Monitoring upper limb movements to detect stroke

Abstract

Methods, systems, and computer-readable media for detecting stroke by monitoring upper limb movement. In some examples, a method for detecting stroke includes receiving, at a stroke detector implemented in at least one processor, movement data from an accelerometer attached to a patient's upper limb over a period of time. The method includes analyzing, at the stroke detector, the movement data using a test statistic that is robust to movement distribution covariate shifts to enable passive monitoring of the patient. The method includes outputting, at the stroke detector, a warning signal in response to detecting a stroke using the movement data.

Description

Technical Field

[0001] Priority Claim This application claims the benefit of U.S. Patent Application No. 63 / 193,053, filed May 25, 2021, the disclosure of which is hereby incorporated by reference in its entirety.

[0002] Government Interests This invention was made with government support under Grant No. 1915398 awarded by the National Science Foundation and Grant No. FA8750-18-C-0090 awarded by the Air Force Research Laboratory. The government has certain rights in this invention.

[0003] Technical Field This specification generally relates to the detection of stroke by monitoring upper limb movement.

Background Art

[0004] Background The eligibility for stroke treatment and the likelihood of a good response to treatment are related to the speed at which a stroke is identified. Strokes in already hospitalized patients are associated with delays in symptom detection and evaluation, less intervention, and worse outcomes compared to strokes in the community.

[0005] Arm strength and movement asymmetry are one of the most common signs of acute stroke. Unfortunately, patients are not always able to detect a stroke early after onset and cannot be examined frequently enough to enable time-limited but proven interventions. Therefore, there is a need for automated methods, systems, and computer-readable media configured to detect stroke quickly.

Summary of the Invention

Means for Solving the Problems

[0006] Summary This specification describes a method, system, and computer-readable medium for detecting a stroke by monitoring the movement of an upper limb, for example, by monitoring asymmetric movements. In some examples, a method for detecting a stroke includes receiving motion data from an accelerometer attached to a patient's upper limb over a period of time in a stroke detector implemented on at least one processor. The method includes analyzing the motion data using a test statistic that is robust to motion distribution covariance shifts in the stroke detector so as to enable passive monitoring of the patient without knowing any specific information (e.g., dominant hand) about each patient. The method includes outputting a warning signal in the stroke detector in response to detecting a stroke using the motion data.

[0007] In some examples, analyzing the motion data using a test statistic includes analyzing the motion data using a parameter-agnostic (PAIN) statistic. The test statistic can be, for example, a Kolmogorov-Smirnov statistic.

[0008] In some examples, receiving the motion data includes receiving the motion data via a first wireless signal from a first wrist-worn accelerometer on the patient's first wrist. Receiving the motion data can include receiving a second wireless signal from a second wrist-worn accelerometer attached to the patient's second wrist. Receiving the motion data can include preprocessing the motion data to remove the effects of accelerometer and bias rotation / sliding.

[0009] In some examples, outputting the warning signal includes displaying a warning message on a display screen. Outputting the warning signal can include sending a message to a caregiver's mobile device.

[0010] The subject matter described herein can be implemented in hardware, software, firmware, or any combination thereof. Therefore, the terms “feature” or “node” as used herein refer to hardware, which may also include software and / or firmware components for implementing the described features. In some exemplary implementations, the subject matter described herein can be implemented using a computer-readable medium that stores computer-executable instructions, which, when executed by a computer processor, control the computer to perform steps.

[0011] Examples of computer-readable media suitable for implementing the subject matter described herein include non-transient computer-readable media such as disk memory devices, chip memory devices, programmable logic devices, and application-specific integrated circuits. Furthermore, computer-readable media implementing the subject matter described herein may be located on a single device or computing platform, or distributed across multiple devices or computing platforms. [Brief explanation of the drawing]

[0012] [Figure 1A] This is a block diagram of an exemplary system for stroke detection. [Figure 1B] This block diagram shows an exemplary configuration of a system for warning caregivers. [Figure 2A] The system displays the median and interquartile range (IQR) of the percentage of stroke cases that trigger a warning as monitoring time increases, using two different alert thresholds. [Figure 2B] The system displays the median and interquartile range (IQR) of the percentage of stroke cases that trigger a warning as monitoring time increases, using two different alert thresholds. [Figure 2C] The system displays the median and interquartile range (IQR) of the percentage of stroke cases that trigger a warning as monitoring time increases, using two different alert thresholds. [Figure 2D]The system displays the median and interquartile range (IQR) of the percentage of stroke cases that trigger a warning as monitoring time increases, using two different alert thresholds. [Figure 3] This is a flowchart illustrating an example method for detecting stroke. [Modes for carrying out the invention]

[0013] Detailed explanation This specification describes a method, system, and computer-readable medium for detecting stroke by monitoring upper limb movement. The stroke detection system described in this document allows for continuous monitoring of patients and can detect strokes with lateral weakness more quickly than usual care, which may lead to earlier stroke intervention and improved outcomes.

