Error Detection Performance Research Articles

Large-Scale Validation of the Feasibility of GPT-4 as a Proofreading Tool for Head CT Reports.

Background The increasing workload of radiologists can lead to burnout and errors in radiology reports. Large language models, such as OpenAI's GPT-4, hold promise as error revision tools for radiology. Purpose To test the feasibility of GPT-4 use by determining its error detection, reasoning, and revision performance on head CT reports with varying error types and to validate its clinical utility by comparison with human readers. Materials and Methods A total of 10 300 head CT reports were retrospectively extracted from the Medical Information Mart for Intensive Care III public dataset. In experiment 1, among the 300 unaltered reports and 300 versions with applied errors, GPT-4 optimization was initially conducted with 200 reports. The remaining 400 were used for evaluation of error type detection, reasoning, and revision, as well as the analysis of reports with undetected errors. The performance was also compared with that of human readers. In experiment 2, the detection performance of GPT-4 was validated on 10 000 unaltered reports that were deemed error-free by physicians, and an analysis of false-positive results was conducted. A permutation test was conducted to assess differences in performance. Results GPT-4 demonstrated commendable performance in error detection (sensitivity, 84% for interpretive error and 89% for factual error), reasoning, and revision. Compared with GPT-4, human readers had worse factual error detection sensitivity (0.33-0.69 vs 0.89; P = .008 for radiologist 4, P < .001 for others) and took longer to review (82-121 seconds vs 16 seconds, P < .001). In 10 000 reports, GPT-4 detected 96 errors, with a low positive predictive value of 0.05, yet 14% of the false-positive responses were potentially beneficial. Conclusion GPT-4 effectively detects, reasons, and revises errors in radiology reports. While it shows excellent performance in identifying factual errors, its ability to prioritize clinically significant findings is limited. Recognizing its strengths and limitations, GPT-4 could serve as a feasible tool. © RSNA, 2025 Supplemental material is available for this article. See also the editorial by Choi in this issue.

Thresholds for the distributed surface code in the presence of memory decoherence

In the search for scalable, fault-tolerant quantum computing, distributed quantum computers are promising candidates. These systems can be realized in large-scale quantum networks or condensed onto a single chip with closely situated nodes. We present a framework for numerical simulations of a memory channel using the distributed toric surface code, where each data qubit of the code is part of a separate node, and the error-detection performance depends on the quality of four-qubit Greenberger–Horne–Zeilinger (GHZ) states generated between the nodes. We quantitatively investigate the effect of memory decoherence and evaluate the advantage of GHZ creation protocols tailored to the level of decoherence. We do this by applying our framework for the particular case of color centers in diamond, employing models developed from experimental characterization of nitrogen-vacancy centers. For diamond color centers, coherence times during entanglement generation are orders of magnitude lower than coherence times of idling qubits. These coherence times represent a limiting factor for applications, but previous surface code simulations did not treat them as such. Introducing limiting coherence times as a prominent noise factor makes it imperative to integrate realistic operation times into simulations and incorporate strategies for operation scheduling. Our model predicts error probability thresholds for gate and measurement reduced by at least a factor of three compared to prior work with more idealized noise models. We also find a threshold of 4×102 in the ratio between the entanglement generation and the decoherence rates, setting a benchmark for experimental progress.

High-Speed Optical Chaotic Data Selection Logic Operations with the Performance of Error Detection and Correction

Based on the chaotic polarization system of optically injected cascaded vertical-cavity surface-emitting lasers (VCSELs), we propose a novel implementation scheme for high-speed optical chaotic data selection logic operations. Under the condition where the slave VCSEL (S-VCSEL) outputs a chaotic laser signal, we calculate the range of the applied electric field and the optical injection amplitude. We also investigate the evolution of the correlation characteristics between the polarized light output from the periodic poled LiNbO3 (PPLN) and the S-VCSEL as a function of the optical injection amplitude under different applied electric fields. Furthermore, we analyze the polarization bistability of the polarized light from the PPLN and S-VCSEL. Based on these results, we modulate the optical injection amplitude as the logic input and the applied electric field as the control logic signal. Using a mean comparison mechanism, we demodulate the polarized light from the PPLN and S-VCSEL to obtain two identical logic outputs, achieving optical chaotic data selection logic operations with an operation speed of approximately 114 Gb/s. Finally, we investigate the influence of noise on the logic outputs and find that both logic outputs do not show any error symbols under the noise strength as high as 180 dBw. The anti-noise performance of logic output O1 is superior to that of optical chaotic logic output O2. For noise strengths up to 185 dBw, error symbols in O2 can be detected and corrected by comparison with O1.

