Systems and methods for biosignal feature detection

Abstract

Systems and methods for biosignal feature detection in accordance with embodiments of the invention are disclosed. In an embodiment of the invention, a method for detecting dicrotic notches (DNs) includes receiving an input signal, pre-processing the received input signal, decomposing the pre-processed input signal into non-stationary and stationary parts using an iterative envelope method, and retrieving the DNs based on the non-stationary part of the input signal.

Description

Systems and Methods for Biosignal Feature Detection CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The current application claims the benefit of and priority under 35 U.S.C. § 119(e) to U.S. Provisional Patent Application No.63 / 617,038 entitled “Systems and Meth- ods for Dicrotic Notch Detection” filed January 2, 2024. The disclosure of U.S. Provisional Patent Application No. 63 / 617,038 is hereby incorporated by reference in its entirety for all purposes. STATEMENT OF FEDERAL SUPPORT

[0002] This invention was made with government support under HL144692, and EB029751 awarded by the National Institutes of Health. The government has certain rights in the invention. FIELD OF THE INVENTION

[0003] The present invention generally relates to signal processing and, more specifically, using signal processing to detect dicrotic notches. BACKGROUND

[0004] A dicrotic notch (DN) is a small dip in the arterial blood pressure (ABP) and photoplethysmography (PPG) waveforms that occur during the relaxation phase of the heart's cycle. This notch is indicative of the closure of the aortic valve and the beginning of the diastolic phase of the cardiac cycle. By accurately detecting the DN, doctors and medical professionals can assess the heart's ability to pump blood effectively and identify potential issues such as arrhythmias, heart valve disorders, and hypertension. Early detection of these conditions can lead to timely interventions and improved patient outcomes. As such, the development of reliable and accurate methods for detecting DNs is an important area of research in cardiovascular medicine. SUMMARY OF THE INVENTION

[0005] Systems and methods for biosignal feature detection in accordance with embodiments of the invention are disclosed. In an embodiment of the invention, a methodfor detecting dicrotic notches (DNs) includes receiving an input signal, pre-processing the received input signal, decomposing the pre-processed input signal into non-stationary and stationary parts using an iterative envelope method, and retrieving the DNs based on the non-stationary part of the input signal.

[0006] In a further embodiment, the decomposing of the pre-processed input signal includes calculating a smoothed input signal, calculating a first derivative of the smoothed input signal, calculating the coordinates of the smoothed input signal at the location of each local maxima and minima of the first derivative of the smoothed input signal, calculating an envelope mean value based on an upper envelope and a lower envelope of the smoothed input signal, and estimating the non-stationary part of the input by subtracting envelope mean value from the input signal.

[0007] In another embodiment, the smoothed input signal is calculated with a Savizky- Golay (SG) family filter.

[0008] In yet another embodiment, the upper envelope is obtained by applying a cubic spline interpolation to connect the coordinates of the local maxima of the first derivative in the smoothed input signal.

[0009] In yet another embodiment again, the lower envelope is obtained by applying a cubic spline interpolation to connect the coordinates of the local minima of the first derivative in the smoothed input signal.

[0010] In several embodiments, the input signal comprises an ABP waveform or a PPG waveform of cardiac cycles.

[0011] In some embodiment, the input signal comprises 4-second windows of ABP or PPG waveforms.

[0012] In more embodiments, each 4-second window is filtered using a 4th-order Butterworth low pass filter with a cut-off frequency of 16 Hz to remove high-frequency noise.

[0013] In yet more embodiments, pre-processing the received input comprises excluding 4-second windows that satisfies at least one criterion selected from the group of: the received input having zero or negative values, the received input having three or fewer peaks exceeding 75% of the window's amplitude, and the received input having more than ten peaks.

[0014] In still more embodiments, the method also includes imposing a plurality of requirements to the non-stationary part of the input signal to identify all key points within a CC. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The description and claims will be more fully understood with reference to the following figures and data graphs, which are presented as exemplary embodiments of the invention and should not be construed as a complete recitation of the scope of the invention.

[0016] FIG. 1 illustrates a cardiac cycle shown on an arterial blood pressure (ABP) waveform in accordance with an embodiment of the invention.

[0017] FIG. 2 illustrates a process for DN detection using an iterative envelope mean (IEM) method in accordance with an embodiment of the invention.

[0018] FIG.3 illustrates a process for decomposing an input signal into stationary and non-stationary parts using an IEM method in accordance with an embodiment of the invention.

[0019] FIG. 4 illustrates the non-stationary and stationary parts of an input signal in accordance with an embodiment of the invention.