[0014] Figure 1A is a block diagram of an exemplary system 100 for stroke detection. System 100 can be deployed, for example, in a hospital or any suitable environment for stroke detection. Caregiver 102 is providing medical treatment to patient 104. However, caregiver 102 is not always with patient 104.

[0015] Two accelerometers 106 and 108 are attached to the upper limb of patient 104. For example, the first accelerometer 106 can be attached around the first wrist of patient 104, and the second accelerometer 108 can be attached around the second wrist of patient 104.

[0016] The caregiver computer system 110 includes at least one processor 112 and memory 114 for storing executable instructions for the processor 112. The caregiver computer system 110 can be, for example, a tablet, laptop, or telephone equipped with a display device and speaker for alerting the caregiver 102. The caregiver computer system 110 includes a stroke detector 116 implemented using the processor 112 and memory 114.

[0017] The stroke detector 116 is configured to passively monitor the patient 104 by receiving motion data from accelerometers 106 and 108. The monitoring is passive in that the caregiver 102 does not need to give any instructions to the patient 104 to perform any specific physical movements. For example, the stroke detector 116 can monitor asymmetric arm movements.

[0018] The stroke detector 106 analyzes motor data using a test statistic robust to motor distribution covariate shifts to enable passive monitoring of the patient. The stroke detector 106 outputs a warning signal in response to the detection of a stroke using the motor data. For example, the stroke detector 106 can display a warning image and / or play a warning sound from the caregiver computer system 110, or the stroke detector 106 can transmit a signal to a remote computer system.

[0019] Figure 1B is a block diagram showing an exemplary configuration 150 of a system for alerting a caregiver 102. In this example, accelerometers 106 and 108 are configured to transmit motion data to a stroke detector 116 via a data communication network 152, for example, by sending it to a wireless router to the internet. The stroke detector 116 can run, for example, on a cloud computing server.

[0020] The stroke detector 116, in response to detecting a stroke, outputs a warning signal by transmitting it to a caregiver device 154, such as a telephone, tablet, or laptop. For example, the stroke detector 116 can send a text message to the caregiver device 154. The caregiver 102 can then provide appropriate medical care to the patient 104.

[0021] In general, the system can be configured in any suitable way to warn the caregiver 102. For example, accelerometers 106 and 108 can be included in a wrist-worn computer system configured to run a stroke detector 116, which can then output a warning signal by playing an audible warning.

[0022] Methods and systems for stroke detection are described further below in relation to the tests conducted on the methodologies.

[0023] Introduction Proven stroke treatments, including intravenous thrombolysis and mechanical thrombectomy, are highly time-dependent. Eligibility for intervention and the probability of favorable outcomes, when performed, decrease sequentially as time elapsed since symptom onset. 1-3 Therefore, rapid detection of stroke symptoms is of paramount importance. 4-6 Of the 800,000 strokes that occur annually in the United States, 5-17% occur in patients who are already hospitalized, and the majority occur in patients who have recently received intervention or treatment. 7-9 Compared to community-based stroke, hospital-acquired stroke is associated with delayed detection and evaluation, less intervention, and worse outcomes. 7-11 Therefore, these complications significantly increase hospital and caregiver costs, length of stay, morbidity, mortality, and forensic liability. 7,8,10-14 Upper limb weakness is one of the most common findings in acute stroke. 15 As a result, asymmetrical arm strength is used in all stroke screening tools. 16,17 In addition, neglect is a common symptom of stroke, and there is a tendency to move the affected arm less. 18,19 This document describes an automated warning system for monitoring asymmetric arm movements, which can be used in conjunction with a wrist-worn accelerometer to rapidly identify strokes in hospitalized patients, facilitate earlier acute stroke treatment, and improve outcomes.

[0024] method To derive and validate a stroke detection algorithm, a prospective case-control study of upper limb movement was conducted in patients admitted to the University of Pennsylvania Hospital. This study was approved by the institutional review board at the University of Pennsylvania Hospital.

[0025] subject All subjects were recruited from hospitalized patients. Controls included patients with transient ischemic attacks (TIAs) who had no history of stroke, were neurologically normal, and did not have acute infarction on magnetic resonance imaging (MRI), patients undergoing examination for transient attacks of uncertain cause with normal MRI, and patients who had recently undergone cardiothoracic or vascular surgery without apparent neurological complications. Cases included patients hospitalized with acute ischemic or hemorrhagic stroke who had a National Institutes of Health Stroke Scale (NIHSS) score ≥ 1 and lateral limb weakness (greater on the affected side, at least 1 for upper limb weakness on NIHSS item 5a or 5b). Before monitoring begins, subjects undergo a neurological assessment including the NIHSS and an intensity assessment using the Medical Research Council Scale, evaluating the deltoid, biceps, triceps, wrist extension, wrist flexion, medial fingers, hip flexors, quadriceps, patellar tendons, ankle extension, and ankle flexion on each side on a scale of 0 (no movement) to 5 (maximum intensity).

