Patient location detection method, computing device, and program

Abstract

A method and corresponding system for detecting a semantic patient location is provided, the method comprising the steps of receiving non-geospatial sensor data from at least one sensor of a wearable monitoring device worn by a patient, obtaining patient activity level data or posture data from the non-geospatial sensor data, extracting one or more activity behavior features from the patient activity level data or one or more posture behavior features from the posture data, classifying the activity behavior features or posture behavior features as belonging to one of at least two predefined patient contexts, and outputting an indication of the detected semantic patient location. Also provided is a method and corresponding system for training a machine learning model, the method comprising training a machine learning model to classify the non-geospatial sensor data, optionally preprocessed, received from at least one sensor of a wearable monitoring device worn by a patient as belonging to one of at least two predefined semantic patient locations.

Description

Technical Field

[0001] The present invention relates to the detection of patient context, and particularly to the detection of the semantic position of a patient.

Background Art

[0002] A wearable remote monitoring system continuously monitors a patient's vital signs and simultaneously shares data wirelessly with remote caregivers. Such systems are becoming increasingly popular for improving patient care, reducing hospital costs, and improving patient outcomes because they are unobtrusive and require minimal or no interaction with the patient. By continuously monitoring vital signs such as heart rate and respiratory rate, an early warning score (EWS) system can be used to detect clinically relevant deteriorations in health status. The EWS system uses an algorithm that applies thresholds to the monitored vital signs to create an early warning score. Based on this score, an alert is sent to the caregiver. However, the alert requirements, including the thresholds applied when the patient is in the hospital, may be different from those applied when the patient is at home after discharge. Known patient monitoring systems detect a patient's location using a GPS system or other geolocation method. However, due to power consumption and privacy limitations, the applicability of these solutions to wearable remote monitoring systems is reduced.

Summary of the Invention

Problems to be Solved by the Invention

[0003] Therefore, there is a need for a cost-effective and privacy-conscious system and method for determining the patient context, particularly whether the patient being monitored is in the hospital or at home, and thereby adjusting the remote monitoring service according to the context. This need is met by the subject matter of the independent claims. Optional features are defined by the dependent claims.

Means for Solving the Problems

[0004] According to the first aspect, a method for detecting patient context is provided. This method is The steps include receiving non-geospatial sensor data from at least one sensor of a wearable monitoring device worn by the patient, A step of obtaining patient activity level data or posture data from non-geospatial sensor data, A step of extracting one or more activity behavioral features from patient activity level data, or one or more postural behavioral features from postural data, A step of classifying activity behavioral characteristics or postural behavioral characteristics as belonging to one of at least two predefined patient contexts, The process includes the step of outputting instructions for the detected patient context.

[0005] Therefore, the present invention enables the detection of patient context without requiring the patient to input information about their location and without the use of geolocation services, thus protecting patient privacy and reducing power consumption.

[0006] As used herein, the term “context” is generally understood to mean situational awareness, but more specifically, it refers to the patient’s location—that is, a semantic location such as “being in a hospital” or “being at home,” rather than a physical or geographical location. “Hospital” should not be understood in a restrictive sense, but rather should be understood to include any medical facility that provides inpatient care.

[0007] "Non-geospatial sensor data" means sensor data obtained from at least one sensor of a wearable monitoring device, other than data obtained from GPS sensors or other geolocation sensors that may be present in the wearable monitoring device. No remote device is connected to determine the subject's location. In one example, non-geospatial sensor data includes accelerometer data or motion data obtained by one or more accelerometers or other motion sensors in the wearable monitoring device. In another example, the patient is monitored using one or more posture sensors, barometric pressure sensors, gyroscopes, or magnetometers or equivalent actigraphy sensors. In addition, other vital signs (such as respiratory rate, blood pressure, PPG waveform, and body temperature) can be used to detect the patient context. The received non-geospatial sensor data may include data for a given period, such as 24 hours. The non-geospatial sensor data may also include time-series data.

[0008] Wearable monitoring devices (also called wearable devices or simply wearables) include any device capable of monitoring a patient's vital signs, such as a chest patch, wristwatch, ear device, or chest strap.

