the system uses tagged training data collected from one or more users while the users are engaged in normal activities (e.g., walking). In one such embodiment, data is collected from a group of users over time, and whenever a fall occurs or is reported, sensor data from the user who has fallen may be tagged as a precursor to a fall. For example, the system may automatically tag data from N minutes ago as a precursor to an imminent fall, and / or tag data from M days ago to indicate that a fall followed several days later. In some embodiments, the system further tags data based on characteristics of the fall, such as fall direction, fall severity, and the like.

[0035] In at least one embodiment, in addition to pressure data, the training data may further indicate various characteristics of the user, such as their age, sex, weight, height, blood pressure, pulse, any mobility limitations, whether the user is using any assistive devices such as a cane or walker, and the like. By including such characteristics in the input to the model, the system may be able to generate more accurate fall predictions, as discussed in more detail below.

[0036] In embodiments, this training data can be used to train one or more machine learning models to predict falls (either imminent falls or potential future falls), predict the severity of such falls, and the like.

[0037] In embodiments, these predictions can be used to drive a wide variety of patient-specific interventions. For example, the system may respond to an imminent risk by dispatching help (e.g., alerting a nearby caregiver), instructing the user to stop, sit, or otherwise remain still, and the like. In some embodiments, the system can additionally or alternatively initiate various preventative measures, such as modifying a physical therapy plan (e.g., to strengthen the patient's muscles that would help prevent certain types of predicted falls), assigning or providing an assistive device, such as a cane or walker, and the like.

[0038] In some embodiments, the pressure data includes a number of individual pressure points on each foot of each user. For example, for each foot, the system may receive and evaluate pressure at: left front foot point, right front foot point, left rear foot point, right rear foot point, center of foot, front center of foot, rear center of foot, left center of foot, right center of foot, and the like.

[0039] In some embodiments, in addition to or instead of predicting falls, the model can be used for diagnostic purposes, to detect when a user is exerting excess pressure on one side, which may indicate pain or other issues on the other side.

[0040] Further, in some embodiments, the model can be used not only to predict impending falls (e.g., falls that may occur in the next few seconds, minutes, or days), but also to assess longer term trends (e.g., to determine whether a user is losing their balance or is currently stumbling more frequently than before, even in the absence of an actual fall).

[0041] In at least one embodiment, the predictive model can also be used to identify and remediate potential hazards within a physical space, such as carpeting, non-uniform floors, cords across aisles, and the like. For example, the system may identify areas within a facility where fall events are more common, where likely fall events occur (e.g., detected trips, stumbles, or trips), where predicted falls are often flagged, and the like. In some embodiments, the system can generate a heat map for the facility, showing average fall risk and / or predicted severity across the floor plan and across a number of users and / or a point in time. By evaluating this data, the system can identify potential hazards that cause (or may cause) falls, thereby initiating remediation that may prevent such occurrences. An Exemplary Environment for Machine Learning to Evaluate Sensor Data

[0042] FIG. 1 depicts an example environment 100 for using machine learning to evaluate sensor data and predict fall events.

[0043] In the illustrated environment 100, a set of one or more pressure sensors 105 is configured to record pressure data from one or both feet of a user. In some embodiments, the pressure sensors 105 are included within a wearable device. For example, the pressure sensors may be included within an insole that is inserted into the user's shoe. In some embodiments, the pressure sensors 105 can be integrated into the shoe itself. In embodiments, there may be any number of pressure sensors 105 used within the environment 100. For example, the system may use four or more pressure sensors 105 per user's foot.

[0044] As shown, the pressure sensor 105 transmits or otherwise provides pressure data 110 to the machine learning system 115. In various embodiments, this pressure data 110 may be provided using any suitable technology, including wired or wireless communication. For example, in one embodiment, the pressure sensor 105 may transmit the pressure data 110 to the machine learning system 115 using cellular communication technology. In some embodiments, the pressure sensor 105 transmits the pressure data 110 using one or more local wireless networks (such as a Wifi network, a Near Field Communication (NFC) network, a Bluetooth network, etc.). In at least one embodiment, the pressure sensor 105 may transmit the pressure data 110 to one or more intermediate devices, which may then automatically forward the data to the machine learning system 115. For example, the pressure sensor 105 may transmit pressure data 110 to a smartphone, tablet, or other device associated with the user (e.g., via Bluetooth), which can automatically forward the data to the machine learning system 115 (e.g., via a Wifi or cellular connection).

[0045] In some embodiments, the pressure data 110 is transmitted over one or more networks, including the Internet. That is, the machine learning system 115 may reside at a location remote from the users and the pressure sensors 105 (e.g., in the cloud). Although a single set of pressure sensors 105 is illustrated for conceptual clarity, in embodiments, data from any number and variety of pressure sensors 105 (and any number of users) may be provided to the machine learning system 115 as well.

[0046] In an embodiment, the machine learning system 115 may generally train one or more machine learning models, analyze the pressure data 110, and / or use the trained models to evaluate the pressure data 110. In the illustrated example, the machine learning system 115 may also optionally receive a set of user records 112. For example, in some embodiments, the user records 112 may include information related to various characteristics of users (e.g., residents in a residential facility) associated with the pressure sensors 105. For example, the user records 112 may indicate, for each user, the pressure sensors 105 associated with a given user, as well as characteristics such as their age, height, weight, whether they use an assistive device such as a cane or walker, whether they have suffered a fall (and, if applicable, details related to the fall, such as time, severity, direction, cause, etc.), and the like.

[0047] In some embodiments, the user records 112 are used as inputs along with the pressure data 110 when training the models and using them to predict fall events. That is, the pressure data 110 from the user may be used as inputs to the model along with one or more of the characteristics indicated in the corresponding user record 112 for the user to generate predicted fall characteristics (e.g., timing, severity, and the like). During inference, this output can be used to drive various interventions, as discussed in more detail below. During training, this output can be compared against known indicators (e.g., indicating whether the user has fallen) to calculate a loss that is used to refine the model.

[0048] As used herein, loss generally corresponds to the difference between the actual model output and the target output. In embodiments, loss may be calculated using various algorithms or techniques, including cross-entropy loss, mean absolute error, and the like. Generally, the magnitude of the difference is directly correlated with the magnitude of the loss. Loss is used to refine the internal parameters of the model, such as weights and / or biases. Generally, the magnitude of the loss is directly correlated with the magnitude of the change required for these parameters. Thus, by iteratively generating losses (using one or more sets of training data), the system can iteratively refine the parameters and generate more accurate predictions.

[0049] In at least one embodiment, the machine learning system 115 can use machine learning to predict various fall-related events, such as the likelihood of a future fall (which may include imminent falls due to stumbling, etc., as well as falls in the distant future due to deterioration of balance over time, etc.), the potential severity of future falls, and the like. In some embodiments, a separate machine learning model is trained for each such output. This may allow the models to be specialized for their particular target output. For example, predicting whether a fall will occur may involve different features (or differently weighted features) compared to predicting the time a fall will occur and / or the severity of the fall that will occur. By training a separate model for each such output, the system allows each model to learn specialized parameters that are best suited to accurately predict a particular output.

[0050] In one embodiment, to train the model, the machine learning system 115 can collect pressure data 110 and / or user records 112 over time from any number of users. For example, each user in a facility (e.g., a long-term care facility) can have an associated set of pressure sensors 105 (e.g., a pair of sensor-equipped insoles or shoes), and the pressure data from each user can be used to train the model. In at least one embodiment, the machine learning system 115 or another component can monitor the environment and tag this pressure data based on various factors, including characteristics of the corresponding user (e.g., age, weight, height, and the like).

[0051] In some embodiments, the data is associated with a "ground truth" indicator based on whether the corresponding user suffered a fall. For example, each time a participating user falls (e.g., as indicated in user record 112), machine learning system 115 (or another system) can retrieve the corresponding record of pressure data 110 from that user for one or more preceding times (e.g., for N seconds, minutes, hours, or days prior to the fall) and flag this data as an indication that a fall is imminent. In at least one embodiment, the indicator further indicates the elapsed time between the sensor reading and the fall. For example, a first record may be flagged to indicate that the user fell 5 minutes after the data was recorded, while a second record indicates that the user fell 30 seconds later.

[0052] In at least one embodiment, the pressure data 110 may additionally or alternatively be tagged to indicate various aspects of the fall, such as the direction of the fall (e.g., left, right, forward, backward, and the like). Similarly, the data may be tagged to indicate the severity of the fall (e.g., whether it resulted in any fractures or other injuries, the extent of these injuries, and the like). In at least one embodiment, the data may additionally or alternatively be tagged with contextual data that may be relevant in predicting similar falls. For example, using video or images of the fall (captured by one or more cameras), the system may determine whether the patient tripped on an object, whether the user was actually using any assistive device, such as a cane, and the like. This information may enable the system to generate more robust training data that more accurately accounts for the particular context of the fall.

