More Like this

1. Bladderworts, the smallest known suction feeders, generate inertia‐dominated flows to capture prey

   https://doi.org/10.1111/nph.16726  

   Müller, Ulrike K. ; Berg, Otto ; Schwaner, Janneke M. ; Brown, Matthew D. ; Li, Gen ; Voesenek, Cees J. ; van Leeuwen, Johan L. ( July 2020 , New Phytologist)

   Aquatic bladderworts (Utricularia gibba and U. australis) capture zooplankton in mechanically triggered underwater traps. With characteristic dimensions less than 1 mm, the trapping structures are among the smallest known to capture prey by suction, a mechanism that is not effective in the creeping‐flow regime where viscous forces prevent the generation of fast and energy‐efficient suction flows.

   To understand what makes suction feeding possible on the small scale of bladderwort traps, we characterised their suction flows experimentally (using particle image velocimetry) and mathematically (using computational fluid dynamics and analytical mathematical models).

   We show that bladderwort traps avoid the adverse effects of creeping flow by generating strong, fast‐onset suction pressures. Our findings suggest that traps use three morphological adaptations: the trap walls' fast release of elastic energy ensures strong and constant suction pressure; the trap door's fast opening ensures effectively instantaneous onset of suction; the short channel leading into the trap ensures undeveloped flow, which maintains a wide effective channel diameter.

   Bladderwort traps generate much stronger suction flows than larval fish with similar gape sizes because of the traps' considerably stronger suction pressures. However, bladderworts' ability to generate strong suction flows comes at considerable energetic expense.

    

   more » « less
2. Thermodynamics of the Bladderwort Feeding Strike—Suction Power from Elastic Energy Storage

   https://doi.org/10.1093/icb/icz144  

   Berg, Otto ; Singh, Krizma ; Hall, Maxwell R. ; Schwaner, M. Janneke ; Müller, Ulrike K. ( August 2019 , Integrative and Comparative Biology)

   The carnivorous plant bladderwort exemplifies the use of accumulated elastic energy to power motion: respiration-driven pumps slowly load the walls of its suction traps with elastic energy (∼1 h). During a feeding strike, this energy is released suddenly to accelerate water (∼1 ms). However, due to the traps’ small size and concomitant low Reynolds number, a significant fraction of the stored energy may be dissipated as viscous friction. Such losses and the mechanical reversibility of Stokes flow are thought to degrade the feeding success of other suction feeders in this size range, such as larval fish. In contrast, triggered bladderwort traps are generally successful. By mapping the energy budget of a bladderwort feeding strike, we illustrate how this smallest of suction feeders can perform like an adult fish.

    

   more » « less
3. A Deep Learning-Based Real-time Seizure Detection System

   https://doi.org/10.1109/SPMB50085.2020.9353623  

   Shawki, N. ; Elseify, T. ; Cap, T. ; Shah, V. ; Obeid, I. ; Picone, J. ( December 2020 , Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium (SPMB))

   Obeid, Iyad Selesnick (Ed.)

