ADAPTIVE SENSOR SELECTION

TECHNICAL FIELD

[0001] This disclosure relates generally to a medical device, and more specifically, but not exclusively, to a medical device for monitoring one or more health conditions.

BACKGROUND

[0002] Efforts are continually being made to address the health concerns of people of any age, but especially for prematurely born babies and infants. Many problems that arise in the first few months of life occur during sleep. Monitoring (term and preterm) baby/infant sleep therefore has been an area of interest for parents and healthcare professionals.

[0003] Various types of unobtrusive devices have been delveoped to acquire vital signs associated with infant sleep. Some devices are passive in that they merely perform a monitoring function. Others produce signals that are analyzed in an attempt to assess sleep quality, e.g., to determine sleep statistics (e.g., sleep efficiency, total sleep time, number of awakenings, and sleep onset latency) and sleep stages/ states (including wake, active sleep and quiet sleep). This information may provide insight into the mental and physical development of an infant.

[0004] One type of passive monitor uses a camera to monitor an infant during sleep, but does not provide any meaningful information beyond what can be visually seen in the camera images.

[0005] Another type of monitor includes a monitoring mat having a top mesh layer, a bottom mesh layer, two flexible sheets, and an optical cable connected to a fiber-optic sensor. The sensor is used to detect infant respiration and movement that may be associated with sudden infant death syndrome.

[0006] Another type of monitor uses load cells in a mattress to measure pressure changes of a sleeper. The pressure changes are converted to signals indicative of sleep states. A data analyzer is also included to analyze heartbeat, respiration, and physical activity data collected from miniature sensors in the bed. These sensors detect physiological and physical data (e.g., heart rate regularity variation, respiration regularity variation, and gross body movement) which is also used to determine sleep states.

[0007] While these devices may provide some limited form of health monitoring, they have many drawbacks. For example, existing devices do not detect many vital factors that directly bear upon the actual health condition of an infant during sleep. Also, existing approaches use of obtrusive sensors, e.g., ones that must be attached to the body of an infant. Also, existing approaches use only one sensor with a very limited coverage area. As a result, signals cannot be detected or accurately monitored, especially when an infant is sleeping in an area not near the sensor. Existing approaches also implement manual scoring techniques which are inefficient.

SUMMARY

[0008] A brief summary of various example embodiments is presented below. Some simplifications and omissions may be made in the following summary, which is intended to highlight and introduce some aspects of the various example embodiments, but not to limit the scope of the invention. Detailed descriptions of example embodiments adequate to allow those of ordinary skill in the art to make and use the inventive concepts will follow in later sections.

[0009] In accordance with one embodiment, a method for processing health information includes receiving signals from sensors in a sleep mat, determining values of a first feature based on the sensor signals, calculating metrics for the sensors based on the first feature values, selecting one of the sensors in the sleep mat based on the metrics, and determining presence of a sleep subject based on signals from the selected sensor.

[0010] Determining the presence of the sleep subject may include extracting values corresponding to at least one feature in a temporal domain from the signals of the selected sensor; extracting values corresponding to at least one feature in a spectral domain from the signals of the selected sensor; and determining the presence of the sleep subject, during at least aportion of an observation period, based on the values corresponding to the at least one feature from in the temporal domain and the values corresponding to the at least one feature in the spectral domain. The method may also include applying a threshold to the values corresponding to the at least one feature from in the temporal domain and the values corresponding to the at least one feature in the spectral domain. The method may also include determining that the sleep subject is absent during at least one time of the observation period based on the values corresponding to the at least one feature from in the temporal domain and the values corresponding to the at least one feature in the spectral domain.

[0011] Determining the presence of a sleep subject may be performed by a machine learning algorithm trained based on information presence detections for the subject or a different subject. The sensor signals may be at least one of pressure signals, acceleration signals, proximity signals, signals from inductive sensors, or signals from capacitive sensors. The pressure signals may correspond to ballistocardiogram signals.

[0012] In accordance with one or more other embodiments, a method for processing health information includes receiving signals from first sensors and second sensors in a sleep mat, calculating metrics based on the signals from the first sensors and the second sensors, selecting one of the first sensors and one of the second sensors based on the metrics, and classifying one or more sleep states of a subject based on signals from the selected one of the first sensors and the selected one of the second sensors, wherein the first sensors have a first gain and the second sensors have a second gain different from the first gain.

[0013] Calculating the metrics may include determining values of a first feature based on the signals from the first sensors, determining values of a second feature based on the signals from the second sensors, calculating first metrics for the first sensors based on the first feature values, calculating second metrics for the second sensors based on the second feature values, and selecting said one of the first sensors based on the first metrics and said one of the second sensors based on the second metrics. The first sensors may detect physical activity of the subject and the second sensors may detect cardiorespiratory activity of the subject.

[0014] The method may include detecting presence of the subject and classifying the one or more sleep states after the presence of the subject is detected. Classifying the one or more sleep states may be performed by a machine learning algorithm trained based on information corresponding to prior classifications of sleep states for the subject or a different subject. The first sensors and the second sensors may be selected from the group consisting of pressure sensors, acceleration sensors, or proximity sensors. The signals from the first sensors and the second sensors may correspond to ballistocardiogram signals. The one or more sleep states may include absent, wake, active sleep, and quiet sleep. At least in the case of infants (and particularly, for neonates), an intermediate (or transitional or indeterminate) sleep state may also be included.

[0015] In accordance with one embodiment, a sleep mat includes a plurality of first sensors, a plurality of second sensors, and a layer of material to support a sleep subject, wherein the first sensors and the second sensors arranged at different locations and are configured to output at least one of pressure signals, acceleration signals, or proximity signals during an observation period, and wherein the layer of material includes or is coupled to the plurality of first sensors and the plurality of second sensors.

BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying figures, where like reference numerals refer to identical or functionally similar elements throughout the separate views, together with the detailed description below, are incorporated in and form part of the specification, and serve to further illustrate example embodiments of concepts found in the claims and explain various principles and advantages of those embodiments.

