AUC Measures Research Articles

Prediction of pancreatic cancer in patients with new onset hyperglycemia: A modified ENDPAC model

Background/Objectives: The Enriching New-Onset Diabetes for Pancreatic Cancer (ENDPAC) model relies primarily on fasting glucose values. Health systems have increasingly shifted practice towards use of glycated hemoglobin (HbA1c) measurement. We modified the ENDPAC model using patients with new onset hyperglycemia. MethodsFour cohorts of patients 50-84 years of age with HbA1c results ≥6.2-6.5% in 2011-2018 were identified. A combine cohort was formed. A widened eligibility criterion was applied to form additional four individual cohorts and one combined cohort. The primary outcome was the diagnosis of pancreatic cancer within 3 years after the first elevated HbA1c testing. The performance of the modified ENDPAC model was evaluated by AUC, sensitivity, positive predictive value, cases detected, and total number of patients screened. ResultsThe individual and combined cohorts consisted of 39,001-79,060 and 69,334-92,818 patients, respectively (mean age 63.5-65.0 years). The three-year PC incidence rates were 0.47%-0.54%. The AUC measures were in the range of 0.75-0.77 for the individual cohorts and 0.75 for the combined cohorts. When the four individual cohorts were combined, more PC cases can be identified (149 by the combined vs. 113-116 by individual cohorts when risk score was 5+). Performance measures were compromised in nonwhites. Asian and Pacific islanders had lower sensitivity compared to other racial and ethnic groups (29% vs. 50-60%) when risk score was 5+. ConclusionsThe modified ENDPAC model targets a broader population and thus identifies more high-risk patients for cancer screening. The differential performance needs to be considered when the model is applied to non-white population.

DRN-CDR: A cancer drug response prediction model using multi-omics and drug features

Cancer drug response (CDR) prediction is an important area of research that aims to personalize cancer therapy, optimizing treatment plans for maximum effectiveness while minimizing potential negative effects. Despite the advancements in Deep learning techniques, the effective integration of multi-omics data for drug response prediction remains challenging. In this paper, a regression method using Deep ResNet for CDR (DRN-CDR) prediction is proposed. We aim to explore the potential of considering sole cancer genes in drug response prediction. Here the multi-omics data such as gene expressions, mutation data, and methylation data along with the molecular structural information of drugs were integrated to predict the IC50 values of drugs. Drug features are extracted by employing a Uniform Graph Convolution Network, while Cell line features are extracted using a combination of Convolutional Neural Network and Fully Connected Networks. These features are then concatenated and fed into a deep ResNet for the prediction of IC50 values between Drug – Cell line pairs. The proposed method yielded higher Pearson’s correlation coefficient (rp) of 0.7938 with lowest Root Mean Squared Error (RMSE) value of 0.92 when compared with similar methods of tCNNS, MOLI, DeepCDR, TGSA, NIHGCN, DeepTTA, GraTransDRP and TSGCNN. Further, when the model is extended to a classification problem to categorize drugs as sensitive or resistant, we achieved AUC and AUPR measures of 0.7623 and 0.7691, respectively. The drugs such as Tivozanib, SNX-2112, CGP-60474, PHA-665752, Foretinib etc., exhibited low median IC50 values and were found to be effective anti-cancer drugs. The case studies with different TCGA cancer types also revealed the effectiveness of SNX-2112, CGP-60474, Foretinib, Cisplatin, Vinblastine etc. This consistent pattern strongly suggests the effectiveness of the model in predicting CDR.

Long-term efficacy, safety and PK data of TH1902 (sudocetaxel zendusortide) in solid tumors: A novel SORT1-targeting peptide-drug-conjugate (PDC).