[0020] FIGS.5A-D illustrate the regression analysis and box plot analysis for the ABP and PPG waveforms in accordance with an embodiment of the invention.

[0021] FIG. 6 illustrates a comparison of detection capabilities of the IEM-based method versus the 2ndderivative method in accordance with an embodiment of the invention.

[0022] FIG.7 illustrates a process for detecting key points using the IEM in accordance with an embodiment of the invention.

[0023] FIG.8 illustrates an example of applying the IEM method to a 4-second window of an ABP input signal in accordance with an embodiment of the invention.

[0024] FIG.9 illustrates a block diagram of a computing device for detecting biosignal features in accordance with an embodiment of the invention.DETAILED DESCRIPTION

[0025] Detecting dicrotic notches (DNs) is an important aspect of maintaining heart health. A DN is a small downward deviation in an arterial pressure waveform that occurs during the relaxation phase of the heart's cycle. The small downward deviation is caused by the closure of the aortic valve, which marks the end of systole and the beginning of diastole. Early and accurate detection of DNs can help medical professionals identify potential issues such as arrhythmia, heart valve disorders, and hypertension. It can also help them assess the heart's ability to pump blood effectively. By identifying these conditions early on, medical professionals can provide timely interventions and improve patient outcomes. Therefore, methods for detecting DNs can be valuable tools for understanding and monitoring cardiovascular health.

[0026] Detecting DNs has been a challenge for medical professionals. DNs are small in size on the arterial blood pressure (ABP) and photoplethysmography (PPG) waveforms, which makes them difficult to identify. The subtle changes in the shape of these waveforms that occur as blood flows away from the heart during the relaxation phase of the heart's cycle can mask the presence of DNs, which presents an additional layer of difficulty in the detection of DNs. The changes in the shape of an ABP or PPG waveform can lead to DNs becoming less pronounced or even disappearing entirely. The presence of noise and artifacts in these waveforms may also complicate the detection process of DNs. Finally, in patients with certain cardiovascular conditions, DNs are not visible. Therefore, it is important to not only identify DNs when they are present but also detect them within waveforms where DNs are not clearly defined.

[0027] Traditional methods, such as visual inspection of waveforms and manual measurement of time intervals, have proven to be subjective and time-consuming and may not always lead to accurate detection. With signal processing techniques becoming modernized, algorithms have been developed to detect DNs in recent years. However, these prior methods are often limited in their ability to handle noises in the measurement waveforms and / or their ability to identify DNs in signals lacking a prominent notch. Small variations or noise in ABP or PPG waveforms can introduce significant challenges in accurately detecting the DN, often resulting in false detections or unreliablemeasurements. These limitations highlight the need for the development of more reliable and objective methods that can detect DNs with high accuracy and efficiency.

[0028] Systems and methods in accordance with embodiments of the invention can remedy the above limitations by utilizing an algorithm that leverages an iterative envelope mean (IEM) method to accurately identify DNs in ABP and PPG waveforms. In many embodiments, systems and methods decompose input signal into stationary and non- stationary parts using the IEM approach, and utilize the non-stationary part of the input signal to accurately detect DNs. In many embodiments, the input signal is either ABP or PPG waveform. In several embodiments, systems and methods using the IEM approach can effectively identify DNs as well as all key points within cardiac cycles (CC) in at least the ABP and PPG waveforms. Systems and methods in accordance with many embodiments have a low computational cost, high noise robustness, and can detect DNs not only in waveforms where they are clearly defined but also in waveforms with less pronounced DN characteristics. Systems and methods in accordance with numerous embodiments can be easily generalized for application in different settings and can outperform current leading methods in DN detection, including visual inspection by ahuman evaluator. While ABP and PPG waveforms are discussed here, the techniquesmay be applicable to other biosignal waveforms that embody or capture an arterial pressure waveform. Biosignal features such as, but not limited to, dicrotic notch, can be detected by systems and methods in accordance with embodiments of the invention discussed herein. Dicrotic Notch (DN) Detection using the IEM Method

[0029] A CC in the context of ABPs and PPGs may be defined by five key points: systolic phase onset (SPO), systolic phase peak (SPP), dicrotic notch (DN), diastolic phase peak (DPP), and diastolic phase endpoint (DPE). These key points can serve as critical landmarks for extracting valuable features that have shown various clinical applications. A CC shown on an ABP in accordance with an embodiment of the invention is illustrated in Fig.1. Fig.1 demonstrates the five key points in an example CC as shown on an ABP, which are: systolic phase onset, systolic phase peak (systolic pressure), dicrotic notch, diastolic phase peak, and diastolic phase endpoint (diastolic pressure),DN amplitude (DNa), DN height (DNh), and systolic decay phase. In numerous embodiments, the DN is the transition point between the systolic and diastolic phases of the cardiac cycle.

