Certain aspects generally pertain to determining stroke risk.
Stroke is a global health concern, with a distressingly high incidence, and substantial morbidity and mortality. Annually, more than ten million people worldwide are impacted by strokes, imposing a heavy toll on affected individuals and their families, with significant health, financial, and quality of life burdens. In the United States alone, strokes affect nearly 800,000 individuals each year. Identifying patients likely to experience stroke may aid in prescribing and/or monitoring lifestyle changes or medications that decrease likelihood of stroke. However, identifying patients likely to experience stroke or determining a patient's stroke risk is difficult or impossible.
Techniques disclosed herein may be practiced with a processor-implemented method, a system comprising one or more processors and one or more processor-readable media, and/or one or more non-transitory processor-readable media.
According to some embodiments, a method for determining stoke risk may involve causing, using one or more light sources disposed on a headset worn by a user, light to be emitted into the head of the user. The method may further involve obtaining, using one or more light detectors disposed on the headset, information indicative of light reflected from one more structures within the head of the user, wherein a portion of the obtained information spans a time period during which the user was holding their breath. The method may further involve based on the obtained information, determining one or more cerebral blood metrics. The method may further involve providing a representation of the one or more cerebral blood metrics as input to a trained machine learning model. The method may further involve determining a likelihood the user will experience a stroke over a predetermined future time period based on an output of the trained machine learning model.
According to some embodiments, a system for determining stroke risk may comprise: a headband configured to encircle a head of a wearer of a headset; a plurality of light sources attached to the headband; a plurality of light detectors attached to the headband; and one or more processors. The one or more processors may be configured to: cause, using the one or more light sources, light to be emitted into the head of the wearer; obtain, using the one or more light detectors, information indicative of light reflected from one more structures within the head of the wearer, wherein a portion of the obtained information spans a time period during which the wearer was holding their breath; based on the obtained information, determine one or more cerebral blood metrics; and determine a likelihood the wearer will experience a stroke over a predetermined future time period based on the one or more cerebral blood metrics.
According to some embodiments, a method for determining stroke risk may involve causing, using one or more light sources disposed on a headset worn by a user, light to be emitted into the head of the user. The method may further involve obtaining, using one or more light detectors disposed on the headset, information indicative of light reflected from one more structures within the head of the user, wherein a portion of the obtained information spans a time period during which the user was holding their breath. The method may further involve based on the obtained information, determining a cerebral blood flow as a function of time and a cerebral blood volume as a function of time, wherein the cerebral blood flow and the cerebral blood volume include the time period during which the user was holding their breath and a baseline time period before the user was holding their breath. The method may further involve determining a likelihood the user will experience a stroke over a predetermined future time period based on the cerebral blood flow and the cerebral blood volume.
According to some embodiments, a method of training a machine learning model to predict stroke risk may involve obtaining training data, the training data comprising, for a group of users, representations of cerebral blood metrics, wherein for each training sample, a portion of the obtained data spans a time period during which the user was holding their breath, and wherein each training sample includes a corresponding ground truth stroke risk for the user. The method may further involve providing the training data to a machine learning model, wherein the machine learning model takes, as input, the representations of the cerebral blood metrics and generates, as an output, a prediction of stroke risk. The method may further involve updating the machine learning model based on differences between the ground truth stroke risk and the predicted stroke risk to generate a trained machine learning model configured to predict stroke risk.
These and other features are described in more detail below with reference to the associated drawings.
These and other features are described in more detail below with reference to the associated drawings.
Different aspects are described below with reference to the accompanying drawings. The features illustrated in the drawings may not be to scale. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the presented embodiments. The disclosed embodiments may be practiced without one or more of these specific details. In other instances, well-known operations have not been described in detail to avoid unnecessarily obscuring the disclosed embodiments. While the disclosed embodiments will be described in conjunction with the specific embodiments, it will be understood that it is not intended to limit the disclosed embodiments.
Stroke is a global health concern, with a distressingly high incidence, and substantial morbidity and mortality. Annually, more than ten million people worldwide are impacted by strokes, imposing a heavy toll on affected individuals and their families, with significant health, financial, and quality of life burdens. In the United States alone, strokes affect nearly 800,000 individuals each year. Identifying patients likely to experience stroke may aid in prescribing and/or monitoring lifestyle changes or medications that decrease likelihood of stroke. However, identifying patients likely to experience stroke or determining a patient's stroke risk is difficult or impossible using conventional techniques. In particular, conventional techniques typically utilize a questionnaire that evaluates health factors, such as hypertension, cholesterol levels, diabetes, and atrial fibrillation, etc. along with lifestyle-related factors such as body mass index, exposure to air pollution, smoking, unhealthy diet, alcohol consumption, and low physical activity, and the like. These risk profiles are based on population-level data and may be useful for a general assessment. However, such questionnaires are not definitive for determining the need for invasive or non-invasive evaluations, nor for guiding surgical or pharmacological interventions. For example, such questionnaires may give rise to a stroke risk prediction that is too general, and therefor, inaccurate. Moreover, use of such questionnaires may not accurately and precisely indicate whether various interventions (whether surgical, pharmacological, or behavioral) have modified a given patient's stroke risk, e.g., to determine whether a given intervention has lowered the patient's risk of stroke.