[0053] By collecting such data over time and from various users, the machine learning system 115 can train or refine various machine learning architectures to automatically build training data sets that can be used to predict future falls, the future proximity in which a fall will occur, the direction in which a user will fall, the severity of the fall that will occur, and the like.

[0054] In at least one embodiment, if a given user does not fall (or does not fall within a defined window), the machine learning system 115 can automatically mark the corresponding records as not indicative of an upcoming fall. For example, the machine learning system 115 may persistently or periodically identify unlabeled pressure records (e.g., records that have not yet been labeled as a precursor to a fall) that are older than a user-defined threshold (e.g., older than one week) and mark these records as not indicative of a fall (if a fall is not reported in the user records 112). In this manner, the machine learning system 115 can generate both positive training samples (e.g., pressure data 110 and / or user characteristics that indicate a fall is imminent) and negative training data (e.g., pressure data 110 and / or user characteristics that indicate the user is stable and / or a fall is not imminent).

[0055] In some embodiments, as discussed above, the machine learning system 115 may train several machine learning models to address various desired predictions. For example, a first model may be trained to predict whether a user will fall, while a second model is used to predict the future proximity that a fall will occur. In some aspects, a first model may be trained to predict an imminent fall (e.g., within the next M seconds based on a narrow input time window), while others are trained using progressively longer intervals for more distant predictions (e.g., evaluating data collected over a period of time to determine whether a fall is likely to occur within the next hour). Other models may be used to predict fall direction, fall severity, and the like.

[0056] In general, training a machine learning model involves providing a set of pressure data 110 from one or more users (e.g., pressure data covering a defined time window) as input to the model and generating a predicted output. In some embodiments, the input also includes one or more characteristics of the users (such as age, whether they use an assistive device, and the like), as discussed above. In the initial stages of training, the output may be relatively random or unreliable (e.g., due to random weights and / or biases used to start the model). The generated output is then compared against relevant indicators for which the model has been trained (e.g., whether a fall occurred, the severity of the fall, and the like), generating losses that can be used to refine the model (e.g., using backpropagation in the case of neural networks). In general, the refinement process may be performed for each user or time window individually (e.g., using stochastic gradient descent) or as a batch (e.g., using batch gradient descent). Further, in embodiments, the machine learning system 115 can train a model to operate on a window of data (e.g., a collection of discrete or sustained pressure values) or based on a single recording (e.g., pressure data at a given moment).

[0057] In an embodiment, once the models are trained, the machine learning system 115 can deploy them for use in real time. In one embodiment, models may be trained on one or more systems and deployed to one or more other systems, as discussed above. In other embodiments, the machine learning system 115 can both train the models and use them to make inferences.

[0058] In some embodiments, during inference, the machine learning system 115 receives pressure data 110 in real time (or near real time) from participating users. In at least one embodiment, the machine learning system 115 also receives user records 112 for specific users. As used herein, inferring refers to a stage of model deployment, where models are trained and deployed for use in evaluating actual inputs during runtime. The model output, in some embodiments, may be referred to as a "inference," and thus this stage may be referred to as the "inference step." During inference, the machine learning system 115 may process this pressure data (in some embodiments, along with user characteristics such as age) using one or more of the trained models to predict potential falls.

[0059] In at least one embodiment, the machine learning system 115 can process the data using the models sequentially, rather than using all the models in parallel. This can significantly reduce the computational cost of evaluating the pressure data 110, thereby reducing the memory space and computational power required, and reducing energy consumption and latency (e.g., when the system is overloaded with a large number of users).

[0060] For example, the machine learning system 115 may initially use an initial model (e.g., to determine whether a fall is imminent). If a fall is determined to be imminent, the machine learning system 115 may immediately initiate corrective or preventative measures, as discussed in more detail below. The fall severity, direction, and other characteristics of a potential fall may be predicted later, or may be bypassed entirely (e.g., as unrelated once a fall occurs or if it does not occur). In one such embodiment, if a fall is not imminent but may occur at some point in the future, the machine learning system 115 may optionally use one or more other models to predict the nature of the fall, such as the time it will occur, the direction in which the user will fall, the severity, and the like.

[0061] If a fall is not predicted, the machine learning system 115 may refrain from initiating any further action. However, in the illustrated example, if a fall is predicted, the machine learning system 115 may initiate a wide variety of corrective and / or preventative actions. For example, if a fall is imminent (e.g., within a defined number of seconds or minutes), the machine learning system 115 may initiate actions such as commanding the user to sit or grasp an object for support, alerting a nearby caregiver to immediately attend to the user, and the like.

[0062] In some embodiments, if the predicted fall is further into the future (e.g., more than a defined number of minutes into the future), the machine learning system 115 can initiate other corrective actions, such as suggesting or prescribing medical devices, including canes, walkers, and wheelchairs. In some embodiments, the machine learning system 115 can suggest or prescribe one or more physical therapy plans or modifications based on the prediction. For example, if the machine learning system 115 predicts that the user may fall toward their left, the machine learning system 115 can determine that physical therapy is needed to strengthen the muscles on the left side of the user's body (or otherwise correct any pain or other concerns on this side). In response, the machine learning system 115 can suggest or initiate physical therapy to correct these concerns.

[0063] In at least one embodiment, based on the aggregated fall predictions or detections, the machine learning system 115 can additionally or alternatively provide corrective or preventative actions 120 throughout the facility. For example, if several falls (or predicted falls) occur within a given area (e.g., if the machine learning system 115 observes that users have tripped at a given location within the facility, and thus, on average, predicted impending falls are increasing at this location), the machine learning system 115 may flag this area for manual review. This may indicate, for example, that a hazard exists within the area (e.g., a raised carpet or a cable across the floor).

[0064] In embodiments, by providing these targeted and user-specific interventions (including immediate assistance as well as more preventative and long-term treatments such as medical devices and physical therapy), the machine learning system 115 can dramatically reduce falls (and fall severity), thereby improving user outcomes. Moreover, by using several machine learning models iteratively, the machine learning system 115 can provide this granular and specific analysis using reduced computational resources and expense only when the output from a prior model meets certain criteria. Exemplary Environment for Evaluating Sensor Data and Predicting Fall Events Using Intermediate Communication Devices and Machine Learning

[0065] FIG. 2 depicts an example environment 200 for using an intermediary communication device and machine learning to evaluate sensor data and predict fall events.

[0066] In the illustrated environment 200, a set of pressure sensors 105 (e.g., one or more devices on each foot, where each device may include one or more physical sensors) are configured to record pressure data from a user's feet. For example, as discussed above, the pressure sensors 105 may be included in a wearable device, such as inserted into the user's shoe, included in an insole, or integrated into the shoe itself.

[0067] In the illustrated environment 200, the pressure sensor 105 is configured to transmit or otherwise provide the pressure data 210 to a user device 215, which can then automatically forward the pressure data 210 (and, in some embodiments, other related data) to the machine learning system 115. That is, while in environment 100, the pressure sensor 105 may transmit the pressure data directly to the machine learning system 115, the environment 200 employs a user device 215 of the user.

[0068] In the illustrated example, the user device 215 is a smartphone. However, in embodiments, other devices may be used. Generally, the user device 215 corresponds to a portable computing device (such as a phone, tablet, smartwatch, and the like) associated with a user. The user (or a support person, such as a caregiver) may use the user device 215 to link the pressure sensor 105 to the corresponding user and user device 215. For example, the user may pair the user device 215 and the pressure sensor 105 using a short-range wireless technology, such as Bluetooth. The pressure sensor 105 may then transmit the pressure data 210 to the user device 215, which may optionally be tagged based on user characteristics, such as the user's identity, the user's age, and the like.

[0069] In the illustrated example, the user device 215 then automatically transfers the pressure data 210 to the machine learning system 115 over a network 220 (e.g., the Internet). In the illustrated example, the user records 112 are separately provided to the machine learning system 115 over the network 220. In at least one embodiment, the user records 112 are alternatively or additionally provided by the user device 215.

[0070] For example, the user device 215 may tag the pressure data 210 with a unique identifier for the user, and the machine learning system 115 may use this identifier to retrieve or determine a corresponding user record 112 that contains information related to the user. Alternatively, the user device 215 may directly contain some or all of this information, such as the user's age, height, weight, and the like.

[0071] In an embodiment, as discussed above, the machine learning system 115 may generally train one or more machine learning models, analyze the pressure data 210 and the user records 112, and / or use the trained models to evaluate the pressure data 110 and the user records 112. For example, as discussed above, the pressure data 210 and the user records 112 may be used as inputs to generate a predicted output (e.g., a fall likelihood or a fall severity), which may be evaluated against ground truth indicators (e.g., determined from the user records 112) to refine the model. Once trained, as discussed above, the machine learning system 115 may use the models to process the pressure data 210 and the user records 112 and predict, for a given user, whether a fall is likely to occur, the potential severity of the fall, and the like. Example Workflow for Generating Labeled Data, Training a Machine Learning Model, and Predicting Fall Events

[0072] FIG. 3 depicts an example workflow 300 for generating labeled data, training a machine learning model, and predicting fall events.

