Error Rate Reduction Research Articles

Comparative study: AutoDPR-SEM for enhancing CNN reliability in SRAM-based FPGAs through autonomous reconfiguration

Convolutional neural networks (CNNs) are widely adopted in safety-critical systems, including space applications and autonomous vehicles. Field-programmable gate arrays (FPGAs) based on SRAM are preferred for accelerating CNN computations due to their unique characteristics. However, the configuration memory of FPGAs is susceptible to single event effects (SEEs), which can corrupt computations and lead to misclassification of CNN outputs. In this study, we investigated the impact of SEEs on SRAM-based FPGAs with Two-Photon Absorption (TPA) laser fault injections through a comparative analysis of two popular CNN acceleration architectures: streaming architecture (SA) and single compute engine (SCE). Experimental results show that SA-based CNNs require more hardware resources but exhibit superior resilience against single event upsets (SEUs). Without any Radiation Hardened by Design (RHBD) protection, SCE has an error rate approximately twice as high as SA. To mitigate errors, the Xilinx IP core - Soft Error Mitigation (SEM) is used for error detection and correction, leading to error rate reductions of up to 50 % in both architectures. Importantly, we propose the AutoDPR-SEM (Autonomous Dynamic Partial Reconfiguration for Soft Error Mitigation) approach, which automatically reconfigures the SEM IP core when it remains idle due to uncorrectable errors. AutoDPR-SEM significantly improves CNN error rates, reducing errors by approximately 17.8 times in SCE and 14.8 times in SA. We also applied software level simulation to validate the TPA experiment, showing similar trends of the testing results across all models. In conclusion, the study confirms the feasibility of AutoDPR-SEM in both architectures, showcasing its potential to improve CNN error rates in safety-critical systems.

Optimization of CPBIS methods applied on enhanced fibrin microbeads approach for image segmentation in dynamic databases

In the empire of image processing and computer vision, the demand for advanced segmentation techniques has intensified with the growing complexity of visual data. This study focuses on the innovative paradigm of fuzzy mountain-based image segmentation, a method that harnesses the power of fuzzy logic and topographical inspiration to achieve nuanced and adaptable delineation of image regions. This research primarily concentrates on determining the age of tigers, a critical and challenging task in the current scenario. The primary objectives include the development of a comprehensive framework for FMBIS and an in-depth investigation into its adaptability to different image characteristics. This research work incorporates those domains of image processing and data mining to predict the age of the tiger using different kinds of color images. Fuzzy mountain-based pixel segmentation arises from the need to capture the subtle gradients and uncertainties present in images, offering a novel approach to achieving high-fidelity segmentations in diverse and complex scenarios. The proposed methods enable image enhancement and filtering and are then assessed during process time, retrieval time, to give a more accurate and reduced error rate for producing higher results for real-time tiger image database.

#2742 Artificial intelligence accurately approximates MGPE dose response relationship in a North American Dialysis Population