Enhanced patient-based real-time quality control using the graph-based anomaly detection.

Patient-based real-time quality control (PBRTQC) is an alternative tool for laboratories that has gained increasing attention. Despite the progress made by using various algorithms, the problems of data volume imbalance between in-control and out-of-control results, as well as the issue of variation remain challenges. We propose a novel integrated framework using anomaly detection and graph neural network, combining clinical variables and statistical algorithms, to improve the error detection performance of patient-based quality control. The testing results of three representative analytes (sodium, potassium, and calcium) and eight independent variables of patients (test date, time, gender, age, department, patient type, and reference interval limits) were collected. Graph-based anomaly detection network was modeled and used to generate control limits. Proportional and random errors were simulated for performance evaluation. Five mainstream PBRTQC statistical algorithms were chosen for comparison. The framework of a patient-based graph anomaly detection network for real-time quality control (PGADQC) was established and proven feasible for error detection. Compared with classic PBRTQC, the PGADQC showed a more balanced performance for both positive and negative biases. For different analytes, the average number of patient samples until error detection (ANPed) of PGADQC decreased variably, and reductions could reach up to approximately 95 % at a small bias of 0.02 taking calcium as an example. The PGADQC is an effective framework for patient-based quality control, integrating statistical and artificial intelligence algorithms. It improves error detection in a data-driven fashion and provides a new approach for PBRTQC from the data science perspective.

Potential of GPT-4 for Detecting Errors in Radiology Reports: Implications for Reporting Accuracy.

Background Errors in radiology reports may occur because of resident-to-attending discrepancies, speech recognition inaccuracies, and large workload. Large language models, such as GPT-4 (ChatGPT; OpenAI), may assist in generating reports. Purpose To assess effectiveness of GPT-4 in identifying common errors in radiology reports, focusing on performance, time, and cost-efficiency. Materials and Methods In this retrospective study, 200 radiology reports (radiography and cross-sectional imaging [CT and MRI]) were compiled between June 2023 and December 2023 at one institution. There were 150 errors from five common error categories (omission, insertion, spelling, side confusion, and other) intentionally inserted into 100 of the reports and used as the reference standard. Six radiologists (two senior radiologists, two attending physicians, and two residents) and GPT-4 were tasked with detecting these errors. Overall error detection performance, error detection in the five error categories, and reading time were assessed using Wald χ2 tests and paired-sample t tests. Results GPT-4 (detection rate, 82.7%;124 of 150; 95% CI: 75.8, 87.9) matched the average detection performance of radiologists independent of their experience (senior radiologists, 89.3% [134 of 150; 95% CI: 83.4, 93.3]; attending physicians, 80.0% [120 of 150; 95% CI: 72.9, 85.6]; residents, 80.0% [120 of 150; 95% CI: 72.9, 85.6]; P value range, .522-.99). One senior radiologist outperformed GPT-4 (detection rate, 94.7%; 142 of 150; 95% CI: 89.8, 97.3; P = .006). GPT-4 required less processing time per radiology report than the fastest human reader in the study (mean reading time, 3.5 seconds ± 0.5 [SD] vs 25.1 seconds ± 20.1, respectively; P < .001; Cohen d = -1.08). The use of GPT-4 resulted in lower mean correction cost per report than the most cost-efficient radiologist ($0.03 ± 0.01 vs $0.42 ± 0.41; P < .001; Cohen d = -1.12). Conclusion The radiology report error detection rate of GPT-4 was comparable with that of radiologists, potentially reducing work hours and cost. © RSNA, 2024 See also the editorial by Forman in this issue.