[0073] In the illustrated workflow 300, the machine learning system 115 receives data from the pressure sensor 105, as discussed above. For example, one or more pressure sensors 105 may transmit pressure data continuously, periodically (e.g., every second, every 5 seconds, etc.), and the like. In one embodiment, as the pressure data is received, the machine learning system 115 may tag or otherwise associate it with a corresponding user. For example, the machine learning system 115 may identify a corresponding user record 305 (e.g., based on an identifier included within the pressure data) such that the pressure data may be associated with relevant user characteristics that are used as inputs to the model, such as age, weight, and the like. In some embodiments, the received pressure data already includes such characteristics, as discussed above. In the illustrated workflow 300, the machine learning system 115 may store this (tagged) pressure data for future labeling, as discussed in more detail below.

[0074] In some embodiments, the machine learning system 115 can evaluate the user records 305 periodically, continuously, or in response to a defined event to flag the pressure data. For example, the machine learning system 115 may evaluate the user records 305 for registered users periodically (e.g., daily or weekly) to determine if any of the users have suffered a fall. In some embodiments, the machine learning system 115 can receive fall information automatically (e.g., as a pushed alert or notification, rather than by actively checking records for fall reports).

[0075] When a fall is indicated in the user records 305, the machine learning system 115 can retrieve one or more corresponding records of pressure data and mark it to indicate a fall. For example, as discussed above, the machine learning system 115 can mark each pressure record to indicate information such as the fact that a fall occurred, the future proximity (relative to the record) that the fall will occur, the direction of the fall, the severity of the fall, and the like. In some embodiments, the machine learning system 115 can mark a sequence of pressure records in this manner. For example, one record can be marked to indicate that it was immediately prior to the fall (e.g., within a defined period of time), while a second record is marked to indicate that the fall occurred one hour after the pressure data was recorded. This can enable the machine learning system 115 to train a set of models to predict the time that a fall will occur.

[0076] In the illustrated workflow 300, the machine learning system 115 can then store the pressure data as labeled training data 310. As discussed above, this labeled training data 310 can be used to train and / or refine one or more machine learning models. In various embodiments, this training may be performed periodically (e.g., weekly) or in response to a defined event. For example, in at least one embodiment, if a fall occurs, the machine learning system 115 can determine whether the model predicted the fall (or would have predicted it if the pressure data had not been used to evaluate). If the model accurately predicted the fall and / or fall characteristics, the machine learning system 115 can label and store the data for subsequent refinement. In some embodiments, if the model fails to predict the fall and / or characteristics, the machine learning system 115 may immediately use the data (and any other labeled training data 310 not yet used to train the model) to refine the model for more accurate predictions.

[0077] In some embodiments, in addition to positively labeled training data 310 (e.g., samples that were predictive of a fall), the machine learning system 115 can also generate labeled training data 310 that corresponds to negative examples. For example, the machine learning system 115 may periodically or persistently identify pressure data that is older than a certain defined threshold (e.g., older than one week). If the user records 305 do not indicate that the user has fallen within the defined window, the machine learning system 115 may automatically label the pressure data as not indicative of an upcoming fall. These negative samples can then be stored as labeled training data 310 for future training and / or refinement, as discussed above.

[0078] As discussed above, these negative examples may also be used periodically (e.g., weekly) or in response to defined events. For example, in at least one embodiment, if the model predicts a fall (or predicts specific fall characteristics) but it does not occur (or the fall characteristics are different), the machine learning system 115 may immediately use the data (and any other labeled training data 310 not already used to train the model) to refine the model for more accurate predictions.

[0079] In this manner, the machine learning system 115 can automatically generate labeled training data 310 for training and refining machine learning models. This automated label generation can result in labeled training data 310 that is efficient and fast, and generally with reduced error (e.g., mislabeled data). Because machine learning models generally require vast amounts of training data, this automated workflow 300 can enable training more accurate models than would be possible using traditional systems (e.g., relying on manually curated data). These more accurate models can then be used for significantly improved fall prediction and intervention. An example floor plan showing the determined fall risk areas.

[0080] 4 depicts an example floor plan 400 with determined fall risk areas indicated. In some embodiments, floor plan 400 may be generated and / or augmented by a machine learning system, such as machine learning system 115 of FIG.

[0081] In some embodiments, as discussed above, a machine learning system (or another system or device) can aggregate fall predictions and / or events to provide deeper insight into potential fall hazards in one or more physical spaces or locations. In the illustrated example, floor plan 400 corresponds to a residential facility (e.g., a nursing home or long-term care facility) in which a user resides. However, in embodiments, the techniques described herein can be readily applied to a wide variety of locations and facilities.

[0082] In the illustrated example, three areas, locations, or regions are indicated by dashed circles 405A-C to indicate that they may be high-risk areas. In some embodiments, to identify these areas, a machine learning system can evaluate reported falls and / or predicted falls for the facility. For example, for each reported (actual) fall, the machine learning system can determine where in the facility the fall occurred. This data may be used to generate an augmented floor plan (e.g., using a heat map) that shows areas where falls occur more frequently.

[0083] In some embodiments, the machine learning system can additionally or alternatively aggregate predicted fall data and indicate problem areas. For example, whenever a machine learning model predicts that a fall is imminent (e.g., due to a user tripping), the machine learning system can determine where the predicted fall occurred (even if an actual fall does not follow). Similarly, by aggregating these predicted falls across the facility, the machine learning system can identify areas where falls are frequently predicted or likely to occur, even if a fall does not occur. This data may also be output, such as via an augmented floor plan (e.g., using a heat map), to indicate these problem areas.

[0084] For example, in the illustrated floor plan 400, area 405A is indicated as potentially problematic (e.g., due to the number or frequency of actual or predicted falls in that area). In embodiments, the particular technique used to indicate area 405 may vary depending on the particular implementation. For example, the machine learning system may generate a heat map for output via a graphical user interface (GUI). In some embodiments, the machine learning system may identify areas where actual and / or predicted falls meet one or more defined criteria (e.g., number of reported or predicted events, frequency of events, and the like) and specifically indicate these areas by highlighting them, outlining them, and the like.

[0085] Based on floor plan 400, it may be inferred that area 405A is potentially dangerous due to, for example, changes in floor level between indoor and outdoor spaces, threshold heights at doorways, and the like. Based on this information, a user or administrator may investigate the area and / or take corrective action if necessary, such as by posting fall warning signs, lowering door threshold heights, adding handrails, providing ramps, the like, etc.

[0086] In the illustrated embodiment, area 405B is also indicated as a potential hazard. For example, this area 405B may correspond to a location where a carpet is lifted or loose, where a cable or other obstruction crosses the room, and the like. As above, a user or administrator may similarly inspect area 405A, identify the hazard, and take appropriate corrective action.

[0087] As shown, the machine learning system has also flagged area 405C as potentially dangerous. Notably, this area 405C is within a room assigned to a specific user. That is, while areas 405A and 405B are generally within public areas that may be accessible to users, area 405C may be within a private room where one or more residents generally reside (e.g., where they sleep and store their items). In some embodiments, as discussed above, the indication of this area may generally enable a user or administrator to inspect area 405C and identify potential dangers (such as furniture or other objects, cords, rugs, and the like).

[0088] In at least one embodiment, a machine learning system or a user may infer that a particular user associated with area 405C may need assistance. That is, in addition to or instead of determining that the area itself is unsafe, the machine learning system may determine or infer that the corresponding user living in that room frequently has difficulty with balance, such as when getting up from a chair or getting out of bed. In this manner, the machine learning system can take preventative measures for the particular user, such as suggesting or commanding a caregiver to be present to assist with standing or sitting, or suggesting or assigning physical therapy tasks, as discussed above.

[0089] In this manner, by aggregating fall events and / or predicted fall events, the machine learning system can derive deeper insights about the facility and its users, thereby meaningfully improving facility safety and reducing potential harm. Exemplary Methods for Generating Labeled Data, Training Machine Learning Models, and Predicting Falls

[0090] 5 is a flow diagram depicting an example method 500 for generating labeled data, training a machine learning model, and predicting falls. In some embodiments, method 500 is performed by a machine learning system, such as machine learning system 115 of FIG. 1. In one embodiment, method 500 provides additional details regarding workflow 300 of FIG.

[0091] At block 505, a machine learning system receives pressure data from one or more pressure sensors (e.g., pressure sensor 105 of FIG. 1). For example, as discussed above, the machine learning system may be configured to receive pressure data from one or more pressure sensors associated with one or more users on a continuous or periodic basis. In some embodiments, the pressure data may also include or be associated with various user data, such as a user identifier, user characteristics, and the like.