   Electroencephalography (EEG) is a popular clinical monitoring tool used for diagnosing brain-related disorders such as epilepsy [1]. As monitoring EEGs in a critical-care setting is an expensive and tedious task, there is a great interest in developing real-time EEG monitoring tools to improve patient care quality and efficiency [2]. However, clinicians require automatic seizure detection tools that provide decisions with at least 75% sensitivity and less than 1 false alarm (FA) per 24 hours [3]. Some commercial tools recently claim to reach such performance levels, including the Olympic Brainz Monitor [4] and Persyst 14 [5]. In this abstract, we describe our efforts to transform a high-performance offline seizure detection system [3] into a low latency real-time or online seizure detection system. An overview of the system is shown in Figure 1. The main difference between an online versus offline system is that an online system should always be causal and has minimum latency which is often defined by domain experts. The offline system, shown in Figure 2, uses two phases of deep learning models with postprocessing [3]. The channel-based long short term memory (LSTM) model (Phase 1 or P1) processes linear frequency cepstral coefficients (LFCC) [6] features from each EEG channel separately. We use the hypotheses generated by the P1 model and create additional features that carry information about the detected events and their confidence. The P2 model uses these additional features and the LFCC features to learn the temporal and spatial aspects of the EEG signals using a hybrid convolutional neural network (CNN) and LSTM model. Finally, Phase 3 aggregates the results from both P1 and P2 before applying a final postprocessing step. The online system implements Phase 1 by taking advantage of the Linux piping mechanism, multithreading techniques, and multi-core processors. To convert Phase 1 into an online system, we divide the system into five major modules: signal preprocessor, feature extractor, event decoder, postprocessor, and visualizer. The system reads 0.1-second frames from each EEG channel and sends them to the feature extractor and the visualizer. The feature extractor generates LFCC features in real time from the streaming EEG signal. Next, the system computes seizure and background probabilities using a channel-based LSTM model and applies a postprocessor to aggregate the detected events across channels. The system then displays the EEG signal and the decisions simultaneously using a visualization module. The online system uses C++, Python, TensorFlow, and PyQtGraph in its implementation. The online system accepts streamed EEG data sampled at 250 Hz as input. The system begins processing the EEG signal by applying a TCP montage [8]. Depending on the type of the montage, the EEG signal can have either 22 or 20 channels. To enable the online operation, we send 0.1-second (25 samples) length frames from each channel of the streamed EEG signal to the feature extractor and the visualizer. Feature extraction is performed sequentially on each channel. The signal preprocessor writes the sample frames into two streams to facilitate these modules. In the first stream, the feature extractor receives the signals using stdin. In parallel, as a second stream, the visualizer shares a user-defined file with the signal preprocessor. This user-defined file holds raw signal information as a buffer for the visualizer. The signal preprocessor writes into the file while the visualizer reads from it. Reading and writing into the same file poses a challenge. The visualizer can start reading while the signal preprocessor is writing into it. To resolve this issue, we utilize a file locking mechanism in the signal preprocessor and visualizer. Each of the processes temporarily locks the file, performs its operation, releases the lock, and tries to obtain the lock after a waiting period. The file locking mechanism ensures that only one process can access the file by prohibiting other processes from reading or writing while one process is modifying the file [9]. The feature extractor uses circular buffers to save 0.3 seconds or 75 samples from each channel for extracting 0.2-second or 50-sample long center-aligned windows. The module generates 8 absolute LFCC features where the zeroth cepstral coefficient is replaced by a temporal domain energy term. For extracting the rest of the features, three pipelines are used. The differential energy feature is calculated in a 0.9-second absolute feature window with a frame size of 0.1 seconds. The difference between the maximum and minimum temporal energy terms is calculated in this range. Then, the first derivative or the delta features are calculated using another 0.9-second window. Finally, the second derivative or delta-delta features are calculated using a 0.3-second window [6]. The differential energy for the delta-delta features is not included. In total, we extract 26 features from the raw sample windows which add 1.1 seconds of delay to the system. We used the Temple University Hospital Seizure Database (TUSZ) v1.2.1 for developing the online system [10]. The statistics for this dataset are shown in Table 1. A channel-based LSTM model was trained using the features derived from the train set using the online feature extractor module. A window-based normalization technique was applied to those features. In the offline model, we scale features by normalizing using the maximum absolute value of a channel [11] before applying a sliding window approach. Since the online system has access to a limited amount of data, we normalize based on the observed window. The model uses the feature vectors with a frame size of 1 second and a window size of 7 seconds. We evaluated the model using the offline P1 postprocessor to determine the efficacy of the delayed features and the window-based normalization technique. As shown by the results of experiments 1 and 4 in Table 2, these changes give us a comparable performance to the offline model. The online event decoder module utilizes this trained model for computing probabilities for the seizure and background classes. These posteriors are then postprocessed to remove spurious detections. The online postprocessor receives and saves 8 seconds of class posteriors in a buffer for further processing. It applies multiple heuristic filters (e.g., probability threshold) to make an overall decision by combining events across the channels. These filters evaluate the average confidence, the duration of a seizure, and the channels where the seizures were observed. The postprocessor delivers the label and confidence to the visualizer. The visualizer starts to display the signal as soon as it gets access to the signal file, as shown in Figure 1 using the “Signal File” and “Visualizer” blocks. Once the visualizer receives the label and confidence for the latest epoch from the postprocessor, it overlays the decision and color codes that epoch. The visualizer uses red for seizure with the label SEIZ and green for the background class with the label BCKG. Once the streaming finishes, the system saves three files: a signal file in which the sample frames are saved in the order they were streamed, a time segmented event (TSE) file with the overall decisions and confidences, and a hypotheses (HYP) file that saves the label and confidence for each epoch. The user can plot the signal and decisions using the signal and HYP files with only the visualizer by enabling appropriate options. For comparing the performance of different stages of development, we used the test set of TUSZ v1.2.1 database. It contains 1015 EEG records of varying duration. The any-overlap performance [12] of the overall system shown in Figure 2 is 40.29% sensitivity with 5.77 FAs per 24 hours. For comparison, the previous state-of-the-art model developed on this database performed at 30.71% sensitivity with 6.77 FAs per 24 hours [3]. The individual performances of the deep learning phases are as follows: Phase 1’s (P1) performance is 39.46% sensitivity and 11.62 FAs per 24 hours, and Phase 2 detects seizures with 41.16% sensitivity and 11.69 FAs per 24 hours. We trained an LSTM model with the delayed features and the window-based normalization technique for developing the online system. Using the offline decoder and postprocessor, the model performed at 36.23% sensitivity with 9.52 FAs per 24 hours. The trained model was then evaluated with the online modules. The current performance of the overall online system is 45.80% sensitivity with 28.14 FAs per 24 hours. Table 2 summarizes the performances of these systems. The performance of the online system deviates from the offline P1 model because the online postprocessor fails to combine the events as the seizure probability fluctuates during an event. The modules in the online system add a total of 11.1 seconds of delay for processing each second of the data, as shown in Figure 3. In practice, we also count the time for loading the model and starting the visualizer block. When we consider these facts, the system consumes 15 seconds to display the first hypothesis. The system detects seizure onsets with an average latency of 15 seconds. Implementing an automatic seizure detection model in real time is not trivial. We used a variety of techniques such as the file locking mechanism, multithreading, circular buffers, real-time event decoding, and signal-decision plotting to realize the system. A video demonstrating the system is available at: https://www.isip.piconepress.com/projects/nsf_pfi_tt/resources/videos/realtime_eeg_analysis/v2.5.1/video_2.5.1.mp4. The final conference submission will include a more detailed analysis of the online performance of each module. ACKNOWLEDGMENTS Research reported in this publication was most recently supported by the National Science Foundation Partnership for Innovation award number IIP-1827565 and the Pennsylvania Commonwealth Universal Research Enhancement Program (PA CURE). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the official views of any of these organizations. REFERENCES [1] A. Craik, Y. He, and J. L. Contreras-Vidal, “Deep learning for electroencephalogram (EEG) classification tasks: a review,” J. Neural Eng., vol. 16, no. 3, p. 031001, 2019. https://doi.org/10.1088/1741-2552/ab0ab5. [2] A. C. Bridi, T. Q. Louro, and R. C. L. Da Silva, “Clinical Alarms in intensive care: implications of alarm fatigue for the safety of patients,” Rev. Lat. Am. Enfermagem, vol. 22, no. 6, p. 1034, 2014. https://doi.org/10.1590/0104-1169.3488.2513. [3] M. Golmohammadi, V. Shah, I. Obeid, and J. Picone, “Deep Learning Approaches for Automatic Seizure Detection from Scalp Electroencephalograms,” in Signal Processing in Medicine and Biology: Emerging Trends in Research and Applications, 1st ed., I. Obeid, I. Selesnick, and J. Picone, Eds. New York, New York, USA: Springer, 2020, pp. 233–274. https://doi.org/10.1007/978-3-030-36844-9_8. [4] “CFM Olympic Brainz Monitor.” [Online]. Available: https://newborncare.natus.com/products-services/newborn-care-products/newborn-brain-injury/cfm-olympic-brainz-monitor. [Accessed: 17-Jul-2020]. [5] M. L. Scheuer, S. B. Wilson, A. Antony, G. Ghearing, A. Urban, and A. I. Bagic, “Seizure Detection: Interreader Agreement and Detection Algorithm Assessments Using a Large Dataset,” J. Clin. Neurophysiol., 2020. https://doi.org/10.1097/WNP.0000000000000709. [6] A. Harati, M. Golmohammadi, S. Lopez, I. Obeid, and J. Picone, “Improved EEG Event Classification Using Differential Energy,” in Proceedings of the IEEE Signal Processing in Medicine and Biology Symposium, 2015, pp. 1–4. https://doi.org/10.1109/SPMB.2015.7405421. [7] V. Shah, C. Campbell, I. Obeid, and J. Picone, “Improved Spatio-Temporal Modeling in Automated Seizure Detection using Channel-Dependent Posteriors,” Neurocomputing, 2021. [8] W. Tatum, A. Husain, S. Benbadis, and P. Kaplan, Handbook of EEG Interpretation. New York City, New York, USA: Demos Medical Publishing, 2007. [9] D. P. Bovet and C. Marco, Understanding the Linux Kernel, 3rd ed. O’Reilly Media, Inc., 2005. https://www.oreilly.com/library/view/understanding-the-linux/0596005652/. [10] V. Shah et al., “The Temple University Hospital Seizure Detection Corpus,” Front. Neuroinform., vol. 12, pp. 1–6, 2018. https://doi.org/10.3389/fninf.2018.00083. [11] F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” J. Mach. Learn. Res., vol. 12, pp. 2825–2830, 2011. https://dl.acm.org/doi/10.5555/1953048.2078195. [12] J. Gotman, D. Flanagan, J. Zhang, and B. Rosenblatt, “Automatic seizure detection in the newborn: Methods and initial evaluation,” Electroencephalogr. Clin. Neurophysiol., vol. 103, no. 3, pp. 356–362, 1997. https://doi.org/10.1016/S0013-4694(97)00003-9. 