[0026] Surveillance The subjects wore wrist straps with integrated accelerometers positioned on both arms. For the algorithm derivation cohort, a commercially available battery-powered Bluetooth®-enabled accelerometer / gyroscope, the Wit Motion BWT901CL Bluetooth output 9-axis accelerometer / gyroscope from Shenzhen, China, was used, along with Android. (Registered trademark)Data was streamed to a cloud-based server (Heroku, San Francisco, California) in synchronization with the tablet. The acceleration measurement device had an expected battery life of 2 - 3 hours. To capture more data and enable comparison of performance between day and night, the inventors needed a longer-lasting acceleration measurement device. Therefore, in the validation cohort, commercially available Samsung Galaxy Watch Active was used to collect acceleration measurement data. The application collected the acceleration measurement data (Laprot, Philadelphia, Pennsylvania) transmitted via WiFi to a cloud-based storage platform (Thingsboard, New York, New York). 20 The expected battery life of this device was 18 - 24 hours. For both stages of the trial, patients and clinical staff were told that they could remove the strap at any time if they were uncomfortable, it interfered with clinical treatment, or for any other reason they chose. To ensure that the conditions represented actual practice, no instructions were given to restrict movement of the treatment or passive range of motion of the affected limb while the patient was being monitored. After the monitoring was completed, a neurological assessment was repeated to confirm that there was no change in the neurological status.

[0027] Algorithm Derivation The algorithm was derived using the Parameter Invariant (PAIN) method designed to maximize diagnostic performance and generalizability. 21-25 This approach has been previously used to develop multiple medical classification algorithms that require high sensitivity and specificity with stable performance across patients without outliers. 22 The PAIN method uses statistical first principle techniques to derive algorithms that are invariant to patient-specific parameters (e.g., being left-handed or right-handed, awake / asleep, restrained / free to move) as well as common system anomalies in accelerometer-based systems (e.g., accelerometer bias / drift or device orientation). As a result, the algorithm achieves stable performance across the population without the need for individual adjustment.

[0028] The algorithm derivation method is described further below. In short, using derived cohort accelerometer data, the inventors identified features invariant to patient-specific parameters and then trained a structured classification tree that combined these features to maximize the stability and accuracy of detecting asymmetric motion patterns observed in stroke patients. Using multiple simultaneous threshold tests with varying durations allows for a balance between the trade-off between accuracy and time to detection. 26 Threshold tests with shorter monitoring periods result in faster detection times, while longer monitoring periods improve accuracy. The algorithm utilizes multiple windows simultaneously (where available) with increasing durations of preceding data, and issues a warning if any of these windows detect a possible presence of a stroke.

[0029] In some cases, a warning that leads to the identification of a stroke triggers a clinical intervention, including the removal of the device. Therefore, if motor data continues to occur after a warning, the algorithm assumes that the previous warning was a false positive and no further warnings are generated for one hour to allow the monitoring window to accumulate new data. For each subsequent warning within four hours of the previous warning, the warning pause is extended by another hour, up to a maximum of four hours. If no warnings are generated within eight hours, the warning pause period is reset to one hour. It should be noted that the proposed strategy results in a maximum false alarm rate of eight warnings in the first 24 hours, and six warnings per day thereafter.

[0030] verification The candidate algorithms were validated using an independent, blinded trial dataset collected separately from the dataset used for algorithm derivation, using different, longer-lasting accelerometers as described above. 27For this analysis, the algorithm evaluated individual patient data and was run every 15 minutes. For control subjects without stroke, the inventors evaluated the algorithm's performance in terms of false alarms per patient per day, defined as the number of alarms divided by the monitoring time (days). The inventors report the median false alarms per day by taking the median of false alarms per day for all control subjects. For each stroke case subject, the algorithm's performance in terms of detection rate was evaluated as the time from the start of monitoring increased. The start time of monitoring was in 15-minute increments throughout the entire monitoring period for each patient. In the inventors' study, data from subjects who progressed from neurologically intact to stroke during monitoring was unavailable, so a conservative evaluation of detection rate versus time to detection, commonly used in the fastest detection literature, was used. 28 As time elapsed since the start of monitoring increases, aggregate trials include only shorter duration windows, and the detection rate is calculated based on the percentage of aggregate trials that identified a stroke.