Performance monitoring and error detection: The role of mid frontal theta and error-related negativity (ERN) among Indian adolescents from different socioeconomic background

The aim of the study was to investigate the relationship between socioeconomic status (SES) and executive functioning, focusing specifically on performance monitoring, error detection, and their association with mid-frontal theta and error-related negativity (ERN). Employing the widely used flanker task, the research involved two phases with participants aged 10-16 years (15 individuals in the pilot phase and 35 in the second phase). Electroencephalogram (EEG) recordings from distinct brain regions were analyzed during various conditions. The study revealed a notable increase in both absolute and relative theta power at Fcz during the flanker task, with a stronger effect observed during incorrect trials. Furthermore, it underscored the influence of socioeconomic status (SES) on mid-frontal theta, highlighting interactions between SES, gender, and experimental conditions impacting both absolute and relative theta. Intriguingly, the research disclosed a positive correlation between parental occupation and error-related negativity (ERN), as well as between age and ERN. These findings underscore the significance of SES, gender, and age in shaping the neural mechanisms associated with performance monitoring and executive functions. The study contributes valuable insights into the intricate interplay between socio-demographic factors and cognitive processes, shedding light on their impact on goal-directed behaviors and brain activity.

Humans Supervising Artificial Intelligence – Investigation of Designs to Optimize Error Detection

ABSTRACT Artificial Intelligence (AI) fundamentally changes the way we work by introducing new capabilities. Human tasks shift towards a supervising role where the human confirms or disconfirms the presented decision. In this study, we utilise the signal detection theory to investigate and explain how the performance of human error detection is influenced by specific information design. We conducted two online experiments in the context of AI-supported information extraction and measured the ability of participants to validate the extracted information. In the first experiment, we investigated the mechanism of information provided prior to conducting the error detection task. In the second experiment, we manipulated the design of the presented information during the task and investigated its effect. Both manipulations significantly impacted the error detection performance of humans. Hence our study provides important insights for developing AI-based decision support systems and contributes to the theoretical understanding of human-AI collaboration.

Climbing Technique Evaluation by Means of Skeleton Video Stream Analysis.

Due to the growing interest in climbing, increasing importance has been given to research in the field of non-invasive, camera-based motion analysis. While existing work uses invasive technologies such as wearables or modified walls and holds, or focuses on competitive sports, we for the first time present a system that uses video analysis to automatically recognize six movement errors that are typical for novices with limited climbing experience. Climbing a complete route consists of three repetitive climbing phases. Therefore, a characteristic joint arrangement may be detected as an error in a specific climbing phase, while this exact arrangement may not considered to be an error in another climbing phase. That is why we introduced a finite state machine to determine the current phase and to check for errors that commonly occur in the current phase. The transition between the phases depends on which joints are being used. To capture joint movements, we use a fourth-generation iPad Pro with LiDAR to record climbing sequences in which we convert the climber's 2-D skeleton provided by the Vision framework from Apple into 3-D joints using the LiDAR depth information. Thereupon, we introduced a method that derives whether a joint moves or not, determining the current phase. Finally, the 3-D joints are analyzed with respect to defined characteristic joint arrangements to identify possible motion errors. To present the feedback to the climber, we imitate a virtual mentor by realizing an application on the iPad that creates an analysis immediately after the climber has finished the route by pointing out the detected errors and by giving suggestions for improvement. Quantitative tests with three experienced climbers that were able to climb reference routes without any errors and intentionally with errors resulted in precision-recall curves evaluating the error detection performance. The results demonstrate that while the number of false positives is still in an acceptable range, the number of detected errors is sufficient to provide climbing novices with adequate suggestions for improvement. Moreover, our study reveals limitations that mainly originate from incorrect joint localizations caused by the LiDAR sensor range. With human pose estimation becoming increasingly reliable and with the advance of sensor capabilities, these limitations will have a decreasing impact on our system performance.

Using Consumer Tweets to Improve Revenue Risk Assessments in Consumer-Oriented Industries

SUMMARY This article summarizes the study of Rozario, Vasarhelyi, and Wang (2022), which examines the use of consumer tweets in improving the prediction and error detection performance of preliminary analytical procedures for the revenue account for firms that belong to consumer-oriented industries. They find that consumer tweets about product or brand interest increase the prediction and error detection ability of analytical procedures compared to analytical procedures that do not include it. These results suggest that this new source of external nonfinancial information is incrementally informative to auditors in developing assessments for the risk of misstated revenue in the planning stage of the audit. The findings of this study have important implications that may be relevant to the audits of other financial statement accounts.