   more » « less
4. Laboratory Wave and Stress Measurements Quantify the Aerodynamic Sheltering in Extreme Winds

   https://doi.org/10.1029/2022JC019505  

   Tan, Peisen ; Smith, Andrew W. ; Curcic, Milan ; Haus, Brian K. ( April 2023 , Journal of Geophysical Research: Oceans)

   In strong winds, air flow detaches from the ocean surface in the lee of wave crests and creates a low‐pressure zone on the wave’s leeward face. The pressure difference between the wave’s rear and front face modulates the momentum input from wind to waves. Numerical wave models parameterize this effect using a so‐called sheltering coefficient. However, its value and dependence on wind speed are not well understood, particularly with background swell waves. To bridge this gap, we conducted laboratory experiments with winds up to Category 4 hurricane force blown over various mechanically generated wave conditions (pure wind sea, mixed waves with directional spreading, and monochromatic unidirectional waves) and measured the wind, waves, and stress at a sufficient frequency to resolve wind‐wave variability over the long‐wave phase. We analyze the results in the context of Jeffreys’s sheltering theory and find two regimes: (a) from low‐to‐moderate wind forcing (10 m s⁻¹ < U₁₀ < 33 m s⁻¹), the aerodynamic sheltering increases with wind speed, consistent with previous studies; (b) in hurricane conditions (U₁₀ > 33 m s⁻¹), the aerodynamic sheltering decreases with wind at a rate depending on wave state. Further, we isolate the short wind waves from the longer paddle waves and find that the aerodynamic sheltering by longer waves leads to a phase‐dependent variability of the short wind‐waves’ local steepness, which is evidenced by the sheltering coefficient’s value derived from wind and wave measurements. Our results emphasize the need for further measurements of aerodynamic sheltering and improving its representation in models.