Disclosed herein are techniques for utilizing physiological metrics to predict stroke risk. In particular, the techniques disclosed herein utilize cerebral perfusion metrics, which may include cerebral blood volume (CBV), cerebral blood flow (CBF), and/or cerebral blood oxygenation (CBO), to predict stroke risk. The cerebral perfusion metrics, sometimes referred to herein as “cerebral blood metrics,” may be obtained during a period of time that includes time a patient is holding their breath. Changes in the cerebral blood metrics during the breath-holding time period relative to a baseline time period, and/or changes in the cerebral blood metrics in the time period after the breath-holding time period may be utilized to predict stroke risk during a future time period. Because patients with a relatively high stroke risk have stiffer blood vessels, their blood vessels may be relatively harder to dilate. Accordingly, patients with a relatively high stroke risk may be observed to have relatively larger increases in cerebral blood flow and relatively lower changes in cerebral blood volume relative to patients with a low stroke risk. Other changes, such as slope of change in cerebral blood metrics, timing characteristics of the return to baseline of cerebral blood metrics after breath-holding terminates, etc. may be characterized. In some embodiments, a trained machine learning model may be provided with either extracted features of cerebral blood flow, or with raw cerebral blood flow metrics data, and may generate a stroke prediction risk based on the cerebral blood flow metrics.
Cerebral blood metrics may be determined using data characterizing light reflected from various structures with a brain of a patient. For example, in some implementations, cerebral blood volume and/or cerebral blood flow metrics may be determined by transmitting light in an infrared wavelength range into the brain of the patient and measuring reflected light. In some implementations, cerebral blood flow may be determined using speckle contrast optical spectroscopy (SCOS), described below in more detail in connection with
In some implementations, light may be transmitted into the brain of the patient using a light emitter, which may be a laser, a light emitting diode (LED), or the like. Scattered light may be captured using a light detector, which may be a camera, a photodetector, a single-photon avalanche diode (SPAD), a SPAD array, or the like. In instances in which cerebral blood oxygenation metrics are determined and therefore, in which light is transmitted at two wavelengths, two light sources (e.g., a laser and an LED) may be used to transmit light, and two detectors (e.g., a camera and a photodetector) may be used to capture reflected light. In some embodiments, two (or more) light sources may be packaged as one light emission package, and two (or more) detectors may be packaged as one light detection package. In some implementations, light sources and/or light detectors may be disposed on a headband configured to encircle the head of the patient (e.g., around the forehead). In some implementations, light sources and light detectors (or light emission packages and light detection packages) may be affixed to the headband at different locations. For example, the light sources and/or light emission packages may be affixed to the headband at positions corresponding to a forehead of the patient, the parietal lobe of the patient, the frontal lobe of the patient, etc. In some implementations, a light source and light detector may be fiber free in that there is no fiber coupling to either the light source or the light detector. Accordingly, all components of the headset may be head-mounted, which may reduce noise emanating from optical fibers' movement. In some implementations, the laser and the camera/detector may be placed directly atop the user's skin. An example of a headset is shown in and described below in connection with
Stroke likelihood data 104 may be a number on a discrete scale (e.g., an integer between 1 and 5), a number on a continuous scale (e.g., a probability value that is a continuous number between 0 and 1), or the like. Stroke likelihood data 104 may represent a likelihood that, given the cerebral blood metrics represented in breath-holding data 104, the patient will have a stroke within a predetermined future time period (e.g., within the next year, within the next five years, within the lifetime of the patient, etc.).
It should be noted that stroke prediction engine 102 may be implemented by one or more computing devices or one or more processors. For example, such a computing device and/or processor may be configured to analyze data from one or more light detectors, generate cerebral blood flow metrics, and/or provide data representative of the cerebral blood flow metrics to a trained machine learning model to generate stroke likelihood data 106. In some implementations, the one or more computing device and/or one or more processors may be disposed on the same headset or headband as the one or more light emitters and one or more light detectors. Additionally, or alternatively, the one or more computing devices and/or one or more processors may be communicatively coupled to the headset or headband, e.g., by a wireless or wired communication channel.
As described above, one or more light emitters and one or more light detectors may be disposed on a head-worn device by a patient. The light may be emitted toward and/or into the head of the patient, thereby probing cerebral blood flow. Reflected light may be captured by the one or more light detectors to assess and/or characterize absorbed and reflected light by structures internal to the patient's brain. In some implementations, a “channel” may be characterized by a light emission package and a corresponding light detector package. In some embodiments, a light emission package may include multiple light emitters, such as a laser and an LED. For example, a laser may emit light in a near infrared or infrared wavelength, and an LED may emit light in a visible, near infrared, or infrared wavelength. Use of multiple wavelengths may allow for cerebral blood oxygenation to be determined. As described below, in some cases, light emitted in an infrared wavelength (e.g., by a laser) may be used to determine cerebral blood volume and/or cerebral blood flow, and light emitted at two different wavelengths (e.g., light emitted in an infrared wavelength and light emitted in a near infrared wavelength) may be used to determined cerebral blood oxygenation. It should be understood that in instances in which two wavelengths of light are used to determined cerebral blood oxygenation, one of the wavelengths may also be used to determine cerebral blood volume and/or cerebral blood flow.