[0017] These and other more detailed and specific features are more fully disclosed in the following specification, reference being had to the accompanying drawings, in which:

[0018] FIG. 1 illustrates an embodiment of a multiple-sensor sleep mat;

[0019] FIG.2 illustrates an embodiment of a system for detecting presence/absence of a sleep subject and for classifying sleep states of the subject;

[0020] FIG. 3 illustrates an embodiment of a method for detecting presence/absence of a sleep subject and for classifying sleep states of the subject;

[0021] FIG. 4 illustrates examples of signals analyzed by the system and method;

[0022] FIG. 5 illustrates examples of signals analyzed by the system and method;

[0023] FIG. 6 illustrates an embodiment of a method for presence/absence detection;

[0024] FIG. 7 illustrates additional processing operations of the system and method;

[0025] FIG. 8 illustrates an embodiment examples of signals analyzed by the system and method;

[0026] FIG. 9 illustrates an embodiment of a method for selecting sensors for classifying sleep states;

[0027] FIG. 10 illustrates an embodiment of a sleep mat with low and high gain sensors;

[0028] FIG. 11 illustrates an embodiment of a method of feature extraction for classifying sleep states;

[0029] FIG. 12 illustrates an embodiment of a method for classifying sleep states; and

[0030] FIG. 13 illustrates examples of signals and waveforms for detecting a sleep subject and classifying sleep states for the subject. DETAILED DESCRIPTION

[0031] It should be understood that the figures are merely schematic and are not drawn to scale. It should also be understood that the same reference numerals are used throughout the figures to indicate the same or similar parts.

[0032] The descriptions and drawings illustrate the principles of various example embodiments. It will thus be appreciated that those skilled in the art will be able to devise various arrangements that, although not explicitly described or shown herein, embody the principles of the invention and are included within its scope. Furthermore, all examples recited herein are principally intended expressly to be for pedagogical purposes to aid the reader in understanding the principles of the invention and the concepts contributed by the inventor(s) to furthering the art and are to be construed as being without limitation to such specifically recited examples and conditions. Additionally, the term,“or,” as used herein, refers to a non-exclusive or (i.e., and/ or), unless otherwise indicated (e.g.,“or else” or “or in the alte native”). Also, the various example embodiments described herein are not necessarily mutually exclusive, as some example embodiments can be combined with one or more other example embodiments to form new example embodiments. Descriptors such as“first,”“second,”“third,” etc., are not meant to limit the order of elements discussed, are used to distinguish one element from the next, and are generally interchangeable. Values such as maximum or minimum may be predetermined and set to different values based on the application.

[0033] Example embodiments describe a system and method for detecting the condition of a sleep subject using a multi-points sensory sleep mat. In some embodiments, the presence or absence of the subject may be determined based on sensor signals from the mat. In these or other embodiments, sleep states may be determined during an observation period by processing the sensor signals. The sleep mat may include sensors of various types (e.g., pressure, accelerometer, and/or proximity sensors) arranged at different locations. The use of multiple sensors increases the effective detection area of the mat and prevents or reduces measurement of invalid or noisy data. The use of multiple sensors also allows for comprehensive monitoring irrespective of the position, movement, or activity of the sleep subject. The mat may be used to monitor the health of a subject of any age, but is particularly useful for prematurely born babies and infants in incubators or cribs.

[0034] In some embodiments, the system and method uses multiple sensors to simultaneously measure signals conveying variations caused by body movement, respiration and/ or heartbeat. The presence or absence of certain states or conditions may be determined by analyzing temporal and spectral features extracted from the signals. In one case, a machine learning algorithm (e.g., linear discriminant (LD) classifier) may be trained on these features and then used to detect and classify the presence and sleep states of a sleep subject. The accuracy of the machine learning algorithm may increase with increasing numbers of uses and detections/classifications, either for the same sleep subject or for different sleep subjects.

[0035] In some embodiments, the temporal and/ or spectral features may be computed for each non overlapping period of time (e.g., 30-s epoch). Such an epoch length maybe used, for example, because it corresponds to the standard of human sleep state analysis according to the guidelines from the American Academy of Sleep Medicine (AASM). However, a different epoch or machine learning algorithm may be used in another embodiment.

[0036] Because an infant can lie on the mat in different positions, different sensors in the mat may detect varying levels of signal quality. For example, sensors closer to a sleeping infant may be more accurate, less noisy, and less susceptible to corruption from outside influences (e.g. parental activity, nurse handling, etc.) than sensors farther away from the infant. A sensor selection algorithm may be implemented to improve the accuracy of the presence detector and sleep state classifier. For example, in one embodiment, signals or channels from sensors closer to the sleep subject (or ones that otherwise produce relatively improved signal quality) may be identified and adaptively selected prior to presence detection and sleep state classification. This may improve performance and, in turn, may alert a care professional to an imminent health threat.

[0037] FIG. 1 illustrates an embodiment of a multi-points sensory sleep mat 110 which includes a plurality of sensors 120 arranged in a predetermined pattern. The mat may be used to detect the presence (or absence) and/or one or more sleep states of a subject. Examples of the sleep states include an awake state, an active sleep state, and a quiet sleep state. In one example, the mat 110 may be used in the incubator, crib, or bed 130 of an infant to be monitored. In other embodiments, the mat maybe used for an older child or adult, for example, in order to measure signals relating to sleep apnea or other heath conditions.

[0038] The sleep mat 110 may be placed between the sleep subject and the top of a mattress pad. In one case, the mat 110 may be coupled to or included within a mattress or bedsheet. In this latter case, the sleep mat 110 may be removable from the mattress, mattress pad, or bedsheet or may be sewn directly into, mounted on, or otherwise made an integral part of the mattress, mattress pad, or bedsheet. The outer covering of the mat may be made, for example, of a fabric, plastic, rubber, or a polymer or composite material. The outer covering may include a surface for supporting the subject.

[0039] The sensors 120 collect various signals indicative of the presence or absence and/ or the sleep states of the subject. In one embodiment, the sensors 120 may include one or more pressure sensors, one or more acceleration sensors, and/or one or more proximity sensors. The pressure sensors may generate signals that may be used as a basis for measuring variations in or patterns of breathing effort and heartbeats. The acceleration sensors may generate signals that measure movement of the sleep subject, e.g., gross body movements as well as breathing and heartbeat movements, all of which correspond to detectable changes in acceleration. The proximity sensors may generate signals (e.g., that measure changes in capacitance or inductance) indicative of the proximity of the body of the sleep subject relative to those sensors. Variations in these signals may be produced, for example, by the breathing and/or presure from the thorax or other body part of the sleep subject. The mat may be powered by batteries and/ or electricity from a wall socket.