[0092] At block 510, the machine learning system retrieves one or more corresponding user records for the user associated with the received pressure data. For example, based on a user ID included with the pressure data, the machine learning system can query one or more repositories and retrieve user records for the user, which can indicate various user characteristics (e.g., their age, height, weight, whether they use an assistive device, and the like). In some embodiments, the user records can additionally or alternatively indicate events, such as falls, that have occurred. For example, each time a user falls (or, in some embodiments, nearly falls), the user record may be updated (e.g., by a caregiver) to indicate the event.

[0093] In some embodiments, the machine learning system performs block 510 when the pressure data is received. For example, if the pressure data is from one or more prior times (e.g., the data is not provided live, but immediately recorded), the machine learning system may immediately retrieve the corresponding record. In some embodiments, if the pressure data is received live (e.g., immediately after it is recorded), the machine learning system may store the data and subsequently retrieve the user record in response to the occurrence of one or more defined criteria (e.g., after a defined period of time has passed, in response to being notified that a fall has occurred, and the like).

[0094] At block 515, the machine learning system evaluates the retrieved user records to determine whether the user suffered a fall. If not, method 500 continues to block 525, where the machine learning system tags the received pressure data. That is, the machine learning system may determine that the user did not fall (during the relevant window after the pressure data was recorded) and tag the data accordingly. In some embodiments, as discussed above, the machine learning system may wait until a defined criterion (e.g., a minimum period of time after the pressure data was recorded) is met before determining that a fall did not occur.

[0095] In at least one embodiment, in addition to determining whether a user has fallen, the machine learning system may determine whether the user has stumbled. Although stumbles may generally be less dangerous than actual falls, they may still indicate a potential concern or problem. By flagging data and predicting stumbles, the system may be able to improve the overall results for the user.

[0096] At block 515, if the machine learning system determines that a fall has occurred, the method 500 continues to block 520. At block 520, the machine learning system determines one or more characteristics of the fall, such as the direction of the fall (e.g., forward, backward, or side), the location of the fall (e.g., physical area within the facility), the severity of the fall (e.g., whether it resulted in a laceration, fracture, sprain, etc.), and the like. In an embodiment, as discussed above, these characteristics are indicated in a user record corresponding to the fall. In some embodiments, these characteristics are stored in a machine-readable format, using defined values ​​or indicators, etc., to indicate the directionality and / or severity of the fall. In other embodiments, the machine learning system may evaluate written records (e.g., written by a caregiver at the time of the fall) by using natural language processing, etc., to determine these fall characteristics.

[0097] Method 500 then continues to block 525, where the machine learning system labels the received pressure data to indicate the presence of a fall and / or fall characteristics. As discussed above, this labeled data can then be used to train or refine one or more machine learning models to predict a fall event. Method 500 then returns to block 505 to receive additional pressure data.

[0098] In this manner, machine learning systems can automatically generate labeled training data. As discussed above, this process can significantly improve the accuracy and reliability of the resulting machine learning models, and significantly reduce the costs and inaccuracies involved in manually labeling data. Exemplary Methods for Training Machine Learning Models

[0099] 6 is a flow diagram depicting an example method 600 for training a machine learning model to predict a fall event based on sensor data. In some embodiments, method 600 is performed by a machine learning system, such as machine learning system 115 of FIG.

[0100] At block 605, the machine learning system receives pressure data from one or more users. As discussed above, this pressure data may generally include pressure exerted by a user's foot or feet at one or more points on the foot (or feet) while the user stands, walks, runs, and the like. In some embodiments, as discussed above, the machine learning system may collect this data from several users (e.g., all patients in a long-term care facility) over a period of time (e.g., over a period of a year). In some embodiments, the machine learning system may further receive historical data for each user, such as their age, weight, height, mobility issues, and the like.

[0101] At block 610, the machine learning system determines a set of indicators for the received data. For example, as discussed above, the machine learning system (or another system) may identify a fall (e.g., recorded or reported in the user's patient data). The machine learning system (or another system) may then read corresponding pressure data records from one or more time points prior to the fall and label the data accordingly. For example, the machine learning system may read records from a 30 second window prior to the fall and label them as "fall imminent," read records starting 5 minutes prior and label them as "fall occurring," etc.

[0102] In some embodiments, as discussed above, the machine learning system may additionally or alternatively flag records based on other characteristics of the fall, such as the direction of the fall, the severity of the fall, and the like. Additionally, in some embodiments, if a sufficient period of time has passed without the user falling (e.g., if the records are older than one day, older than one week, and the like), the machine learning system may flag these records as not indicative of an upcoming fall. One example of determining pressure data flags is described in further detail above with reference to FIG. 5.

[0103] Once the labels have been determined, method 600 continues at block 615, where the machine learning system trains one or more machine learning models based on the labeled data, as discussed above. Although the illustrated example depicts data collection, labeling, and training as occurring sequentially for conceptual clarity, in some embodiments, the machine learning system may first collect and label data for a period of time and move on to model training after the period has elapsed (or after sufficient data has been collected).

[0104] In an embodiment, as discussed above, training the model generally involves providing the pressure data (as well as user characteristics or other patient data) as input to the model and using the labels to calculate a loss (based on the generated output). In this way, machine learning systems can iteratively refine models based on the labeled data to improve their predictions.

[0105] At block 620, the machine learning system determines whether training is complete. In embodiments, this determination may be made based on various termination criteria. For example, the machine learning system may determine whether a defined number of training cycles have been completed, whether a minimum number of training records have been used, whether a defined period of time has elapsed during training, whether a minimum preferred accuracy has been met (e.g., determined using the tagged data as test data), and the like. If training is not complete, method 600 returns to block 605 to receive a new set of pressure data.

[0106] If training is complete, method 600 continues at block 625, where the machine learning system deploys the trained model to make inferences. As discussed above, this may include using the model locally (on the machine learning system) to process new data, providing the model to one or more other systems or devices to predict falls, or the like. Exemplary Methods for Training Likelihood and Severity Models and Predicting Fall Events

[0107] 7 is a flow diagram depicting an example method 700 for training likelihood and severity models to predict fall events. In some embodiments, method 700 is performed by a machine learning system, such as machine learning system 115 of FIG 1. In the illustrated example, method 700 provides additional details relative to block 615 of FIG 6.

[0108] At block 705, the machine learning system trains a likelihood model based on the pressure data. In embodiments, the likelihood model is generally a machine learning model that learns to predict the likelihood of a user falling within a defined time period. In some embodiments, the machine learning system predicts whether a user will fall within a fixed time period. For example, a first model may predict an imminent fall (e.g., within 30 seconds), while a second model predicts a future fall (e.g., within hours or days), etc. In at least one embodiment, a single model predicts both the likelihood or probability of a fall and the expected time or delay to a fall.

[0109] In some embodiments, training the likelihood model involves processing the pressure data (and, in some embodiments, various user characteristics such as age and whether the user is using a cane, walker, or other assistive device) with the likelihood model to generate an output fall probability or prediction. This output can then be compared against ground truth indicators for the pressure data (e.g., indicating whether a fall actually occurred and, if applicable, the time). This comparison can be used to calculate a loss (e.g., cross-entropy loss) based on the difference between the prediction and the indicator, and the loss may be used to refine the likelihood model (e.g., using backpropagation). Although stochastic gradient descent (e.g., refining a model for each individual example individually) is described for conceptual clarity, in embodiments the machine learning system may refine the likelihood model using batch gradient descent (based on a set of examples).

[0110] At block 720, the machine learning system determines whether the indications in the pressure data indicate that a fall has occurred. If not, the method 600 ends and returns to block 620 of Figure 6. That is, if the pressure data is not related to a fall, the machine learning system does not train or refine any other models and the (negative) examples are used only to train the likelihood model.

[0111] At block 720, if the machine learning system determines that the data is predictive of a fall, method 700 continues to block 725, where the machine learning system trains a severity model based on the data. Generally, the severity model is a machine learning model that learns to predict the severity of a fall for a given patient, assuming that a fall will occur. In some embodiments, the machine learning system refines the severity model using only user data (e.g., age, weight, height, assistive devices, and the like). In others, the machine learning system uses pressure data about the user along with user characteristics to train the severity model.

[0112] For example, a severity model may be trained to receive pressure data (and in some embodiments, additional prior pressure data that may not have been used to train the likelihood model) along with one or more user characteristics and generate a prediction regarding the severity of a fall, if one occurs. In some embodiments, this prediction includes considering the direction of the fall (e.g., forward or backward). In at least one embodiment, the severity model may be trained to predict the direction in which a user will fall, and this direction can be used to predict fall severity. In another embodiment, the fall direction can be learned implicitly internally by the model and then used to predict severity. That is, in some embodiments, the predicted severity depends on the predicted fall direction, regardless of whether the severity model itself outputs a predicted direction of the fall.