Hsiao Codes Properties in Discrete Devices Technical Diagnostics Systems

The error detection features by modified Hamming codes — codes with odd column weights, or Hsiao codes — are analyzed. The error detection properties of Hsiao codes are considered separately for data symbols, as well as for all code word symbols. It is shown that Hsiao codes detect any errors with odd multiplicities and errors with multiplicities d &lt; 4. The Hsiao codes error detection performance is compared with the error detection performance by classic Hamming codes. The experiments results showing the Hsiao codes high efficiency in organizing the calculation testing at the automation device outputs are given. Hsiao codes can be very effectively used in solving problems of automation devices synthesis with the malfunctions and errors detection in calculations, including the devices synthesis with controllable structures and with concurrent calculation testing circuits.

Morphology-Based Spell Checker for Dawurootsuwa Language

Processing of textual information by using word-processing tools is extremely increased due to the presence of misspelled or erroneous words. In order to minimize these misspelled words from digital information, different spellchecker tools are needed. A plenty of works are performed in technological favored languages like English and European languages but not for an underresourced language like Dawurootsuwa. The primary idea behind a morphology-based spellchecker is to use a dictionary lookup approach with morphological properties of the language to reduce dictionary size while also handling word inflection, derivation, and compounding. Two distinct tests were carried out in this work to evaluate the performance of a morphology-based spellchecker: error detection and error correction. The Hunspell dictionary format was utilized to construct the root words in this study, which included a total of 5,000 root words and more than 2,500 morphological rules along with 3,156 unique words for testing. The experimental result showed the overall spell error detection performance of 90.4% and the overall spell error correction performance of 79.31%. Moreover, we are working further towards developing a real word spelling checker that incorporate more numbers of language rules.

The DetectDeviatingCells algorithm was a useful addition to the toolkit for cellwise error detection in observational data

ObjectivesWe evaluated the error detection performance of the DetectDeviatingCells (DDC) algorithm which flags data anomalies at observation (casewise) and variable (cellwise) level in continuous variables. We compared its performance to other approaches in a simulated dataset. Study Design and SettingWe simulated height and weight data for hypothetical individuals aged 2–20 years. We changed a proportion of height values according to predetermined error patterns. We applied the DDC algorithm and other error-detection approaches (descriptive statistics, plots, fixed-threshold rules, classic, and robust Mahalanobis distance) and we compared error detection performance with sensitivity, specificity, likelihood ratios, predictive values, and receiver operating characteristic (ROC) curves. ResultsAt our chosen thresholds error detection specificity was excellent across all scenarios for all methods and sensitivity was higher for multivariable and robust methods. The DDC algorithm performance was similar to other robust multivariable methods. Analysis of ROC curves suggested that all methods had comparable performance for gross errors (e.g., wrong measurement unit), but the DDC algorithm outperformed the others for more complex error patterns (e.g., transcription errors that are still plausible, although extreme). ConclusionsThe DDC algorithm has the potential to improve error detection processes for observational data.

Reliable Constructions for the Key Generator of Code-based Post-quantum Cryptosystems on FPGA

Advances in quantum computing have urged the need for cryptographic algorithms that are low-power, low-energy, and secure against attacks that can be potentially enabled. For this post-quantum age, different solutions have been studied. Code-based cryptography is one feasible solution whose hardware architectures have become the focus of research in the NIST standardization process and has been advanced to the final round (to be concluded by 2022–2024). Nevertheless, although these constructions, e.g., McEliece and Niederreiter public key cryptography, have strong error correction properties, previous studies have proved the vulnerability of their hardware implementations against faults product of the environment and intentional faults, i.e., differential fault analysis. It is previously shown that depending on the codes used, i.e., classical or reduced (using either quasi-dyadic Goppa codes or quasi-cyclic alternant codes), flaws in error detection could be observed. In this work, efficient fault detection constructions are proposed for the first time to account for such shortcomings. Such schemes are based on regular parity, interleaved parity, and two different cyclic redundancy checks (CRC), i.e., CRC-2 and CRC-8. Without losing the generality, we experiment on the McEliece variant, noting that the presented schemes can be used for other code-based cryptosystems. We perform error detection capability assessments and implementations on field-programmable gate array Kintex-7 device xc7k70tfbv676-1 to verify the practicality of the presented approaches. To demonstrate the appropriateness for constrained embedded systems, the performance degradation and overheads of the presented schemes are assessed.