[0040] In one embodiment, the sensors 120 may detect the body movement, respiration, and/ or cardiac activity (e.g., electrocardiogram or heart rate) of the sleep subject after the subject is detected to be present. Certain time and spectral features may be extracted from the detected signals and processed to determine physical and/or physiological activities of the subject. These features are described in greater detail below in connection with the system and method embodiments.

[0041] FIG. 2 illustrates an embodiment of a system 200 for determining the presence/absence and the sleep state classification of a subject based on sensor signals from the sleep mat 110. The system includes a processor 210, a memory 220, a storage area 230, and a display 240. The processor 210 processes the sensor signals from the sleep mat using one or more algorithms in order to generate detection and classification information. The memory 220 may store logic (e.g., hardware, software modules, or a combination thereof) for implementing the algorithms. The storage area 230 may store processing results for detecting and monitoring operations performed by the processor. These results may be used as a basis for training the detectors, classifiers, and machine learning algorithms of the disclosed embodiments. The display 240 may display the results generated by the processor, for example, in the form of text, graphics, and/ or images.

[0042] FIG. 3 illustrates an embodiment which may use the system of FIG. 2 to implement two methods: a presence/ absence detection method and a sleep state classification method. Both methods may be performed based on sensor signals output from the mat of FIG. 1. In one case, the presence/absence detection method may be implemented without performing the sleep state classification method. In another case, the sleep state classification method maybe performed without performing the presence/absence detection method. In yet another case, the sleep state classification method may be performed only after the presence/absence detection method is performed. Presence /Absence Detection Method

[0043] In accordance with one embodiment, the method for detecting the presence or absence of a sleep subject is performed during an observation period, which, for example, may be determined by a user turning on power of the sleep mat and initiating a recording session or automatically based on programming of a controller of the sleep mat.

[0044] Referring to FIG. 3, the method includes, in operation 1305, receiving signals from the sensors 120 in the sleep mat. As previously indicated, the sensors may include one or more pressure sensors, one or more acceleration sensors, one or more proximity sensors, one or more inductive sensors, and/ or one or more capacitive sensors. For illustrative purposes, the sensors may be pressure sensors that generate (or may be converted to) ballistocardiogram (BCG) signals. In one embodiment, BCG signals are received from all k sensors in the sleep mat.

[0045] In operation 1310, the signals are pre-processed (e.g., normalized, filtered, etc.) prior to use in training the algorithm (e.g., classifier) for detecting presence/absence of the sleep subject. In one embodiment, the pre-processing may involve processing the BCG signals to generate BCG-derived respiratory (BDR) signals. In another embodiment, the sensor signals may be used in subsequent operations without pre-processing, and/ or signals different from BCG or BDR signals may be used.

Adaptive Sensor Selection

[0046] One of the algorithms stored in memory 220 may adaptively select one or more sensors to be used in detecting the presence or absence of the sleep subject, which in this case is an infant. In operation 1315, this algorithm extracts one or more features from the sensors. The feature(s) to be extracted may be, for example, the one(s) indicated in Table 1, described in greater detail below. [0047] In operation 1320, data quality index (DQI) metrics are calculated for the k sensors based on the extracted feature(s). Based on the DQI metrics, a predetermined number of the k sensors is selected by the algorithm for performing presence/ absence detection.

[0048] In one embodiment, only one of the sensors 120 is selected, e.g., the sensor producing the highest DQI value. In another embodiment, the sensor generating a DQI value that falls within a predetermined range may be selected, even if the DQI metric is not the highest one produced by the sensors in the mat. Calculation of the DQI metrics and subsequent selection of a sensor may be performed, for example, based on Equations 1 and 2 indicated below.

[0049] During the observation period (during which the sensor signals may be recorded), it is possible that the selected sensor may malfunction or conditions may change that may make another one of the sensors more suitable for performing presence/absence detection. In accordance with one embodiment, the algorithm may continue to receive signals, continuously or periodically, from all sensors, generate corresponding DQI metrics, and then select another sensor in the same observation period should that other sensor produce, for example, a better quality signal (e.g., a higher DQI metric). In this additional sense, the sensor selection algorithm may be considered adaptive in nature.

Presence/ Absence Detection

[0050] In operation 1325, the sensor selected in operation 1320 is input into a feature extraction module, which extracts one or more predetermined features to be used for detecting the presence (or absence) of the infant on the mat. In one embodiment, multiple features are extracted in temporal and/or spectral domains based on the sensor (e.g., BCG signals) output from the selected sensor.

[0051] In operation 1330, the features extracted in operation 1325 are input into a classifier or detector (e.g,. machine learning algorithm). The features are analyzed by the classifier/detector to generate a result indicating whether the infant is present or absent. The classifier may output signals, continuously or periodically, to indicate the presence or absence of the infant throughout the observation period. The data and features generated during feature extraction, sensor selection, and presence/absence detection may be stored for purposes of training the classifier in performing future sensor selection and presence detection operations in subsequent observation periods, either for the same sleep subject or a different sleep subject.

[0052] In operation 1335, the presence/absence detection result from the classifier training may be smoothed, for example, to filter out spurious or unwanted signals. Once smoothed and a sleep subject is detected to be present or absent, a control signal 1410 is output to the classifier training module to initiate sleep state classification. Example embodiments of algorithms that may be used to perform the operations of the presense/absence detection method are discussed in greater detail below.

Sleep State Classification Method

[0053] The sleep state classification method may be performed alone or in tandem with (e.g., after) the presence/absence detection method. The sleep states to be classified may include, for example, absent, wake, active sleep, and quiet sleep, etc. In order to detect these states, an adaptive sensor selection algorithm is performed, which is followed by a feature extraction algorithm, based on signals from the sensors in the sleep mat. In this example, it is assumed that the sensors are piezo (touch) pressure sensors, but other types of sensors may be used.