[0113] In an embodiment, similar to the likelihood model, the severity model can be trained by generating a predicted severity using the relevant input data and comparing the predicted severity to the ground truth severity (indicated by the label of the pressure data). This difference can be used to calculate a loss, which may be used to refine the severity model (e.g., using stochastic gradient descent and / or batch gradient descent). The method 700 then ends and returns to block 620 of FIG. 6. Exemplary Method for Evaluating Sensor Data Using a Trained Machine Learning Model

[0114] 8 is a flow diagram depicting an example method 800 for evaluating sensor data using a trained machine learning model. In some embodiments, method 800 is performed by a machine learning system, such as machine learning system 115 of FIG 1. In some embodiments, method 800 is performed using a machine learning model trained using method 600 of FIG 6.

[0115] At block 805, the machine learning system receives pressure data. In an embodiment, as discussed above, this pressure data may generally include pressure exerted by a user's foot or feet at one or more points on the foot (or feet) while the user stands, walks, runs, and the like. In an embodiment, the pressure data received at block 805 corresponds to data from a single user at a single point in time or over a time window (e.g., over a 30 second window). In some embodiments, the machine learning system may further receive historical data of the user, such as their age, weight, height, athletic concerns, and the like, as discussed above.

[0116] At block 810, the machine learning system processes the received data (which may include the user's current pressure data as well as historical data) using one or more trained machine learning models as discussed above. In some embodiments, this includes processing the data using all of the available models in parallel. For example, the machine learning system may use one model to predict whether a fall will occur imminently (e.g., within a defined (short) period of time), another model to predict whether a fall will occur in the future but is not imminent, another model to predict the direction and / or severity of the fall, and the like.

[0117] In some embodiments, as discussed above, the machine learning system may use some models sequentially, using only some models if the output from one or more preceding models meets defined criteria. For example, the machine learning system may only use a fall direction and / or severity model if a fall occurrence model indicates that a future fall is likely to occur.

[0118] At block 815, based on the generated prediction, the machine learning system determines whether one or more defined criteria are met. In some embodiments, the criteria include determining whether a future fall is predicted with a sufficient level of confidence. In at least one embodiment, the criteria include determining whether the predicted fall severity exceeds a defined threshold (even if the fall likelihood is not). That is, the machine learning system may consider whether a fall would be particularly severe even if the fall is not likely to occur. If the criteria are not met, the method 800 returns to block 805 to receive new pressure data for evaluation. If one or more of the criteria are met, the method 800 continues to block 820.

[0119] At block 820, the machine learning system generates and / or initiates one or more corrective and / or preventative actions based on the generated prediction. For example, if the machine learning system determines that a fall is imminent, a caregiver can be alerted and the patient can be instructed to sit or stay still. If the prediction indicates that a fall is likely to occur within the next few hours or days, the machine learning system may prescribe a medical device, such as a cane or walker. If the fall is predicted to be severe (even if unlikely), more proactive interventions may include instructions such as additional physical therapy for the user.

[0120] In at least one embodiment, the particular preventative action selected may be determined, at least in part, based on using the model itself. For example, to determine whether to suggest a cane, walker, improved physical therapy, and the like, the machine learning system may modify one or more of the user's characteristics and use this modified data as input to the model to determine whether the predicted risk or likelihood of falling has decreased. This may enable the machine learning system to not only determine whether an intervention is needed, but also to identify the best intervention most likely to reduce the risk of falling (or fall severity). One example of this prediction process is discussed in more detail below with reference to FIG. 12.

[0121] In at least one embodiment, as discussed above, the machine learning system can also provide facility-wide suggestions to aggregate predicted falls and / or severity and identify potential hazards such as cords in hallways, non-uniform ground surfaces, and the like. These suggestions can be used to improve the facility as a whole. Exemplary Methods for Predicting Fall Risk and Severity Using Machine Learning Models

[0122] 9 is a flow diagram depicting an example method 900 for predicting fall risk and potential severity using a machine learning model. In some embodiments, method 900 is performed by a machine learning system, such as machine learning system 115 of FIG 1. In some embodiments, method 900 provides additional details relative to block 810 of FIG 8.

[0123] At block 905, the machine learning system generates a fall likelihood score by processing the user data (which may include pressure data as well as historical data) using a first machine learning model. In some embodiments, the likelihood score indicates the probability or likelihood that the user will suffer a fall at some point in the future. Depending on the particular implementation, the timeline of this prediction may vary. For example, the first model may be trained to predict whether a fall is likely to occur within the next 5 minutes or 5 days. In some embodiments, rather than a single model, to generate the likelihood score, the machine learning system uses a set of models, each configured to predict the likelihood of a fall over a corresponding future time window.

[0124] In some embodiments, as discussed above, the machine learning system may selectively use these models depending on the output of previous models. For example, the machine learning system may use a model that predicts falls over a 5 minute window only if a model that predicts falls within the next 30 seconds does not indicate that a fall is likely to occur. Similarly, the machine learning system may only use a model that predicts falls over the next 5 hours only if a 5 minute model does not indicate that a fall is likely to occur.

[0125] At block 910, the machine learning system generates a severity score by processing the user data (e.g., pressure data and historical data) using a second machine learning model. In an embodiment, the severity score indicates the likely, probable, or possible severity of a future fall (e.g., in terms of likely types of injuries, the severity or intensity of those injuries, and the like). In some embodiments, the severity is based in part on the predicted fall direction (which may be predicted using a separate model or using a second machine learning model). Similarly, because the models include user data such as age, weight, and height as inputs, they essentially learn to adjust the predicted severity based on these contributing factors.

[0126] In some embodiments, as discussed above, the machine learning system may selectively use the severity model depending on the output of the risk model. For example, if the likelihood model does not indicate that a fall is likely to occur, the machine learning system may not use the severity model to reduce computational costs and latency. Similarly, in one embodiment, if the likelihood model indicates that a fall is imminent (e.g., within 30 seconds), the machine learning system may not use the severity model. In contrast, if the likelihood model predicts that a fall is likely to occur but not imminent, the machine learning system may use the severity model to enable improved selection of available targeted interventions. Furthermore, in some embodiments, the machine learning system may use the severity model regardless of the output of the likelihood model (e.g., to enable interventions for users who are generally stable but may suffer catastrophic injury in the event of a fall). Additionally, in at least one embodiment, the likelihood model and the severity model may be collectively referred to as a fall risk model. Exemplary Methods for Predicting Fall Events and Severity Using Machine Learning

[0127] 10 is a flow diagram depicting an example method 1000 for predicting fall events and severity using machine learning. In some embodiments, method 1000 is performed by a machine learning system, such as machine learning system 115 of FIG. 1. In the illustrated example, method 1000 provides additional details relative to block 810 of FIG. 8. In one embodiment, method 1000 differs from method 900 in that using method 1000, a machine learning system may determine whether to apply different models based on the output of a prior model.

[0128] At block 1005, the machine learning system generates a likelihood score for the user using a first machine learning model. For example, as discussed above, the machine learning system may use one or more likelihood models to process the user's pressure data (e.g., from a 30 second window) and / or the user's characteristics or historical data to determine whether there is a likelihood that the user will fall at some point in the future (e.g., within one or more defined time windows).

[0129] At block 1010, the machine learning system determines whether the predicted probability or likelihood of a fall meets a defined threshold. For example, the machine learning system may determine whether a fall is predicted to occur, an generated confidence in a possible future fall, and the like. If the probability criteria are not met, method 1000 ends and returns to block 815 (the machine learning system will likely determine that the criteria are not met and that intervention should not be initiated).

[0130] At block 1010, if the fall probability criterion is met, method 1000 continues to block 1015, where the machine learning system determines whether one or more temporal criteria are met. In some embodiments, these temporal criteria relate to the urgency of a potential fall (e.g., the urgency or closeness with which it is predicted to occur). For example, if a fall is predicted to occur within some defined "imminent" timeframe (e.g., within seconds or minutes), the machine learning system may determine that the temporal criteria are met and thus bypass the use of the severity model.

[0131] That is, if a fall is predicted to occur imminently, the machine learning system may determine that using a severity model to process the data is not helpful because the fall will likely occur before the predicted severity is relevant. For example, if a user is expected to fall within the next few seconds or minutes, the severity of the potential fall may not generally be relevant to appropriate intervention, i.e., immediate response by one or more caregivers. By not using the severity model, the machine learning system can therefore save significant computational resources that would not otherwise need to be expended.

[0132] At block 1015, if the machine learning system determines that the temporal criteria is not met (e.g., because a fall is predicted to occur over the next few hours or days, rather than immediately), method 1000 continues to block 1020, where the machine learning system uses the second trained machine learning model to generate a severity score. As discussed above, this severity score may generally indicate the likely severity of any future falls, such as whether they may result in fractures (and, if applicable, which bones), lacerations, or other complications. In an embodiment, the potential severity may be an important factor in selecting appropriate interventions and implementing other measures. Exemplary Method for Initiating Preventive and Corrective Actions Based on Machine Learning Models

[0133] 11 is a flow diagram depicting an example method 1100 for initiating preventative and / or corrective actions based on machine learning model output. In some embodiments, method 1100 is performed by a machine learning system, such as machine learning system 115 of FIG 1. In some embodiments, method 1100 provides additional details relative to block 820 of FIG 8.