[0009] Processing involves preprocessing raw non-geospatial sensor data in preparation for classification. Preprocessing involves obtaining patient activity level data and / or posture data from the non-geospatial sensor data. Activity level data is based on, for example, an ordered list of values ​​(e.g., 0, 1, 2, 3, 4, ...). Posture data is based on, for example, an unordered list of classes (e.g., lying down, sitting, standing, upright, ...). While motor data can also be constructed by combining activity level data and posture data, it will be understood that separate forms of activity level data or posture data can be used for classification of the patient context. Preprocessing involves extracting one or more activity behavior features and / or postural behavior features from the patient activity level data and / or posture data. Preprocessing may involve low-pass filtering of the patient activity level data and / or posture data before feature extraction. Furthermore, preprocessing may include one or more of the following techniques: data interpolation to address missing values ​​in time series, data fusion to integrate multiple data sources, data cleaning to detect and correct corrupted data, such as data augmentation to prevent overfitting, data normalization, and data transformation.

[0010] Feature extraction may involve extracting one or more activity-behavioral or postural-behavioral features that are suitable for distinguishing patient contexts. For example, activity-behavioral or postural-behavioral features may include features that describe the variance (variability, dispersion, or spread) in patient activity level data or postural data. In particular, activity-behavioral features may include one of the following in patient activity level data: the standard deviation of the activity level signal, the range of the activity level signal, the area around its mean, the duration for which the activity level signal is below the low activity threshold, the sum of the area of ​​the activity level signal below the low activity threshold and the area of ​​the activity level signal above the high activity threshold, or the sum of the duration of the activity level signal above the high level threshold and the duration of the activity level signal below the low level threshold. Postural-behavioral features may include one or more of the following in postural data: the mean of the postural signal, the median of the postural signal, the standard deviation of the postural signal, the range of the postural signal, and the area of ​​the postural signal. Furthermore, composite postural and activity-behavioral features such as the mean, median, and variability of the activity level data for each specific patient posture may be extracted. Further high-level composite features include, for example, data on circadian rhythms (e.g., amplitude, peak phase, period, intermediate phase).

[0011] In one favorable example, where the first predefined patient context includes the semantic patient location "in the hospital" and the second predefined patient context includes the semantic patient location "at home," the activity behavioral features include an area of ​​activity levels above the high activity threshold and below the low activity threshold. In this way, based on the insightful recognition that hospitalized patients are rarely inactive or as highly active as they are at home, patients are readily located as if they were in the hospital or at home.

[0012] Processing non-geospatial sensor data and classifying it as belonging to one of at least two predefined patient contexts involves performing the classification using any computational algorithm or a trained machine learning model.

[0013] For example, processing non-geospatial sensor data and classifying it as belonging to one of at least two predefined patient contexts involves using a comparison-based algorithm to identify similarities between an input dataset (i.e., non-geospatial sensor data in raw or preprocessed form, e.g., activity level data or activity behavioral features extracted therefrom, or posture data or posture behavioral features extracted therefrom) and at least one reference dataset representing expected (raw or preprocessed) data for a particular patient context. For example, one reference dataset is provided for the semantic patient location "in the hospital," and another reference dataset is provided for the semantic patient location "at home." Thus, the comparison-based algorithm identifies correspondences between measured behavioral patterns in the input dataset and expected behavioral patterns in the reference dataset for one or more predefined patient contexts (e.g., "in the hospital," "at home"). Thus, the comparison-based algorithm can distinguish temporal patterns in the dataset that characterize the input dataset as belonging to one of the predefined patient contexts. The comparison-based algorithm is, for example, a matching algorithm such as a template matching algorithm. In one example, a template matching algorithm multiplies patient activity level data or posture data by a template of weights representing expected patient activity level patterns or posture patterns for a particular patient context. Alternatively, the template matching algorithm can cross-correlated the input dataset with a reference dataset, which could be obtained, for example, from a data repository representing a typical patient population.