[0031] As described above, false alarms result in a temporary pause in alarm generation. To illustrate how this feature affects the time to stroke case detection, we calculated the duration of the alarm pause per day based on the algorithmic performance of control subjects. Next, we added the median and interquartile range of the delay due to the pause to the time to stroke case detection. For example, if the median false alarm rate in the control group was 1 per day, the alarm would be paused for 1 hour out of 24 hours. Assuming that a stroke could occur at any time during the 24-hour period, there are 23 hours with no additional delay and 1 hour where the alarm is paused (median delay is 30 minutes), resulting in (23 / 24)*0 minutes + (1 / 24)*30 minutes = an additional expected delay of 1.25 minutes per day. 27 Finally, we assessed whether patient-specific factors contributed to variability in algorithm performance by comparing the median detection time with the false alarm rate due to handedness and non-dominant hemisphere involvement, and we used the Wilcoxon rank-sum test to assess whether monitoring occurred at night or during the day.

[0032] result From May 8, 2018, to November 23, 2021, the inventors enrolled 405 patients, including 200 in the derivation cohort and 205 in the validation cohort. Due to technical difficulties, accelerometer data were unavailable for 5 control subjects in the validation cohort and were excluded from the analysis. The algorithm derivation cohort included 77 patients with acute stroke and lateral arm weakness, as well as 123 neurologically intact control subjects. A total of 540 hours of bilateral arm accelerometer data was obtained at this stage. The algorithm validation cohort included 33 patients with acute stroke and 167 controls, with a total of 4169 hours of bilateral arm accelerometer data. Table 1 shows the clinical and demographic characteristics of the controls for the derivation and validation cohorts, and Table 2 provides these data for stroke cases. Among the non-stroke controls, subjects in the validation cohort were less likely to be female and more likely to have recently undergone surgery. For stroke cases, the validation cohort was similar to the derivation cohort, except for a greater difference in arm strength between the affected and unaffected sides as measured by the sum of the Medical Research Council upper extremity motor scores, although the differences in NIHSS upper extremity motor scores were similar.

[0033] Within the validation cohort, stroke cases had similar age (mean 68 vs. 65 years, P=0.25), proportion of females (45% vs. 32%, p=0.15), and proportion of right-handedness (94% vs. 84%, p=0.13) compared to controls, but stroke patients were more likely to be non-white (45% vs. 11%, p<0.001) and more likely to be in the ICU or antihypertensive unit (85% vs. 34%, p<0.001). Overall, stroke cases in the validation cohort were predominantly ischemic (73%) and moderately severe (NIHSS median 14, IQR 9–18). Wrist straps were well tolerated in both the algorithm derivation cohort and the validation cohort. None of the patients in the derivation cohort and neither of the two patients in the validation cohort removed the device, and the trial was terminated prematurely (after 1 hour and 22 hours of monitoring, respectively). Nurses did not report any issues with the straps interfering with clinical care. There was no change in the patient's upper limb strength or the presence of neglect when comparing tests at baseline and at test completion.

[0034] [Table 1]

[0035] [Table 2]

[0036] *The upper limb muscle groups evaluated using the Medical Research Council's muscle strength score (range 0-5) included the deltoid, biceps, triceps, wrist extension, wrist flexion, and inner finger strength. Maximum strength was achieved in all six muscles tested, resulting in a score of 30.

[0037] Algorithm performance Figures 2A–2D show the median and interquartile range (IQR) of the percentage of stroke cases that trigger a warning as monitoring time increases using two different warning thresholds. Figures 2A–2D show the stroke detection rate and the false alarm rate per day over time.

[0038] Figure 2A shows the median (solid line) and interquartile range (dashed line) of the percentage of patients who received a stroke warning as the monitoring period lengthened. Figure 2B shows the distribution of false alarms per patient per day in non-stroke controls. The black line represents the cumulative percentage of patients. Figure 2C shows the effect of lowering the warning threshold on the time to detection. Figure 2D shows the effect of lowering the warning threshold on the false alarm rate.

[0039] Sensitivity (i.e., the proportion of stroke patients detected as having a stroke) was positively correlated with the monitoring period. Comparing the results for two different target false alarm rates demonstrates a correlation between sensitivity and false alarm rate.

[0040] As the false alarm rate increases, the time to detection decreases. When the median false alarm rate among non-stroke controls was 1.1 alarms per patient per day (IQR 0–2.2 alarms per patient per day), the median time to alarm in stroke cases was 29 minutes (IQR 11–58 minutes). At 60 minutes, the algorithm is expected to detect 76% of strokes. When the median false alarm rate was 3.6 alarms per patient per day (IQR 2.1–5.0 false alarms), the median time to detection in stroke cases was 15 minutes (IQR 8–74 minutes). In this setting, the algorithm is expected to detect 91% of strokes at 60 minutes. Importantly, the algorithm was unaffected by patient-specific factors that could theoretically lead to variable performance. Specifically, using a lower sensitivity threshold, there was no significant difference in false alarm rates (median 1.2 vs. 1.0 alarms / day, p=0.17) or time to detection (median 29 vs. 27 minutes, p=0.83) between right-handed and left-handed patients. Similarly, there was no difference in time to detection when the stroke involved a dominant or non-dominant hemisphere (median 29 vs. 29 minutes, p=1.0). Most importantly, there was no difference in detected false alarms (median 0 vs. 0 alarms, p=0.57) or time to detection (median 28 vs. 26 minutes, p=0.79) between daytime and nighttime. These results were similar when evaluated using a higher sensitivity threshold.

