Comparing ROC Curve based Thresholding Methods in Online Transactions Fraud Detection System using Deep Learning

Fraudulent transactions and the methods to detect them are an an important issue for financial organizations globally. The requirement for progressive fraud detection systems to protect properties and maintain customer trust is predominant for financial organizations, but particular factors make the development of efficient and effective fraud detection models a challenge. Deep Learning (DL) has greatly improved fraud detection accuracy by detecting intrinsic patterns, whereas interpretability techniques improve transparency and build trust by making predictions understandable to experts. This study presents a Fraud Detection System using Recursive Feature Elimination and Waterwheel Plant Optimization (FDS-RFEWPO) model for financial transactions. The aim is to perform a comprehensive evaluation of fraud detection in high-dimensional financial transactions using advanced techniques. Initially, the FDS-RFEWPO technique follows min-max-based data pre-processing to normalize the input data. For the feature selection process, the FDS-RFEWPO model employs the Recursive Feature Elimination (RFE) technique to select the most relevant features from the dataset. Furthermore, the Variational Autoencoder/Wasserstein Autoencoder (VAE/WAE) model is employed for fraud detection and classification. To further enhance model performance, the Waterwheel Plant Optimization (WPO) technique is employed for hyperparameter tuning, ensuring the selection of optimal parameters that contribute to improved accuracy. Finally, the Explainable Artificial Intelligence (XAI) technique applies Local Interpretable Model-Agnostic Explanations (LIME) to improve the transparency, interpretability, and trustworthiness of Artificial Intelligence (AI) methods by making their decision-making procedures clear to humans. To evaluate the performance of the FDS-RFEWPO model, a comprehensive experimental analysis is conducted using a financial fraud detection dataset. The comparison study of the FDS-RFEWPO model demonstrates a superior accuracy of 97.41% over existing techniques.

This article has been retracted. A retraction notice can be found at https://doi.org/10.3233/JIFS-219433.

Objective To study the sensitivity and specificity of parameter Corvis biomechanical index (CBI)— a new biomechanical index of Corvis ST in the diagnosis of keratoconus and evaluate the role of CBI in the diagnosis of keratoconus and the change of biomechanic. Methods A diagnostic trial study was adopted, and 66 eyes from 49 keratoconic patients (keratoconus group) and 91 right eyes from 91 myopic patients (control group) from April in 2018 to August in 2018 in Henan Eye Hospital were enrolled.Pentacam and Corvis ST examinations were performed by the same operator after the basic eye examinations.Using the evaluation of diagnostic test, consistency test, and receiver operating characteristic (ROC) curve analysis, gain the outcome of the sensitivity, specificity, consistency, Youden index, and area under the ROC curve of the parameter CBI.This study protocol was approved by Ethic Committee of Henan Eye Hospital and followed the Declaration of Helsinki.Written informed consent was obtained from each patient prior to any medical examination. Results Sixty eyes were diagnosed as keratoconus by the parameter CBI of Corvis ST.Evaluation of diagnosis test: the sensitivity was 97.0%, and the specificity was 97.8%; consistency check: Kappa=0.948, P<0.05.ROC curve analysis: the sensitivity was 98.5%, the specificity was 96.8%, Youden index was 96.3%, P<0.000 1 and AUC was 0.996. Conclusions CBI can separate healthy eyes from keratoconic eyes with highly sensitivity and specificity, which was highly consistency with results of Rabinowitz keratoconus diagnostic criteria.CBI could be used as a new biomechanical indicator for the diagnosis of keratoconus. Key words: Corneal visualization scheimpflug technology; Corneal biomechanics; Keratoconic diagnosis; Corvis biomechanical index; Sensitivity; Specificity