[0014] Furthermore, a machine learning model can be trained to classify input datasets (i.e., non-geospatial sensor data in raw or preprocessed form) as belonging to one of a predefined patient context. The training data includes, for example, one or more reference datasets, obtained from a repository or specifically created for training the model, representing expected raw or preprocessed sensor data for patients in each patient context. In one example, a logistic regression classifier is trained to classify one or more extracted activity or postural behavioral features as belonging to one of two predefined contexts. In another example, classification can be performed using any other machine learning technique for processing features, such as support vector machines, random forests, or Bayesian classifiers. Additionally, a neural network, such as a recurrent neural network, can be used to process raw activity level and / or postural data to output the probability that the data was generated in a particular predefined context.

[0015] Depending on how the classifier is implemented, the step of outputting an indication of the detected patient context includes outputting the probability that the current patient context belongs to a particular predefined patient context, and / or a decision about the predefined patient context to which the current patient context belongs. "Current patient context" here refers to the patient context captured by the non-geospatial sensor data, i.e., the context in which the non-geospatial sensor data was generated.

[0016] For example, if the classifier is a probabilistic classifier, outputting an instruction involves outputting a probability distribution across a set of available predefined contexts. Furthermore, or otherwise, outputting an instruction involves running a decision algorithm to select one of the predefined contexts (e.g., the one with the highest probability or the "best" predefined context) as the detected patient context. Thus, the instruction may include the selected predefined context and / or one or more probabilities associated with each predefined context.

[0017] If the classifier is a deterministic classifier, outputting an indication involves outputting a predefined context that is determined to be the "best" predefined context as an indication.

[0018] According to a second aspect, a method is provided for training the machine learning model of the first aspect to classify an input dataset (i.e., non-geospatial sensor data in raw or preprocessed form) as belonging to one of a predefined patient context.

[0019] According to a third aspect, a computing device is provided which includes a processor that performs the method of the first or second aspect. The methods of the first and second aspects are, of course, performed by a computer.

[0020] According to a fourth aspect, a computer-readable medium containing instructions is provided. When executed by a computing device, the instructions enable the computing device to perform the method of the first or second aspect, or cause the computing device to perform the method of the first or second aspect.

[0021] The present invention may include one or more aspects, embodiments, or features, either separately or in combination (whether or not the combination or separation being specifically disclosed). Any optional features or sub - aspects of one of the above - mentioned aspects are applied as necessary to any of the other aspects.

[0022] These and other aspects of the present invention will become apparent from the embodiments described below and will be described with reference to those embodiments.

Brief Description of the Drawings

[0023] A detailed description is given below by way of example only, with reference to the accompanying drawings.

[0024] [Figure 1] FIG. 1 shows a method for detecting patient context. [Figure 2A] FIG. 2A shows data of a first exemplary activity level. [Figure 2B] FIG. 2B shows data of a second exemplary activity level. [Figure 3A] FIG. 3A shows a receiver operating characteristic curve showing the relative prediction values of various behavioral characteristics when distinguishing patient context when using a computational classification algorithm. [Figure 3B] FIG. 3B shows a receiver operating characteristic curve showing the prediction values of the behavioral characteristics of FIG. 3A when classification is performed using a trained machine learning model. [Figure 4] FIG. 4 relates to an exemplary dataset showing differences in sensor data and resulting behavioral characteristics for two different patient contexts. [Figure 5] FIG. 5 relates to an exemplary dataset showing differences in sensor data and resulting behavioral characteristics for two different patient contexts. [Figure 6] FIG. 6 relates to an exemplary dataset showing differences in sensor data and resulting behavioral characteristics for two different patient contexts. [Figure 7]Figure 7 shows an exemplary dataset illustrating the differences in sensor data and behavioral characteristics derived from it for two different patient contexts. [Figure 8] Figure 8 shows a computing device that can be used according to the systems and methods disclosed herein. [Modes for carrying out the invention]

[0025] This specification discloses a method and system for classifying patient contexts (such as hospital or home) using data acquired from wearable devices. Body movement is detected using non-geospatial sensors, such as accelerometers and other motion sensors embedded in the wearable device. Body movement measurements are performed continuously day and night to determine whether the patient's activity behavior represents a typical hospital pattern or a home pattern. The present invention is based on the insight that activity level characteristics in hospitals differ significantly from those observed at home. In particular, the present invention is based on the surprising realization that hospitalized patients are less physically active than when at home, and at the same time, low activity levels are rarely maintained. Furthermore, circadian rhythm patterns are less pronounced in hospitals than at home. By observing activity levels over time, it is possible to distinguish between time patterns that typically occur in hospitals and time patterns that occur at home.