Adaptive Sensor Selection

[0054] In operation 1340, the method includes adaptively selecting a predetermined number of sensors of the sleep mat to be used for sleep state classification. In one embodiment, the mat has four sensors: two high gain sensors and two low gain sensors. Different features may be extracted from the sensor signals output from the low gain and high gain sensors. Example methods for performing adaptive sensor selection for sleep state classification will be discussed below. Feature Extraction

[0055] In operation 1345, DQI metrics are calculated for the high and low gain sensors. Based on the the DQI metrics, one high gain sensor and one low gain sensor are selected. These sensors may be the ones having, for example, the highest respective DQI metrics.

[0056] In operation 1350, feature extraction is performed to detect the sleep state(s) of the infant during the presence period based on the signals output from the two sensors selected in operation 1345. This may be performed, for example, based on embodiments previously described, e.g., extracting multiple features in time and spectral domains based on BCG and BDR signals only from the two selected (low/high) sensors. The features extracted for sleep state classification may be the same or different from the feature(s) selected for presence/absence detection. Examples of the features extracted for sleep state classification are set forth in Table 2 below.

Sleep State Classification

[0057] In operation 1355, the sleep state(s) of the infant during the observation period are determined based on the extracted features from the selected high gain and selected low gain sensors in operation 1350. The detection may be performed by a classifier (e.g., machine learning algorithm) different from the classifier/detector used to perform presence/absence detection. The data and features extracted during these operations may be used to train the sleep state classifier for purposes of detecting sleep states for future observation periods, for the same or a different sleep subject. The output of the classifier in 1355 may be smoothed in operation 1360 being being output, for example, on a display for viewing by a user. Smoothing

[0058] In applying the embodiments described herein, it is possible that the method would produce very short periods of infant presence or absence (e.g., periods of one or two epochs). However, heuristically, these short periods should not occur very often. A post-processing operation may be performed to smooth the detection results (e.g., see operations 1335 and 1360) in order to filter out those very short periods. In one embodiment, a median filter with a sliding window (e.g., a window of predetermined size w) may be used for smoothing. Such a filter takes an average over time, which may reduce or eliminate the short periods mentioned above. This window size may, for example, be optimized in various ways, e.g., the window size w = 10 min or another appropriate value may be used to smooth presence/absence detection results. In one embodiment, a median filter with a window size of 12.5 min may be used, for example, to smooth the sleep state classification results.

[0059] Example embodiments for performing the algorithms of the method and system of FIG. 3 will now be discussed. In these embodiments, signals are received from four piezo pressure sensors at predetermined locations (e.g., at respective corners) of the mat in oder to detect BCG signals from the sleep subject. The piezo pressure sensors may be load cells in the form of piezo touch plates, for example, of the type disclosed in WO 2015078937.

Presence/ Absence Detection: Adaptive Sensor Selection

[0060] Sensor selection may be adaptively performed based on one or more features indicated, for example, in Table 1. In one embodiment, only one of the features in Table 1 maybe used. The feature to be used in adaptive sensor selection for this example is spectral vlf but another feature (including any of the ones from Table 1) may be used in another embodiment. Once this feature is determined (which, for example, may be preselected and/ or preprogrammed into the algorithm), a data quality index (DQI) metric may be calculated based on the values of feature spectmlvlf during the observation period.

Table 1

[0061] FIG. 4 illustrates examples of values generated for feature spectml vlf throughout an observation period. In this case, the observation period was 40 minutes (e.g., which may translate into 80 epochs). The signal segments in (a) and (b) are 120 seconds in duration and are provided to show differences in signal values of feature spectml vlf in periods that will be detected as absence and presence periods.

[0062] In the enlarged section of waveform (b), the values (or signal pattern) of feature spectml vlf may be correlated to infant physiological and physical activities when the infant is present. The values (or signal pattern) in section (a) may be correlated to the absence of the infant. More specifically, in the enlarged section of waveform (a), the signal pattern of feature spectml vlf may be indicative of slow signal drift of sensor 1 (e.g., see the mat in FIG. 1). Although the output of sensor 1 is illustrated in FIG. 4, it is understood that the values for feature spectml vlf are collected and processed so that only one of the sensors is selected for presence/ absence detection. For example, the values for feature spectral vlf may be calculated and compared to one another to determine the signal quality of each sensor, and then one of the sensors is selected for determining presence or absence of the infant.

[0063] In some cases, slow signal drift may not exist, for example, depending upon the hardware. In that case, the values for feature spectral vlf may still be useful for sensor selection due to its characterization on physiological activity. In other embodiments, a different feature (e.g,. from Table 1) may be preselected for use (e.g,. as a backup parameter) for performing adaptive sensor selection.

[0064] The signal quality of the sensors may be determined, for example, based on DQI metrics calculated for the sensors. An example of how DQI metrics may be specifically calculated will now be given. For a period of time (with n epochs), during which a target sensor is to be selected, the DQI metric for each of the four sensors 120 may be computed according to Equation 1.

DQI = percentile (Fx, a) / percentile (Fx, 100-a), (1) where Fx corresponds to a series of continuous values x of feature F during the observation period and percentile (Fx, a) is the a-th percentile of those feature values (0 < a < 100). In other words, DQI indicates the certain inter-percentile range of the values of one of the features in Table 1, which in this case is spectral vlf In one case, the value of a = 99.5, which indicates that a higher feature quality DQI corresponds to a larger range between the high (99.5-th) percentile and the low (0.5-th) percentile. When an infant is absent, low frequency (LF) and high frequency (HF) powers may be very small, for example, due to the absence of physical/physiological information. At the same time, VTF power may be relatively large because of the presence of slow signal drift, which leads to large feature values.

[0065] When an infant is present, the TF and HF frequency powers are expected to substantially increase due to the presence of physical/physiological information conveyed in the signals (e.g., BCG) output from the sensors. Therefore, the values of the selected feature (e.g., spectral ylj) may be small. In one case, larger differences in feature values between presence and absence states should translate into a higher feature quality (associated with a higher DQI value), which, in turn, may provide a better basis for distinguishing between the two states (presence/absence).

[0066] FIG. 5 illustrates an examples of spectral vlf values calculated based on signals from all four sensors of the sleep mat 110. The signals may be different from one another because the sensors are at different locations on the mat and because the infant may shift position during the observation period. As shown in waveform (a) of FIG. 5, sensor 1 produced the greatest difference in the feature value spectral vlf and thus the greatest DQI value (as determined by Equation 1). Accordingly, this sensor may be selected as the sensor to be used in determining the presence or absence of a sleep subject, either for the same subject or for different subjects, in future applications of the sleep mat.

