More Like this

1. Assessing Replication of Pyrogenic Organic Compounds in Coeval Tropical Stalagmites

   Denniston, R.F. ; Azenon, J. ; Argiriadis, E. ; Ondei, S. ; Genuzio, G. ; Baltieri, M. ; Nguyen, H. ; Thompson, J. ( January 2022 , Transactions American Geophysical Union)

   Polycyclic aromatic hydrocarbons (PAHs) are produced by the burning of biomass, with molecular weights reflecting combustion conditions. After being formed, PAHs are transported downward through soil and bedrock by infiltrating rainwater (Perrette et al., 2013), and in karst areas can become incorporated into stalagmites as they crystallize from dripwater in underlying caves (Perrette et al., 2008; Denniston et al., 2018). Thus, when stalagmite growth is high, infiltration times short, and fluid mixing minimized, there exists the potential for PAHs in stalagmites to preserve evidence of the presence and intensity of fire through time. We have previously reported a high-resolution analysis of PAH distributions in two non-overlapping aragonite stalagmites from cave KNI-51, tropical Western Australia, that together span the majority of the last 900 years. The geologic conditions of this site make it well suited for the transmission of discrete pulses of fire-derived compounds from the land surface to the stalagmite. Soils are thin to absent above the stalagmite chamber and the cave is shallow. As a result, homogenization of infiltrated water (and thus PAHs) is expected to be small on interannual time scales. In addition, intense summer monsoon rains flush fire debris from the hillsides over the cave. These characteristics, coupled with the fast growth rates (1-2 mm/yr) and precise radiometric dates (±1-30 years 2 s.d. over the last millennium) of KNI-51 stalagmites suggest that they hold the potential for extremely high resolution paleofire reconstruction. Here we provide the first test of replication of PAH abundances, ratios, and trends in coeval stalagmites. Samples were analyzed at Ca’ Foscari University using methods of Argiriadis et al. (2019) and the results validated by comparing them with fire activity detected through satellite images. Stalagmites KNI-51-F and -G overlap in age from CE 1310-1630, allowing an examination of the consistency of the PAH signal along different infiltration pathways. References Argiriadis, E. et al. (2019) European Geosciences Union Annual Meeting, Vienna, Australia. Denniston, R.F. et al. (2018) American Geophysical Union Annual Meeting, Washington, D.C. Perrette, Y. et al. (2008) Chemical Geology, 251, 67-76. Perrette, Y. et al. (2013) Organic Geochemistry, 65, 37-45. 

   more » « less
2. Speleothem organic biomarkers trace last millennium fire history at near-annual resolution in northwestern Australia

   Argiriadis, E. ; Denniston, R.F. ; Ondei, S. ; Bowman, D. ( January 2023 , European Geosciences Union)

   Recent developments in speleothem science are showing their potential for paleofire reconstruction through a variety of inorganic and organic proxies including trace metals (1) and the pyrogenic organic compound levoglucosan (2). Previous work by Argiriadis et al. (2019) presented a method for the analysis of trace polycyclic aromatic hydrocarbons (PAHs) and n -alkanes in stalagmites (3). These compounds reflect biogeochemical processes occurring at the land surface, in the soil, and in the cave. PAHs are primarily related to combustion of biomass while n-alkanes, with their potential for vegetation reconstruction (4), provide information on fuel availability and composition, as well as fire activity. These organic molecules are carried downward by infiltrating water and incorporated into speleothems (5), thereby creating the potential to serve as novel paleofire archives. Using this approach, we developed a high-resolution stalagmite record of paleofire activity from cave KNI-51 in tropical northwestern Australia. This site is well suited for high resolution paleofire reconstruction as bushfire activity in this tropical savanna is some of the highest on the continent, the cave is shallow and overlain by extremely thin soils, and the stalagmites are fast-growing (1-2 mm yr-1) and precisely dated. We analyzed three stalagmites which grew continuously in different time intervals through the last millennium - KNI-51-F (CE ~1100-1620), KNI-51-G (CE ~1320-1640), and KNI-51-11 (CE ~1750-2009). Samples were drilled continuously at 1-3 mm resolution from stalagmite slabs, processed in a stainless-steel cleanroom to prevent contamination. Despite a difference in resolution between stalagmites KNI-51-F and -G, peaks in the target compounds show good replication in the overlapping time interval of the two stalagmites, and PAH abundances in a portion of stalagmite KNI-51-11 that grew from CE 2000-2009 are well correlated with satellite-mapped fires occurring proximally to the cave. Our results suggest an increase in the frequency of low intensity fire in the 20th century relative to much of the previous millennium. The timing of this shift is broadly coincident with the arrival of European pastoralists in the late 19th century and the subsequent displacement of Aboriginal peoples from the land. Aboriginal peoples had previously utilized “fire stick farming”, a method of prescribed, low intensity burning, that was an important influence of ecology, biomass, and fire. Prior to the late 1800s, the period with the most frequent low intensity fire activity was the 13th century, the wettest interval of the entire record. Peak high intensity fire activity occurred during the 12th century. Controlled burn and irrigation experiments capable of examining the transmission of pyrogenic compounds from the land surface to cave dripwater represent the next step in this analysis. Given that karst is present in many fire-prone environments, and that stalagmites can be precisely dated and grow continuously for millennia, the potential utility of a stalagmite-based paleofire proxy is high. (1) L.K. McDonough et al., Geochim. Cosmochim. Acta. 325, 258–277 (2022). (2) J. Homann et al., Nat. Commun., 13:7175 (2022). (3) E. Argiriadis et al., Anal. Chem. 91, 7007–7011 (2019). (4) R.T. Bush, F. A. McInerney, Geochim. Cosmochim. Acta. 117, 161–179 (2013). (5) Y. Sun et al., Chemosphere. 230, 616–627 (2019). 