[0041] Consideration This study demonstrates that arm accelerometer data can be used to distinguish between patients with weakness caused by acute stroke and neurologically intact hospitalized patients. The algorithm's diagnostic performance achieves high sensitivity and specificity, enabling it to provide clinically useful monitoring for rapid detection of stroke onset while maintaining a low false alarm rate. The warning threshold is adjustable; lowering the threshold increases sensitivity and shortens the time to detection, but consequently increases the false alarm rate. Importantly, the time to detection estimates are conservative and may be faster in clinical use. In the time to detection analysis, the inventors included only motor data from stroke patients.

[0042] In reality, patients may change from non-stroke to stroke during the evaluation window, and even in such cases, an alert may be triggered, so the time from onset to detection is shorter than reported by the inventors. Furthermore, stroke cases were cared for in real-world, routine clinical practice while being monitored, and the care team or family may have moved the patient's limp arm. These periods were not censored for the validation analysis. In clinical use, the inventors expect the algorithm to detect more than half of strokes within 30 minutes of onset, while adequately maintaining fewer than two false alarms per day for the majority of patients. Most importantly, there was no significant variation in algorithm performance based on handedness, involvement of the non-dominant hemisphere, or whether monitoring was daytime versus nighttime. This latter finding suggests that strokes can be detected equally well during sleep or wakefulness, which is a key feature of a useful stroke monitor.

[0043] In-hospital stroke is a major public health issue, accounting for a significant portion of all strokes, and is associated with delays in assessment and treatment, poor outcomes, and dramatically increased costs and length of hospital stay. 7-12 Importantly, the majority of cases in most series are perioperative strokes, and the stroke rate for common procedures such as aortic valve surgery is much higher than is generally reported when prospective evaluations are performed. 7-9,29-31Given that the algorithm detects asymmetry and is not based on changes in motor patterns during the baseline period, it is particularly well-suited for detecting perioperative strokes in which patients may awaken from anesthesia while still muscular. Previous studies of in-hospital stroke have reported time from last known normal to symptom detection ranging from approximately 2 to 10 hours. 29-31 While proven stroke treatments can offer robust benefits, the likelihood of receiving these treatments and the response to them steadily decline over time. 1-3 Therefore, rapid detection of stroke onset remains critically important. Devices incorporating this algorithm for continuous monitoring of stroke onset can dramatically reduce the time to evaluation, leading to faster intervention and better patient outcomes.

[0044] Upper limb weakness is one of the most common symptoms of acute stroke, occurring in approximately 75% of patients. 15 Therefore, all pre-hospital stroke screening tools and scales aimed at identifying patients most likely to have a major vascular occlusion include arm strength. 16,17,32 Furthermore, attentional neglect is present in 20–70% of strokes, and studies of stroke patients using wrist-worn accelerometers have demonstrated that neglect is associated with asymmetrical movements. 18,19 Importantly, both weakness and neglect are strongly associated with long-term physical disability following a stroke. 33,34 Therefore, while upper limb acceleration monitoring does not capture all strokes, it can identify most strokes, including those most likely to result in disability and those for which thrombectomy, which has been proven to dramatically improve outcomes, is most appropriate.

[0045] Physiological monitoring of patients is common in hospitals and especially in intensive care units, with multimodal monitoring being the standard of care. Unfortunately, the prevalence of such monitoring can lead to alarm fatigue, which can result in delayed or absent responses. 35Fatigue is more likely to occur when there are far more unmanageable warnings than manageable warnings that require both clinical recognition and intervention. Stroke is a manageable and extremely time-sensitive patient event. 6 A study of 461 adults treated in the intensive care unit annotated a total of 381,560 unique audible warnings over a 31-day study period. 36 Accelerated ventricular arrhythmia warnings, potentially serious patient abnormalities, occurred at an average of 4.5 warnings per patient per day of monitoring, of which only 12 (0.3%) were clinically relevant and manageable events. This stroke detection algorithm offers a much lower false alarm rate while significantly reducing the time from symptom onset to stroke detection compared to current clinical practice.

[0046] This study demonstrates that arm accelerometer data can distinguish patients with acute stroke from neurologically normal hospitalized patients at risk of stroke. In particular, the inventors performed validation analyses on a separate prospectively acquired patient cohort using a different accelerometer than the one used to collect data for deriving the algorithm. The performance of the algorithm under these conditions reflects its robustness and generalizability. The validation cohort included control patients who underwent cardiothoracic or vascular surgery, reflecting a population at high stroke risk and who would benefit from continuous stroke monitoring.