Abstract Background and Aims The Anemia Control Model (ACM) is a certified medical device aimed at optimizing the management of CKD-related anemia among patients on dialysis. ACM is based on an artificial neural network (ANN) trained on over 900’000 patients treated in the European Nephrocare Network. Over the years, ACM has offered physicians hundreds of thousands of erythropoietin stimulating agent (ESA) and intravenous iron dosing suggestions and ensured high hemoglobin target achievement rate while minimizing drug use. However, evidence concerning its validity in non-European samples is still limited. Apollo Dial DB is the largest, fully anonymized, collaborative, international database of dialysis patients. The database includes detailed clinical information of more than 500,000 patients over millions of patient-months of observation with an extensive representation over geographies as well as ethnic groups. Given its detailed medical information and size, the database may offer unprecedented opportunities to develop and test equitable and fair decision support systems based on artificial intelligence. In this study we sought to re-train and validate and assess the accuracy of the ACM in an ethnically diverse sample of North American dialysis patients in Apollo Dial DB. Method We constructed the cohort for training and validation of this new ACM version by including a random sample of patients treated in the North American Fresenius Medical Care Dialysis Network. The sample accounted for 20% of the entire North American Section of the Apollo Dial DB dataset. Each patient history has been segmented in a collection of patient-months, which was the unit of analysis. The initial cohort was split in 3 datasets for training (70%), validation (20%) and test (10%). The input variables included laboratory test results, body weight, treatment data, age, sex, and previous ESA and iron prescription. The outcome of interest was the variation in serum hemoglobin concentration over 1 month of follow up. We compared the accuracy of the original ACM based on ANN (baseline model) with a newly trained algorithm based on high-performance light GBM. Results We included 44,220 patients followed by 2,277,276 patient-months overall. The mean age=60.9 ± 13.4, 59.2% were male, 56.4% were Caucasians, 32.2% African American, 3.2% Asians and 8.2% were others. At baseline, mean Hb= 10.9 ± 1.2. The original EMEA version of the Anemia Control Model (ACM) was still accurate in North American patients (mean absolute error, MAE = 0.65 g/dL). The fine-tuned light GBM model achieved a 16% reduction in error rate (MAE = 0.55 g/dL; Fig. 1). No difference in MAE was observed across sex and ethnic groups. Conclusion This study demonstrates the excellent predictive performance of the original ANN-based ACM in the North American dialysis population. Fine tuning by re-training ACM on the dataset further improved the accuracy of the model with substantial equivalence in performance across different demographics, thus ensuring algorithmic fairness. ACM is a promising tool for anemia management in ethnically diverse dialysis population and different healthcare settings.

Development of Fused Recurrent Neural Network‐Gated Recurrent Unit with Improved Sandpiper Optimization‐based Intelligent Node Localization Framework in WSN

SummaryEmploying mobile nodes or Sensor Nodes (SN) in the Wireless Sensor Networks (WSNs) is a critical task within the coverage area. However, these techniques produce huge errors in the evaluated distance between source nodes and unknown nodes, which increases the average NL error of the unknown node. The usage of a Global Positioning System (GPS) for all the SNs needs high cost for implementation in large‐scale WSNs. The designing of NL in WSN with less energy requirements and less cost is very significant. Enhanced NL is performed with deep learning techniques together with the heuristic strategy for performance enhancement. The developed localization model is mainly utilized for providing fast, scalable, and easily implementable results over heterogeneous and homogeneous networks. Moreover, it is applicable for sparse and dense deployment of nodes. Here, a hybridized deep structured approach is employed by integrating the Recurrent Neural Network (RNN) with a Gated Recurrent Unit (GRU) for calculating the distance among the source nodes to facilitate determining the optimal location for the unidentified node by analyzing the node information present in the received packets. Here, the optimization of constraints executed in the deep learning techniques with the Modified Exploration‐based Sandpiper Optimization Algorithm (ME‐SOA), and also it is utilized to find the optimal location of unknown nodes based on computed distance from the received packets that minimize the average localization error. The experimental results are validated and show the effectiveness of the proposed algorithm with a reduced error rate than other algorithms.

CAREx: context-aware read extension of paired-end sequencing data

BackgroundCommonly used next generation sequencing machines typically produce large amounts of short reads of a few hundred base-pairs in length. However, many downstream applications would generally benefit from longer reads.ResultsWe present CAREx—an algorithm for the generation of pseudo-long reads from paired-end short-read Illumina data based on the concept of repeatedly computing multiple-sequence-alignments to extend a read until its partner is found. Our performance evaluation on both simulated data and real data shows that CAREx is able to connect significantly more read pairs (up to 99%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$99\\%$$\\end{document} for simulated data) and to produce more error-free pseudo-long reads than previous approaches. When used prior to assembly it can achieve superior de novo assembly results. Furthermore, the GPU-accelerated version of CAREx exhibits the fastest execution times among all tested tools.ConclusionCAREx is a new MSA-based algorithm and software for producing pseudo-long reads from paired-end short read data. It outperforms other state-of-the-art programs in terms of (i) percentage of connected read pairs, (ii) reduction of error rates of filled gaps, (iii) runtime, and (iv) downstream analysis using de novo assembly. CAREx is open-source software written in C++ (CPU version) and in CUDA/C++ (GPU version). It is licensed under GPLv3 and can be downloaded at (https://github.com/fkallen/CAREx).