3081 Background: TH1902 is a novel SORT1 targeting PDC. It exploits the natural function of SORT1, a scavenger receptor, which rapidly internalizes its natural ligands via endocytosis (internalization half-life ≤4 min). SORT1 is highly expressed in multiple solid tumors compared to normal healthy tissues making it an attractive target for rapid delivery of anti-cancer therapeutics. In this updated analysis, further data on long-term efficacy, safety and PK is presented from Parts 1&2 of Ph1 with focus on the 300 mg/m2 q3w. Due to safety observations in 6 patients (pts) enrolled at 420 mg/m2 dose level (Gr 3 neuropathy [n=2], Gr 4 neutropenia [n=2], Gr 3 ocular [n=1] and Gr 2 skin [n=3] toxicities, the dose was reduced to 300 mg/m2 in Part 2. Methods: Primary objective of the study was to characterize the safety/tolerability of TH1902. Part 1 (modified intrapatient dose escalation) included pts with recurrent/refractory advanced solid tumors (all comers) with no limit on number of previous therapies, including taxanes. Part 2 (dose expansion) included pts with known high SORT1 expression (e.g. OVC, endometrial, TNBC, melanoma). Results: Twenty-five heavily pretreated pts were enrolled in the 300 mg/m2 group. At least one TRAE was observed in 80% of pts. Grade 3 (Gr 3) events of interest were neuropathy (12%), keratitis (8%), anemia (8%), and neutropenia (4%), for an overall incidence of 32%. All grade neuropathy was 28%. Most TRAEs were mild to moderate severity and manageable with standard supportive care or dose reductions. Safety profile of TH1902 was different from that of docetaxel. PK measures of Cmax and AUC showed that exposure to free docetaxel was much lower than that of TH1902; Cmax = 0.58 μM for free docetaxel vs 30.4 μM for TH1902 and AUC24 = 3.1 h.nmol/mL for free docetaxel vs 74.8 h.nmol/mL for TH1902. Three pts exhibited RECIST 1.1 confirmed long-term stabilizations of disease, even after drug discontinuation, which ranged from 8 to 19 mos from treatment initiation, of which one OvC pt had overall PR (with RECIST 1.1 confirmed CR in target lesions) and remained on treatment for a total of 5 months. In addition, one endometrial cancer (part 1) was dose escalated from 60-360 mg/m2 and completed 11 cycles in total. Pt remained in SD during the 8 mos of treatment, up to time of consent withdrawal. All 4 pts had prior taxane exposure. Conclusions: TH1902 induces durable disease stabilization that lasts beyond treatment completion, suggesting a unique, multimodal mechanism of action (MOA) that differs from other cancer therapeutics. TH1902 has a manageable safety profile at 300mg/m2 with few Gr3 AEs. Low levels of free docetaxel in human plasma may in part explain the low rate of taxane related AEs (i.e. neutropenia, no alopecia). Next phase of the study involves dose optimization to further limit toxicity and improve efficacy. Clinical trial information: NCT04706962 .

Block-wise imputation EM algorithm in multi-source scenario: ADNI case

Alzheimer’s disease is the most common form of dementia and the early detection is essential to prevent its proliferation. Real data available has been of paramount importance in order to achieve progress in the automatic detection despite presenting two major challenges: Multi-source observations containing Magnetic resonance (MRI), Positron emission tomography (PET) and Cerebrospinal fluid data (CSF); and also missing values within all these sources. Most machine learning techniques perform this predictive task by using a single data modality. Nevertheless, the integration of all these sources of evidence could possibly bring a higher performance at different stages of disease progression. The Expectation Maximization (EM) algorithm has been successfully employed to handle missing values, but it is not designed for typical Machine Learning scenarios where an imputation model is created over training data and subsequently applied on a testing set. In this work, we propose EMreg-KNN, a novel supervised and multi-source imputation algorithm. Based on the EM algorithm, EMreg-KNN builds a regression ensemble model for the imputation of future data thus allowing the further utilization of any vector-based Machine Learning method to automatically assess the Alzheimer’s disease diagnosis. Using the ADNI database, the proposed method achieves significant improvements on F1, AUC and Accuracy measures over classical imputation methods for this database using four classification algorithms. Considering these classifiers in four different classification scenarios, our algorithm is experimentally superior in terms of the F measure, in nearly 82% of the cases under evaluation.

Exploring various measures of the area under the curve for the assessment of dose-proportionality and estimation of bioavailability.