[0047] conclusion Hospital-acquired stroke is a major public health problem, and monitors that can rapidly detect stroke onset and facilitate rapid assessment and treatment would significantly improve patient outcomes. The inventors derived a stroke detection algorithm using upper limb acceleration measurement data from hospitalized patients, demonstrating promising diagnostic performance in a prospective validation cohort. A prospective trial to monitor patients at risk of stroke is necessary to demonstrate clinical utility and tolerability.

[0048] References:

[0049] [Table 3]

[0050] [Table 4]

[0051] [Table 5]

[0052] [Table 6]

[0053] Figure 3 is a flowchart illustrating an exemplary method 300 for detecting stroke. Method 300 includes a training phase 302 and a detection phase 304. Typically, the training phase 302 is first performed by a first computer system to generate a stroke detector, i.e., a model containing data to be delivered to other computer systems. The other computer systems can then independently perform the detection phase 304, and the model is used to detect stroke in patients.

[0054] Method 300 includes collecting or acquiring training data (306). The training data includes accelerometer data obtained from patients and stroke data indicating whether the corresponding accelerometer data is from a patient who has experienced a stroke.

[0055] Method 300 includes training a stroke detector using training data and a test statistic robust to motor distribution covariate shifts (308). Training the stroke detector generates a model, which may include data that can be stored in individual computer systems or cloud systems, as described above with reference to Figures 1A and 1B, for example. Training the stroke detector is further described below with reference to an example.

[0056] Method 300 includes receiving motion data from one or more patient accelerometers (310). In some examples, receiving motion data includes receiving motion data from radio signals from wrist-worn accelerometers on both wrists of the patient. Receiving motion data may include preprocessing the motion data to remove the effects of rotation / sliding of the accelerometer and bias. While accelerometer data alone may be sufficient for stroke detection, in some examples other data, such as data from other sensors for detecting motion, such as gyroscopes and magnetometers, are collected.

[0057] Method 300 includes analyzing motor data using a trained stroke detector (312) and determining whether the patient is experiencing a stroke (314). If the stroke detector determines that the patient is experiencing a stroke, Method 300 includes triggering an alert to the caregiver (316). Otherwise, Method 300 continues to receive motor data (returning to 310) until a stroke is detected or patient monitoring is terminated.

[0058] Stroke detection training example Data preprocessing To design a low-cost, lightweight, and comfortable wrist-worn device for stroke detection, the inventors attempted to utilize only accelerometer data in their stroke detection analysis. While incorporating additional sensors such as gyroscopes and magnetometers would theoretically allow for estimation of device orientation and arm position, they would also increase the device's cost, power consumption, battery size, and weight. Consequently, this study and the preprocessing considered herein, aimed at detecting stroke using commercially available low-power accelerometers, assumed that only accelerometer data was available.

[0059] A typical low-cost, low-power accelerometer worn on the wrist collects 3D data a at time k. x (k), a y (k), and a z(k) is generated, but is susceptible to wrist bias and rotation / sliding. The inventors have found a constant bias c x (k), c y (k), and c z This is shown as (k), and these effects were eliminated by using the first derivative of acceleration (known as the "jerk"). Because J x (k) = d / dk(a x (k)+c x ) = d / dk(a x (k) is the same for the y and z dimensions. Once the bias is removed, the inventors have found that

[0060]

number

[0061] By considering only the magnitude of the jerk described for the motion data of the left arm, the effect of rotation / sliding on the wrist is removed, where X represents the feature space and a similar formula applies to the right arm x R (k) exists. This preprocessing step is intended to remove system-specific biases that are likely to occur during real-world deployments and is consistent with other data preprocessing techniques for accelerometer data where gyroscopes and magnetometers are unavailable.

[0062] Test and statistical processing This is a space of probability distributions that maps the feature space to probabilities in order to design a test statistic to distinguish between stroke subjects and neurologically uninjured subjects.

[0063]

number

[0064] We began by describing the following: In a controlled evaluation environment (an idealized version) in which the subject performs a predetermined series of actions / movements,

[0065]

number

[0066] This allows for the distinction between the two. While this idealized scenario could result in highly sensitive and specific stroke detection, in practice it is too invasive, requiring frequent neurological evaluations to detect strokes in a timely manner.

[0067] Rather than requiring patients to perform a predetermined set of tasks at set intervals, the inventors sought to design a test statistic suitable for passive monitoring scenarios. Such a test statistic must be robust to changes in the underlying patient movement distribution, known in the statistical literature as covariate shift. 1 Motor distribution covariate shifts are common in passive monitoring scenarios and incorporate the influence of any patient-specific tendencies in the data (e.g., handedness, comorbidities, etc.). However, the influence of motor covariate shifts is limited by the patient's neurological state, which is presumed to be unknown at the time of the study. As a result, we have defined a family of motor covariate shifts as being influenced by the patient's (unknown) neurological state, i.e.