Objective To investigate features of fasting plasma glucose (FPG) during different trimesters and hemoglobin A1c (HbA1c) in the third trimester of pregnant women in Zhujing Region of Shanghai. Methods A total of 606 cases of pregnant women who received prenatal examination and 75 g oral glucose tolerance test (OGTT) in Jinshan Branch of Shanghai Sixth People′s Hospital from January 1 to December 31, 2014 were collected as research subjects. According to the results of 75 g OGTT, they were divided into gestational diabetes mellitus (GDM) group (n=136) and non-GDM group (n=470). The clinical data, such as age, body mass index (BMI), as well as the concentration of FPG during different pregnancy trimesters and level of HbA1c in the third trimester were collected retrospectively in both two groups by retrospective analysis method. According to the quartile method, level of HbA1c in the third trimester of all subjects were divided into four ranges: Q1 (HbA1c level<5.1%), Q2 (5.1%≤HbA1c level<5.2%), Q3 (5.2%≤HbA1c level<5.5%), and Q4 (HbA1c level≥5.5%), respectively. The age, BMI, concentrations of FPG and the level of HbA1c between two groups were compared by independent-samples t test. Chi-square test was used to compare the GDM incidence of all subjects with different ranges of HbA1c level in the third trimester, and further comparison was conducted by adjusting inspection level. Then receiver operator characteristic (ROC) curve of HbA1c level in the third trimester for predicting the incidence of GDM was drawn, and the area under ROC curve (ROC-AUC) was calculated. The optimal critical value of HbA1c level in the third trimester for predicting the incidence of GDM was obtained when the Youden index reaching the maximum value. And its sensitivity and specificity were calculated. This study met the requirements of the World Medical Association Declaration of Helsinki revised in 2013. Results ①The age and BMI of GDM group were higher than those in non-GDM group, and there were statistically significant differences (t=5.306, 5.250; P<0.001). Among the subjects in GDM group, the concentration of FPG in second trimester was the highest among 3 trimesters, which was (4.9±0.6) mmol/L, and the concentration of FPG in second trimester was higher than that in first and third trimester, respectively, and both the differences were statistically significant (t=2.087, 1.960; P=0.039, 0.041). Among the subjects in non-GDM group, the concentration of FPG in first trimester was the highest among 3 trimesters, which was (4.6±0.3) mmol/L, and the concentration of FPG in first trimester was higher than that in second and third trimester, respectively, and both the differences were statistically significant (t=15.230, 5.613; P<0.001). The concentration of FPG in first, second and third trimesters of GDM group were higher than those of non-GDM group, respectively, and there were statistically significant differences (t=5.416, 15.526, 4.471; P<0.001). Besides, HbA1c level in third trimester of GDM group was (5.6±0.4)%, which was higher than that of non-GDM group (5.4±0.4)%, and there was significant difference (t=5.845, P<0.01). ②There was statistical difference in GDM incidences in all subjects with different ranges of HbA1c levels in third trimester (χ2=22.707, P<0.001). Multiple comparison results showed that the incidences of GDM in all subjects with Q1 and Q2 ranges of HbA1c levels in third trimester were higher than that in subjects with Q4 range of HbA1c levels in third trimester, respectively, and both the differences were statistically significant (χ2=15.071, 16.785; P<0.001). ③The results of ROC curve analysis of HbA1c level in third trimester in predicting the incidence of GDM showed that the ROC-AUC was 0.647 (95%CI: 0.625-0.672, P<0.001), and the optimal cut-off value of HbA1c level in third trimester to predict GDM incidence was 5.5%, and the sensitivity of HbA1c level in predicting the incidence of GDM was 63.4%, and the sensitivity was 69.7%. Conclusions The concentrations of FPG in GDM pregnant women in Zhujing Region of Shanghai begin to increase from the first trimester and peak during the second trimester, and decrease during the third trimester after lifestyle intervention. It is advisable for GDM pregnant women to control HbA1c levels below 5.5% during the third trimester. Key words: Diabetes, gestational; Fasting plasma glucose; Hemoglobin A, glycosylated; Glucose tolerance test; ROC curve; Forecasting; Pregnant women