[0030] Systems and methods in accordance with numerous embodiments of the invention utilize an IEM method to detect DNs. The IEM method is a signal processing technique that can be used to accurately identify DNs in ABP and PPG waveforms. The IEM method involves iteratively estimating the upper and lower envelopes of a smoothed input signal. In many embodiments, the input signal is either an ABP or PPG waveform. This estimation may be based on the coordinates of the smoothed input signal at the location of each of the first derivative’s local maxima and minima, respectively. In each iteration, the envelope mean can be computed using the estimated envelopes. The estimated envelope mean can be subtracted from the input signal and the resulting signal can serve as the input for the next iteration. In many embodiments, the iteration process continues until it satisfies a stopping criterion after a number of iterations, I. At the end of the iteration process, the subtraction between the input signal and the envelope mean can provide the non-stationary part, while the summation of the envelope means from each iteration can yield the stationary part of the original input signal. A process for DN detection using the IEM method in accordance with an embodiment of the invention is illustrated in Fig. 2. Process 200 receives (210) an input signal. In many embodiments, the received signal is either an ABP waveform or a PPG waveform.

[0031] Process 200 pre-processes (220) the received input signal. Pre-processing the received input can eliminate windows of signals that have artifacts. Several embodiments use the 'find_peaks' function from the scipy Python package for peak extraction. In some embodiments, systems can exclude any 4-second window that meets one or more of the following criteria: containing zero or negative values, having three or fewer peaks exceeding 75% of the window's amplitude, or having more than ten peaks. In certain embodiments, each selected 4-second window can be filtered using a 4th-order Butterworth low pass filter with a cut-off frequency of 16 Hz to remove high-frequency noise. The amplitude of the filtered 4-second window signal can be normalized using:^^^^^^^^^௫ ^^^where y(n) represents the 4- is the sample index,^^^is the minimum value of the input^^௫value, and^^^^denotes the normalized 4-second input signal.

[0032] Process 200 decomposes (230) the input signal into non-stationary and stationary parts. In numerous embodiments, systems and methods utilize the IEM method to iteratively estimate the envelope mean, which is subtracted from the input signal, where the resulting signal is then used as input for the next iteration. In several embodiments, the IEM method includes a stopping condition that can stop the iteration process when the stopping condition is met. Subtraction of the envelope mean from the input signal can provide the non-stationary part of the original input signal. Additional details regarding the IEM method are provided below.

[0033] Process 200 identifies (240) the DNs and CC features and within the input signal. Systems and methods in accordance with numerous embodiments of the invention can utilize the non-stationary part of the input signal to identify the locations of DNs within the input waveform.

[0034] While specific processes for DN detection using the IEM method are described above, any of a variety of processes can be utilized to detect DNs as appropriate to the requirements of specific applications. In certain embodiments, steps may be executed or performed in any order or sequence not limited to the order and sequence shown and described. In a number of embodiments, some of the above steps may be executed or performed substantially simultaneously where appropriate or in parallel to reduce latency and processing times. In some embodiments, one or more of the above steps may be omitted.

[0035] A process for decomposing an input signal into stationary and non-stationary parts using the IEM method in accordance with an embodiment of the invention is illustrated in Fig. 3. Process 300 calculates (310) a smoothed version of an input signal. In many embodiments, the input signal is smoothed using a Savizky-Golay (SG) family filter. Parameters of the SG family filter may include a polynomial fitting degree^andcoefficients^. In some embodiments,^may have a value of 4.^may be approximately one to two times the half-width of the shortest-duration feature of interest in the input signal. In ABP and PPG CC signals, the shortest duration may occur between any two consecutive key points within a CC. A CC typically includes five key points: systolic phase onset, systolic phase peak, DN, diastolic phase peak, and diastolic phase endpoint. The duration of a single CC may be approximately 0.8 seconds. Therefore, the duration between two consecutive key points may be around 0.2 seconds. In some embodiments, the SG family filter has parameters of^,^, and order of derivation^. Systems and methods in accordance with certain embodiments of the invention calculate the first and second derivatives of the smoothed input signal and can identify the locations of all local extrema of the first derivative^ᇱ. In many embodiments, the local extrema may be classified as either maxima or minima byexamining sign changes in the second derivative^ᇱᇱ of the smoothed input signal^.