[0068]

number

[0069] As a result, the inventors sought a test statistic that can evaluate neurological conditions robust to motor distribution covariate shifts.

[0070] One exemplary method for obtaining a robust test statistic is to utilize a parameter-invariant (PAIN) statistic that has been previously applied across multiple domains. 2-5 When considering a group of troublesome transformations, the PAIN statistic t eliminates information affected by the troublesome transformations (i.e., maximum

[0071]

number

[0072]

number

[0073]

number

[0074] He proved this. Furthermore, t(f L ) and t(f R ) has attractive properties, namely, f (as in the case of a neurologically intact subject in an idealized scenario) L= f R If that is the case, t(f L )=t(f R ) and

[0075]

number

[0076] It should be noted that this is formally stated as follows: This means that if, in an idealized surveillance scenario, the subjects are neurologically intact, then in a passive surveillance scenario they should also appear neurologically intact.

[0077] Therefore, the inventors have found a neurologically intact subject (i.e., t(f L )=t(f R )) and stroke subjects (i.e.,

[0078]

number

[0079] The goal was to generate a test statistic to distinguish between ) and ). In this scenario, t L = t(f L ) and t R = t(f R ) and Test statistic

[0080]

number

[0081] By writing the Kolmogorov-Smirnov (KS) statistic, 6、7 This is the probability mass function t L and t R It represents a nonparametric statistic of the distribution equality equal to the maximum absolute deviation of the corresponding cumulative distribution function. The KS statistic is a widely used test of distributional equivalence when the underlying family of test distributions is unknown or unparalleled (i.e., nonparametric).

[0082] Test generation Next, the inventors developed a threshold test for the test statistic s derived in the previous section. The test statistic is the probability mass function t L and t R This requires a corresponding cumulative distribution function. Unfortunately, these are not generally known, and preprocessed sampled data X(k){(x L (k), x R (k)),(x L (k-1), x R} = (k-2)),...} must be estimated from the recent history (1 hour). Using sampled data estimates instead of the actual distribution presents two potential concerns. First, if there is significant missing data, the amount of information contained in the sampled data is reduced. Second, whenever the patient is not moving (i.e., completely still) while there is no technically missing data, it does not provide identifying information for a stroke vs. neurologically intact examination. Therefore, we write s(k) as the test statistic estimated using X(k), write r1(k)=X(k) as the number of data points in X(k), and write r1(k)=X(k) as the proportion of X(k) when there is patient movement.

[0083]

number

[0084] It is written as follows. To derive the threshold test, the inventors utilize r1 and r2, and the resulting test has a constant false alarm rate.

[0085]

number

[0086] The threshold was fitted to have the following characteristics. To achieve this, the inventors grouped the data using k=100 kmmeans for [(r1(1),r2(1)),(r1(2),r2(2)),...] and generated corresponding thresholds for each group to achieve a constant false alarm rate α. To achieve the maximum distributional accuracy when adjusting the false alarm rate, the threshold test was calibrated before threshold selection. 8 During execution, a new s(k) was generated using the corresponding r1(k) and r2(k). The decision threshold used to test s(k) corresponds to the group containing (r1(k), r2(k)). Below, the above threshold test is performed.

[0087]

number

[0088] These terms refer to the absence of strokes, where φ=0 predicts the absence of strokes and φ=1 predicts the presence of strokes. To improve susceptibility to stroke, the inventors have developed different monitoring periods d1, ...d L Using this method, multiple threshold tests φ1, ...φ are performed simultaneously. L This was executed. For example, each at time t

[0089]

number

[0090] For φ1, data within the time range [t-d1, t] is used. By utilizing multiple threshold tests, the inventors have developed an aggregated threshold test that predicts the presence of stroke only if one of the monitoring periods predicts the presence of stroke.

[0091]

number

[0092] We defined α. Note that the false alarm rate of the aggregate test is always greater than α. As a result, we select α in the threshold test design such that the aggregate test is small enough to achieve our desired false alarm rate.

[0093] References:

[0094] [Table 7]

[0095] While specific examples and features have been described above, these examples and features are not intended to limit the scope of this disclosure, even if only a single example is given for a particular feature. The examples of features provided in this disclosure are intended to be illustrative, not restrictive, unless otherwise specified. The above descriptions are intended to include alternatives, modifications, and equivalents that would be apparent to those skilled in the art who are interested in this disclosure.

[0096] The scope of this disclosure includes any features or combinations of features disclosed herein (expressly or implicitly), or any generalizations of the disclosed features, regardless of whether such features or generalizations mitigate any or all of the issues described herein. Therefore, new claims can be formulated for any combination of such features during the execution of this application (or an application claiming priority to this application). In particular, with reference to the appended claims, features from dependent claims can be combined with features from independent claims, and features from each independent claim can be combined in any suitable manner, not just in the specific combinations enumerated in the appended claims.