The escalating threat of fraud in financial institutions is a global issue, with the Malaysian sector being no exception. This study focuses on the implementation and efficacy of Data Mining and Machine Learning methodologies in identifying and mitigating fraudulent activities within these institutions. The paper critically reviews existing literature, bridging the gap between advanced technology application and fraud management. Fraudulent transactions in the financial sector are dynamic and sophisticated, requiring advanced detection techniques. Traditional approaches often struggle to manage this complexity effectively, demonstrating a need for more advanced and adaptive strategies. This is where Data Mining and Machine Learning techniques, renowned for their predictive and analytical prowess, can significantly contribute. Data Mining, the process of uncovering patterns and correlations within large datasets, is a useful tool for detecting anomalies that may suggest fraud. The study assesses various data mining techniques, such as clustering, classification, and association, and explores their application in detecting fraudulent transactions. Findings indicate that these techniques can substantially enhance fraud detection rates while minimizing false positives. Furthermore, Machine Learning, an artificial intelligence subset, has shown immense potential in fraud detection. Its ability to learn from and make decisions based on data makes it a viable solution for fraud detection. This paper explores both supervised and unsupervised learning algorithms and their efficacy in identifying fraud in the Malaysian financial sector. Results suggest that machine learning models, when correctly implemented, can significantly improve the accuracy of fraud detection. The review underscores the importance of employing advanced technologies like Data Mining and Machine Learning to combat financial fraud effectively. It also suggests future research directions, emphasizing the need for context-specific, localized models considering Malaysia's unique socio-economic environment. Moreover, the development of hybrid models, integrating both data mining and machine learning, could offer improved results. In conclusion, this study sets a precedent for further exploration into the application of advanced analytical tools in fraud detection in the Malaysian financial sector. The potential these technologies offer for improving accuracy and adaptability in fraud detection systems is substantial and warrants thorough investigation. Keywords: Fraud Detection, Malaysian Financial Institutions, Data Mining, Machine Learning, Financial Fraud Management

This study aims to compare deep learning explainability (DLE) with explainable artificial intelligence and causal artificial intelligence (Causal AI) for fraud detection, emphasizing their distinct methodologies and potential to address critical challenges, particularly in finance. An empirical evaluation was conducted using the Bank Account Fraud datasets from NeurIPS 2022. DLE models, including deep learning architectures enhanced with interpretability techniques, were compared against Causal AI models that elucidate causal relationships in the data. DLE models demonstrated high accuracy (95% for Model A and 96% for Model B) and precision (97% for Model A and 95% for Model B) but exhibited reduced recall (98% for Model A and 97% for Model B) due to opaque decision-making processes. By contrast, Causal AI models showed balanced but lower performance with accuracy, precision, and recall, all at 60%. These findings underscore the need for transparent and reliable fraud detection systems, highlighting the trade-offs between model performance and interpretability. This study addresses a significant research gap by providing a comparative analysis of DLE and Causal AI in the context of fraud detection. The insights gained offer practical recommendations for enhancing model interpretability and reliability, contributing to advancements in AI-driven fraud detection systems in the financial sector.