Multi-view Self-supervised Learning and Multi-scale Feature Fusion for Automatic Speech Recognition

To address the challenges of the poor representation capability and low data utilization rate of end-to-end speech recognition models in deep learning, this study proposes an end-to-end speech recognition model based on multi-scale feature fusion and multi-view self-supervised learning (MM-ASR). It adopts a multi-task learning paradigm for training. The proposed method emphasizes the importance of inter-layer information within shared encoders, aiming to enhance the model’s characterization capability via the multi-scale feature fusion module. Moreover, we apply multi-view self-supervised learning to effectively exploit data information. Our approach is rigorously evaluated on the Aishell-1 dataset and further validated its effectiveness on the English corpus WSJ. The experimental results demonstrate a noteworthy 4.6%\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\%$$\\end{document} reduction in character error rate, indicating significantly improved speech recognition performance . These findings showcase the effectiveness and potential of our proposed MM-ASR model for end-to-end speech recognition tasks.

Demonstration of D-band 1×2 SIMO millimeter-wave wireless delivery over 1.2 km employing MRC technology

We experimentally demonstrate the transmission of 20-Gbaud QPSK millimeter-wave signals within a 1.2-km wireless range at 128-GHz using constant-modulus-algorithm (CMA) and the maximum ratio combining (MRC) technology with over 2.2-dB gain in a photonics-assisted millimeter wave (MMW) 1 × 2 Single Input Multi Output (SIMO) communication system. MRC is applied for SIMO signal merging which leads to Signal to Noise Ratio (SNR) enhancement and significant Bit Error Rate (BER) reduction. The comparison which is performed for Single Input Single Output (SISO) and SIMO system, demonstrates the BER improvement. To the best of our knowledge, this is the first experimental demonstration of high gain in a D-band MMW 1 × 2 SIMO system by using these technologies.

Planning the implementation of a new oncology compounding information system in an aseptic compounding unit: a mixed methods study

Abstract Introduction Systemic anti-cancer therapy (SACT) compounding is a high-risk process. Traditional mixed method systems for workflow management in aseptic compounding units (ACUs) are inherently prone to risk. Workflow management systems (WFMSs) are recognised internationally to improve safety, reduce costs, and increase ACU capacity.[1] Planning of the implementation of a WFMS has not been characterised before. In our ACU, Clinichemo® will be replaced with BD-CATO®. BD-CATO® is a WFMS and is the vendor for the National Cancer Information System (NCIS), implemented by the National Cancer Control Programme (NCCP).[2] Aim The aim of this study was to demonstrate how the implementation of a new WFMS can be safely planned, managed, and assessed. This aim was achieved using the following objectives: 1. Describe and risk assess the current workflows for SACT compounding. 2. Investigate agreed and published standards for SACT compounding. 3. Define the change management process. 4. Describe and evaluate the benefits, challenges and risks associated with the introduction of a new SACT compounding system. 5. Interview adopters of NCIS/BD-CATO® to assess their experience of the barriers encountered, how challenges were overcome, risks identified, mitigated, and eliminated through the change management process and ongoing. Methods The current workflow and future workflow for SACT compounding were characterised and current process risk assessed. Semi-structured interviews were carried out to evaluate pharmacist and pharmacy technicians’ attitudes and opinions towards the implementation of BD-CATO® in their organisations and the system itself. Participants were recruited by means of purposive and snowball sampling from ACUs across the country that had implemented/were in the process of implementing BD-CATO®/NCIS. Themes identified in interviews were compared to those found in the literature and guidelines. Interviews, a review of guidelines, published standards and the literature were used to guide and risk assess future process development. The change management plan was also defined. Results Nine participants (8 pharmacists and 1 pharmacy technician) were interviewed in June-August 2023. The interviews were conducted across 7 ACUs in Ireland. Thematic analysis of interview data yielded four key themes: 1. Participants discussed the importance of planning activities, such as communication with key stakeholders, discussions around phased implementation, and development of training material. 2. Having a designated individual tasked with change implementation improved outcomes of the implementation project. 3. Participants also discussed benefits and risks associated with BD-CATO®. 4. Participants discussed their reasons for perceived complexity of the system. A combination of interview data, review of guidelines, published standards and the literature were used to devise future process workflow and the change management plan. The future process workflow was also risk assessed using bowtie analysis. Conclusion The introduction of WFMSs has been interpreted both in the literature and within the Irish healthcare setting to increase efficiencies, reduce error rates, and improve the accuracy of SACT compounding. A formal change management strategy devised with resourcing, training, and obtaining stakeholder engagement is key to successful implementation.