[0036] Process 300 calculates (320) the coordinates of the smoothed input signal at the location of each of the first derivative local maxima and minima. A cubic spline interpolation can be used to connect the coordinates in the smoothed input signalcorresponding to the local maxima of the first derivative to define the upper envelope^^௩ , and correspondingly, the coordinates corresponding to the local minima of thefirst derivative can be connected with each other to obtain the lower envelope^^௩.

[0037] Process 300 calculates (330) the envelope mean value based on the obtained coordinates of the smoothed input signal at the location of each of the local maxima and minima of the first derivative. In several embodiments, the envelope mean value is calculated using the estimated upper and lower envelopes: ^^௩^ ^^௩^^

[0038] Process 300of the input by subtracting the envelope mean value from the input signal. In selected embodiments, the non- stationary part of the input can be calculated by:^ ^ ^

[0039] Process 300 determines (350) whether the resulting input signal satisfies a stopping criterion. The stopping criterion may be defined as: ଶଶ^ ^ି^ ^where E denotes^ି^. If the resultingsignal satisfies the part of the input signal. Otherwise, a new input signal is assigned using^ା^ ^, and the process may be repeated from step 310 to further attempt to estimate the non-stationary part of the input signal.

[0040] The non-stationary part of the signal may be defined as: ூ whereas the stationary part of the signal may be defined as: ூ^The stationary part maymeans obtained across all iterations of the IEM method.

[0041] While specific processes for decomposing an input signal into stationary and non-stationary parts using the IEM method are described above, any of a variety of processes can be utilized to decompose an input signal as appropriate to the requirements of specific applications. In certain embodiments, steps may be executed or performed in any order or sequence not limited to the order and sequence shown and described. In a number of embodiments, some of the above steps may be executed or performed substantially simultaneously where appropriate or in parallel to reduce latency and processing times. In some embodiments, one or more of the above steps may be omitted.

[0042] Fig. 4 illustrates the non-stationary and stationary parts of an input signal in accordance with an embodiment of the invention. DNs can be located in valleys within aCC in the non-stationary component generated by the IEM method. In numerous embodiments, the valleys can be located between the systolic peak and the diastolic endpoint (the endpoint of the CC). In some situations, there may be several valleys generated by other secondary waves caused by pressure reflections within the arterial system between the systolic peak and the diastolic endpoint in the ABP waveform. Additionally, non-physiological oscillations in both ABP and PPG waveforms can also contribute to the emergence of multiple valleys within the CC. Therefore, systems and methods in accordance with many embodiments of the invention utilize several conditions to filter the potential valleys to locate the true DNs. In many embodiments, systems only deem valleys that are at least 0.1 seconds or 25 samples away, at an example sampling frequency of 256 Hz, from the systolic peak as DNs. In selected embodiments, the selected values should have a value corresponding to this valley in the non-stationarycomponent of less than zero. The first valley that satisfies both conditions may bedesignated as the DN location. In Fig. 4, the vertical dashed lines represent the DN positions located within a CC by the IEM-based algorithm.

[0043] Several embodiments utilize a large perioperative dataset (MLORD) that include physiological waveforms collected from 17,327 patients who underwent surgeries between 2019 and 2022 at the David Geffen School of Medicine at the University of California Los Angeles to assess the performance of the methods disclosed above. The MLORD dataset includes waveform data totaling over 72,264 hours in duration and 7.6 terabytes in size. This waveform data contains continuous physiological waveforms obtained from vital sign monitoring devices, such as Electrocardiogram (ECG) and PPG, as well as data recorded by other devices, including invasive ABP monitors. The continuous waveform data was collected at a sampling rate of 256 Hz.

[0044] Out of the 17,327 patients in the MLORD dataset, 4901 patients have ABP waveforms, and 17,170 patients have PPG waveforms. A total of 352,2574-s windows from ABP waveforms including 1,171,288 full cardiac cycles, and 1100,6894-s windows from PPG waveforms representing 3424,975 full cardiac cycles, were used for the evaluation. Marking all the temporal locations of DN on all the cardiac cycles to serve as a reference was extremely difficult due to the dataset’s large size. Additionally, in certain instances, the DN or diastolic phase peak may be less distinct, especially in PPGwaveforms, making it challenging to establish their temporal location as a reference. In many embodiments, 25,000 4-second windows from each ABP and PPG waveforms where DN could be observed within a CC were selected to evaluate performance. Selected ABP windows contained a total of 87,345 CCs, while PPG windows contained 86,766 CCs. An experienced researcher visually inspected these windows and marked the temporal location of DN, as well as the onset of the systolic phase within each cardiac cycle, using the ’find_peaks’ function from the SciPy Python package.