[0026] Figure 1 illustrates the method for detecting patient context according to this disclosure. Broadly speaking, the method includes, in step 101, receiving non-geospatial sensor data from at least one sensor of a wearable monitoring device worn by the patient; in step 104, detecting the patient context by processing the non-geospatial sensor data and classifying it into one of at least two predefined patient contexts; and in step 106, outputting an indication of the detected patient context. In this way, the method and system disclosed herein advantageously provide automated detection of whether a patient is in a hospital or at home without requiring patient input or a geolocation positioning system. These steps of the method will be described in more detail, along with various optional features. The method will be performed by any suitable computing device as described below.

[0027] In the method described with respect to Figure 1, the data received in step 101 includes waveform 101a or time series received from a sensor of a wearable device worn by the patient for the purpose of monitoring physical activity. The use of a wearable sensor provides inconspicuousness and convenience for long-term use. In other embodiments, in addition to or instead of acceleration data, posture, barometric pressure, gyroscope, magnetometer, or equivalent actigraphy sensor data can be used. An example of a wearable device suitable for use with the methods and systems disclosed herein is the Healthdot device offered by Philips.

[0028] The data received in step 101 is preprocessed in steps 102 and 103 to prepare it for classification.

[0029] In step 102, the method includes determining the patient's activity level, for example, by calculating an activity level metric from the acceleration signal waveform 101a. For example, the activity level can be expressed as the standard deviation or interquartile range of the acceleration signal at intervals of a predefined length (e.g., 5 seconds). Thus, patient activity level data 102a is obtained from the received acceleration waveform 101a. The activity level data 102a, which includes a series of activity level metrics, is also called an activity level sequence.

[0030] Figure 2A shows the first exemplary activity level data recorded for a patient who was hospitalized for two days and then discharged home. Trace 202 shows the activity level data, and trace 204 shows the patient context using values ​​1 and 2, which represent the semantic positions "in hospital" and "at home," respectively. As can be seen, the activity level pattern observed for the patient context "at home" is significantly different from that measured for "in hospital." In particular, the activity level pattern for "at home" is characterized by increased variance, including both high and low values, compared to "in hospital."

[0031] Figure 2B shows second exemplary activity level data recorded for a patient who was discharged after a three-day hospital stay. The activity level pattern follows a circadian rhythm, such as alternating periods of wakefulness and sleep, as reflected by activity levels that are far more pronounced in the patient context "at home" compared to "in the hospital."

[0032] Referring again to Figure 1, the method includes obtaining an activity level profile 103a using the activity level data 102a in step 103. For example, feature extraction is performed to extract one or more activity behavior features from the patient activity level data 102a. The activity level data 102a is low-pass filtered over a long time window (e.g., 24 hours) (e.g., using a moving average filter with a 30-minute window) to capture the patient's daily activity level trends. The activity levels are then processed to extract features that characterize or distinguish the patient's activity behavior. The extracted features are combined to form the patient's activity level profile 103a. These features include, for example, the standard deviation of activity levels over 24 hours, the range, the area of ​​activity levels around the mean activity level, the duration of activity levels below the minimum threshold, and the area of ​​activity levels below the low activity threshold and above the high activity threshold. This last feature is particularly useful because it is rare for hospitalized patients to be completely inactive or extremely active (which usually occurs at home instead).