Symbol error rate minimization using deep learning approaches for short-reach optical communication networks

Short-reach optical communication networks have various applications in areas where high-speed connectivity is needed, for example, inter- and intra-data center links, optical access networks, and indoor and in-building communication systems. Machine learning (ML) approaches provide a key solution for numerous challenging situations due to their robust decision-making, problem-solving, and pattern-recognition abilities. In this work, our focus is on utilizing deep learning models to minimize symbol error rates in short-reach optical communication setups. Various channel impairments, such as nonlinearity, chromatic dispersion (CD), and attenuation, are accurately modeled. Initially, we address the challenge of modeling a nonlinear channel. Consequently, we harness a deep learning model called autoencoders (AEs) to facilitate communication over nonlinear channels. Furthermore, we investigate how the inclusion of a nonlinear channel within an autoencoder influences the received constellation as the optical fiber length increases. Another facet of our work involves the deployment of a deep neural network-based receiver utilizing a channel influenced by chromatic dispersion. By gradually extending the optical length, we explore its impact on the received constellation and, consequently, the symbol error rate. Finally, we incorporate the split-step Fourier method (SSFM) to emulate the combined effects of nonlinearities, chromatic dispersion, and attenuation in the optical channel. This is accomplished through a neural network-based receiver. The outcome of this work is an evaluation and reduction of the symbol error rate as the length of the optical fiber is varied.

Cox Regressive Outlier Robust Incremental Extreme Learning Machine based IoT aware Smart Car Parking and Route Optimization for Intelligent Transportation

An Intelligent Transport System (ITS) is a wide-ranging framework that influences advanced technologies to improve efficiency of transportation networks. An ITS integrated with an IoT enhances the capabilities of transportation networks. Owing to quick development of vehicles, traffic congestion during road network is major problems. Therefore, intelligent traffic control system is required to create smart transportation through optimization of vehicle route, and so on. In this paper, Cox Regressive Fuzzified Outlier Robust Incremental Extreme Learning Machine (CRFORIELM) is developed for smart ITS. The IoT technology in IT’S is used to collect traffic data, vehicle information, Parking slot occupancy data. After data collection process, proposed CRFORIELM technique includes two major processing steps namely parking slot availability detection and traffic aware route optimization. The extreme learning machine includes three types of layers such as input, three hidden layers as well as output layer. Initial hidden layer predicts the parking slot availability using Cox regression is performed. With the parking lot availability predictions, integrate this information into a route optimization algorithm. This algorithm considers the real-time traffic conditions as well as parking availability for detecting optimal route by applying fuzzy triangular membership function in the second hidden layer. Finally, nearest route path is effectively determined to minimize travel time in intelligent transport system. Experimental analysis indicates that the CRFORIELM technique achieved 2%, 5%, 10% improvement in accuracy, sensitivity, specificity and reduced error rates, route detection time by 39%, and 36%, compared to existing methods.

Complex transmission matrix retrieval for a highly scattering medium via regional phase differentiation