[0067] In one embodiment, a may be set to a fixed predetermined value or may be adaptively set, for example, depending on the signal length n. Assuming an infant sleep monitoring system with k piezo pressure sensors with corresponding data quality DQE (k = 1, 2, ...), the sensor S that is selected may be the one having the largest DQI as determined by Equation 2.

S = argmax _f c (DQI DQI ₂ , ... , DQI*) (2)

Presense/Absence Detection: Feature Extraction

[0068] FIG. 6 illustrates an embodiment of a method for determining the presence or absence of a sleep subject based on signals from the selected sensor, e.g., sensor 1 indicated above. In operation 610, signals are received from the selected sensor in the observation period, which, for example, may be 2400 seconds or 40 minutes. In operation 620, the signals from the selected sensor are converted or otherwise correspond to BCG signals. [0069] In operation 630, the BCG signals are processed to provide an indication of the absence or presence of a sleep subject. The signals may be processed by a machine learning algorithm as described herein. The processing may involve, for example, amplifying, filtering, and extracting information from the sensor signals indicative of whether the sleep subject is on the mat (and thus present) or whether the sleep subject is off the mat or otherwise absent. Among other indicia, the information may correspond to the amplitude of the signals and/ or certain recognizable patterns in the signals over the observation period. The machine learning algorithm may be trained to recognize and identify these patterns as corresponding to presence and absence detection.

[0070] In operation 640, a decision on whether the sleep subject is absent or present is made by the machine learning algorithm. This may involve, for example, comparing the sensor signals, either with themselves, to one or more predetermined thresholds, and/or to sensor data received from the sensors in the past used to train or update the machine learning algorithm. In one embodiment, the location of the sleep subject relative to different locations of the mat may be determined based on a comparison of the BGC sensor signals.

[0071] The machine learning algorithm may be based, for example, on a Bayesian TD classifier, e.g., a single-epoch classifier which uses Bayes’ rule to minimize the probability of error by selecting, based on the observations (features) separately for each epoch, the most likely class (sleep stage), e.g., the class that maximizes the following discriminant function: where X is the feature vector, /i; is the average feature vector for class i, and å is the pooled covariance matrix for all classes. P(tU _j ) maybe a fixed term with the prior probability of each class. All parameters m _ΐ , å, and P(tU _j ) may be determined during training. For example, in the case of two classes a and b, the classifier may choose class a if

9a(x) - 9b (x) > D, where D is a decision threshold. Assuming the prior probabilities of all classes as constant, the decision boundary D may be given by the following equation:

[0072] While an LD classifier may provide to be efficient and straightforward to interpret for sleep state classification, a different (linear or nonlinear) classifier may be used to perform sleep state classification in another embodiment. Examples of other classifiers include logistic regression, decision tree, random forest, support vector machines, Adaboost, extreme gradient boosting, conditional random fields, and deep neural networks.

[0073] In one embodiment, values corresponding to more than one feature may be extracted from the sensor signals in order to determine whether the infant is present or absent. In one embodiment, six features may be extracted from the sensor signals (e.g., as set forth in Table 1), each detected in continuous non-overlapping 30 s epochs.

[0074] FIG. 7 illustrates an embodiment of a method for performing feature extraction from the sensor signals in order to determine the presence or absence of the sleep subject. In describing this method, the sensor signals are assumed to be BCG signals in order to be consistent with the example previously discussed. However, all or a portion of the sensor signals may be different from BCG signals in another embodiment. [0075] Referring to FIG. 7, an initial operation 710 includes normalizing the BCG signals received during the observation period. In one case, the BCG signals may be normalized to have zero mean and unit variance. The signal may be normalized in a different manner in another embodiment.

[0076] In operation 720, the standard deviation (time std) and the zero-crossing rate ( time zcr ) are computed for each of the BCG signals in the temporal domain.

[0077] In operation 730, the power spectrum density (PSD) is calculated for each of the sensor signals using fast Fourier transform (FFT). This results in transforming the BCG signals from the temporal domain to the frequency (or spectral) domain.

[0078] In operation 740, the power spectrum generated in opeation 730 is divided into a plurality of frequency bands. In one embodiment, the power spectrum is divided into three different frequency bands: (i) very-low-frequency (VLF) band from 0 to 0.3 Hz, (ii) low-frequency (LF) band from 0.3 to 3 Hz, and (iii) high-frequency (HF) band between 3 and 10 Hz. The relative VTF spectral power (spectral _vlf) may be computed, for example, as the ratio between the frequency power in the VTF band and that in the other two bands. The relative TF spectral power (spectral If) may be computed, for examlpe, as the ratio between the frequency power in the TF band and that in the VTF and HF bands. When the sleep subject is an infant, the physiological information associated with respiration and heartbeat may be reflected in the TF band, while spectral components of body movements may likely occur in both the TF and the HF band.

[0079] In operation 750, an analysis may be performed to extract features indicative of the presence or absence of the infant during the observation period. The analysis may include, for example, performing linear regression to fit the PSD curve and then calculating the corresponding intercept (spectral intercept) and regression coefficient or slope (spectral slope). These features are summarized and described in Table 1. By analyzing the manner in which the spectral power varies with increasing frequency in each of the frequency bands, features may be extracted that are indicative of presence/absence and sleep states of the infant. These feature may be determined by the machine learning algorithm to correspond to specific sleep states based on past historical patterns and data.

[0080] FIG. 8 illustrates an example of BCG signals received and processed from respective ones of the sensors 120 according to one example. In FIG. 8, section (a) corresponds to BCG signals output from sensor 1, section (b) corresponds to BCG signals output from sensor 2, section (c) corresponds to BCG signals output from sensor 3, and section (d) corresponds to BCG signals output from sensor 4. The four sensors are arranged in the manner illustrated in FIG. 1.