[0045] The evaluation of the algorithms was based on the systolic phase duration, which is defined as the duration from the onset of the systolic phase to the occurrence of DN in the CC. The first minima point before the systolic peak, which is the largest peak ina CC, can be used as an onset point of the systolic phase. The calculated systolic phaseduration, based on the researcher's markings, served as a reference to assess the performance of the algorithm. To ensure accuracy, the marking was validated by an engineer and an anesthesiologist, who conducted a visual examination of the marked ABP and PPG windows.

[0046] To evaluate the performance of the IEM-based algorithm, many embodiments utilize the systolic phase duration (SPD), which is defined as the time difference between the onset of the systolic phase and the occurrence of DN in the cardiac cycle. The measured SPD, based on the marking, served as a reference. As the systolic phase onset marked by the researcher can also be used as the onset point for the SPD measured by the algorithm, any difference in SPD between the reference and the estimate from the algorithm can be attributed to differences in the location identified as the DN only.

[0047] To evaluate the algorithm's robustness in detecting DN in the presence of noise, all 352,257 4-second windows of ABP waveforms and 1,100,689 4-second windows of PPG waveforms were included in the analysis. In various embodiments, the algorithm's noise robustness was tested across a range of signal-to-noise ratios (SNR) ranging from -30 dB to -5 dB in steps of 1 dB. Firstly, each 4-second window was decomposed into its non-stationary and stationary components using the IEM Method. Secondly, DN was estimated within each CC in a 4-second window using the non- stationary component of the IEM method. Using the detected DN and the systolic onset point, the systolic phase duration was calculated and used as a reference. The non-stationary component of the IEM method was embedded in the stationary component of the IEM method, simulating varying SNR levels from -30 dB to -5 dB in steps of 1 dB. The IEM-based algorithm was then applied again, and DN was detected within each CC in a 4-second window after adding the noise. Using the detected DN and the systolic phase onset point after adding the noise, the systolic phase duration was calculated and compared with the reference systolic phase duration at a given SNR to test the noise robustness of the proposed algorithm.

[0048] The method’s performance was assessed by comparing its results with those marked by an experienced researcher. The evaluation was conducted using regressionanalysis, box plots, and the rate of detectability ( ோ). Box plots were used to visuallycompare the difference between the reference systolic phase duration (as marked by the researcher) and the measured systolic phase duration (obtained using an algorithm) between the proposed IEM-based algorithm and the existing 2ndderivative method. A systolic phase duration difference closer to zero indicates a greater accuracy in locating the position of the DN within a CC by the algorithm. The regression analysis demonstrates the ability of the proposed algorithm to track changes in the reference systolic phaseduration, whereas the rate of detectability ( ோ was used to assess the DN detectionoutcome of an algorithm.

[0049] In the context of noise robustness analysis, many embodiments use metrics, including the cross-correlation index (CCI) to measure the similarity between the non- stationary component of the IEM method before and after adding noise, rate ofdetectability ( ோ to measure the DN detection outcome of an algorithm in the presence ofnoise. Differences in systolic phase duration before and after introducing noise at a givenSNR were also computed. The Rate of Detectability ( ோ) measures the ability of the IEMbased algorithm to detect the DN within a CC.ோcan be calculated using: ^ ோ % ோ where^is the number of DNs detected by the proposed algorithm andோis the number of reference DNs.

[0050] For SPD difference and error of DN detection, the difference in SPD betweenthe reference and the measured, or between before and after scaling can be calculated using the following equation:^ ோ∨^ௌ ெ∨^ௌ, whereோ∨^ௌis the reference SPD or SPD before scaling in the context of robustness analysis andெ∨^ௌis the measured SPD by the algorithm or SPD after scaling in the context of robustness analysis. Any difference in SPD between the reference and the estimate from the algorithm or in SPD between before and after scaling may be due to differences in the location identified as the DN only. Therefore, the absolute value of the difference in SPDcan be referred to as the error of DN detection, described by the following:^ே^ .

[0051] Several embodiments use the cross-correlation index (CCI) to indicate the ability of the IEM-based algorithm to detect DN using the non-stationary component of the IEM method for a given SNR: ே^ୀ^ ^^ ^^ ^^ ^^%the noise,^^is the non-stationary component of the IEM method after adding the noise for a given SNR, n is the sample index (n = 1, 2…N), and^^and^^are the average values of the non-stationary component of the IEM method before and after adding the noise, respectively.