In one particular example, a laser may emit light in a near-infrared wavelength, and the reflected light may be captured by a camera. The reflected light in the near-infrared wavelength range may be used to determine cerebral blood flow and/or cerebral blood volume. Concurrently, an LED may emit light in another near infrared wavelength and the reflected light may be captured by a photodetector. The reflected light from both wavelengths may be used to determined cerebral blood oxygenation. Note that the combination of laser, LED, camera, and photodetector of this example may be considered one “channel,” and a headset may have multiple such channels (e.g., two, four, ten, etc.) disposed at various locations around the headset, each probing a different region of the brain.
In some embodiments, a light detection package may include multiple light detectors, such as multiple cameras, a camera and a photodetector, or the like. In some implementations, a distance between a light emitter and a corresponding light detector (or between a light emission package comprising two or more light emitters and a corresponding light detection package) may be within a range of about 2.5 cm-4.0 cm. This distance is generally referred to as the “source-detector distance,” and example values may include 2 cm, 2.5 cm, 3 cm, 3.5 cm, 4 cm, etc. In instances in which a headset includes multiple channels (each comprising at least one light emitter or at least one light emission package, and corresponding light detectors or light detection packages), the distance between light emitter and light detector for different channels may be different. In some implementations, positions of light emitters and/or light detectors (or a light emission package and or a light detection package) may be modifiable. For example, a light emission package and/or a light detection package may be affixed to a headband or a headset via screws, Velcro, or other hardware at a position that is modifiable. This may allow for distances between a light source and a light detector to be modified, which may in turn allow the depth of imaging to be modified as different source-detector distances may impact the depth the emitted light can penetrate within the brain, as shown in and described below in connection with
Referring to
Note that a headset or headband may include one or more channels (e.g., one, two, four, five, ten, etc.). For example, in some implementations, a headset or headband may include four channels. By way of example, the four channels may be configured to measure cerebral blood metrics on a forehead (or frontal lobe region), a parietal lobe region, or the like. In some embodiments, the channels may be symmetrically disposed with respect to one another. For example, a first channel may be disposed proximate the left parietal lobe, and a second channel may be disposed proximate the right parietal lobe.
In some implementations, cerebral blood metrics may include one or more of cerebral blood flow, cerebral blood volume, and/or cerebral blood oxygenation. In some embodiments, cerebral blood volume may be determined based on the intensity of reflected light as measured by a light detector. Note that cerebral blood volume may be determined based on reflected light of a single wavelength (e.g., emitted from a laser, from an LED, etc.). In some implementations, cerebral blood oxygenation may be determined based on the intensity of reflected light at two wavelengths (e.g., emitted by a laser and an LED), and determining oxygenation based on the differential optical transmission changes for the two different wavelengths. For example, the two wavelengths may be one where oxyhemoglobin absorbs more light and one where deoxyhemoglobin absorbs more light. The two wavelengths may be in the visible, near-infrared, or infrared wavelength range. In some embodiments, known formulas for calculating oxygen saturation (e.g., based on the Beer-Lambert Law, or other similar techniques) may be utilized to determine cerebral blood oxygenation metrics using two wavelengths.
In some embodiments, cerebral blood flow may be determined using collected scattered light at a single wavelength (e.g., light emitted in an infrared wavelength range). Note that wavelength choice may affect penetration depth of the light. For example, light emitted in the infrared wavelength range may penetrate to deeper structures relative to near infrared light. In some implementations, cerebral blood flow may be based on diffuse correlation spectroscopy (DCS), or on speckle visibility spectroscopy (SVS), which is also referred to as speckle contrast optical spectroscopy (SCOS). In general, a “speckle” refers to the pattern of bright and dark spots in an image resulting from scattering of illuminated laser light (e.g., scattered by the scalp, skull, and/or brain) resulting from constructive and destructive interference of the light. The speckle pattern dynamics change with blood flow dynamics. The time that it takes one speckle pattern to change to a different speckle pattern is referred to as the decorrelation time, which is correlated with cerebral blood flow rate. Both DCS and SCOS are techniques for measuring how fast the speckle pattern changes, and hence, estimating cerebral blood flow rate. DCS may determine changes in speckle decorrelation time based on a single pixel of the speckle image, i.e., requiring a fast photodetector to detect the light. In contrast, SCOS may determine speckle decorrelation time using a relatively slow camera with a large number of pixels. In particular, the camera exposure time, typically in the range of 0.3 milliseconds to 15 milliseconds, may be set to be substantially longer than the decorrelation time of the speckle field (typically a few tens of microseconds), which may result in multiple different speckle patterns summed up into a single camera frame. As the speckle field fluctuates, the speckle pattern recorded by the camera is smeared and washed out within the exposure time. Accordingly, the cerebral blood flow may be quantified from the degree of blurring of the captured frame, which is generally referred to herein as the speckle contrast. Note that use of SCOS to measure cerebral blood flow may have the advantage that a relatively inexpensive camera may be used as a light detector, because a high frame rate is not needed, unlike detectors used for DCS. Moreover, the camera may be directly mounted on a headset or headband, which may eliminate optical loss associated with a light guide running from the head to the camera, which may introduce its own noise and motion artifacts. In some examples, a camera used for SCOS may have an integration time of within a range of about 0.3 milliseconds-15 milliseconds. The camera may use a frame rate of within a range of about 30 frames per second-150 frames per second. In one example, the camera may have a frame rate of about 80 frames per second.