    

   more » « less
5. One‐Photon Excitation Followed by a Three‐Step Sequential Energy–Energy–Electron Transfer Leading to a Charge‐Separated State in a Supramolecular Tetrad Featuring Benzothiazole–Boron‐Dipyrromethene–Zinc Porphyrin–C 60

   https://doi.org/10.1002/chem.202004262  

   Badgurjar, Deepak ; Seetharaman, Sairaman ; D'Souza, Francis ; Chitta, Raghu ( December 2020 , Chemistry – A European Journal)

   A panchromatic triad, consisting of benzothiazole (BTZ) and BF₂‐chelated boron‐dipyrromethene (BODIPY) moieties covalently linked to a zinc porphyrin (ZnP) core, has been synthesized and systematically characterized by using¹H NMR spectroscopy, ESI‐MS, UV‐visible, steady‐state fluorescence, electrochemical, and femtosecond transient absorption techniques. The absorption band of the triad, BTZ‐BODIPY‐ZnP, and dyads, BTZ‐BODIPY and BODIPY‐ZnP, along with the reference compounds BTZ‐OMe, BODIPY‐OMe, and ZnP‐OMe exhibited characteristic bands corresponding to individual chromophores. Electrochemical measurements on BTZ‐BODIPY‐ZnP exhibited redox behavior similar to that of the reference compounds. Upon selective excitation of BTZ (≈290 nm) in the BTZ‐BODIPY‐ZnP triad, the fluorescence of the BTZ moiety is quenched, due to photoinduced energy transfer (PEnT) from¹BTZ^*to the BODIPY moiety, followed by quenching of the BODIPY emission due to sequential PEnT from the¹BODIPY* moiety to ZnP, resulting in the appearance of the ZnP emission, indicating the occurrence of a two‐step singlet–singlet energy transfer. Further, a supramolecular tetrad, BTZ‐BODIPY‐ZnP:ImC₆₀, was formed by axially coordinating the triad with imidazole‐appended fulleropyrrolidine (ImC₆₀), and parallel steady‐state measurements displayed the diminished emission of ZnP, which clearly indicated the occurrence of photoinduced electron transfer (PET) from¹ZnP* to ImC₆₀. Finally, femtosecond transient absorption spectral studies provided evidence for the sequential occurrence of PEnT and PET events, namely,¹BTZ^*‐BODIPY‐ZnP:ImC₆₀→BTZ‐¹BODIPY^*‐ZnP:ImC₆₀→BTZ‐BODIPY‐¹ZnP^*:ImC₆₀→BTZ‐BODIPY‐ZnP^.+:ImC₆₀^.−in the supramolecular tetrad. The evaluated rate of energy transfer,k_EnT, was found to be 3–5×10¹⁰ s⁻¹, which was slightly faster than that observed in the case of BODIPY‐ZnP and BTZ‐BODIPY‐ZnP, lacking the coordinated ImC₆₀. The rate constants for charge separation and recombination,k_CSandk_CR, respectively, calculated by monitoring the rise and decay of C₆₀^.−were found to be 5.5×10¹⁰and 4.4×10⁸ s⁻¹, respectively, for the BODIPY‐ZnP:ImC₆₀triad, and 3.1×10¹⁰and 4.9×10⁸ s⁻¹, respectively, for the BTZ‐BODIPY‐ZnP:ImC₆₀tetrad. Initial excitation of the tetrad, promoting two‐step energy transfer and a final electron‐transfer event, has been successfully demonstrated in the present study.

    

   more » « less