Automatic Speech Recognition and Pronunciation Error Detection of Dutch Non-native Speech: cumulating speech resources in a pluricentric language

• Improving the performance of Automatic Speech Recognition (ASR) on learner's speech in a non-dominant variety of a pluricentric language by cumulating speech resources from different varieties of the same language. • Through the transfer learning approach, the knowledge from a dominant variety can be transferred to non-dominant variety and this benefits non-native speech recognition. • Introducing plausible pronunciation errors in a native corpus of the non-dominant variety based on knowledge from the dominant variety to evaluate the performance of Pronunciation Error Detection (PED) algorithms. The shortage of large-scale learners’ speech corpora and precise manual annotations are two major challenges for automatic L2 speech recognition and error detection in L2 speech, especially for non-dominant varieties of pluricentric languages. In these cases, collecting and annotating large non-native (L2 learner) corpora for all language varieties is often unattainable. In this study, we investigated ways of addressing these problems through conventional and transfer learning Deep Neural Network (DNN) based Automatic Speech Recognition (ASR) and ASR-based pronunciation error detection (PED) by cumulating Netherlandic Dutch and Flemish Dutch speech resources. First, we show that for ASR the baseline system can be improved by combining the Netherlandic Dutch and Flemish Dutch datasets. Next, through the knowledge learned from models trained on the Netherlandic Dutch data, the Flemish Dutch learners' ASR model can be further improved. In order to evaluate the performance of the PED algorithms in the absence of learner speech data with pronunciation error annotations, we introduced plausible pronunciation errors in the native corpora based on knowledge from Flemish learner speech, in order to simulate non-native speech errors. For PED we found that the results are much better for a GOP classifier trained on Flemish Dutch data than for one trained on Netherlandic Dutch data. PED produced worse results when the Netherlandic Dutch data were merged with the Flemish Dutch data, while for ASR, lower WERs were attained. Whether adding Netherlandic Dutch data to Flemish Dutch data is beneficial, thus seems to depend on the specific task the data are used for. We discuss these results, compare them to those of related research and suggest avenues for future research.

Explainability does not improve biochemistry staff trust in artificial intelligence-based decision support.

Explainability, the aspect of artificial intelligence-based decision support (ADS) systems that allows users to understand why predictions are made, offers many potential benefits. One common claim is that explainability increases user trust, yet this has not been established in healthcare contexts. For advanced algorithms such as artificial neural networks, the generation of explanations is not trivial, but requires the use of a second algorithm. The assumption of improved user trust should therefore be investigated to determine if it justifies the additional complexity. Biochemistry staff completed a wrong blood in tube (WBIT) error identification task with the help of an ADS system. One-half of the volunteers were provided with both ADS predictions and explanations for those predictions, while the other half received predictions alone. The two groups were compared in terms of their rate of agreement with ADS predictions, as an index of user trust, and WBIT error detection performance. Since the AI model used to generate predictions was known to out-perform laboratory staff, increased trust was expected to improve user performance. Volunteers reviewed 1590 sets of results. The volunteers provided with explanations demonstrated no difference in their rate of agreement with the ADS system compared to volunteers receiving predictions alone (83.3% versus 81.8%, p = 0.46). The two volunteer groups were also equivalent in accuracy, sensitivity and specificity for WBIT error identification (p-values >0.78). For a WBIT error identification task, there was no evidence to justify the additional complexity of explainability on the grounds of increased user trust.

Cross Domain Iterative Detection for Orthogonal Time Frequency Space Modulation

Recently proposed orthogonal time frequency space (OTFS) modulation has been considered as a promising candidate for accommodating various emerging communication and sensing applications in high-mobility environments. In this paper, we propose a novel cross domain iterative detection algorithm to enhance the error performance of OTFS modulation. Different from conventional OTFS detection methods, the proposed algorithm applies basic estimation/detection approaches to both the time domain and delay-Doppler (DD) domain and iteratively updates the extrinsic information from two domains with the unitary transformation. In doing so, the proposed algorithm exploits the time domain channel sparsity and the DD domain symbol constellation constraints. We evaluate the estimation/detection error variance in each domain for each iteration and derive the state evolution to investigate the detection error performance. We show that the performance gain due to iterations comes from the non-Gaussian constellation constraint in the DD domain. More importantly, we prove that the proposed algorithm can indeed converge and, in the convergence, the proposed algorithm can achieve almost the same error performance as the maximum-likelihood sequence detection even in the presence of fractional Doppler shifts. Furthermore, the computational complexity associated with the domain transformation is low, thanks to the structure of the discrete Fourier transform (DFT) kernel. Simulation results are consistent with our analysis and demonstrate a significant performance improvement compared to conventional OTFS detection methods.