Speckle contrast may be determined at each camera frame as:
In the equation given above, Ij represents the instantaneous intensity recorded at the camera at pixel J, or represents the standard deviation of I, and Ī represents the mean of I. The speckle decorrelation time is determined as:
The speckle decorrelation time is directly correlated with the cerebral blood flow rate. The relative cerebral blood flow, measured in units of blood flow index (BFI), is inversely correlated with t and therefore inversely correlated with βK2. The correction factor β is a constant depending on system setup, e.g. speckle size, pixel size, polarization of the laser light.
Note that each time period may be any suitable duration of time. Example duration of breath holding may be 15 seconds, 30 seconds, 45 seconds, 60 seconds, etc. In some embodiments, the duration of breath holding may be however a long a patient can hold their breath relatively comfortably (e.g., the breath holding time period may vary for different people). In some implementations, the baseline time period and/or the recovery time period may be at least as long as the breath holding time period. In some embodiments, the recovery time period may be longer than the breath holding time period. In some embodiments, the recovery time period may be a time duration long enough that cerebral blood flow metrics return to within a predetermined range of the corresponding values during the baseline time period. In some such embodiments, the recovery time period may be dynamically adjusted. For example, the recovery time period may be stopped responsive to determined that cerebral blood metrics have returned to baseline values.
As illustrated in graphs 602 and 604, both cerebral blood flow and cerebral blood volume increase during breath holding time period 608. This is attributed to the brain's increased demand for blood to transport oxygen and carbon dioxide until breath holding stops. During the breath holding time period 608, the brain enters a heightened state of alert which triggers a sequence of protective mechanisms to ensure stable regulation of carbon dioxide and oxygen, which is achieved through accelerated circulation of blood leading to increased blood flow together with an increase in blood volume in the brain via dilation of blood vessels.
Additionally, note that both cerebral blood flow rates and cerebral blood volume remain elevated at the beginning of recovery time period 610 prior to recovery to the baseline level of each metric.
Plots 612 and 614 illustrate subsets of cerebral blood flow graph 602 over short time scales. Similarly, plots 616 and 618 illustrate subsets of cerebral blood volume graph 604 over short time scales. Note that the pulsations evident in each of plots 612, 614, 616, and 618 are not noise but represent blood pulsations. As illustrated in frequency domain graph 620, the pulsations may be used to determine a heart rate of the patient (e.g., based on the periodicity of the pulsations). For example, heart rate may be determined by taking a Fourier transform of time domain data. Additionally, note that cerebral blood flow and cerebral blood volume may capture different details in blood flow. For example, the dicrotic notch and peak pressure are discernible in plot 614 of the cerebral blood flow.
In some implementations, raw data traces of cerebral blood metrics (e.g., cerebral blood volume, cerebral blood flow, and/or cerebral blood oxygenation as a function of time) may be provided to a trained machine learning model configured to output a stroke likelihood prediction. Such a machine learning model that accepts raw data traces may be a DNN or other suitable architecture. Alternatively, in some embodiments, one or more features associated with the cerebral blood metrics may be extracted. The extracted features may then be provided to a trained machine learning model, which in turn may generate a predicted stroke likelihood. The extracted features may be considered “a representation of the one or more cerebral blood flow metrics.” In instances in which one or more extracted features are provided to a machine learning model, the model may be a perceptron, a random forest, or any other suitable architecture. Note that techniques for generating predicted stroke likelihoods using a machine learning model are shown in and described below in connection with
In some implementations, extracted features may include information a duration of time a patient was able to hold their breath (generally referred to herein as TBH).
In some implementations, extracted features may include percentage change in a cerebral blood metric at a maximum or minimum after initiation of breath holding compared to a baseline value. For example, referring to
In some implementations, the extracted features may include a rate of change (e.g., a slope) in the change in a cerebral blood metric during breath holding. For example, rate of change of cerebral blood flow may be determined by dividing a percentage change of cerebral blood flow (e.g., as indicated in
In some implementations, the extracted features may include a ratio of a percentage change of one cerebral blood metric to a percentage change of another cerebral blood metric. In one example, an extracted feature may be a ratio of the percentage change of cerebral blood flow (e.g., CBF change in
In some implementations, extracted features may include fine grained features from within a cardiac cycle as observed within a cerebral blood flow trace. For example, referring to
Note that extracted features may be extracted and/or determined autonomously (e.g., without user input) upon collection of the one or more cerebral blood flow metrics. The extracted features may then be provided to a trained machine learning model configured to generate the prediction of stroke likelihood for the patient. Note that any suitable number or combination of extracted features may be utilized.