[0033] Figure 3A shows a receiver operating characteristic curve 300 that shows the predicted values ​​or areas under the receiver operating characteristic curve for various behavioral characteristics when distinguishing the patient context "in the hospital" from "at home" using a single feature classifier based on a threshold rule that, for example, if feature > threshold, the patient is "at home," and otherwise, the patient is "in the hospital." The activity behavioral characteristics shown are as follows: standard deviation of activity level 302 (Act_LP_std), range of activity level 304 (Act_LP_range), area around its mean value of activity level 306 (Act_LP_area), duration of activity level signal below low activity (rest) threshold 308 (Act_LP_re_duration), duration of activity level signal below rest threshold 310 plus duration of activity level above high activity threshold (Act_LP_reha_duration), and area of ​​activity level below rest threshold 312 plus area above high activity threshold (Act_LP_reha_strength). Postural behavioral features 314-322 are described below. Note that some or all of these behavioral features include features that describe variance. Here, "LP" refers to features obtained from low-pass filtered data.

[0034] Returning to Figure 1, the method, in step 104, executes a computational algorithm (or computational model) to analyze the input dataset and classify whether it belongs to the patient context "in the hospital" or "at home."

[0035] In one implementation, the computational algorithm takes activity behavior features defined by the activity level profile 103a as input, processes them, and determines whether the patient is "in the hospital" or "at home." For example, the activity behavior features are processed using the computational algorithm to determine their correspondence with reference datasets 104b for home and hospital stays. Here, reference dataset 104b contains the expected activity level profiles for "in the hospital" and / or "at home," for example, in the form of one or more thresholds associated with each activity behavior feature. Next, a single feature classifier is used to perform binary classification of the input dataset by comparing features to thresholds based on a threshold rule, for example, feature > threshold (if so, "at home," otherwise "in the hospital"). This represents one form of comparison algorithm. Alternatively, multiple feature classifiers can be used to apply a combination of such rules.

[0036] In addition to or instead of taking activity level profiles 103a as input data, the computation algorithm may process activity level data 102a. In another implementation, the computation algorithm includes a template matching algorithm that multiplies the input activity level data 102a by a template of weights (w[i]) that describe the expected activity level while staying at home (e.g., captured before hospitalization or obtained from a data repository of a representative patient population). For example, activity level data 102a (ACL[i] consisting of Ni sample values) and an activity level template (W[i] of Ni sample values) for the patient context "at home" HOME [i]) Matching value M HOME The (or matching level) can be calculated according to the following formula: M HOME =(Σi:0···Ni(W HOME [i]×ACL[i])) / Ni

[0037] In this case, reference dataset 104b contains activity level templates for "at home" and / or "in hospital". HOMEThe decision of, and / or the corresponding M HOSPITAL The decision is taken as the probability 10⁴a for hospital / home, as shown in Figure 1.

[0038] In another implementation, a computational algorithm is executed to classify the input dataset as belonging to the patient context "in the hospital" or "at home" by measuring the correlation or cross-correlation between measured activity level data and expected activity level data for "in the hospital" and / or "at home," or by multiplying the measured activity level data by the expected activity level data. A larger sum of the multiplications indicates a stronger correlation or a greater cross-correlation, which indicates a higher probability that the measured activity level is typical for each patient context.

[0039] The use of computational algorithms is optional, and it should be understood that machine learning methods may be applied instead or additionally to classify the input dataset as belonging to either "at home" or "in the hospital."

[0040] In one example, the input dataset contains activity behavioral features from activity level profile 103a. In this example, classification is performed using a logistic regression classifier. Figure 3B shows the classification accuracy in distinguishing between input data for "being in the hospital" and input data for "being at home" when using a logistic regression classifier acting on the activity behavioral features shown in Figure 3A. The classifier is trained based on expected activity behavioral or postural behavioral features, such as those that form part of the reference dataset 104b.

[0041] In another example, the input data includes activity level data 102a. In this example, the activity level data 102a is processed using a deep learning algorithm such as a recurrent neural network (long- and short-term memory layers, gated recurrent units, etc.). This type of neural network layer is particularly well-suited to discovering inherent temporal patterns within the input dataset to represent output values ​​such as the probabilities of being "at home" and "in the hospital." The neural network is trained on expected activity level or posture data for at least one of a predefined context.