Research Type: Level 3 - Retrospective cohort study, Case-control study, Meta-analysis of Level 3 studies Introduction/Purpose: The recently proposed Progressive Collapsing Foot Deformity (PCFD) classification system categorizes different deformity patterns into five classes (A, Hindfoot Valgus; B, Midfoot/Forefoot Abduction; C, Arch Collapse; D, Peritalar Subluxation; E, Ankle Valgus Instability). Currently, the decision-making process to determine the presence or absence of each deformity class relies on the surgeon's experience and interpretation of clinical and radiographic assessments, without established cut-off threshold values in the literature for PCFD measurements. This limits the reliability and reproducibility of the classification system. The goal of the current study was to establish cut-off threshold values for the most commonly used PCFD measurements assessing Class A to D deformities, using a large cohort of PCFD patients and controls. Methods: Two patient cohorts were included: a prospective cohort consisting of normally aligned and asymptomatic control volunteers (n=197 patients, 103 females/94 males, average age 41.7 years, average BMI 28.9) and a retrospective cohort of PCFD patients (n=321 patients, 136 females/185 males, average age 50.7 years, average BMI 29.8). All patients underwent Weight-Bearing Computed Tomography (WBCT) scans of the foot and ankle. Foot bone segmentation was performed semi-automatically using specialized software, followed by automatic measurements of various parameters including: Class A (Hindoot Moment Arm – HMA), Class B (Talonavicular Coverage Angle – TCA; Talus-First Metatarsal Angle Axial – TFMA-A), Class C (Talus-First Metatarsal Angle Sagittal – TFMA-S; Forefoot Arch Angle - FAA), Class D (Coverage Maps of the Subtalar Joint Middle-MF, Posterior Facets-PF; and Sinus Tarsi-ST). Receiver Operating Characteristic (ROC) curves, Youden's Indexes and Areas Under the Curve (AUC) were calculated for each measurement. P-values of 0.05 or less were considered significant. Results: The Cut-Off Threshold Values identified by the ROC curves using Youden’s Index for different PCFD measurements were as follows (all p-values &lt; 0.0001): HMA of 13.9mm or more (AUC 85.3%) was considered diagnostic for Class A PCFD deformity. For Class B PCFD deformity, a TNCA of 38.7o or higher (AUC 84.4%) and TFMA-A of 20.3o or higher (AUC 82.1%) were considered diagnostic. For Class C PCFD deformity, a FAA of 8.7o or lower (AUC 83.9%) and TFMA-S of 18.7o or higher (AUC 82.9%) were diagnostic. For Class D Deformity, Coverage values for Middle Facet of 73.5% or lower (AUC 73.4%), Posterior Facet of 84.5% or lower (AUC 82.7%), and Sinus Tarsi of 25.7% or higher (AUC 84.4%) were diagnostic. Conclusion: In this diagnostic study involving 312 PCFD patients and 197 volunteer controls, we successfully determined the cut-off threshold values for key PCFD measurements related to Hindfoot Valgus (Class A), Midfoot/Forefoot Abduction (Class B), Arch Collapse (Class C), and Peritalar Subluxation (Class D) deformities. By utilizing a large cohort of patients and controls, we achieved high diagnostic accuracy for most measurements assessed (AUC of 80% or more) in identifying the presence or absence of these PCFD deformity classes. Our findings can potentially serve as normative threshold values to enhance the utilization, reliability, and reproducibility of the PCFD classification system. Figure 1 Scatter plots and ROC (Receiver Operating Characteristic) curves used to analyze parameters for the various deformity classifications in Progressive Collapsing Foot Deformity (PCFD). For the scatter plots, red indicates a normal control, green indicates an asymptomatic PCFD, and blue indicates a symptomatic PCFD.

Credit card fraud has caused a huge loss to both banks and consumers in recent years. Thus, an effective Fraud Detection System (FDS) is important to minimize the loss for banks and cardholders. Based on our analysis, the credit card transaction dataset is very skewed, there are much fewer samples of frauds than legitimate transactions. Furthermore, due to the data security and privacy, different banks are usually not allowed to share their transaction datasets. These problems make FDS difficult to learn the patterns of frauds and also difficult to detect them. In this paper, we propose a framework to train a fraud detection model using behavior features with federated learning, we term this detection framework FFD (Federated learning for Fraud Detection). Different from the traditional FDS trained with data centralized in the cloud, FFD enables banks to learn fraud detection model with the training data distributed on their own local database. Then, a shared FDS is constructed by aggregating locally-computed updates of fraud detection model. Banks can collectively reap the benefits of shared model without sharing the dataset and protect the sensitive information of cardholders. Furthermore, an oversampling approach is combined to balance the skewed dataset. We evaluate the performance of our credit card FDS with FFD framework on a large scale dataset of real-world credit card transactions. Experimental results show that the federated learning based FDS achieves an average of test AUC to 95.5%, which is about 10% higher than traditional FDS.

Area under the curve may hide poor generalisation to external datasets

This study aims to develop an attention-based deep learning model for distinguishing benign from malignant breast lesions on CESM. Preoperative CESM images of 1239 patients, which were definitely diagnosed on pathology in a multicentre cohort, were divided into training and validation sets, internal and external test sets. The regions of interest of the breast lesions were outlined manually by a senior radiologist. We adopted three conventional convolutional neural networks (CNNs), namely, DenseNet 121, Xception, and ResNet 50, as the backbone architectures and incorporated the convolutional block attention module (CBAM) into them for classification. The performance of the models was analysed in terms of the receiver operating characteristic (ROC) curve, accuracy, the positive predictive value (PPV), the negative predictive value (NPV), the F1 score, the precision recall curve (PRC), and heat maps. The final models were compared with the diagnostic performance of conventional CNNs, radiomics models, and two radiologists with specialised breast imaging experience. The best-performing deep learning model, that is, the CBAM-based Xception, achieved an area under the ROC curve (AUC) of 0.970, a sensitivity of 0.848, a specificity of 1.000, and an accuracy of 0.891 on the external test set, which was higher than those of other CNNs, radiomics models, and radiologists. The PRC and the heat maps also indicated the favourable predictive performance of the attention-based CNN model. The diagnostic performance of two radiologists improved with deep learning assistance. Using an attention-based deep learning model based on CESM images can help to distinguishing benign from malignant breast lesions, and the diagnostic performance of radiologists improved with deep learning assistance.