   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. Geochemical Evidence for the Control of Fire by Middle Palaeolithic Hominins

   https://doi.org/10.1038/s41598-019-51433-0  

   Brittingham, Alex ; Hren, Michael T. ; Hartman, Gideon ; Wilkinson, Keith N. ; Mallol, Carolina ; Gasparyan, Boris ; Adler, Daniel S. ( October 2019 , Scientific Reports)

   The use of fire played an important role in the social and technological development of the genusHomo. Most archaeologists agree that this was a multi-stage process, beginning with the exploitation of natural fires and ending with the ability to create fire from scratch. Some have argued that in the Middle Palaeolithic (MP) hominin fire use was limited by the availability of fire in the landscape. Here, we present a record of the abundance of polycyclic aromatic hydrocarbons (PAHs), organic compounds that are produced during the combustion of organic material, from Lusakert Cave, a MP site in Armenia. We find no correlation between the abundance of light PAHs (3–4 rings), which are a major component of wildfire PAH emissions and are shown to disperse widely during fire events, and heavy PAHs (5–6 rings), which are a major component of particulate emissions of burned wood. Instead, we find heavy PAHs correlate with MP artifact density at the site. Given that hPAH abundance correlates with occupation intensity rather than lPAH abundance, we argue that MP hominins were able to control fire and utilize it regardless of the variability of fires in the environment. Together with other studies on MP fire use, these results suggest that the ability of hominins to manipulate fire independent of exploitation of wildfires was spatially variable in the MP and may have developed multiple times in the genusHomo.

    

   more » « less
5. Climate variability in the northern Levant from the highly resolved Qadisha record (Lebanon) during the Holocene optimum

   https://doi.org/10.1017/qua.2023.24  

   Nehme, Carole ; Verheyden, Sophie ; Kluge, Tobias ; Nader, Fadi H. ; Edwards, R. Lawrence ; Cheng, Hai ; Eiche, Elisabeth ; Claeys, Philippe ( January 2023 , Quaternary Research)

   Abstract New stalagmites from Qadisha Cave (Lebanon) located at 1720 m above sea level provide a high-resolution and well-dated record for northern Mount Lebanon. The stalagmites grew discontinuously from 9.2 to 5.7 and at 3.5 ka, and they show a tendency to move from a more negative oxygen isotope signal at ~9.1 ka to a more positive signal at ~5.8 ka. Such a trend reflects a change from a wetter to a drier climate at high altitudes. The δ 13 C signal shows rapid shifts throughout the record and a decreasing trend toward more negative values in the mid-Holocene, suggesting enhanced soil activity. In the short-term trend, Qadisha stalagmites record rapid dry/wet changes on centennial scales, with a tendency to more rapid dry events toward the mid-Holocene. Such changes are characterized by overall good agreement between both geochemical proxies and stalagmite growth and might be affected by the seasonal variations in snow cover. The Qadisha record is in good agreement with other Levantine records, showing more humid conditions from 9 to 7 ka. After 7 ka, a drier climate seems to affect sites at both low- and high-altitude areas. The Qadisha record reflects uniquely mountainous climate characteristics compared with other records, specifically the effect of snow cover and its duration regulating the effective infiltration. 

   more » « less