[0081] Referring to FIG. 8, from the amplitudes of the BCG signals from all four sensors, it is evident that the sleep subject (infant) was absent from the mat for the first 20 minutes of the observation period and present during the last 20 minutes of this period. During the period when the infant was absent, the signal from sensor 1 had lower amplitude disturbances or noises compared to the other sensors. During the period when the sleep subject was present on the mat, the quality of the signals from sensors 1, 2, and 4 was better than the signal from sensor 3. The signals from sensors 1, 2, and 4 may therefore provide a more reliable indication for separating the periods when the infant was present and absent. In particular, the signal from sensor 1 had a lower amplitude signal when infant is present relative to sensors 2 and 4. This is probably because the sensor 1 signal may have been less influenced by other activities, e.g., parent preparation of putting the infant into the bed or nurse handling during which they might press the mat. In each section, enlarged windows are provided to illustrate the specific waveforms from respective ones of the sensors over an approximately 10 second time frame.

[0082] In accordance with one embodiment, both sensor (signal channel) selection and presence/absence detection may be performed by thresholding signal power at a frequency range which contains vital sign signals. In this case, the presence of vital signs that exist when the infant is in bed will boost power in the frequency range between 0 and 7 Hz to a level higher compared to times when bed is empty.

[0083] In one embodiment, in selecting of the best sensor(s), the gain of four signal pre-amplifiers may be set to the same level for corresponding ones of the sensors. A sum of the signal power in the range of 0 to 7 Hz may then be performed to select the channel showing the highest power value. The closer the infant (signal source) is to the sensor(s), the stronger the manifestation of vital signs in the sensor signals. Therefore, the channel(s) may be selected where the presence of the vital signs is strongest. After that, an unsupervised classifier is used to classify sleep states.

Sleep State Classification: Adaptive Sensor Selection

[0084] FIG. 9 illustrates an embodiment of a method for adaptively selecting one or more sensors for performing sleep state classification. This method maybe performed without detecting beforehand the presence or absence of the sleep subject, or may be performed only after the presence of a sleep subject is detected. As with other embodiments described herein, the signals may processed in an on going and continuous manner during the observation period (e.g., in real-time or near real-time) or may be processed after the observation period is over.

[0085] In operation 910, signals are received from all sensors of the sleep mat. However, in accordance with one embodiment, only a certain number of sensors are selected. In order to achieve improved sleep state classification, body movements and cardiorespiratory activity may be determined and optimally characterized. Thus, two sensors may be selected, one optimized, sensitive to, or tuned for body movements and one optimized, sensitive to, or tuned for cardiorespiratory activity. In the persent embodiment, the sleep mat has four sensors, two with low amplification gain (sensor 1 or 3) and two with high amplification gain (sensor 2 or 4). Thus, one low gain sensor and one high gain sensor may be selected. [0086] In operation 920, specific DQI metrics are computed for the high gain sensors based on features listed, for example, in Table 2.

Table 2

[0087] For the high gain sensors, DQI_high metrics may be computed based on Equation 3.

DQI_high— min(std(Flx)), (3) where Fix is a series of continuous values x of a first feature FI during that period and std is standard deviation. Here, the feature spectral hf from Table 2 was used because it is expected that a smaller variation in the HF spectral power of the BCG signal indicates less influence from body movements or other noise. In another embodiment, DQI high may be calculated using an equation different from Equation (3).

[0088] In operation 930, one of the high gain sensors is selected based on the values of the DQI metric using Equation 4, wherein k is the number of high gain sensors in the mat (in this case k = 2).

S_high = argmaxi (DQI_high DQI_high ₂ , . . . , DQI_higlu) (4)

[0089] In operation 940, the values of the DQI metrics are computed for the low gain sensors in the mat based on another feature in Table 2. For the low gain sensors, values of the DQI_low metrics may be computed based on Equation 5.

DQI_low— min (mean (F2x)), (5) where F2x is a series of continuous values x from a second feature F2 during that period and mean represents an average. Flere, the DQI metrics for the low gain sensors may be computed based on the spectral spread feature from Table 2, because this feature is expected to have less spread over of spectral power in the frequency domain of the BCG signal. Low gain sensors with higher DQI metrics may be expected to be less noisy and thus produce clearer signals than low gain sensors with lower DQI metrics. For example, low gain sensors with higher values of DQI metrics may be considered more informative with regard to physiological components resulting from physiological activity. In another embodiment, DQI low may be calculated using an equation different from Equation (5).

[0090] In operation 950, one of the low gain sensors is selected based on the values of the DQI metrics using Equation 6, wherein k is the number of high gain sensors in the mat (in this case k = 2).

S_low = argmaxi (DQI_low _l5 DQI_low ₂ , ... , DQI_low*) (6)

Thus, the present embodiments select one high gain sensor and one low gain sensor optimized for different features FI and F2 and for different conditions, e.g., body movements and cardiorespiratory activity, for purposes of classifying sleep state(s) during an observation period.

[0091] FIG. 10 illustrates an example arrangement of the four sensors in FIG. 1 set to improve or optimize detection of physical activity (e.g., body movements) and physiological activity (e.g., heart rate variability and respiration) simultaneously, and thus to better determine the sleep states, of the infant during the observation period. In one embodiment, the sensors 120 of the sleep mat 110 may be designed differently. For example, the amplification factor (gain) of all or a portion of the sensors may be optimized or set to predetermined levels in order to capture both body movements and cardiorespiratory activity. In the example illustrated in FIG. 10, sensors 1 and 3 are designed with a relatively low gain (compare with sensors 2 and 4) to better characterize body movements. Sensors 2 and 4 are designed with relatively high gain (compared with sensors 1 and 3) to better characterize physiological information (e.g., cardiorespiratory activity).

Sleep State Classification: Feature Extraction

[0092] FIG. 11 illustrates a method for classifying the sleep state(s) of a subject (e.g,. infant) based on sensor signals output from the sleep mat. In operation 1110, the presence of the infant is detected on the mat placed in the bed, crib, or incubator as described herein. Also, in operation 1120, one low amplification gain sensor (e.g., sensor 1 or 3) and one high amplification gain sensor (e.g., sensor 2 or 4) are selected to perform the sleep state classification, as previously described.

[0093] In operation 1130, one or more features may be extracted from the signals output from the selected sensors. These features may be ones in Table 2 and may be the same or different from one(s) used to detect the presence or absence of the infant. For example, the one or more features extracted from the selected sensor signals may characterize body movements and cardiorespiratory activity of the infant due to regulation of autonomic nervous system. In one embodiment, a plurality of features may be selected for performing sleep state classification, e.g., a total of 22 features may selected and calcualted: 12 features calculated from a high gain sensor and 10 features calculated from a low gain sensor. Only one high gain sensor and only one low gain sensor may be selected for each recording based on adaptive sensor selection, according to one embodiment.