Accurately measuring the complex transmission matrix (CTM) of the scattering medium (SM) holds critical significance for applications in anti-scattering optical imaging, phototherapy, and optical neural networks. Non-interferometric approaches, utilizing phase retrieval algorithms, can robustly extract the CTM from the speckle patterns formed by multiple probing fields traversing the SM. However, in cases where an amplitude-type spatial light modulator is employed for probing field modulation, the absence of phase control frequently results in the convergence towards a local optimum, undermining the measurement accuracy. Here, we propose a high-accuracy CTM retrieval (CTMR) approach based on regional phase differentiation (RPD). It incorporates a sequence of additional phase masks into the probing fields, imposing a priori constraints on the phase retrieval algorithms. By distinguishing the variance of speckle patterns produced by different phase masks, the RPD-CTMR can effectively direct the algorithm towards a solution that closely approximates the CTM of the SM. We built a prototype of a digital micromirror device modulated RPD-CTMR. By accurately measuring the CTM of diffusers, we achieved an enhancement in the peak-to-background ratio of anti-scattering focusing by a factor of 3.6, alongside a reduction in the bit error rate of anti-scattering image transmission by a factor of 24. Our proposed approach aims to facilitate precise modulation of scattered optical fields, thereby fostering advancements in diverse fields including high-resolution microscopy, biomedical optical imaging, and optical communications.

Role of artificial intelligence in customer engagement: a systematic review and future research directions

Role of artificial intelligence in customer engagement: a systematic review and future research directions

The Proposal of Countermeasures for DeepFake Voices on Social Media Considering Waveform and Text Embedding

In recent times, advancements in text-to-speech technologies have yielded more natural-sounding voices. However, this has also made it easier to generate malicious fake voices and disseminate false narratives. ASVspoof stands out as a prominent benchmark in the ongoing effort to automatically detect fake voices, thereby playing a crucial role in countering illicit access to biometric systems. Consequently, there is a growing need to broaden our perspectives, particularly when it comes to detecting fake voices on social media platforms. Moreover, existing detection models commonly face challenges related to their generalization performance. This study sheds light on specific instances involving the latest speech generation models. Furthermore, we introduce a novel framework designed to address the nuances of detecting fake voices in the context of social media. This framework considers not only the voice waveform but also the speech content. Our experiments have demonstrated that the proposed framework considerably enhances classification performance, as evidenced by the reduction in equal error rate. This underscores the importance of considering the waveform and the content of the voice when tasked with identifying fake voices and disseminating false claims.

DNN-based Multilingual Acoustic Modeling for Four Ethiopian Languages

In this paper, we present the results of experiments conducted on multilingual acoustic modeling in the development of an Automatic Speech Recognition (ASR) system using speech data of phonetically much related Ethiopian languages (Amharic, Tigrigna, Oromo and Wolaytta) with multilingual (ML) mix and multitask approaches. The use of speech data from only phonetically much related languages brought improvement over results reported in a previous work that used 26 languages (including the four languages). A maximum Word Error Rate (WER) reduction from 25.03% (in the previous work) to 21.52% has been achieved for Wolaytta, which is a relative WER reduction of 14.02%. As a result of using multilingual acoustic modeling for the development of an automatic speech recognition (ASR) system, a relative WER reduction of up to 7.36% (a WER reduction from 23.23% to 21.52%) has been achieved over a monolingual ASR. Compared to the ML mix, the multitask approach brought a better performance improvement (a relative WERs reduction of up to 5.9%). Experiments have also been conducted using Amharic and Tigrigna in a pair and Oromo and Wolaytta in another pair. The results of the experiments showed that languages with a relatively better language resources for lexical and language modeling (Amharic and Tigrigna) benefited from the use of speech data from only two languages. Generally, the findings show that the use of speech corpora of phonetically related languages with the multitask multilingual modeling approach for the development of ASR systems for less-resourced languages is a promising solution.

DRANetSplicer: A Splice Site Prediction Model Based on Deep Residual Attention Networks.