The state-of-the-art Credit Card Fraud Detection (CCFD) systems utilize supervised rule-based schemes to identify fraudulent behaviors in static data. Designing an efficient framework for CCFD is particularly challenging due to the imbalanced distribution of positive and negative classes in data. Conventional works overlook hidden transaction patterns and the performance evaluation methods of algorithm or system. The early models have also been criticized for over-fitting, i.e. models having less prediction accuracy than training or classification accuracy. However, the categorical accuracy is more important than overall accuracy for extremely imbalanced data; most of the previous works do not evaluate their model on the basis of categorical accuracy. In this work, we propose a predictive framework (DEAL) based on novel Deep Ensemble Learning for detecting fraudulent transaction in real-time data stream. The proposed framework is adaptive against data imbalance and robust against latent transaction patterns such as spending behavior. To build the model and predict the fraud, the transactions from real data from a major bank are passed as tensors to ensemble of Deep Learning Network. The adaptive optimization is suggested to minimize the objective function and improve fraud prediction. All Experiments are conducted in Python using scikit-learn, Google TensorFlow, and Keras Deep Learning libraries. The results demonstrate its superiority over state-of-the-art methods in catching frauds.

Objective To analyze the expression of interleukin (IL)-31 in tuberculous pleural effusion, and to evaluate its diagnostic value of tuberculous effusion. Methods Seventy-one patients with pleural effusion were enrolled, including 40 cases of tuberculous pleural effusion and 31 cases of malignant pleural effusion. Luminex method was applied to detect the IL-31 expression in pleural effusion. IL-31 levels were compared using non-parametric Mann-Whitney U test, and the receiver operator characteristic (ROC) curve was used to elvaluate the diagnostic value of IL-31. Results IL-31 expression in tuberculous pleural effusion was significantly higher than that in malignant pleural effusion with statistical significance (529.4 ng/L vs 13.8 ng/L, U=62, P<0.01). Based on the level of IL-31 expression, area under the ROC curve was 0.95 with the optimum cut-off value of 67.5 ng/L. Thus, the sensitivity and specificity of IL-31≥67.5 ng/L for diagnosis of tuberculous pleurisy were 82.5% (95%CI: 73.3%-94.2%) and 100.0% (95%CI: 91.4%-100.0%), respectively. Conclusion IL-31 is highly sensitive and specific for the diagnosis of tuberculous pleural effusion, which favors the differentiation of tuberculosis from malignance. Key words: Interleukin-31; Pleural effusion, malignant; Tuberculosis, pleural; Differential diagnosis

BackgroundIn Denmark, a nationwide mandatory infliximab biosimilar-to-biosimilar switch, from CT-P13 to GP1111, was recently conducted in patients with rheumatoid arthritis (RA) to save costs (ref). It is of interest to...

Identifying final pathology of FNA-indeterminant nodules before surgical resection could decrease the number of unnecessary surgeries and total cost to patients. This project explores how radiomics (RM) and deep learning (DTLM) models may be combined to improve the potential for clinical interpretability of machine learning models in the task of classifying indeterminant thyroid nodules on ultrasound. Two radiomic and deep learning combination models were created: a simple classifier combination model (SCM) and an interpretability-driven combination model (ICM). SCM provided a nodule malignancy score. ICM merged radiomic and deep learning features through correlation and provided echogenicity-related, composition-related, and shape/margin-related malignancy scores which were averaged to yield an overall nodule malignancy score. Models were trained and tested on a de-identified dataset of 476 grayscale ultrasound images collected under IRB approval containing 222 images from 69 indeterminant nodules with a final pathology of malignant and 254 images from 82 indeterminant nodules with a final pathology of benign. Models were tested using 5- fold cross-validation by nodule over 100 iterations. Receiver-operating characteristic (ROC) analysis was conducted with area under the ROC curve (AUC) serving as the statistic of merit for model performance. Models yielded mean AUC [95%CI] of 0.75 [.67,.83], 0.70 [.62,.78], 0.77 [0.70,0.84], 0.76 [.69,.84] for RM, DTLM, SCM, and ICM respectively. This work failed to demonstrate a statistically significant difference in model performances. However, the ICM presents a novel method for combining radiomics and deep learning features focused on improving interpretability for clinical implementation in the task of indeterminant thyroid nodule classification.