[0094] The features for performing sleep state classification may be calculated based on BCG signals and/or based on respiratory signals (BCG-derived respiratory (BDR) signals). BDR signals may be obtained, for example, by filtering BCG signals with a 5*-order Chebyshev filter at a passband of 0.25Hz to lHz. The BCG signals and BDR signals from the sensors for each recording may be normalized to have zero mean and unit variance. [0095] In operation 1140, the sleep state(s) occurring during the observation period are classified based on the extracted features. For BDR signals, peaks and troughs may be detected to extract the features of interest, and the values or amplitudes of the peaks and troughs may be compared to predetermined thresholds to indicate the sleep state(s) in the presence period. The classification may be performed using, for example, any of the machine learning algorithms previously described.

Sleep State Classification Algorithm: Example

[0096] FIG. 12 illustrates a detailed example embodiment of a method for determining the presence or absence of a sleep subject and then classifying sleep stages of the subject during an observation period based on signals from the sleep mat 110. The sensor maybe, for example, a predetermined one of the pressure sensors 120 in the sleep mat, but in another embodiment may be any of the other types of sensors mentioned herein.

[0097] In operation 1210, a recording session (observation period) is initiated for purposes of receiving signals from all the sensors 120 of the sleep mat 110. The length of the recording session may be automatically set or set based manual turn on/ turn off operations performed by a user. In this example, the setting may be for a 24-hour period. Signals output from one or more of the sensors are illustrated in FIGS. 13A to 13C, which plot the sensor signals against time, where the leftmost point on the horizontal axis corresponds to midnight for Day 1 and the rightmost point corresponds to midnight for Day 2. The vertical axis indicates the frequency of the sensor signals.

[0098] In operation 1220, a power spectrum is generated based on one or more of the sensor signals, as illustrated in FIG. 4A. In this figure, the strongest power signals 1310 occur from midnight of Day 1 to morning hours, with intermittent low- or no-amplitude periods corresponding, for example, to times when the sleep subject (in this case an infant) is removed from the incubator or crib. Tighter shade areas 1320 of the power signals represent values of power that fall within a predetermined frequency range. The predetermined frequency range may include, for example, a vital sign range.

[0099] In operation 1230, the power signals in FIG. 13A are summed in a predetermined spectral range, which in this case is between 0 Hz and 7 Hz to produce power curve 1330 in FIG. 13B. The power curve is plotted against time epochs indicated on the horizontal axis.

[00100] In operation 1240, a threshold 1340 is applied to the power curve 1330. The threshold 1340 may be determined beforehand, for example, based on data corresponding to multiple babies observed over multiple nights for the particular sensor type(s) used by the sleep mat. In this example, the threshold 1340 is set as an infant presence/absence threshold at a value (vertical axis) of minus 320.

[00101] In operation 1250, the presence or absence of the sleep subject (infant) is determined at various times throughout the 24-hour observation period based on the applied threshold 1340. Values of the power curve 1330 below the threshold indicate times when the infant is absent from the incubator or crib, e.g., not on the sleep mat). Values of the power curve 1330 above the threshold indicate presence of the infant in the incubator or crib, e.g., on the sleep mat.

[00102] In operation 1260, an unsupervised sleep state classification algorithm is performed based on the values generated by applying the threshold 1340 to the power curve 1330. As illustrated in FIG. 5C, a curve 1350 is generated which shows a plurality of stepped areas 1360 throughout the observation period. The stepped areas correspond to one of four values. A value of 0 indicates an absence of the sleep from the sleep mat. A value of 1 indicates an awake state. A value of 2 indicates a quiet sleep state. And, a value of 3 indicates an active sleep state.

[00103] Once the sensor(s) are selected for presence/absence detection and sleep state classification, various features may be used to train a classifier using machine learning techniques. Once trained, the classifier may detect the presence/absence of a sleep subject during an observation period and then may classify sleep state(s) during that period using features computed from new incoming data. Validation

[00104] To examine the validity of the processing performed by the embodiments described herein, BCG signals acquired from a four-sensor sleep mat were analyzed from five term infants aged less than 1.5 years. The BCG signals were collected for 10 nocturnal sleep episodes and 20 daytime naps. An TD classifier was used in combination with leave-one-subject-out cross validation to test the method. Overall accuracy and Cohen’s Kappa coefficient of agreement (mean ± standard deviation over subjects) were used to evaluate classification and detection performance.

[00105] Table 3 shows infant presence/absence detection results generated for validation. In generating these results, the feature used for sensor selection was spectral vlf and the features used for presence and absence detection were those listed in Table 1. For comparison purposes, Table shows detection results using each of the four sensors and a combination of all four sensors. As expected, the adaptive sensor selection corresponds to the best results.

Table 3

[00106] Table 4 shows infant sleep state classification results generated for validation. For sleep state classification, the features used to sensor selection were spectral _hf and spectral spread for DQI_high and DQI_low, respectively. For comparison, Table 4 presents the classification results using each of the sensors and the combination of all sensors. As expected, the adaptive sensor selection corresponds to the best results.

Table 4

[00107] In accordance with one or more of the aforementioned embodiments, a system and method is provided for performing presence/ absence detection and sleep state classification based on adaptive sensor selection methods. The adaptive sensor selection methods may select one or more sensors of the mat that demonstrate a signal quality of a predetermined level, e.g,. a signal quality considered better than the quality produced from other sensors. The signal quality may be determined, for example, based on whether features of the sensor signals fall within one or more predetermined ranges. The sensor(s) selected by the method maybe considered, for example, to be the one(s) having the highest data/ signal quality for classifying the presence or absence of the sleep subject, relative to the other sensors.

[00108] In one embodiment, one or more sensors 120 may be designed with a higher amplification factor (gain) to better characterize a first type of information, e.g., physiological information. One or more other sensors may be designed with a lower amplification factor (gain) to better capture a second type of information, e.g., physical activity (e.g., gross body movements). In one embodiment, the adaptive sensor selection method may choose the sensors (e.g., one from high gain sensors and one from low gain sensors) that are producing signals that fall within predetermined ranges or demonstrate a certain or best signal quality, so that the features derived from the signals from these sensor have the highest data/ signal quality for classifying sleep states (e.g., wake, active sleep, and quiet sleep), compared with the other sensors.