[0052] Figs. 5A-D illustrate the regression analysis and box plot analysis for the ABP and PPG waveforms in accordance with an embodiment of the invention. Fig. 5A illustrates a regression analysis of systolic phase duration estimated by the IEM method for the ABP waveforms. Fig. 5B illustrates a regression analysis of the systolic phase duration estimated by the IEM method for the PPG waveforms. Fig. 5C illustrates a comparison of box plots for the IEM method versus the 2nd derivative for ABP waveforms. Fig.5D illustrates a comparison of box plots for the IEM method versus the 2nd derivative for PPG waveforms.

[0053] Fig. 6 illustrates a comparison of the detection capabilities of the IEM-based method versus the 2ndderivative method in accordance with an embodiment of theinvention. The DN detection ability of the IEM-based algorithm was compared with the 2ndderivative method in terms of the rate of detectability ( ோ) and the difference in systolicphase duration (reference-measured). As illustrated in Fig. 6, both methods achieved a 100%ோfor both ABP and PPG CCs, but the IEM-based algorithm outperforms the 2ndderivative method in correctly estimating DN within a CC in terms of the difference between the reference and measured systolic phase duration. These findings clearly show that the proposed algorithm not only detects the DN within a CC but also provides a more accurate DN location compared to the 2ndderivative method in both ABP and PPG waveforms. Signal Processing Toolkit

[0054] CCs in ABPs and PPGs can display five key points that indicate the conditions of a patient. For example, the duration from SPO to DN, known as SPD in the ABP waveform, can be used to monitor cardiac function. Changes in DN-based features, such as DN amplitudes and the SPD, can often occur early in the vascular disease course and may help with early recognition. In PPG waveforms, the stiffness index, which is the ratio of the subject's height to the duration between SPP and DPP, has demonstrated predictive abilities for the risk of coronary heart diseases, such as hypertension and diabetes. Moreover, the reflection index, which is the ratio of DPP amplitude to SPP amplitude relative to SPO, may serve as a valuable indicator for vascular assessment. The augmentation index, which is the ratio of the difference between SPP and DPP amplitudes to the amplitude of SPP (with amplitudes relative to SPO), tends to increase in older individuals and those with cardiovascular disease.

[0055] In many embodiments, systems and methods include a signal processing tool based on the IEM method that can detect all key points within CC in ABP and PPG wave- forms and can be used for extracting features from these waveforms. Several embodi- ments can create a PPG feature library that includes 632 features per CC, along with simultaneous SBP, DBP, and MAP values that were extracted from corresponding CCs in the ABP waveform by extracting from the large perioperative MLORD dataset compris- ing 17,327 patients.

[0056] In various embodiments, the key points are detected using the iterativeenvelope mean method, where the IEM method is adapted for detecting the temporal location of all five key points within a CC in ABP and PPG waveforms that can be used as critical landmarks for feature extraction in these waveforms. A process for detecting key points using the IEM in accordance with an embodiment of the invention is illustrated in Fig. 7. Many embodiments implement the signal processing tool using the PYTHON programming language.

[0057] Several embodiments utilize input waveforms of 4-second windows containing an ABP or PPG waveform. Windows that exhibit artifacts may be excluded based on the following criteria: any window containing '0' or negative values and having a number of peaks less than 3 or more than 10, exceeding 75% of the window's amplitude. After the artifacts are removed, many embodiments apply a 4th-order Butterworth low-pass filter with a cutoff frequency of 16 Hz to the 4-s input window to eliminate high-frequency noise. The filtered signal may be normalized using the following: ^^^^^^^^^^where x(n) represents the index,^is the minimum value^^of the input signal,^^௫is its maximum value, and^^^^denotes the normalized input signal.

[0058] In many embodiments, the IEM method decomposes the signal into its non- stationary and stationary components. An example of applying the IEM method to a 4- second window of an ABP input signal in accordance with an embodiment is illustrated in Figs.8A-C. Initially, the input signal is smoothed and its first and second derivatives are calculated using the SG family filter, which has parameters similar to the SG family filter described above.

[0059] The upper and lower envelopes can be calculated using the coordinates of the smoothed input signal at the location of the first derivative local maxima and minima, respectively. The first derivative local extrema in accordance with various embodiments is calculated and classified as maxima and minima using the sign changes over the second derivative of the smoothed input signal. An envelop mean value may be determined by averaging the upper and lower envelopes of a smoothed input signal usingEquation 8: ^^௩^ ^^௩^^where n is the sample 2, …., N and l is the iteration number where l = 1, 2,…,

[0060] In many embodiments, the estimated envelope mean value is subtracted from the original input signal using Equation 9: ^^ ^and the resulting signal^is subsequent iterations, where^is the input signal at iteration l. process when the stopping criterion is met at iteration L, as illustrated by Equation 10: ଶଶ^ ^ି^ ^where E{.} of^. Severalି^embodiments use accuracy level .