[0134] Method 1100 generally illustrates a variety of optional interventions that may be selected and implemented by a machine learning system depending on the particular implementation and the particular likelihood and / or severity scores generated.

[0135] At block 1105, the machine learning system may optionally select one or more assistive or preventative devices for the user. For example, the machine learning system may suggest a cane, a walker, a wheelchair, an electric scooter, and the like. This selection may be determined using machine learning (e.g., by processing modified user data with a model as discussed above), using a rule-based system (e.g., based on the user's age, weight, exercise status, and the like), and the like. In some embodiments, the machine learning system may further consider a potential severity score to determine the best preventative device. For example, if the predicted fall severity is below a threshold, the machine learning system may indicate that the user can try a cane to reserve exercise without significant inconvenience. However, if the severity is above a certain threshold, the machine learning system may indicate that an electric scooter should be used because the harm from a potential fall may be significant (even with other devices such as a cane).

[0136] At block 1110, the machine learning system can optionally generate one or more therapy plans for the user. For example, as discussed above, if the machine learning system predicts that the user will fall toward one side, the machine learning system can suggest a therapy or exercise that will improve the user's strength on this side, thereby reducing the chance of falling in that direction. Similarly, if the fall severity is predicted to exceed a certain threshold, additional physical therapy may be useful to help reduce the likely harm from the fall. In an embodiment, as discussed above, a particular therapy may be selected using machine learning (e.g., by processing the modified pressure data or user data using a model), a set of defined rules (e.g., when a fall is more likely to occur toward one side, prescribe strength training on that side), and the like.

[0137] At block 1115, the machine learning system optionally alerts one or more users of a potential upcoming fall. This may include, for example, alerting the user that they may be about to suffer a fall (e.g., via the user's smartphone, via one or more output devices in the vicinity, such as a smart TV or speaker, and the like). For example, the machine learning system may suggest that the user grasp an object for support, sit down, slow down, and the like. In some embodiments, the alert can be provided to one or more caregivers (e.g., a caregiver responsible for the user or a caregiver identified as being physically near the user). The alert may indicate various data depending on the risk and / or severity score generated. For example, the alert may indicate the urgency of the risk (e.g., how quickly the user needs to respond and / or the severity of a possible fall).

[0138] At block 1120, the machine learning system can optionally indicate problem areas within the space. For example, as discussed above, in some embodiments, the machine learning system can aggregate predicted falls across users and / or time and identify physical areas of the facility that are particularly at risk (e.g., using heat maps of user locations at the time of predicted or actual falls). These locations may contain hazards that can or should be mitigated, as discussed above.

[0139] In general, each of the possible interventions in method 1100 is optional, and the actual interventions initiated by the machine learning system may vary according to a wide variety of concerns and implementation details. Exemplary Methods for Evaluating and Selecting Personalized Interventions Using Machine Learning

[0140] Figure 12 is a flow diagram depicting an example method 1200 for evaluating and selecting personalized interventions using machine learning. In some embodiments, method 1200 is performed by a machine learning system, such as machine learning system 115 of Figure 1. In the illustrated example, method 1200 provides additional details relative to block 820 of Figure 8 (after the machine learning system has determined that a fall is likely to occur and has decided to initiate one or more interventions).

[0141] At block 1205, the machine learning system identifies a set of potential interventions for the user. For example, as discussed above with reference to FIG. 11, the machine learning system may select various assistive or preventative medical devices (such as a cane or walker), generate new or revised therapy plans, alert nearby users, and the like. In some embodiments, the potential interventions are selected based at least in part on the current characteristics of the user. For example, if the user is not currently using any assistive devices, the machine learning system may select one or more suggestions. If the user is using a device such as a cane, the machine learning system may suggest a more robust device such as a walker. In some embodiments, the machine learning system may determine whether the user has assistance standing and / or walking and suggest such assistance.

[0142] At block 1210, the machine learning system modifies the user data based on the potential interventions. That is, the machine learning system may artificially adjust the user data to reflect at least one of the proposed interventions. At block 1215, the machine learning system determines a revised fall risk for the user based on the modified user data. As discussed above, the fall risk may generally correspond to the likelihood of a fall and / or the severity of a potential fall.

[0143] For example, if the user is not using a cane, the machine learning system may reprocess the user's data (and, in some aspects, the pressure data) and use a machine learning model to indicate that the user has a cane. This can allow the machine learning system to simulate the effects of various interventions. In at least one embodiment, the machine learning system repeats blocks 1210 and 1215 for each potential intervention determined in block 1205. That is, the machine learning system can separately determine a revised fall risk for each possible intervention (and, in some embodiments, for combinations of interventions).

[0144] At block 1220, the machine learning system determines whether the potential intervention reduced fall risk compared to the user's unmodified data. If not, method 1200 continues to block 1225, where the machine learning system initiates one or more default interventions. For example, if the simulation does not indicate that a walker would reduce fall risk, the machine learning system may still assign, provide, or facilitate the provision of a walker for the user.

[0145] At block 1220, if the machine learning system determines that one or more potential interventions would reduce fall risk, method 1200 continues to block 1230, where the machine learning system initiates one or more of these personalized interventions. Because these interventions are specifically designed for a particular user and validated using machine learning and via one or more simulations, they may be better able to reduce risk to the user and improve overall results. Exemplary Method for Determining Aggregate Fall Risk Using Machine Learning

[0146] 13 is a flow diagram depicting an example method 1300 for determining aggregate fall risk using machine learning. In some embodiments, the method 1300 is performed by a machine learning system, such as the machine learning system 115 of FIG.

[0147] At block 1305, the machine learning system generates a set of fall risk scores for the set of users. As discussed above, the fall risk scores may generally quantify, for each user, the likelihood of a future fall and / or the severity of such a potential fall. In an embodiment, the machine learning system may generate a fall risk score for each user by processing data associated with each user (e.g., pressure data and / or user characteristics) using one or more trained machine learning models, as discussed above.

[0148] In some embodiments, the machine learning system may generate multiple risk scores for each user based in part on the time at which pressure data is collected. That is, the machine learning system may generate risk scores for a single user at multiple time points, with the score at each time point being based on pressure data from the user at or near that time. For example, the machine learning system may generate a first score at a first time (e.g., noon) based on pressure data from one or more windows prior to noon, and a second score at a second time (e.g., 3:00 p.m.) based on pressure data from a window prior to 3:00 p.m.

[0149] At block 1310, the machine learning system aggregates the generated fall risks based on the locations within the facility where the user was when the pressure data used to generate the risk scores was collected. That is, for each generated fall risk score, the machine learning system may identify the physical locations of the user when the pressure data was collected and aggregate the scores based on these locations. For example, the machine learning system may identify a set of risk scores associated with each location for one or more physical spaces or locations within the facility (e.g., pressure data was collected within a defined distance from that location).

[0150] In some embodiments, this aggregation includes determining the number of likely falls in each space (e.g., the number of times the machine learning system predicted a fall with sufficient probability), determining the severity of potential falls in each space (e.g., average or maximum predicted severity, among other risk scores), or the like.

[0151] At block 1315, the machine learning system outputs the aggregated data. For example, the machine learning system may provide a text summary of the aggregated risk (e.g., indicating the predicted number of falls, the predicted severity, identifying users and / or regions with a high number of predicted falls, and the like).

[0152] At block 1320, the machine learning system may optionally display the aggregated data on a facility map. For example, as discussed above with reference to Figure 4, the machine learning system may generate and output a heat map showing the distribution of fall probabilities, fall severity, or a combination of the two. Such an augmented floor plan or map may be particularly useful to help a user quickly identify problem or hazardous areas.

[0153] At block 1325, the machine learning system determines whether any hot spots exist within the facility. That is, the machine learning system may determine whether any of the locations or spaces meet one or more defined criteria related to the expected number and / or frequency of falls, potential fall severity, and the like. If no such hot spots are found, the method 1300 returns to block 1305.

[0154] If at least one hotspot is identified, method 1300 continues to block 1330, where the machine learning system initiates one or more interventions. For example, as discussed above, the machine learning system may indicate the location to one or more users or caregivers and instruct them to investigate the area and identify potential hazards. In some embodiments, if a hotspot is identified within a private or semi-private area (e.g., a user's private room), the machine learning system may identify the user (or users) associated with the space to enable more targeted interventions for these users. Method 1300 then returns to block 1305. Exemplary Methods for Training One or More Machine Learning Models to Predict Fall Risk

[0155] 14 is a flow diagram depicting an example method 1400 for training one or more machine learning models to predict fall risk. In some embodiments, method 1400 is performed by a machine learning system, such as machine learning system 115 of FIG.

[0156] At block 1405, pressure data corresponding to one or both feet of a user is received.

[0157] At block 1410, an indicator is determined, the indicator indicating whether the user has fallen within some defined time window after the pressure data was collected.

[0158] In block 1415, one or more machine learning models are trained to predict fall risk based on the pressure data and the indicators.