Optimization and Validation of Limit Check Error-Detection Performance Using a Laboratory-Specific Data-Simulation Approach: A Prerequisite for an Evidence-Based Practice.

Autoverification procedures based on limit checks (LCs) provide important support to preanalytical, analytical, and postanalytical quality assurance in medical laboratories. A recently described method, based on laboratory-specific error-detection performances, was used to determine LCs for all chemistry analytes performed on random-access chemistry analyzers prior to application. Using data sets of historical test results, error-detection simulations of limit checks were performed using the online MA Generator system (www.huvaros.com). Errors were introduced at various positions in the data set, and the number of tests required for an LC alarm to occur was plotted in bias detection curves. Random error detection was defined as an LC alarm occurring in 1 test result, whereas systematic error detection was defined as an LC alarm occurring within an analytical run, both with ≥97.5% probability. To enable the lower limit check (LLC) and the upper limit check (ULC) to be optimized, the simulation results and the LC alarm rates for specific LLCs and ULCs were presented in LC performance tables. Optimal LLCs and ULCs were obtained for 31 analytes based on their random and systematic error-detection performances and the alarm rate. Reliable detection of random errors greater than 60% was only possible for analytes known to have a rather small variation of results. Furthermore, differences for negative and positive errors were observed. The used method brings objectivity to the error-detection performance of LCs, thereby enabling laboratory-specific LCs to be optimized and validated prior to application.

PGE

Although product graphs (PGs) have gained increasing attentions in recent years for their successful applications in product search and recommendations, the extensive power of PGs can be limited by the inevitable involvement of various kinds of errors. Thus, it is critical to validate the correctness of triples in PGs to improve their reliability. Knowledge graph (KG) embedding methods have strong error detection abilities. Yet, existing KG embedding methods may not be directly applicable to a PG due to its distinct characteristics: (1) PG contains rich textual signals, which necessitates a joint exploration of both text information and graph structure; (2) PG contains a large number of attribute triples, in which attribute values are represented by free texts. Since free texts are too flexible to define entities in KGs, traditional way to map entities to their embeddings using ids is no longer appropriate for attribute value representation; (3) Noisy triples in a PG mislead the embedding learning and significantly hurt the performance of error detection. To address the aforementioned challenges, we propose an end-to-end noise-tolerant embedding learning framework, PGE, to jointly leverage both text information and graph structure in PG to learn embeddings for error detection. Experimental results on real-world product graph demonstrate the effectiveness of the proposed framework comparing with the state-of-the-art approaches.

Melody, not Beat Perception, Predicts Rhythmic Error Detection

The purpose of this study was to examine the relationships among beat perception, error detection, and musical experience. We presented monophonic rhythms using a piano timbre along with two measures of beat perception (Harvard Beat Finding and Interval Test [BFIT] and Goldsmiths Beat Alignment Test) and a measure of melodic error detection. College musicians’ ( N = 43) ability to detect rhythm errors was not significantly correlated to their ability to perceive beat alignment (Goldsmiths test) or tempo change (BFIT). Age was related to performance on only one of the measures, the BFIT test. A regression model yielded pitch error detection as the only significant predictor of rhythmic error detection. We suggest that college musicians already possess a requisite ability for beat processing that allows them to perform error detection. The lack of relationship between beat perception and rhythmic error detection is explained by this requisite ability in the population, and we promote future research for pitch and rhythm processing as it relates to rhythm perception or performance.

Artificial Neural Network Assisted Error Correction for MLC NAND Flash Memory.

The multilevel per cell technology and continued scaling down process technology significantly improves the storage density of NAND flash memory but also brings about a challenge in that data reliability degrades due to the serious noise. To ensure the data reliability, many noise mitigation technologies have been proposed. However, they only mitigate one of the noises of the NAND flash memory channel. In this paper, we consider all the main noises and present a novel neural network-assisted error correction (ANNAEC) scheme to increase the reliability of multi-level cell (MLC) NAND flash memory. To avoid using retention time as an input parameter of the neural network, we propose a relative log-likelihood ratio (LLR) to estimate the actual LLR. Then, we transform the bit detection into a clustering problem and propose to employ a neural network to learn the error characteristics of the NAND flash memory channel. Therefore, the trained neural network has optimized performances of bit error detection. Simulation results show that our proposed scheme can significantly improve the performance of the bit error detection and increase the endurance of NAND flash memory.