Claims

1. A method for detecting a stroke, wherein the method is A stroke detector implemented in at least one processor receives motion data from an accelerometer attached to the patient's upper limb over a certain period of time, The stroke detector analyzes the motor data using a test statistic based on passive monitoring of the patient, and the analysis of the motor data using the test statistic includes evaluating the motor data using an aggregated threshold test, the aggregated threshold test is configured to output a first value that predicts the presence of a stroke if the accelerometer data in one of a plurality of different monitoring periods indicates a stroke, and to output a second value that predicts the absence of a stroke if the accelerometer data in all of the monitoring periods indicates the absence of a stroke. A method comprising: outputting a warning signal in the stroke detector in response to detecting a stroke using the motion data.

2. The method according to claim 1, wherein analyzing the exercise data using the test statistic includes analyzing the exercise data using a parameter-invariant (PAIN) statistic.

3. The method according to claim 2, wherein the test statistic is the Kolmogorov-Smirnov statistic.

4. The method according to claim 1, wherein receiving the motion data includes receiving the motion data by a first wireless signal from a first wrist-worn accelerometer located on the patient's first wrist.

5. The method according to claim 4, wherein receiving the motion data includes receiving a second wireless signal from a second wrist-worn accelerometer located on the patient's second wrist.

6. The method according to claim 1, wherein receiving the motion data includes preprocessing the motion data to remove the effects of rotation / sliding of the accelerometer and bias.

7. The method according to claim 1, wherein outputting the warning signal includes displaying a warning message on a display screen.

8. A system for detecting stroke, wherein the system is At least one processor, A stroke detector implemented in at least one of the aforementioned processors, Over a certain period, motion data is received from an accelerometer attached to the patient's upper limb. The method involves analyzing the exercise data using a test statistic based on passive monitoring of the patient, wherein the analysis of the exercise data using the test statistic includes evaluating the exercise data using an aggregated threshold test, the aggregated threshold test being configured to output a first value predicting the presence of a stroke if the accelerometer data in one of several different monitoring periods indicates a stroke, and a second value predicting the absence of a stroke if the accelerometer data in all of the monitoring periods indicates the absence of a stroke. A system comprising: a stroke detector configured to perform an action that outputs a warning signal in response to detecting a stroke using the aforementioned motion data.

9. The system according to claim 8, wherein analyzing the exercise data using the test statistic includes analyzing the exercise data using a parameter-invariant (PAIN) statistic.

10. The system according to claim 9, wherein the test statistic is the Kolmogorov-Smirnov statistic.

11. The system according to claim 8, wherein receiving the motion data includes receiving the motion data by a first wireless signal from a first wrist-worn accelerometer located on the patient's first wrist.

12. The system according to claim 11, wherein receiving the motion data includes receiving a second wireless signal from a second wrist-worn accelerometer located on the patient's second wrist.

13. The system according to claim 8, wherein receiving the motion data includes preprocessing the motion data to remove the effects of rotation / sliding of the accelerometer and bias.

14. The system according to claim 8, wherein outputting the warning signal includes displaying a warning message on a display screen.

15. When executed by at least one processor of a computer, Over a certain period, motion data is received from an accelerometer attached to the patient's upper limb. The method involves analyzing the exercise data using a test statistic based on passive monitoring of the patient, wherein the analysis of the exercise data using the test statistic includes evaluating the exercise data using an aggregated threshold test, the aggregated threshold test being configured to output a first value predicting the presence of a stroke if the accelerometer data in one of several different monitoring periods indicates a stroke, and a second value predicting the absence of a stroke if the accelerometer data in all of the monitoring periods indicates the absence of a stroke. A non-temporary computer-readable medium that stores executable instructions that control the computer to perform an action, which outputs a warning signal in response to the detection of a stroke using the aforementioned motor data.

16. The non-temporary computer-readable medium according to claim 15, wherein analyzing the motion data using the test statistic includes analyzing the motion data using a parameter-invariant (PAIN) statistic.

17. The non-temporary computer-readable medium according to claim 16, wherein the test statistic is the Kolmogorov-Smirnov statistic.

18. The non-temporary computer-readable medium according to claim 15, wherein receiving the motion data includes receiving the motion data by a first wireless signal from a first wrist-worn accelerometer located on the patient's first wrist.

19. The non-temporary computer-readable medium according to claim 18, wherein receiving the motion data includes receiving a second wireless signal from a second wrist-worn accelerometer located on the patient's second wrist.

20. The non-temporary computer-readable medium according to claim 15, wherein receiving the motion data includes preprocessing the motion data to remove the effects of rotation / sliding of the accelerometer and bias.

Images