[00109] The adaptive sensor selection method may perform sensor selection recurrently and in a manner adapted to a certain period of time for each sleep subject (infant), where the period length (e.g., time interval) is pre-defmed. The period length may be, for example, one day (24 hrs), half day (e.g. daytime or night-time), one recording (e.g., depending on the switching ON and OFF of the device), or every epoch of time (e.g., 30s).

[00110] Referring to FIG. 2, the processor 210 may calculate various features based on the signals generated by the selected sensors 120. In one embodiment, the processor 210 may calculate one or more data quality index (DQI) metrics. The DQI metrics may be compared and or otherwise analyzed by a machine learning algorithm to select the best performing sensor for each period. The DQI metrics may be specific for different tasks, including detection of the presence/absence of the sleep subject and sleep state classification.

[00111] The DQI metric(s) may quantify the signal or feature quality. In one embodiment, the DQI metrics may be computed based on a combination of multiple features and/ or based on raw sensor signals or derived respiratory/ cardiac signals using, for example, signal-to-noise ratio, signal entropy, or another measure. In addition to an adaptive sensor selection algorithm, memory 220 may also store a machine learning algorithm that may be used to train a classifier, for example, based on features extracted from the selected sensor signals. The classifier may be used to detect the presence or absence of the sleep subject and/or to classify sleep states of the subject. The results of the algorithms maybe smoothed using, for example, a low-pass filter. Various techniques may be used to perform the smoothing, e.g., moving average, spline fitting, or another technique.

[00112] Additionally, while one or more features of the embodiments may involve the use of a mathematical formula, the embodiments are in no way restricted solely to a mathematical formula. Nor are they directed to a method of organizing human activity or a mental process. Rather, the complex and specific approach taken by the embodiments, combined with the amount of information processing performed, negate the possibility of the embodiments being performed by human activity or a mental process. Moreover, while a computer or other form of processor may be used to implement one or more features of the embodiments, the embodiments are not solely directed to using a computer as a tool to otherwise perform a process that was previously performed manually.

[00113] The methods, processes, and/or operations described herein may be performed by code or instructions to be executed by a computer, processor, controller, or other signal processing device. The code or instructions may be stored in a non-transitory computer-readable medium in accordance with one or more embodiments. Because the algorithms that form the basis of the methods (or operations of the computer, processor, controller, or other signal processing device) are described in detail, the code or instructions for implementing the operations of the method embodiments may transform the computer, processor, controller, or other signal processing device into a special-purpose processor for performing the methods herein.

[00114] The modules, stages, models, algorithms, processors, and other information generating, processing, and calculating features of the embodiments disclosed herein maybe implemented in logic which, for example, may include hardware, software, or both. When implemented at least partially in hardware, the modules, models, algorithms, processors, and other information generating, processing, or calculating features may be, for example, any one of a variety of integrated circuits including but not limited to an application-specific integrated circuit, a field-programmable gate array, a combination of logic gates, a system-on-chip, a microprocessor, or another type of processing or control circuit.

[00115] When implemented in at least partially in software, the modules, models, algorithms, processors, and other information generating, processing, or calculating features may include, for example, a memory or other storage device for storing code or instructions to be executed, for example, by a computer, processor, microprocessor, controller, or other signal processing device. Because the algorithms that form the basis of the methods (or operations of the computer, processor, microprocessor, controller, or other signal processing device) are described in detail, the code or instructions for implementing the operations of the method embodiments may transform the computer, processor, controller, or other signal processing device into a special-purpose processor for performing the methods herein.

[00116] It should be apparent from the foregoing description that various exemplary embodiments of the invention may be implemented in hardware. Furthermore, various exemplary embodiments may be implemented as instructions stored on a non-transitory machine-readable storage medium, such as a volatile or non-volatile memory, which may be read and executed by at least one processor to perform the operations described in detail herein. A non-transitory machine-readable storage medium may include any mechanism for storing information in a form readable by a machine, such as a personal or laptop computer, a server, or other computing device. Thus, a non-transitory machine- readable storage medium may include read-only memory (ROM), random-access memory (RAM), magnetic disk storage media, optical storage media, flash-memory devices, and similar storage media and excludes transitory signals.

[00117] It should be appreciated by those skilled in the art that any blocks and block diagrams herein represent conceptual views of illustrative circuitry embodying the principles of the invention. Implementation of particular blocks can vary while they can be implemented in the hardware or software domain without limiting the scope of the invention. Similarly, it will be appreciated that any flow charts, flow diagrams, state transition diagrams, pseudo code, and the like represent various processes which may be substantially represented in machine readable media and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.

[00118] Accordingly, it is to be understood that the above description is intended to be illustrative and not restrictive. Many embodiments and applications other than the examples provided would be apparent upon reading the above description. The scope should be determined, not with reference to the above description or Abstract below, but should instead be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled. It is anticipated and intended that future developments will occur in the technologies discussed herein, and that the disclosed systems and methods will be incorporated into such future embodiments. In sum, it should be understood that the application is capable of modification and variation.

[00119] The benefits, advantages, solutions to problems, and any element(s) that may cause any benefit, advantage, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential features or elements of any or all the claims. The invention is defined solely by the appended claims including any amendments made during the pendency of this application and all equivalents of those claims as issued.

[00120] All terms used in the claims are intended to be given their broadest reasonable constructions and their ordinary meanings as understood by those knowledgeable in the technologies described herein unless an explicit indication to the contrary in made herein. In particular, use of the singular articles such as“a,”“the,”“said,” etc. should be read to recite one or more of the indicated elements unless a claim recites an explicit limitation to the contrary.

[00121] The Abstract of the Disclosure is provided to allow the reader to quickly ascertain the nature of the technical disclosure. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims. In addition, in the foregoing Detailed Description, it can be seen that various features are grouped together in various embodiments for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the claimed embodiments require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Thus the following claims are hereby incorporated into the Detailed Description, with each claim standing on its own as a separately claimed subject matter.