[0159] In some embodiments, the pressure data is collected via one or more wearable devices comprising multiple pressure sensors, and the pressure data is collected while a user is walking within a physical space.

[0160] In some embodiments, determining whether the user has fallen includes determining a time when the pressure data was collected, retrieving a set of user records corresponding to a defined time window after the determined time, the set of user records indicating patient data regarding the user, and evaluating the set of user records to determine whether a fall has been reported.

[0161] In some embodiments, the method 1400 further includes receiving pressure data corresponding to a plurality of users and training one or more machine learning models based on the pressure data for the plurality of users.

[0162] In some embodiments, training the one or more machine learning models includes training a first machine learning model to predict the likelihood that a user will fall, and training a second machine learning model to predict the severity of the potential fall.

[0163] In some embodiments, the predicted severity is based, at least in part, on the predicted direction of the potential fall.

[0164] In some embodiments, the method 1400 further includes receiving patient data indicative of characteristics of the user, and training one or more machine learning models based on the characteristics. Exemplary Methods for Predicting Fall Risk Using Machine Learning

[0165] 15 is a flow diagram depicting an example method 1500 for predicting fall risk using machine learning. In some embodiments, the method 1500 is performed by a machine learning system, such as the machine learning system 115 of FIG.

[0166] At block 1505, pressure data corresponding to one or both feet of a user is received.

[0167] At block 1510, a predicted fall risk is generated by processing the pressure data using one or more trained machine learning models.

[0168] At block 1515, in response to determining that the predicted fall risk meets one or more defined criteria, an intervention is selected for the user and application of the intervention is initiated.

[0169] In some embodiments, the pressure data is collected via one or more wearable devices comprising multiple pressure sensors, and the pressure data is collected while a user is walking within a physical space.

[0170] In some embodiments, generating a predicted risk of falling includes using a first machine learning model to generate a likelihood that the user will fall by processing the pressure data.

[0171] In some embodiments, the method 1500 further includes the steps of: not predicting the fall severity in response to determining that the likelihood that the user will fall meets one or more temporal criteria; and alerting one or more nearby caregivers.

[0172] In some embodiments, the method 1500 further includes using a second machine learning model to generate a predicted severity of the potential fall by processing the pressure data.

[0173] In some embodiments, the step of generating the predicted severity is performed in response to determining that the likelihood that the user will fall does not satisfy one or more temporal criteria.

[0174] In some embodiments, the predicted severity is based, at least in part, on the predicted direction of the potential fall.

[0175] In some embodiments, the method 1500 further includes receiving patient data indicative of characteristics of a user, and generating a predicted fall risk by processing the characteristics of the user using one or more trained machine learning models.

[0176] In some embodiments, method 1500 further includes generating a plurality of predicted fall risks for each individual one of a plurality of users in a physical facility by processing the individual pressure data using one or more trained machine learning models; determining locations within the physical facility where the predicted fall risks occurred for at least a subset of the plurality of predicted fall risks; selecting a facility intervention for the physical facility in response to determining that the physical locations satisfy one or more location criteria; and initiating application of the facility intervention. Exemplary Processing System for Improved Machine Learning Models

[0177] 16 depicts an exemplary computing device 1600 configured to implement various aspects of the disclosure. Although depicted as a physical device, in embodiments, the computing device 1600 may be implemented using a virtual device and / or across several devices (e.g., in a cloud environment). In one embodiment, the computing device 1600 corresponds to the machine learning system 115 of FIG.

[0178] As shown, computing device 1600 includes a CPU 1605, memory 1610, storage 1615, network interface 1625, and one or more I / O interfaces 1620. In the illustrated embodiment, CPU 1605 reads and executes programming instructions stored in memory 1610, and stores and reads application data resident in storage 1615. CPU 1605 generally represents a single CPU and / or GPU, multiple CPUs and / or GPUs, a single CPU and / or GPU with multiple processing cores, and the like. Memory 1610 is included as generally representing random access memory. Storage 1615 may be any combination of disk drives, flash-based storage devices, and the like, and may include fixed and / or removable storage devices such as fixed disk drives, removable memory cards, caches, optical storage devices, network attached storage (NAS), or storage area networks (SAN).

[0179] In some embodiments, I / O devices 1635 (keyboard, monitor, etc.) are connected via I / O interface 1620. Additionally, via network interface 1625, computing device 1600 can be communicatively coupled to one or more other devices and components (e.g., via a network, which may include the Internet, a local network, and the like). As shown, CPU 1605, memory 1610, storage device 1615, network interface 1625, and I / O interface 1620 are communicatively coupled by one or more buses 1630.

[0180] In the illustrated embodiment, memory 1610 includes a training component 1650, a prediction component 1655, and an action component 1660, which may implement one or more embodiments discussed above. Although depicted as discrete components for conceptual clarity, in an embodiment, the operations of the depicted components (and others not shown) may be combined or distributed across any number of components. Additionally, while depicted as software resident in memory 1610, in an embodiment, the operations of the depicted components (and others not shown) may be implemented using hardware, software, or a combination of hardware and software. In one embodiment, training component 1650 may be used to train a machine learning model, such as by using method 600 of FIG. 6, prediction component 1655 may be configured to predict falls and / or severity using the model, such as by using method 800 of FIG. 8, and action component 1660 may be configured to generate and / or initiate various responsive actions to risk, such as by using method 1100 of FIG. 11.

[0181] In the illustrated example, storage 1615 includes historical data 1670 (which may correspond to labeled pressure data and / or user data used to train and / or evaluate the model), as well as one or more machine learning models 1675. Although depicted as residing in storage 1615, historical data 1670 and machine learning models 1675 may be stored in any suitable location, including memory 1610. Generally, historical data 1670 includes previously received (and labeled) pressure data, as well as any associated user data (e.g., user age and weight data) used to train the machine learning models 1675. Example notes

[0182] Appendix 1: A method for training a machine learning model, the method comprising: receiving pressure data corresponding to one or both feet of a user; determining an indicator indicating whether the user has fallen within a defined time window after the pressure data is collected; and training one or more machine learning models to predict risk of falling based on the pressure data and the indicators.

[0183] Appendix 2: The method of Appendix 1, wherein the pressure data is collected via one or more wearable devices comprising a plurality of pressure sensors, and the pressure data is collected while the user is walking within a physical space.

[0184] Appendix 3: The method of any one of Appendixes 1-2, wherein the step of determining whether the user has fallen includes the steps of determining a time when the pressure data was collected, retrieving a set of user records corresponding to a defined time window after the determined time, the set of user records indicating patient data relating to the user, and evaluating the set of user records to determine whether a fall has been reported.

[0185] Addendum 4: The method of any one of Addendums 1-3, further comprising receiving pressure data corresponding to a plurality of users, and training one or more machine learning models based on the pressure data relating to the plurality of users.

[0186] Appendix 5: The method of any one of Appendixes 1-4, wherein the step of training one or more machine learning models includes training a first machine learning model to predict the likelihood that a user will fall, and training a second machine learning model to predict the severity of the potential fall.

[0187] Appendix 6: The method of any one of Appendixes 1-5, wherein the predicted severity is based, at least in part, on a predicted direction of the potential fall.

[0188] Addendum 7: The method of any one of Addendums 1-6, further comprising receiving patient data indicative of characteristics of a user, and training one or more machine learning models based on the characteristics.

[0189] Appendix 8: A method for predicting an event using a machine learning model, comprising: receiving pressure data corresponding to one or both feet of a user; generating a predicted risk of falling by processing the pressure data using one or more trained machine learning models; and, in response to determining that the predicted risk of falling satisfies one or more defined criteria, selecting an intervention for the user; and initiating application of the intervention.

[0190] Appendix 9: The method of Appendix 8, wherein the pressure data is collected via one or more wearable devices comprising a plurality of pressure sensors, and the pressure data is collected while the user is walking within a physical space.

[0191] Addendum 10: The method of any one of Addendums 8-9, wherein the step of generating a predicted risk of falling includes a step of generating a likelihood that the user will fall by processing the pressure data using a first machine learning model.

[0192] Appendix 11: The method of any one of appendices 8-10, further comprising the steps of: not predicting fall severity in response to determining that the likelihood that the user will fall meets one or more temporal criteria; and alerting one or more nearby caregivers.

[0193] Appendix 12: The method of any one of appendices 8-11, further comprising a step of generating a predicted severity of a potential fall by processing the pressure data using a second machine learning model.

[0194] Appendix 13: The method of any one of appendices 8-12, wherein the step of generating the predicted severity is performed in response to determining that the likelihood that the user will fall does not satisfy one or more temporal criteria.

[0195] Appendix 14: The method of any one of Appendixes 8-13, wherein the predicted severity is based, at least in part, on a predicted direction of the potential fall.

[0196] Addendum 15: The method of any one of Addendums 8-14, further comprising the steps of receiving patient data indicative of characteristics of a user, and generating a predicted fall risk by processing the characteristics of the user using one or more trained machine learning models.