Objective To evaluate the relationship between levels of serum angiopoietin (Ang)-2, interleukin (IL)-6 and tumor necrosis factor (TNF)-α, and the severity of sepsis in children, and to provide evidence for clinical diagnosis and treatment of children with sepsis. Methods A total of 64 cases of children with fever who were treated in Department of Pediatrics, Jinhua Hospital of Zhejiang University and The Children′s Hospital, Zhejiang University School of Medicine were selected as research subjects. Their age ranged from 1 month after birth to 12 years old. They were divided into three groups according to whether they suffered from sepsis or not and the grading standards of sepsis: group A (n=22, children with sepsis), group B [(n=24, children with infection and without systemic inflammatory response syndrome (SIRS)], and group C (n=18, children with severe sepsis). In the same period, another 21 children underwent health examination in Department of Child Healthcare, Jinhua Hospital of Zhejiang University were included into control group. Their age ranged from 3 months after birth to 12 years old. Enzyme-linked immunosorbent assay (ELISA) was used to detect the levels of serum Ang-2, IL-6 and TNF-α. The receiver operating characteristic (ROC) curve of level of serum Ang-2 for diagnosing children with severe sepsis among 4 groups were drawn, and the area under the ROC curve (ROC-AUC) was calculated. The optimal critical value of serum Ang-2 level for diagnosing children with severe sepsis was determined according to the principle of maximum Youden index, and its sensitivity and specificity were calculated. Linear correlation analysis was used to analyze the correlation between levels of serum Ang-2 level and serum IL-6 and TNF-α. This study was in line with the requirements of World Medical Association Declaration of Helsinki revised in 2013. The guardians of all children signed the clinical research informed consents. Results ① There were no significant differences in age, gender composition ratio and other general clinical data among 4 groups (P>0.05). ②There were significant differences among 4 groups in the levels of serum Ang-2, TNF-α and IL-6 (F=261.24, 33.82, 28.00; P all<0.001). Further comparison results showed that the levels of serum Ang-2 and TNF-α in group C were significantly higher than those in group A, group B and control group, and all the differences were statistically significant (level of serum Ang-2: LSD-t=-5.35, -15.98, 25.31, P all<0.001; level of serum TNF-α: LSD-t=-2.24, P=0.030, LSD-t=-4.91, P<0.001, LSD-t=8.59, P<0.001). The level of serum IL-6 in group C was higher than that in group B and control group, respectively, and both the differences were statistically significant (LSD-t=-4.93, 8.50; P all<0.001). ③The results of ROC curve analysis showed that ROC-AUC of serum Ang-2 level in diagnosis of children with severe sepsis was 0.998 (95%CI: 0.946-1.000, P<0.001). According to the principle of maximum Youden index, the optimal critical value of serum Ang-2 level to predict children with severe sepsis was 163.4 ng/L, and its sensitivity and specificity for diagnosis of children with severe sepsis were 100.0% and 98.5%, respectively. ④Among the children in 4 groups, serum Ang-2 level was positively correlated with serum IL-6 and TNF-α levels, respectively (r=0.606, 0.556; P all<0.001). Conclusions The level of serum Ang-2 is significantly elevated with the severity of children with sepsis, which can be used to evaluate the severity of children with sepsis. The levels of serum IL-6 and TNF-α are correlated with serum Ang-2 level, which can assist in the diagnosis of children with sepsis. Key words: Sepsis; Angiopoietin-2; Interleukin-6; Tumor necrosis factor-alpha; ROC curve; Child