[0042] Referring again to Figure 1, in step 105, post-processing of one or more calculated probabilities is optionally performed. For example, post-processing may include adjusting probabilities to account for temporal trends. For example, post-processing may include further modification of probabilities if an increasing or decreasing trend in probabilities (such as daily changes) is measured. The rationale for this is that an increasing trend is less observable than an irregularly high probability and may indicate that the patient is transitioning to a different context. The output of post-processing will include probability values ​​but will be adjusted by some correction factor that depends on the temporal trend, such as daily changes or the absolute value of the probability compared to a specific threshold.

[0043] Finally, the binary decision between the patient context "in the hospital" and "at home" is made in step 106 by processing probability values ​​and outputting decision 106a. For example, the probability is compared to a predefined threshold indicating the certainty that the patient context is either "in the hospital" or "at home". The decision threshold can also be modified based on past classification outputs. For example, if the patient context has previously been classified as "at home", the next decision is more likely to also be "at home", and the home decision threshold for the following day can be lowered accordingly.

[0044] In the above description, activity level data 102a, and therefore activity level profile 103a, is obtained using acceleration waveforms 101a acquired using the accelerometer of a wearable device. However, it will be understood that the same or similar features can be extracted from posture data acquired using posture sensors, for example, the same motion sensor or accelerometer used to acquire the acceleration waveforms 101a. For example, Figure 3A shows posture behavior features in posture data, including the mean value 314 of the posture signal, the median value 316 of the posture signal, the standard deviation 318 of the posture signal, the range 320 of the posture signal, and the area 322 of the posture signal.

[0045] Next, referring to Figures 4-7, the following disclosure relates to an exemplary dataset illustrating the differences in patterns between two patient contexts: "at home" and "in the hospital."

[0046] Figures 4A and 4B show exemplary activity level data 102a for patient contexts of "at home" and "in the hospital," respectively. "BARI" represents mean accelerometer data obtained from patients after bariatric surgery, and "ONCO" represents mean accelerometer data obtained from cancer patients. "BARI_std" and "ONCO_std" represent activity level data in the form of the standard deviation of the respective accelerometer signals over predefined intervals, as described above. As can be seen, the activity patterns for "at home" for both patients show a larger variance over 24 hours than the activity patterns for "in the hospital."

[0047] Figures 5A and 5B show exemplary posture data for the same two patient groups. Here, "BARI_iqr" and "ONCO_iqr" represent the interquartile range in their respective posture type distributions, which were used to generate the posture data. It can be seen that both patients spend most of their day in a "leaning backward" posture when in the hospital, and their interquartile range is much higher when at home.

[0048] The datasets shown in Figures 4A, 4B, 5A, and 5B are used, for example, as expected activity level data or expected posture data that form part of the reference dataset 104b, for performing calculations as described above and / or training machine learning models.

[0049] Figure 6 shows numerous examples of expected activity behavioral features and expected postural behavioral features that form the basis of the above classification. These features form part of the reference dataset 104b and are used, for example, to obtain the thresholds for the classification rules and / or to train machine learning models. In addition to the above features, features include "ACT_acor_peakiness" (the difference between the peak amplitude and mean amplitude of the autocorrelation function), "ACT_sin_rmse" (the root mean square error of the sine function fitting that fits the activity level data, along with the corresponding "ACT_cos_rmse"), and "sin_amp" and "cos_amp" (amplitude coefficients of the sine or cosine fitting function applied to the activity level data or postural data). In all cases, a clear distinction is observed between the patient contexts of "at home" and "in the hospital".

[0050] Figure 7 shows the classification accuracy of the logistic regression model using the features shown in Figure 6. In the following table, "test" and "training" represent separate datasets used for validation and development of the logistic regression classifier (LogReg), respectively. [Table 1] The term CV refers to the classification results obtained on the training dataset after using a patient-by-patient one-miss cross-validation approach. The various parameters listed in the table are, in this case, classic performance metrics of a binary classifier aimed at distinguishing between home and hospital contexts using features from activity level data or posture data.

[0051] Next, with reference to Figure 8, a high-level schematic diagram of an exemplary computing device 800 that can be used in accordance with the systems and methodologies disclosed herein is shown. The computing device 800 includes at least one processor 802 that executes instructions stored in memory 804. Instructions include, for example, instructions for implementing functions described above as being performed by one or more components, or instructions for performing one or more of the above methods. The processor 802 can access memory 804 via a system bus 806. In addition to storing executable instructions, memory 804 can also store conversation inputs, scores assigned to conversation inputs, and the like.