Alternatively, in some embodiments, rather than utilizing a trained machine learning model, stroke risk may be predicted based on extracted features, e.g., by comparing an extracted feature to one or more predetermined thresholds. By way of example, stroke risk may be classified as “high,” responsive to determining that a value of a particular extracted feature (e.g., the percentage change in cerebral blood flow) exceeds a predetermined threshold. As another example, stroke risk may be classified as “high” responsive to determining that a value of a particular extracted feature (e.g., the percentage change in cerebral blood volume) is below a predetermined threshold. In one example, stroke likelihood may be determined based on a ratio of a percentage change in cerebral blood flow (between a peak cerebral blood flow value after the user begins holding their breath and a baseline cerebral blood flow to a percentage change in cerebral blood volume (between a peak cerebral blood volume after the user begins holding their breath and a baseline cerebral blood volume), as shown in and described below in connection with
In one example, extracted features may be based on a breath holding index (BHI), generally defined herein as the maximal change from baseline during breath-holding divided by the duration of breath-holding TBH for a given cerebral blood metric. For example, BHI for cerebral blood flow may be defined by:
Similarly, the BHI for cerebral blood volume may be defined by:
In some examples, the flow to volume ratio may be an extracted feature used to predict stroke risk. The flow to volume ratio may be determined based on the ratio of the BHI for cerebral blood flow to the BHI for cerebral blood volume:
In some cases, BHICBF may have a value within a range of about 0% to 5%, and BHICBV may have a value within a range of about 0% to 3%. The flow to volume ratio may have a value within a range of about 0.9 to 2. A threshold for categorizing a patient as having relatively high stroke risk may be based on the flow to volume ratio (e.g., the patient may be categorized as high risk if the flow to volume ratio is greater than 1.2, greater than 1.3, greater than 1.4, etc.)
Process 700 can begin at 702 by causing, using one or more light sources disposed on a headset worn by a user, light to be emitted into the head of the user. An example of such a headset is shown in and described above in connection with
At 704, process 700 may obtain, using one or more light detectors disposed on the headset, information indicative of light reflected from one or more structures within the head or brain of the user, wherein a portion of the obtained information spans a time period during which the user was holding their breath. An example of one or more light detectors disposed on a headset is shown in and described above in connection with
Note that, as described above in connection with
At 706, process 700 can, based on the obtained information, determine one or more cerebral blood metrics. As described above, the cerebral blood metrics may include cerebral blood volume, cerebral blood flow, and/or cerebral blood oxygenation. Process 700 may additionally determine a duration of time the user held their breath. The duration of time may be determined by, e.g., thresholding any of the cerebral blood metrics to determine a time point at which the cerebral blood metric began deviating from baseline to the time point at which the cerebral blood metric reached a minimum or maximum value. Note that, as described above, cerebral blood volume may be determined based on intensity of the reflected light signal at a single wavelength. Cerebral blood flow may be determined using DCS and/or SCOS (as shown in and described above in connection with
At 708, process 700 can provide a representation of the one or more cerebral blood metrics as input to a trained machine learning model. Note that the representation of the one or more cerebral blood metrics may include the raw data of the cerebral blood metrics as a function of time, or, alternatively, may include one or more extracted features extracted from the one or more cerebral blood metrics. Examples of extracted features are described above.
At 710, process 700 can determine a likelihood the user will experience a stroke over a predetermined future time period based on an output of the trained machine learning model. For example, the likelihood may correspond to a time period of the next year, the next five years, the next ten years, the remainder of their lifetime, etc. The likelihood may be provided as a discrete classification (e.g., low risk, medium risk, high risk), a continuous value (e.g., a continuous probability value), or in any other suitable format.
In some implementations, the stroke likelihood may be presented in any suitable manner. For example, the stroke likelihood may be audibly presented (e.g., by speakers associated with the headset or headband, or by speakers of an associated or paired device). The stroke likelihood may be visibly presented, e.g., on a display of an associated or paired device. In some implementations, the stroke likelihood may be automatically (e.g., without user input) stored in an electronic medical record associated with a patient, e.g., such that a physician can review the stroke likelihood. In cases in which the stroke likelihood is stored, e.g., as medical data, the stroke likelihood may be stored in conjunction with timestamp information indicating a date and/or time the stroke likelihood prediction was made. This may allow a physician or other healthcare provider to monitor changes in the stroke likelihood for a given patient over time. This may allow the healthcare provider to determine whether various interventions are modifying stroke likelihood (e.g., lowering the risk of stroke) over time. In some implementations, an updated stroke likelihood may be determined after a first stroke likelihood is determined (e.g., two months later, six months later, a year later, two years later, etc.) A difference between the first stroke likelihood and the updated stroke likelihood may be determined, e.g., to determine if the patient's likelihood of experiencing stroke remains the same over time, is increasing over time, or is decreasing over time. In some embodiments, a recommendation may be generated based on the change in stroke likelihood over time. For example, a recommendation to continue or implement particular lifestyle modifications may be made, a recommendation to initiate a particular medical treatment may be made, etc. In some embodiments, an indication of the change in stroke likelihood over time may be transmitted to a user device associated with the patient, with a healthcare provider, etc.