In pharmacokinetics, the area under the concentration versus time curve (AUC) extrapolated to infinity (AUC0-∞) is the preferred metric but it is not always possible to have a reliable estimate of the terminal phase half-life. Here we sought to explore the accuracy of three different area measures to accurately identify dose proportionality and bioavailability. One to three compartment model simulations with different doses for dose-proportionality or different rates and/or extents of bioavailability. Area measures evaluated were AUC0-∞, to the last quantifiable concentration (AUCtlast), and to a common time value (AUCt'). Under linear pharmacokinetics, AUCt' provided the most accurate measure of dose proportionality. Except for the one compartment model where AUC0-∞ provided the best predictor of the true measure, there was no clear advantage to the use of either of the three measures of AUC. With uncertainty about the terminal phase half-life, the use of AUCt' can be a very useful and even the preferred measure of exposure for use in assessing proportionality in exposure between doses. The choice of AUC measure in bioavailability is less clear and may depend on compartmental nature of the drug, and study parameters including assay sensitivity and sampling protocols.

Evaluation of Statistical Approaches in Developing a Predictive Model of Severe COVID-19 during Early Phase of Pandemic with Limited Data Resources.

As evidence of risk factors for severe cases of coronavirus disease 2019 (COVID-19) was uncertain in early phases of the pandemic, the development of an efficient predictive model for severe cases to triage high-risk individuals represented an urgent yet challenging issue. It is crucial to select appropriate statistical models when available data and evidence are limited. This study was conducted to assess the accuracy of different statistical models in predicting severe cases using demographic data from patients with COVID-19 prior to the emergence of consequential variants. We analyzed data from 929 consecutive patients diagnosed with COVID-19 prior to March 2021, including their age, sex, body mass index, and past medical histories, and compared areas under the receiver operating characteristic curve (ROC AUC) between different statistical models. The random forest (RF) model, deep learning (DL) models with not too many neurons, and naïve Bayes model exhibited AUC measures of > 0.70 with the validation datasets. The naïve Bayes model performed the best with the AUC measures of > 0.80. The accuracies in RF were more robust with narrower distribution of AUC measures compared to those in DL. The benefit of performing feature selection with a training dataset before building models was seen in some models, but not in all models. In summary, the naïve Bayes and RF models exhibited ideal predictive performance even with limited available data. The benefit of performing feature selection before building models with limited data resources depended on machine learning methods and parameters.

PU‐GNN: A Positive‐Unlabeled Learning Method for Polypharmacy Side‐Effects Detection Based on Graph Neural Networks

The simultaneous use of multiple drugs, known as polypharmacy, heightens the risks of harmful side effects due to drug‐drug interactions. Predicting these interactions is crucial in drug research due to the rising prevalence of polypharmacy. Researchers employ a graphical structure to model these interactions, representing drugs and side effects as nodes and their interactions as edges. This creates a multipartite graph that encompasses various interactions such as protein‐protein interactions, drug‐target interactions, and side effects of polypharmacy. In this study, a method named PU‐GNN, based on graph neural networks, is introduced to predict drug side effects. The proposed method involves three main steps: (1) drug features extraction using a novel biclustering algorithm, (2) reducing uncertainity in input data using a positive‐unlabeled learning algorithm, and (3) prediction of drug’s polypharmacies by utilizing a graph neural network. Performance evaluation using 5‐fold cross‐validation reveals that PU‐GNN surpasses other methods, achieving high scores of 0.977, 0.96, and 0.949 in the AUPR, AUC, and F1 measures, respectively.

Identifying the Most Important Factors in Determining the Osteoporosis in Women Using Data Mining Techniques

Osteoporosis is one of the primary causes of disability and mortality in the elderly. If osteoporosis's significant features can be identified, the risk of developing this disease will be reduced. In recent years, data mining approaches have become a suitable tool for medical researchers. This study applied data mining methods to identify osteoporosis’s significant features. This study applied data from women having osteoporosis or osteopenia in the period 2011-2019 in the Osteoporosis Diagnosis Center, Isfahan, Iran. Data mining methods such as linear regression, naïve bayes, decision tree, support vector machine, random forest, and neural network were implemented on the dataset. This study consisted of 8258 patients’ information, of which 1482 had osteoporosis. The results showed that the support vector machine, decision tree, neural network are the best method based on accuracy, precision, and AUC measures. Six candidate features were age, weight, back pain, low activity, menopause date, and previous fracture. Support vector machine, decision tree, and neural network are the best candidate techniques for predicting osteoporosis. Thin older people are more at risk of osteoporosis than other people. Yet, people with middleweight and middle age are at lower risk of osteoporosis.