[0052] Furthermore, the computing device 800 also includes a datastore 808 accessible to the processor 802 via the system bus 806. The datastore 808 may contain executable instructions, log data, and the like. The computing device 800 also includes an input interface 810 that allows external devices to communicate with the computing device 800. For example, the input interface 810 can be used to receive instructions from external computer devices or users. The computing device 800 also includes an output interface 812 that provides an interface between the computing device 800 and one or more external devices. For example, the computing device 800 can display text, images, and the like via the output interface 812.

[0053] External devices communicating with the computing device 800 via the input interface 810 and output interface 812 are intended to be included in an environment that provides virtually any type of user interface with which the user can interact. Examples of user interface types include graphical user interfaces and natural user interfaces. For example, a graphical user interface can accept input from a user using input devices such as a keyboard, mouse, or remote control, and provide output to an output device such as a display. Furthermore, a natural user interface allows the user to interact with the computing device 800 without being constrained by input devices such as a keyboard, mouse, or remote control. Rather, a natural user interface relies on voice recognition, touch and stylus recognition, gesture recognition both on and in the immediate vicinity of the screen, air gestures, head and eye tracking, voice and speech, vision, touch, gestures, artificial intelligence, etc.

[0054] Furthermore, although described as a single system, it should be understood that computing device 800 may be a distributed system. Therefore, for example, multiple devices can communicate via a network connection and collectively perform the tasks described as being performed by computing device 800.

[0055] The various functions described herein can be implemented in hardware, software, or any combination thereof. If implemented in software, the functions can be stored or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable medium includes computer-readable storage media. Computer-readable storage media can be any available storage medium accessible to a computer. Examples of such computer-readable storage media include, but are not limited to, RAM, ROM, EEPROM®, CD-ROM, or other optical disk storage devices, magnetic disk storage devices, or other magnetic storage devices, or any other medium accessible from a computer that can be used to hold or store desired program code in the form of instructions or data structures. As used herein, the terms "disk" and "disc" include Compact Disc (CD), LaserDisc (disc), Optical Disc (disc), Digital Multipurpose Disc (disc) (DVD), Floppy Disk (disk), and Blu-ray Disc (disc) (BD), where a disk typically reproduces data magnetically, and a disc typically reproduces data optically using a laser. Furthermore, propagated signals are not included in the scope of computer-readable storage media. Computer-readable media also include communication media, which include any medium that facilitates the transfer of computer programs from one place to another. For example, a connection can be a communication medium. For example, if software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, or microwave, then coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, or microwave are included in the definition of communication media. The above combinations should also be included within the scope of computer-readable media.

[0056] Alternatively, or furthermore, the functions described herein can also be performed, at least in part, by one or more hardware logic components. For example, but not limited to, exemplary types of usable hardware logic components include field-programmable gate arrays (FPGAs), program-specific integrated circuits (ASICs), program-specific standard products (ASSPs), system-on-chip systems (SoCs), and complex-programmable logic devices (CPLDs).

[0057] It will be understood that the above circuit can have other functions in addition to the functions described above, and that these functions can be performed by the same circuit.

[0058] The applicant discloses each feature described herein, and any combination of two or more such features, separately, to the extent that such features or combinations can be done as a whole under this specification, in light of common sense of those skilled in the art, regardless of whether such features or combinations of features solve the problems disclosed herein and without limitation to the claims. The applicant shows that aspects of the present invention may consist of such individual features or combinations of features.

[0059] The embodiments of the present invention are described by reference to different categories. Specifically, some embodiments are described by reference to methods, while others are described by reference to apparatus. However, those skilled in the art will be able to infer from the above description that, unless otherwise specified, any combination of features belonging to one category, as well as any combination of features relating to different categories, are disclosed in this application. However, all features can be combined to provide a synergistic effect beyond a mere collection of features.