A machine learning model may be trained using a training set that includes representations of cerebral blood metrics (which may include cerebral blood flow, cerebral blood volume, and/or cerebral blood oxygenation as a function of time, or extracted features as discussed above), and corresponding ground truth data. The training set may include data associated with multiple patients, who may be of varying demographics (e.g., different lifestyles, different ages, different sexes, etc.). In some embodiments, ground truth data may be questionnaire data, where the questionnaire predicts a stroke likelihood based on demographic data, lifestyle data, physiological metrics (e.g., blood pressure, resting heart rate, etc.), and the like. Alternatively, in some embodiments, ground truth data may be actual stroke occurrence data, e.g., obtained from longitudinal data that collects representations of cerebral blood metrics from a set of patients who are followed over time. Continuing with this example, the ground truth data may indicate that the patient did not have a stroke if no stroke was recorded during the duration of time the longitudinal study occurred, or, conversely, may indicate that the patient did have a stroke if such stroke occurred. Regardless of how ground truth data is obtained, a machine learning model may be trained by providing representations of cerebral blood flow metrics as input, obtaining a stroke risk prediction based on the input, and updating weights of the model based on a difference between the predicted stroke risk and the ground truth stroke risk. This procedure may be implemented to train a DNN that operates on traces of cerebral blood metrics as a function of time, and/or a perceptron, random forest, or other type of network that operates on extracted features associated with cerebral blood metrics or associated with aspects of the breath holding task (such as duration of breath holding).
Process 800 can begin at 802 by obtaining training data, the training data comprising, for a group of users, representations of cerebral blood metrics, wherein for each training sample, a portion of the obtained data spans a time period during which the user was holding their breath, and where each training sample includes a corresponding ground truth stroke risk for the user. As described above, the representations of cerebral blood metrics may include one or more cerebral blood metrics (e.g., cerebral blood flow, cerebral blood volume, and/or cerebral blood oxygenation) as a function of time, and/or extracted features associated with the cerebral blood metrics and/or the breath holding task. Note that cerebral blood metrics for the training data may have been collected using one or more light sources and/or one or more light detectors disposed on a headset or headband similar to the one shown in and described above in connection with
At 804, process 800 can provide the training data to a machine learning model, where the machine learning model takes, as input, the representations of the cerebral blood metrics and generates, as an output, a prediction of stroke risk.
At 806, process 800 can update the machine learning model based on differences between the ground truth stroke risk and the predicted stroke risk. For example, weights of the model may be updated based on a loss function that considers the difference between the ground truth stroke risk and the predicted stroke risk. Note that model updating may be performed for each training sample, or for a batch of training samples.
Note that data (e.g., cerebral blood metric data) included in the training set may be obtained using a headset or head-worn device, and the trained machine learning model resulting from such a training set may be used to generate stroke risk predictions based on cerebral blood metrics obtained using the same headset or head-worn device, or a different but similar headset or head-worn device. For example, a similar device may be one in which light emitted by one or more light emitters is at about the same wavelength, where source-detector distance is about the same, or the like.
As illustrated in
Conversely, as shown by curves 906 and 908, the cerebral blood volume for the low risk group peaks at a higher level than for the high risk group. The difference in peaks in cerebral blood flow and cerebral blood volume (i.e., the percentage change in cerebral blood flow from baseline and the percentage change in cerebral blood volume from baseline) were found to be significant across the low risk and high risk groups, and accordingly, may be strong predictors of stroke risk. Additionally, note that because the peak change is in opposite directions for cerebral blood flow versus cerebral blood volume (in other words, the low risk group had a higher cerebral blood volume and a lower cerebral blood flow, and vice versa for the high risk group), the ratio of the percentage change of cerebral blood flow to percentage change of cerebral blood volume was found to be a highly significant predictor of stroke risk.
The techniques described above may be implemented using one or more computing devices. For example, a machine learning model may be trained and/or utilized at inference time using a computational device such as a server device, a laptop computer, a desktop computer, or the like.
In
Processor(s) 1060 is in electronic communication with CRM 1070 (e.g., memory). Processor(s) 1060 is also in electronic communication with display(s) 1080, e.g., to display image data, text, etc. on display 1080.
Processor(s) 1060 may retrieve and execute instructions stored on the CRM 1070 to perform one or more functions described above. For example, processor(s) 1060 may execute instructions to perform one or more operations to analyze collected data (e.g., light reflection/absorption data), provide collected data to a trained model, train a machine learning model to generate a stroke risk prediction, etc.
The CRM (e.g., memory) 1070 can store instructions for performing one or more functions of the described above. These instructions may be executable by processor(s) 1070. CRM 1070 can also store raw images, e.g., speckle images, or the like.
Embodiment 1: A method of determining stroke risk, the method comprising: causing, using one or more light sources disposed on a headset worn by a user, light to be emitted into the head of the user; obtaining, using one or more light detectors disposed on the headset, information indicative of light reflected from one more structures within the head of the user, wherein a portion of the obtained information spans a time period during which the user was holding their breath; based on the obtained information, determining one or more cerebral blood metrics; providing a representation of the one or more cerebral blood metrics as input to a trained machine learning model; and determining a likelihood the user will experience a stroke over a predetermined future time period based on an output of the trained machine learning model.