Hybrid Feature Selection based on BTLBO and RNCA to Diagnose the Breast Cancer

Feature selection is a feasible solution to improve the speed and performance of machine learning models. Optimization algorithms are doing a significant job in searching for optimal variables from feature space. Recent feature selection methods are purely depending on various meta heuristic algorithms for searching a good combination of features without considering the importance of individual features, which makes classification models to suffer from local optima or overfitting problems. In this paper, a novel hybrid feature subset selection technique is introduced based on Regularized Neighborhood Component Analysis (RNCA) and Binary Teaching Learning Based Optimization (BTLBO) algorithms to overcome the above problems. RNCA algorithm assigns weights to the attributes based on their contribution in building the learning models for classification. BTLBO algorithm computes the fitness of individuals with respect to the weights of features and selects the best ones. The results of similar feature selection methods are matched with the proposed hybrid model and proved better performance in terms of classification accuracy, recall and AUC measures over breast cancer datasets.

Features Extraction Based on Probability Weighting for Fake News Classification on Social Media

Fake news is a massive problem globally, especially on social media. Most people spend a lot of time consuming social media every day, and it is very possible for people as social media users to receive fake news without realizing it. Primarily due to this situation, we developed a machine learning tool to detect fake news that operates with the aid of various algorithms such as Decision Tree, K-Nearest Neighbor, and Naïve Bayes. Our experiement is tested based on machine learning that selected only one technique used to classify the data by finding the model set. In addition, the performance of the set describes the classification of the model and the inconsistency solution for each iteration. This study proposed a model which used the probability weighting of the model in features extraction processing for data classification. The concept is the enhancement of probability weighting features that converge exactly the class labels of classification. Our work was also implemented based on traditional Count Vectorizer and TF-IDF Vectorizer sentiment analysis and combined probability weighting features for fake news articles. The experimental results of the work illustrate that the best accuracy achieved by a proposed model used probability weighting features to find out the impact of classifiers models. In addition, the results of experimental information is represented by enhancing the overall performance of Decision Tree, K-Nearest Neighbor, and Naïve Bayes with various datasets. In addition, the measures of precision, recall, F1-measure, AUC, and accuracy for each class and deep in each class were achieved and reached the highest performance of the proposed model.

Stuttering Disfluency Detection Using Machine Learning Approaches

Stuttering is a neurodevelopmental speech disorder wherein people suffer from disfluency in speech generation. Recent research has applied machine learning and deep learning approaches to stuttering disfluency recognition and classification. However, these studies have focussed on small datasets, generated by a limited number of speakers and within specific tasks, such as reading. This paper rigorously investigates the effective use of eight well-known machine learning classifiers, on two publicly available datasets (FluencyBank and SEP-28k) to automatically detect stuttering disfluency using multiple objective metrics, i.e. prediction accuracy, recall, precision, F1-score, and AUC measures. Our experimental results on the two datasets show that the Random Forest classifier achieves the best performance, with an accuracy of 50.3% and 50.35%, a recall of 50% and 42%, a precision of 42% and 46%, and an F1 score of 42% and 34%, against the FluencyBank and SEP-28K datasets, respectively. Moreover, we show that the machine learning-based approaches may not be effective in accurate stuttering disfluency evaluation, due to diverse variations in speech rate, and differences in vocal tracts between children and adults. We argue that the use of deep learning approaches and Automatic Speech Recognition (ASR) with language models may improve outcomes, specifically for large scale and imbalanced datasets.

Applications of Bayesian shrinkage prior models in clinical research with categorical responses