[0197] Appendix 16: The method of any one of appendices 8-15, further comprising: generating a plurality of predicted fall risks for each individual one of a plurality of users in a physical facility by processing the individual pressure data using one or more trained machine learning models; determining locations within the physical facility where the predicted fall risks occurred for at least a subset of the plurality of predicted fall risks; selecting a facility intervention for the physical facility in response to determining that the physical locations satisfy one or more location criteria; and initiating application of the facility intervention. Additional Considerations

[0198] The foregoing description is provided to enable any person skilled in the art to practice the various embodiments described herein. The examples discussed herein are not intended to limit the scope, applicability, or embodiments described in the claims. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments. For example, changes may be made to the functions and arrangements of the elements discussed without departing from the scope of the disclosure. Various examples may omit, substitute, or add various procedures or components, as appropriate. For example, the methods described may be performed in different orders than those described, and various steps may be added, omitted, or combined. Also, features described with respect to some examples may be combined in some other examples. For example, an apparatus may be implemented or a method may be practiced using any number of aspects described herein. Additionally, the scope of the disclosure is intended to cover such apparatus or methods practiced using other structures, functionality, or structures and functionality in addition to or other than the various aspects of the disclosure described herein. It should be understood that any aspect of the disclosure disclosed herein may be embodied by one or more elements of a claim.

[0199] As used herein, the word "exemplary" means "serving as an example, instance, or illustration." Any aspect described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects.

[0200] As used herein, phrases referring to "at least one of" a list of items refer to any combination of the items, including single members. By way of example, "at least one of a, b, or c" is intended to cover a, b, c, ab, ac, bc, and abc, as well as any combination with multiple identical elements (e.g., aa, aaa, aab, aac, abb, acc, bb, bbb, bbc, cc, and ccc, or any other permutation of a, b, and c).

[0201] As used herein, the term "determining" encompasses a wide variety of actions. For example, "determining" may include calculating, computing, processing, deriving, investigating, looking up (e.g., looking up a table, database, or another data structure), ascertaining, and the like. "Determining" may also include receiving (e.g., receiving information), accessing (e.g., accessing data in a memory), and the like. "Determining" may also include resolving, selecting, choosing, establishing, and the like.

[0202] The methods disclosed herein include one or more steps or actions for achieving the method. The method steps and / or actions may be interchanged without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of specific steps and / or actions may be modified without departing from the scope of the claims. Furthermore, various operations of the methods described above may be performed by any suitable means capable of performing the corresponding functions. The means may include various hardware and / or software components and / or modules, including, but not limited to, circuits, application specific integrated circuits (ASICs), or processors. Generally, where there are operations illustrated in the figures, those operations may have corresponding counterparts, means-plus-function components, with similar numbering.

[0203] An embodiment of the present invention may be provided to end users through a cloud computing infrastructure. Cloud computing generally refers to the provisioning of scalable computing resources as a service over a network. More formally, cloud computing may be defined as a computing capability that provides an abstraction between computing resources and their underlying technical architecture (e.g., servers, storage, networks), enabling opportunistic, on-demand network access to a shared pool of configurable computing resources that can be rapidly provisioned and released with minimal administrative effort or service provider interaction. Thus, cloud computing allows users to access virtual computing resources (e.g., storage, data, applications, and even fully virtualized computing systems) in the "cloud," regardless of the underlying physical systems (or the location of those systems) used to provide the computing resources.

[0204] Typically, cloud computing resources are provided to users on a pay-per-use basis, and users are charged only for the computing resources they actually use (e.g., the amount of storage space consumed by the user or the number of virtualized systems instantiated by the user). Users can access any of the resources residing in the cloud at any time and from any location across the Internet. In the context of the present invention, users may access applications or systems (e.g., machine learning system 115) or related data available in the cloud. For example, machine learning system 115 may run on a computing system in the cloud and train and / or use a machine learning model. In such a case, machine learning system 115 may train a model to predict a user fall and store the model in a storage location in the cloud. Doing so allows users to access this information from any computing system attached to a network connected to the cloud (e.g., the Internet).

[0205] The following claims are not limited to the embodiments set forth herein, but are intended to be accorded the full scope consistent with the language set forth in the claims. Within the claims, reference to an element in the singular is intended not to mean "one and only one" unless specifically so stated, but rather "one or more." The term "several" refers to one or more, unless specifically stated otherwise. No claim element is to be construed under the provisions of 35 U.S.C. § 112(f) unless the element is expressly recited using the phrase "means for," or, in the case of a method claim, the element is recited using the phrase "step for." All structural and functional equivalents of the elements of the various aspects described throughout this disclosure that are known or later become known to those of ordinary skill in the art are expressly incorporated herein by reference and are intended to be encompassed by the claims. Furthermore, nothing disclosed herein is intended to be made available to the public, regardless of whether such disclosure is expressly recited in the claims.

Claims

1. 1. A method for training a machine learning model, comprising: Receiving pressure data corresponding to one or both feet of a user; determining an indication of whether the user has fallen within a defined time window after the pressure data has been collected; and training one or more machine learning models to predict risk of falling based on the pressure data and the indicators; A method comprising:

2. the pressure data is collected via one or more wearable devices comprising a plurality of pressure sensors; The pressure data is collected while the user is walking within a physical space. The method of claim 1.

3. Determining whether the user has fallen includes: determining a time when the pressure data was collected; and retrieving a set of user records corresponding to a defined time window after the determined time, the set of user records indicating patient data relating to the user; evaluating the set of user records to determine whether a fall has been reported; and The method of claim 1 , comprising:

4. receiving pressure data corresponding to a plurality of users; training the one or more machine learning models based on pressure data for the plurality of users; The method of claim 1 further comprising:

5. Training the one or more machine learning models includes: Training a first machine learning model to predict a likelihood that a user will fall; Training a second machine learning model to predict potential fall severity; The method of claim 1 , comprising:

6. The method of claim 5 , wherein the predicted severity is based, at least in part, on a predicted direction of the potential fall.

7. receiving patient data indicative of a characteristic of the user; training the one or more machine learning models based on the characteristics; The method of claim 1 further comprising:

8. 1. A method of predicting an event using a machine learning model, comprising: Receiving pressure data corresponding to one or both feet of a user; generating a predicted risk of falling by processing the pressure data using one or more trained machine learning models; In response to determining that the predicted risk of falling satisfies one or more defined criteria, selecting an intervention for the user; Initiating application of said intervention; A method comprising:

9. the pressure data is collected via one or more wearable devices comprising a plurality of pressure sensors; The pressure data is collected while the user is walking within a physical space. The method according to claim 8.

10. Generating the predicted risk of falling includes: The method of claim 8 , comprising generating a likelihood that the user will fall by processing the pressure data using a first machine learning model.

11. In response to determining that a likelihood that the user will fall satisfies one or more temporal criteria, Avoid predicting fall severity and Alerting one or more nearby caregivers; The method of claim 10 further comprising:

12. 11. The method of claim 10, further comprising generating a predicted severity of a potential fall by processing the pressure data using a second machine learning model.

13. 13. The method of claim 12, wherein generating the predicted severity is performed in response to determining that a likelihood that the user will fall does not meet one or more temporal criteria.

14. The method of claim 12 , wherein the predicted severity is based, at least in part, on a predicted direction of the potential fall.

15. receiving patient data indicative of a characteristic of the user; generating the predicted fall risk by processing characteristics of the user using one or more trained machine learning models; The method of claim 8 further comprising:

16. generating a plurality of predicted fall risks for each individual user of a plurality of users in a physical facility by processing the individual pressure data using the one or more trained machine learning models; determining, for at least a subset of the plurality of predicted fall risks, locations within the physical facility where the predicted fall risks occurred; In response to determining that the physical location satisfies one or more location criteria, selecting a facility intervention for the physical facility; Initiating the application of said institutional intervention; The method of claim 8 further comprising:

17. A non-transitory computer readable storage medium comprising computer readable program code that, when executed using one or more computer processors, performs: Receiving pressure data corresponding to one or both feet of a user; generating a predicted risk of falling by processing the pressure data using one or more trained machine learning models; In response to determining that the predicted risk of falling satisfies one or more defined criteria, selecting an intervention for the user; Initiating application of said intervention; performing an operation to predict an event using a machine learning model comprising: A non-transitory computer readable storage medium.

18. Generating the predicted risk of falling includes:

20. The non-transitory computer readable storage medium of claim 17, comprising generating a likelihood that the user will fall by processing the pressure data using a first machine learning model.

19. In response to determining that a likelihood that the user will fall satisfies one or more temporal criteria, Avoid predicting fall severity and Alerting one or more nearby caregivers; 20. The non-transitory computer readable storage medium of claim 18, further comprising:

20. 20. The non-transitory computer readable storage medium of claim 18, further comprising generating a predicted severity of a potential fall by processing the pressure data using a second machine learning model.

Images