Embodiment 2: The method of embodiment 1, further comprising: determining an updated likelihood the user will experience a stroke based on updated cerebral blood metrics; and determining a change between the likelihood and the updated likelihood.
Embodiment 3: The method of embodiment 2, further comprising providing at least one recommendation based on the change between the likelihood and the updated likelihood.
Embodiment 4: The method of any one of embodiments 1-3, wherein the one or more light sources comprise at least two light sources, each configured to emit light in a different wavelength.
Embodiment 5: The method of claim 4, wherein the at least two light sources comprise a laser and a light emitting diode packaged together.
Embodiment 6: The method of any one of embodiments 4 or 5, wherein the at least two light sources comprise a laser configured to emit light in an infrared wavelength range.
Embodiment 7: The method of any one of embodiments 4-6, wherein the at least two light sources comprise a light emitting diode configured to emit light in a near-infrared wavelength range.
Embodiment 8: The method of any one of embodiments 1-7, wherein the one or more light detectors comprise at least two detectors of different types.
Embodiment 9: The method of embodiment 8, wherein the at least two detectors of different types comprise at least one camera.
Embodiment 10: The method of embodiment 9, wherein the at least one camera is configured to capture images comprising a speckle pattern.
Embodiment 11: The method of any one of embodiments 1-10, wherein the one or more cerebral blood metrics comprise at least one of: cerebral blood flow, cerebral blood oxygenation, or cerebral blood volume.
Embodiment 12: The method of embodiment 11, wherein the cerebral blood flow is determined based on a decorrelation time associated with a series of speckle patterns obtained from a series of images captured by at least one camera included in the one or more light detectors.
Embodiment 13: The method of any one of embodiments 11 or 12, wherein the representation of the obtained information comprises a ratio of change in cerebral blood flow during the time period the user was holding their breath to a baseline period to a change in cerebral blood volume during the time period the user was holding their breath to a baseline period.
Embodiment 14: The method of any one of embodiments 1-13, wherein providing the representation of the one or more cerebral blood metrics as input comprises providing raw traces of the one or more cerebral blood metrics as a function of time to the trained machine learning model, and wherein the trained machine learning model is a deep neural network (DNN).
Embodiment 15: The method of any one of embodiments 1-14, wherein providing the representation of the one or more cerebral blood metrics as input comprises providing features of the one or more cerebral blood metrics as an input to trained machine learning model.
Embodiment 16: A system for determining stroke risk, the system comprising: a headband configured to encircle a head of a wearer of a headset; a plurality of light sources attached to the headband; a plurality of light detectors attached to the headband; and one or more processors. The one or more processors may be configured to: cause, using the one or more light sources, light to be emitted into the head of the wearer; obtain, using the one or more light detectors, information indicative of light reflected from one more structures within the head of the wearer, wherein a portion of the obtained information spans a time period during which the wearer was holding their breath; based on the obtained information, determine one or more cerebral blood metrics; determine a likelihood the wearer will experience a stroke over a predetermined future time period based on the one or more cerebral blood metrics.
Embodiment 17: The system of embodiment 16, wherein the plurality of light sources comprise a plurality of light emission packages, each light emission package comprising at least two light sources configured to emit light in different wavelengths.
Embodiment 18: The system of embodiment 17, wherein the at least two light sources comprise a laser configured to emit light in an infrared wavelength range, and a light emitting diode (LED) configured to emit light in a near infrared wavelength range.
Embodiment 19: The system of any one of embodiments 16-18, wherein the plurality of light detectors comprise a plurality of light detection packages, each light detection package comprising at least two light detectors.
Embodiment 20: The system of embodiment 19, wherein a light detector of the at least two light detectors of a light detection package comprises a camera.
Embodiment 21: The system of embodiment 20, wherein the obtained information comprises a speckle pattern obtained using images captured by the camera.
Embodiment 22: The system of embodiment 21, wherein the one or more cerebral blood metrics comprise a cerebral blood flow determined based on the speckle pattern.
Embodiment 23: The system of any one of embodiments 16-22, wherein a distance between a light source of the one or more light sources and a light detector of the one or more light detectors is adjustable by changing a position of the light source and/or a light detector on the headband.
Embodiment 24: A method of determining stroke risk, the method comprising: causing, using one or more light sources disposed on a headset worn by a user, light to be emitted into the head of the user; obtaining, using one or more light detectors disposed on the headset, information indicative of light reflected from one more structures within the head of the user, wherein a portion of the obtained information spans a time period during which the user was holding their breath; based on the obtained information, determining a cerebral blood flow as a function of time and a cerebral blood volume as a function of time, wherein the cerebral blood flow and the cerebral blood volume include the time period during which the user was holding their breath and a baseline time period before the user was holding their breath; and determining a likelihood the user will experience a stroke over a predetermined future time period based on the cerebral blood flow and the cerebral blood volume.
