More Like this

1. Modeling in Science Education: A Synthesis of Recent Discovery Research PreK-12 Projects

   Margolin, J. ; Pinerua, I. ; Gerdeman, D. ( April 2022 , American Institutes for Research)

   The report summarizes the results from recent research and development projects that focused on modeling and simulations in science education. The Discovery Research PreK-12 (DRK-12) program of the National Science Foundation funded these projects as part of its mission to support the teaching and learning of science, technology, engineering, and mathematics (STEM) in grades PreK12 through innovative educational approaches.1 This report synthesizes findings from 33 articles produced by 18 DRK-12 grants awarded from 2011 to 2015, all of which funded development of resources or instructional practices to support student modeling in PreK-12 science education. This synthesis had two broad purposes: to describe 18 modeling-focused DRK-12 projects with respect to the resources they studied and the methods they used, and to summarize the new knowledge these projects produced related to modeling instruction. Link to PDF: https://www.air.org/sites/default/files/2022-05/Modeling-in-Scientific-Education-Synthesis-April-2022.pdf 

   more » « less
2. Mathematical and Scientific Argumentation in PreK-12: A Cross-Disciplinary Synthesis of Recent DRK-12 Projects

   Witherspoon, E. ; Miller, D. I. ; Pinerua, I. ; Gerdeman, D. ( April 2022 , American Institutes for Research)

   Argumentation is a core disciplinary practice in mathematics and science that is important for both content understanding and everyday reasoning. In this report, we investigate how the National Science Foundation’s (NSF’s) recent research investments have advanced understanding and supported the development of interventions that improve the teaching and learning of argumentation in mathematics and science education. In the 5 years spanning 2011 to 2015, NSF’s Discovery Research PreK–12 (DRK-12) program funded or cofunded 23 projects relating to argumentation, with more than $40 million awarded. These 23 DRK-12 projects primarily focused on argumentation in high school and middle school and applied correlational/observational and longitudinal methods (rather than quasiexperimental or experimental methods), often reporting on the design and implementation of technological supports for the teaching and learning of argumentation. Our synthesis of empirical findings focused on how these projects studied both teacher- and student-facing interventions that improved the teaching and learning of argumentation, as well as naturalistic observations of argumentation in classroom settings that helped inform the design and development of future argumentation interventions. Link to PDF: https://www.air.org/sites/default/files/2022-05/Mathematical-and-Scientific-Argumentation-in-PreK-12-April-2022.pdf 

   more » « less
3. Teachers’ Pedagogical Content Knowledge in Mathematics and Science: A Cross-Disciplinary Synthesis of Recent DRK-12 Projects

   Miller, D. I. ; Pinerua, I. ; Margolin, J. ; Gerdeman, D. ( April 2022 , American Institutes for Research)

   Teachers’ pedagogical content knowledge (PCK) is a complex, multifaceted construct that is widely seen as foundational to the act of teaching. In this synthesis, we investigated how the National Science Foundation’s (NSF’s) recent research investments have advanced understanding and supported the development of teachers’ PCK in PK–12 mathematics and science education. In the 5 years from 2011 to 2015, NSF’s Discovery Research PK–12 program (DRK-12) funded or cofunded 27 projects relating to PCK, totaling $62 million awarded. These 27 DRK-12 projects primarily applied correlational/observational and longitudinal methods (rather than quasi-experimental or experimental methods), often targeting teaching in the middle school grades. Our synthesis of empirical findings focused on how these projects studied PCK, including its measurement, development, and relationship to teaching and student learning. Link to PDF: https://www.air.org/sites/default/files/2022-05/Teachers-Pedagogical-Content-Knowledge-in-Math-and-Science-April-2022.pdf 

   more » « less
4. 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
5. Impacts of the National Science Foundation-funded Mentor-Connect Project on Two-Year Colleges

   Craft, Elaine L. ; Hata, David M. ; Dewitt, Emery ; Ritchie, Liesel ; Campbell, Nnenia ; Vickery, Jamie ( July 2020 , ASEE annual conference)

   Applying for grants from the National Science Foundation (NSF) requires a paradigm shift at many community and technical colleges, because the primary emphasis at two-year colleges is on teaching. This shift is necessary because of the NSF expectation that a STEM faculty member will lead the project as Principal Investigator. Preparing successful NSF grant proposals also requires knowledge, skills, and strategies that differ from other sources from which two-year colleges seek grant funding. Since 2012, the Mentor-Connect project has been working to build capacity among two-year colleges and leadership skills among their STEM faculty to help them prepare competitive grant proposals for the National Science Foundation’s Advanced Technological Education (NSF-ATE) program. NSF-ATE focuses on improving the education of technicians for advanced technology fields that drive the nation’s economy. As an NSF-ATE-funded initiative, Mentor-Connect has developed a three-pronged approach of mentoring, technical assistance, and digital resources to help potential grantees with the complexities of the proposal submission process. Grant funding makes it possible to provide this help at no cost to eligible, two-year college educators. Mentor-Connect support services for prospective grantees are available for those who are new to ATE (community or technical colleges that have not received an NSF ATE award in 7 or more years), those seeking a larger second grant from the ATE Program after completing a small, new-to-ATE project, and for those whose first or second grant proposal submission to the NSF ATE Program was declined (not funded). The Mentor-Connect project has succeeded in raising interest in the NSF-ATE program. Over a seven-year period more than 80% of the 143 participating colleges have submitted proposals. Overall, the funding rate among colleges that participated in the Mentor-Connect project is exceptionally high. Of the 97 New-to-ATE proposals submitted from Cohorts 1 through 6, 71 have been funded, for a funding rate of 73%. Mentor-Connect is also contributing to a more geographically and demographically diverse NSF-ATE program. To analyze longer-term impacts, the project’s evaluator is conducting campus site visits at the new-to-ATE grantee institutions as their initial ATE projects are being completed. A third-party researcher has contributed to the site-visit protocol being used by evaluators. The researcher is also analyzing the site-visit reports to harvest outcomes from this work. This paper shares findings from seven cohorts that have completed a grant cycle with funding results known, as well as qualitative data from site visits with the first two cohorts of grantees. Recommendations for further research are also included. 

   more » « less