[0060] The present invention is illustrated and described in detail in the drawings and the above description, but such illustrations and descriptions should be considered illustrative and not restrictive. The present invention is not limited to the disclosed embodiments. Other variations of the disclosed embodiments can be understood and practiced by those skilled in the art from a consideration of the drawings, disclosure and appended claims.

[0061] The word "includes" does not exclude other elements or steps.

[0062] The singular form of an element does not preclude plurality. Furthermore, as used herein, the singular form of an element should generally be interpreted as meaning "one or more" unless otherwise specifically designated or unless the context makes it clear that it is singular. A single processor or other unit may perform the functions of several items described in the claims. Means described in mutually different dependent claims may be combined advantageously. A computer program may be stored / distributed on any suitable medium, such as optical storage media or solid-state media, supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless communication systems. Any reference numeral in the claims should not be interpreted as limiting the scope. Unless otherwise specifically designated or unless the context makes it clear, as used herein, the expression "A and / or B" is intended to mean all possible sorts of one or more of the listed items. That is, the expression "X includes A and / or B" is satisfied by any of the following cases: X includes A, X includes B, or X includes both A and B.

Claims

1. A method for detecting the semantic position of a patient, Computers The steps include receiving non-geospatial sensor data from at least one sensor of a wearable monitoring device worn by the patient, The steps include obtaining patient activity level data or posture data from the aforementioned non-geospatial sensor data, The steps include extracting one or more activity behavioral features from the patient activity level data, or extracting one or more postural behavioral features from the posture data, The steps include classifying the activity behavioral characteristics or postural behavioral characteristics as belonging to at least two predefined semantic patient positions, The steps include outputting an indication of the detected semantic patient position, How to do it.

2. The method according to claim 1, further comprising the step of low-pass filtering the patient activity level data or the posture data before extracting the features.

3. The method according to claim 1, wherein the activity behavioral feature or the postural behavioral feature includes a feature that describes the variance in the patient activity level data or the postural data.

4. The activity behavioral characteristics include one of the following: the standard deviation of the activity level signal in the patient activity level data, the range of the activity level signal, the area around the mean value of the activity level signal, the duration for which the activity level signal is below the low activity threshold, the area of ​​the activity level signal that is below the low activity threshold and above the high activity threshold, and the sum of the activity level signals that are above the high level threshold or below the low level threshold. The method according to claim 1, wherein the posture behavior feature includes one or more of the mean value of the posture signal in the posture data, the median value of the posture signal, the standard deviation of the posture signal, the range of the posture signal, and the area of ​​the posture signal.

5. The method according to claim 1, wherein the first predefined semantic patient location includes the semantic patient location "in the hospital", and the second predefined semantic patient location includes the semantic patient location "at home".

6. The method according to claim 1, wherein the step of processing the non-geospatial sensor data and classifying it as belonging to one of the at least two predefined semantic patient locations includes the step of using a comparison-based algorithm to identify similarities between the non-geospatial sensor data and at least one reference dataset representing expected data for the corresponding predefined semantic patient locations.

7. The method according to claim 1, wherein the step of processing the non-geospatial sensor data and classifying it to belong to one of the at least two predefined semantic patient locations includes using a template matching algorithm that multiplies the patient activity level data or the posture data by a template of weights representing the expected patient activity level pattern or expected posture pattern for the corresponding predefined semantic patient location.

8. The method according to claim 1, wherein the step of processing the non-geospatial sensor data and classifying it as belonging to one of the at least two predefined semantic patient locations includes the step of using a trained machine learning model to perform the classification.

9. The method according to claim 8, wherein the machine learning model includes a logistic regression classifier trained to classify the extracted one or more activity behavioral features and / or postural behavioral features as belonging to one of two predefined semantic patient positions.

10. The method according to claim 1, wherein the step of outputting the indication of the detected semantic patient location includes the step of outputting the probability that the current semantic patient location belongs to the predefined semantic patient location, and / or a determination relating to the predefined semantic patient location to which the current semantic patient location belongs.

11. A computing device comprising a processor that performs the method according to any one of claims 1 to 10.

12. A program that performs the method according to any one of claims 1 to 10.

Images