Embodiment 25: The method of embodiment 24, wherein the likelihood the user will experience the stroke is based on a ratio of a percentage change of cerebral blood flow from a peak after the user began holding their breath to a baseline cerebral blood flow during the baseline time period to a percentage change of cerebral blood volume from a peak after the user began holding their breath to a baseline cerebral blood volume during the baseline time period.
Embodiment 26: A method of training a machine learning model to predict stroke risk, the method comprising: obtaining training data, the training data comprising, for a group of users, representations of cerebral blood metrics, wherein for each training sample, a portion of the obtained data spans a time period during which the user was holding their breath, and wherein each training sample includes a corresponding ground truth stroke risk for the user; providing the training data to a machine learning model, wherein the machine learning model takes, as input, the representations of the cerebral blood metrics and generates, as an output, a prediction of stroke risk; and updating the machine learning model based on differences between the ground truth stroke risk and the predicted stroke risk to generate a trained machine learning model configured to predict stroke risk.
Embodiment 27: The method of embodiment 26, wherein the representations of cerebral blood metrics are determined based on light reflectance data obtained using one or more light emitters and one or more light detectors disposed on a first head-worn device worn by users in the group of users, and wherein the trained machine learning model is provided to a computing device of a second head-worn device on which one or more light emitters and one or more light detectors are disposed.
Embodiment 28: The method of any one of embodiments 26 or 27, wherein the ground truth stroke risk is obtained based on questionnaire data.
Embodiment 29: The method of any one of embodiments 26-28, wherein the ground truth stroke risk is obtained based on longitudinal stroke occurrence data for users of the group of users.
Modifications, additions, or omissions may be made to any of the above-described embodiments without departing from the scope of the disclosure. Any of the embodiments described above may include more, fewer, or other features without departing from the scope of the disclosure. Additionally, the steps of described features may be performed in any suitable order without departing from the scope of the disclosure. Also, one or more features from any embodiment may be combined with one or more features of any other embodiment without departing from the scope of the disclosure. The components of any embodiment may be integrated or separated according to particular needs without departing from the scope of the disclosure.
It should be understood that certain aspects described above can be implemented in the form of logic using computer software in a modular or integrated manner. Based on the disclosure and teachings provided herein, a person of ordinary skill in the art will know and appreciate other ways and/or methods to implement the present invention using hardware and a combination of hardware and software.
Any of the software components or functions described in this application, may be implemented as software code using any suitable computer language and/or computational software such as, for example, Java, C, C#, C++ or Python, Matlab, or other suitable language/computational software, including low level code, including code written for field programmable gate arrays, for example in VHDL; embedded artificial intelligence computing platform, for example in Jetson. The code may include software libraries for functions like data acquisition and control, motion control, image acquisition and display, etc. Some or all of the code may also run on a personal computer, single board computer, embedded controller, microcontroller, digital signal processor, field programmable gate array and/or any combination thereof or any similar computation device and/or logic device(s). The software code may be stored as a series of instructions, or commands on a CRM such as a random-access memory (RAM), a read only memory (ROM), a magnetic media such as a hard-drive or a floppy disk, or an optical media such as a CD-ROM, or solid stage storage such as a solid state hard drive or removable flash memory device or any suitable storage device. Any such CRM may reside on or within a single computational apparatus, and may be present on or within different computational apparatuses within a system or network. Although the foregoing disclosed embodiments have been described in some detail to facilitate understanding, the described embodiments are to be considered illustrative and not limiting. It will be apparent to one of ordinary skill in the art that certain changes and modifications can be practiced within the scope of the appended claims.
The terms “comprise,” “have” and “include” are open-ended linking verbs. Any forms or tenses of one or more of these verbs, such as “comprises,” “comprising,” “has,” “having,” “includes” and “including,” are also open-ended. For example, any method that “comprises,” “has” or “includes” one or more steps is not limited to possessing only those one or more steps and can also cover other unlisted steps. Similarly, any composition or device that “comprises,” “has” or “includes” one or more features is not limited to possessing only those one or more features and can cover other unlisted features.
All methods described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. The use of any and all examples, or exemplary language (e.g. “such as”) provided with respect to certain embodiments herein is intended merely to better illuminate the present disclosure and does not pose a limitation on the scope of the present disclosure otherwise claimed. No language in the specification should be construed as indicating any non-claimed element essential to the practice of the present disclosure.
Groupings of alternative elements or embodiments of the present disclosure disclosed herein are not to be construed as limitations. Each group member can be referred to and claimed individually or in any combination with other members of the group or other elements found herein. One or more members of a group can be included in, or deleted from, a group for reasons of convenience or patentability. When any such inclusion or deletion occurs, the specification is herein deemed to contain the group as modified thus fulfilling the written description of all Markush groups used in the appended claims.
This application claims priority to and benefit of U.S. Provisional Patent Application No. 63/547,033, filed on Nov. 2, 2023, which is hereby incorporated by reference in its entirety and for all purposes.
This invention was made with government support under Grant No. EY033086 awarded by the National Institutes of Health. The government has certain rights in the invention.
Number | Date | Country | |
---|---|---|---|
63547033 | Nov 2023 | US |