[0061] The IEM method in accordance with many embodiments provides an estimateof the non-stationary component ( ) of the input signal after the last iteration (L),according to Equation 11. Additionally, by summing up the envelope means from eachiteration, IEM methods can yield an estimate of the stationary component ( ) of thesignal as illustrated in Equation 12. ^ ^

[0062] The non-stationary components of the input signal after applying the IEM method are illustrated in Fig.8B and 8C, respectively. While the IEM method can reveal all key points within a CC, challenges arise due to pressure reflections in the arterial system and non-physiological oscillations in PPG and ABP waveforms thatintroduce multiple valleys within the CC in NSTS that may lead to misinterpretation.

[0063] To obtain all key points within a CC in its non-stationary component, systems and methods in accordance with numerous embodiments impose the following conditions in locating the key points. DN and DPP key points in accordance with several embodiments are at least 0.1 s (25 samples) away from SPP and DPE key points. In the NSTS, the y-axis value for SPP point may be greater than zero. DPP in many embodiments is zero. SPO, DN, and DPE valleys in accordance with various embodiments are less than zero. Several embodiments use the 'find_peaks' function from the scipy PYTHON package for peak extraction. The second zero crossing after the SPP point in the NSTS may be assigned as the DPP point.

[0064] The effectiveness of the IEM method signal processing toolkit may also similarly be evaluated using the MLORD dataset as described above. In the MLORD dataset, out of 17,327 patients, 4901 patients have ABP waveforms, 17170 patients have PPG waveforms, and 4893 patients have both ABP and PPG waveforms. Various embodiments utilize the 4893 patients having both ABP and PPG waveforms where the sampling frequency of the ABP and PPG waveforms was 256 Hz. Due to the large size of the data set, 10004-s windows were randomly selected from both the ABP and PPG waveforms to assess the performance of the toolkit in terms of detecting key points within a cardiac cycle, where all five key points could be observed within a cardiac cycle.

[0065] The ABP windows contained 3420 cardiac cycles, and the PPG windows consisted of 3440 cardiac cycles. An experienced researcher marked the key points within these cardiac cycles using the 'find_peaks' function from the scipy Python package. To ensure accuracy, the marking was validated by an engineer and an anesthesiologist. They conducted a visual examination of marked ABP and PPG windows. In certain embodiments, 1,487,955 PPG cardiac cycles and an equal number of ABP cardiac cycles from 4893 patients with having both PPG and ABP waveforms in the MLORD dataset were utilized to create a feature library.

[0066] In many embodiments, three parameters are used to evaluate the performance of the signal processing toolkit in terms of key point detection. The three parametersinclude sensitivity (SE), positive predictive value (PPV), and F-score ( ^), as shown inEquations 13-15, respectively:TP refers to the number of true s for the number of false negatives, and FP for the number of false positives. SE can indicate the fraction of true key points detected by the tool compared to the key points marked by the experienced researcher. PPV refers to the fraction of points assigned as key points by the tool, which are true key points, and F1 is the harmonic mean of SE and PPV, commonly referred to as a measure of the overall performance. In selected embodiments, a threshold of 8 ms (2 samples) was used to admit the proposed tool results as TP or reject them as FP or FN.

[0067] FIG.9 illustrates a block diagram of a computing device for detecting biosignal features in accordance with an embodiment. The biosignal features detection device 900 includes a processor 905, memory 920, and a data interface 915. In the illustrated em- bodiment, memory 920 can include a DN and feature detection application 925 and pa- tient data 930. The DN and feature detection application 925 can be machine-readable instructions that can configure the processor 905 to execute the instructions, thereby de- tecting DN and other biosignal features in CCs using the IEM method based on patient data such as according to processes discussed above. In several embodiments, patient data and / or output data from the DN and feature detection application 925 can be input and / or output by data interface 915. Data interface 915 can be any of a variety of inter- faces, such as, but not limited to, removeable memory or a network interface.

[0068] Although a specific example of a computing device for detecting biosignal fea- tures is illustrated in this figure, any of a variety of computing devices can be utilized to perform processes for detecting biosignal features similar to those described herein as appropriate to the requirements of specific applications in accordance with embodiments of the invention.