The precise identification of splice sites is essential for unraveling the structure and function of genes, constituting a pivotal step in the gene annotation process. In this study, we developed a novel deep learning model, DRANetSplicer, that integrates residual learning and attention mechanisms for enhanced accuracy in capturing the intricate features of splice sites. We constructed multiple datasets using the most recent versions of genomic data from three different organisms, Oryza sativa japonica, Arabidopsis thaliana and Homo sapiens. This approach allows us to train models with a richer set of high-quality data. DRANetSplicer outperformed benchmark methods on donor and acceptor splice site datasets, achieving an average accuracy of (96.57%, 95.82%) across the three organisms. Comparative analyses with benchmark methods, including SpliceFinder, Splice2Deep, Deep Splicer, EnsembleSplice, and DNABERT, revealed DRANetSplicer's superior predictive performance, resulting in at least a (4.2%, 11.6%) relative reduction in average error rate. We utilized the DRANetSplicer model trained on O. sativa japonica data to predict splice sites in A. thaliana, achieving accuracies for donor and acceptor sites of (94.89%, 94.25%). These results indicate that DRANetSplicer possesses excellent cross-organism predictive capabilities, with its performance in cross-organism predictions even surpassing that of benchmark methods in non-cross-organism predictions. Cross-organism validation showcased DRANetSplicer's excellence in predicting splice sites across similar organisms, supporting its applicability in gene annotation for understudied organisms. We employed multiple methods to visualize the decision-making process of the model. The visualization results indicate that DRANetSplicer can learn and interpret well-known biological features, further validating its overall performance. Our study systematically examined and confirmed the predictive ability of DRANetSplicer from various levels and perspectives, indicating that its practical application in gene annotation is justified.

Reading between the Lines: Image-Based Order Detection in OCR for Chinese Historical Documents

Chinese historical documents, with their unique layouts and reading patterns, pose significant challenges for traditional Optical Character Recognition (OCR) systems. This paper introduces a tailored OCR system designed to address these complexities, particularly emphasizing the crucial aspect of Reading Order Detection(ROD). Our system operates through a threefold process: text detection using the Differential Binarization++ model, text recognition with the SVTR Net, and a novel ROD approach harnessing raw image features. This innovative method for ROD, inspired by human perception, utilizes visual cues present in raw images to deduce the inherent sequence of ancient texts. Preliminary results show promising reductions in page error rates. By preserving both content and context, our system contributes meaningfully to the accurate and contextual digitization of Chinese historical manuscripts.

Deep Quantum Error Correction

Quantum error correction codes (QECC) are a key component for realizing the potential of quantum computing. QECC, as its classical counterpart (ECC), enables the reduction of error rates, by distributing quantum logical information across redundant physical qubits, such that errors can be detected and corrected. In this work, we efficiently train novel end-to-end deep quantum error decoders. We resolve the quantum measurement collapse by augmenting syndrome decoding to predict an initial estimate of the system noise, which is then refined iteratively through a deep neural network. The logical error rates calculated over finite fields are directly optimized via a differentiable objective, enabling efficient decoding under the constraints imposed by the code. Finally, our architecture is extended to support faulty syndrome measurement, by efficient decoding of repeated syndrome sampling. The proposed method demonstrates the power of neural decoders for QECC by achieving state-of-the-art accuracy, outperforming for small distance topological codes, the existing end-to-end neural and classical decoders, which are often computationally prohibitive.

Dynamic Reserve Calculation Method of Fractured-Vuggy Reservoir Based on Modified Comprehensive Compression Coefficient

The low comprehensive compressibility coefficient characteristic of fracture-vuggy reservoirs often leads to imprecise dynamic reserve calculations. This study introduces a novel method for estimating dynamic reserves, which incorporates a modified comprehensive compressibility coefficient to enhance accuracy. This methodology has been applied to 23 wells in the Tahe Oilfield, resulting in error rates substantially lower than those associated with traditional techniques, thereby markedly enhancing the accuracy of dynamic reserve estimations. Specifically, for karst cave and fracture-vuggy reservoirs, the error rate in dynamic reserve calculations is reduced to under 10%, surpassing conventional methods by more than fivefold. In the case of fractured reservoirs, despite minor fluctuations in error rates due to stress sensitivity, diversion capacity, and channel variations, the proposed method still demonstrates a significant reduction in error rates compared to standard practices.