[0069] Although specific methods of biosignal feature detection using the IEM method are discussed above, many different monitoring methods can be implemented in accordance with many different embodiments of the invention. It is therefore to be understood that the present invention may be practiced in ways other than specifically described, without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive. Accordingly, the scope of the invention should be determined not by the embodiments illustrated, but by the appended claims and their equivalents.

Claims

WHAT IS CLAIMED IS:

1. A method for detecting dicrotic notches (DNs), the method comprising: receiving an input signal; pre-processing the received input signal; decomposing the pre-processed input signal into non-stationary and stationary parts using an iterative envelope method; and retrieving the DNs based on the non-stationary part of the input signal.

2. The method of claim 1, wherein the decomposing of the pre-processed input signal further comprises: calculating a smoothed input signal; calculating a first derivative of the smoothed input signal; calculating the coordinates of the smoothed input signal at the location of each local maxima and minima of the first derivative of the smoothed input signal; calculating an envelope mean value based on an upper envelope and a lower envelope of the smoothed input signal; and estimating the non-stationary part of the input by subtracting envelope mean value from the input signal.

3. The method of claim 2, wherein the smoothed input signal is calculated with a Savizky-Golay (SG) family filter.

4. The method of claim 2, wherein the upper envelope is obtained by applying a cubic spline interpolation to connect the coordinates of the local maxima of the first derivative in the smoothed input signal.

5. The method of claim 2, wherein the lower envelope is obtained by applying a cubic spline interpolation to connect the coordinates of the local minima of the first derivative in the smoothed input signal.

6. The method of claim 1, wherein the input signal comprises an ABP waveform or a PPG waveform of cardiac cycles.

7. The method of claim 1, wherein the input signal comprises 4-second windows of ABP or PPG waveforms.

8. The method of claim 7, wherein each 4-second window is filtered using a 4th-order Butterworth low pass filter with a cut-off frequency of 16 Hz to remove high-frequency noise.

9. The method of claim 1, wherein pre-processing the received input comprises excluding 4-second windows that satisfies at least one criterion selected from the group of: the received input having zero or negative values; the received input having three or fewer peaks exceeding 75% of the window's amplitude; and the received input having more than ten peaks.

10. The method of claim 1, further comprising imposing a plurality of requirements to the non-stationary part of the input signal to identify all key points within a CC.

11. A non-transitory machine readable medium containing processor instructions for detecting dicrotic notches (DNs), where execution of the instructions by a processor causes the processor to perform a process that comprises: receiving an input signal; pre-processing the received input signal; decomposing the pre-processed input signal into non-stationary and stationary parts using an iterative envelope method; and retrieving the DNs based on the non-stationary part of the input signal.

12. The non-transitory machine readable medium of claim 11, wherein the decomposing of the pre-processed input signal further comprises: calculating a smoothed input signal; calculating a first derivative of the smoothed input signal; calculating the coordinates of the smoothed input signal at the location of each local maxima and minima of the first derivative of the smoothed input signal. calculating an envelope mean value based on an upper envelope and a lower envelope of the smoothed input signal; and estimating the non-stationary part of the input by subtracting envelope mean value from the input signal.

13. The non-transitory machine readable medium of claim 12, wherein the smoothed input signal is calculated with a Savizky-Golay (SG) family filter.

14. The non-transitory machine readable medium of claim 12, wherein the upper envelope is obtained by applying a cubic spline interpolation to connect the coordinates of the local maxima of the first derivative in the smoothed input signal.

15. The non-transitory machine readable medium of claim 12, wherein the lower envelope is obtained by applying a cubic spline interpolation to connect the coordinates of the local minima of the first derivative in the smoothed input signal.

16. The non-transitory machine readable medium of claim 11, wherein the input signal comprises an ABP waveform or a PPG waveform of cardiac cycles.

17. The non-transitory machine readable medium of claim 11, wherein the input signal comprises 4-second windows of ABP or PPG waveforms.

18. The non-transitory machine readable medium of claim 17, wherein each 4-second window is filtered using a 4th-order Butterworth low pass filter with a cut-off frequency of 16 Hz to remove high-frequency noise.

19. The non-transitory machine readable medium of claim 11, wherein pre-processing the received input comprises excluding 4-second windows that satisfies at least one criterion selected from the group of: the received input having zero or negative values; the received input having three or fewer peaks exceeding 75% of the window's amplitude; and the received input having more than ten peaks.

20. The non-transitory machine readable medium of claim 11, further comprising imposing a plurality of requirements to the non-stationary part of the input signal to identify all key points within a CC.

Images