Abstract 337: Balanced-strand sequencing for highly efficient duplex variant calling in circulating tumor DNA

Abstract Several approaches for the detection of circulating tumor DNA (ctDNA) are anchored in the accurate identification of cancer-specific single nucleotide variants (SNVs) in plasma. Such approaches are limited in ultra-low tumor fraction (TF) contexts, such as early cancer detection or the identification of postoperative minimal residual disease, where ctDNA SNVs are difficult to distinguish against background sequencing error. ctDNA detection in these critical low TF contexts is further limited by the sparse amount of cell free DNA (cfDNA) in plasma (Zviran et al., 2020). To reduce single-stranded artifact and base calling errors, duplex sequencing can be used to jointly analyze corresponding Watson and Crick strands to detect true, double-stranded variants. Recent work has demonstrated that duplex sequencing can decrease error rates to 10−7 (Hoang et al., 2016) and can be deployed in the whole-genome sequencing (WGS) setting (Cheng et al., 2023). A significant limitation for existing duplex techniques is the need to analytically pair matching DNA strands through exhaustive sequencing, resulting in inefficient use of scarce input material, particularly in plasma cfDNA. These techniques typically capture 1-5% of sequenced molecules, restricting duplex coverage depth. We present balanced-strand sequencing, a PCR-free approach that leverages the hydrogen bonds of complementary DNA to carry matched native strands directly into sequencing, enabling massively scalable duplex sequencing. The approach relies on Ultima Genomics sequencing, where DNA denaturation is not required prior to clonal amplification. In balanced strand sequencing, double-stranded DNA is partitioned and clonally amplified on sequencing beads through emulsion-PCR. Each double-stranded DNA molecule contributes to a single sequencing read, allowing for a linear increase in duplex recovery with increasing sequencing depth. The approach does not require redundant sequencing and therefore maximizes unique molecule throughput. We applied balanced strand sequencing to cfDNA, input as low as 5 ng, and achieved 7-35x whole-genome duplex coverage, roughly 20-fold higher than our prior duplex WGS. We further developed a machine-learning guided single read mutation calling framework, which when combined with duplex error suppression, reduced error rates to 10−7 and enabled accurate ctDNA detection at concentrations <10−5. Use of our analysis framework in APOBEC-mutant bladder cancer plasma samples allowed us to identify APOBEC signatures in plasma and platinum exposure signatures indicative of prior chemotherapy treatment. Altogether, balanced strand sequencing radically enhanced duplex efficiency and has broad applications across cancer. In cfDNA applications, balanced strand sequencing enhances error-correction in tumor-informed ctDNA detection and paves the way for non-tumor informed WGS ctDNA detection. Citation Format: Adam Widman, Alexandre P. Cheng, Aaron Sossin, Srinivas Rajagopalan, Majd Al Assaad, Sunil Deochand, Zoe Steinsnyder, Dina Manaa, Melissa Marton, Catherine Reeves, Itai Rusinek, Eti Meiri, Omer Barad, Zohar Shipony, Shlomit Gilad, Ariel Jaimovich, Michael Sigouros, Jyothi Manohar, Abigail King, David Wilkes, John Otilano, Olivier Elemento, Bishoy M. Faltas, Juan M. Mosquera, Dan A. Landau. Balanced-strand sequencing for highly efficient duplex variant calling in circulating tumor DNA [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 337.

Sampleformer: An efficient conformer-based Neural Network for Automatic Speech Recognition

The Convolution-augmented Transformer (Conformer) model, which was recently introduced, has attained state-of-the-art(SOTA) results in Automatic Speech Recognition (ASR). In this paper, a series of methodical investigations uncover that the Conformer’s design decisions may not represent the most efficient choices when operating within the constraints of a limited computational budget. After a thorough re-evaluation of the Conformer architecture’s design choices, we propose Sampleformer which reduces the Conformer architecture complexity and has more robust performance. We introduce downsampling to the Conformer Encoder, and to exploit the information in the speech features, we incorporate an additional downsampling module to enhance the efficiency and accuracy of our model. Additionally, we propose a novel and adaptable attention mechanism called multi-group attention, effectively reducing the attention complexity from O⁢(n2⁢d) to O⁢(n2⁢d⋅f/g). We performed experiments on the AISHELL-1 corpora, our 13.3 million-parameter CTC model demonstrates a 3.0%/2.6% relative reduction in character error rate (CER) on the dev/test sets, all without the utilization of a language model (LM). Additionally, the model exhibits a 30% improvement in inference compared to our CTC Conformer baseline and trains 27% faster.