BackgroundPrediction and classification algorithms are commonly used in clinical research for identifying patients susceptible to clinical conditions such as diabetes, colon cancer, and Alzheimer’s disease. Developing accurate prediction and classification methods benefits personalized medicine. Building an excellent predictive model involves selecting the features that are most significantly associated with the outcome. These features can include several biological and demographic characteristics, such as genomic biomarkers and health history. Such variable selection becomes challenging when the number of potential predictors is large. Bayesian shrinkage models have emerged as popular and flexible methods of variable selection in regression settings. This work discusses variable selection with three shrinkage priors and illustrates its application to clinical data such as Pima Indians Diabetes, Colon cancer, ADNI, and OASIS Alzheimer’s real-world data.MethodsA unified Bayesian hierarchical framework that implements and compares shrinkage priors in binary and multinomial logistic regression models is presented. The key feature is the representation of the likelihood by a Polya-Gamma data augmentation, which admits a natural integration with a family of shrinkage priors, specifically focusing on Horseshoe, Dirichlet Laplace, and Double Pareto priors. Extensive simulation studies are conducted to assess the performances under different data dimensions and parameter settings. Measures of accuracy, AUC, brier score, L1 error, cross-entropy, and ROC surface plots are used as evaluation criteria comparing the priors with frequentist methods as Lasso, Elastic-Net, and Ridge regression.ResultsAll three priors can be used for robust prediction on significant metrics, irrespective of their categorical response model choices. Simulation studies could achieve the mean prediction accuracy of 91.6% (95% CI: 88.5, 94.7) and 76.5% (95% CI: 69.3, 83.8) for logistic regression and multinomial logistic models, respectively. The model can identify significant variables for disease risk prediction and is computationally efficient.ConclusionsThe models are robust enough to conduct both variable selection and prediction because of their high shrinkage properties and applicability to a broad range of classification problems.

Explainable Artificial Intelligence Approach for the Early Prediction of Ventilator Support and Mortality in COVID-19 Patients

Early prediction of mortality and risk of deterioration in COVID-19 patients can reduce mortality and increase the opportunity for better and more timely treatment. In the current study, the DL model and explainable artificial intelligence (EAI) were combined to identify the impact of certain attributes on the prediction of mortality and ventilatory support in COVID-19 patients. Nevertheless, the DL model does not suffer from the curse of dimensionality, but in order to identify significant attributes, the EAI feature importance method was used. The DL model produced significant results; however, it lacks interpretability. The study was performed using COVID-19-hospitalized patients in King Abdulaziz Medical City, Riyadh. The dataset contains the patients’ demographic information, laboratory investigations, and chest X-ray (CXR) findings. The dataset used suffers from an imbalance; therefore, balanced accuracy, sensitivity, specificity, Youden index, and AUC measures were used to investigate the effectiveness of the proposed model. Furthermore, the experiments were conducted using original and SMOTE (over and under sampled) datasets. The proposed model outperforms the baseline study, with a balanced accuracy of 0.98 and an AUC of 0.998 for predicting mortality using the full-feature set. Meanwhile, for predicting ventilator support a highest balanced accuracy of 0.979 and an AUC of 0.981 was achieved. The proposed explainable prediction model will assist doctors in the early prediction of COVID-19 patients that are at risk of mortality or ventilatory support and improve the management of hospital resources.

Rough–Granular Approach in Imbalanced Bankruptcy Data Analysis

During the global economic crisis data were collected on the financial condition of Polish enterprises. Correct analysis of this type of data gives a chance to draw conclusions and avoid the problem of bankruptcy in the future. Unfortunately, the collected data have imbalanced characteristics. Hence, they require a dedicated processing strategy. In this paper, we propose to use an optimized version of our rough–granular approach (RGA) to improve the classification efficiency of bankruptcy data. The RGA oversampling algorithm generates new minority class samples in specific areas of the feature space taking into consideration additional difficulty factors. Moreover, the method in the filtering step provides removal of generated ambiguities. This approach has been successful with bankruptcy data. The study showed in most cases an improvement in the detection rate of the minority class compared to the classification without preprocessing and the SMOTE technique. The RGA algorithm was confirmed to be effective in terms of AUC and recall measures on imbalanced data with complex distribution. The accuracy metric proved to be unreliable for data with these characteristics. XGBoost was used as classifier. In our experiments, we focused on an in-depth analysis of the types of information granules present in the dataset and their impact on classification performance.

SIZE OF KNEE OSTEOPHYTES is under‐estimated as Assessed by MRI USING CT AS A REFERENCE STANDARD

The radiographic diagnosis of knee osteoarthritis (OA) is based on the presence of a definite osteophyte (OP). Epidemiologic studies and clinical trials commonly employ MRI using standard clinical sequences and OP assessment is one of the features included. Commonly used sequence protocols in such studies are based on triplanar fat suppressed T2- or proton density-weighted sequences with bony features as well as fibrous and lipomatous tissues appearing hypointense. This may potentially result in underestimation of such features including osteophyte assessment. Computed tomography (CT) is considered as the modality of choice for the analysis of OPs and may therefore be considered a reference standard. To explore the diagnostic performance of routine MRI as used in observational studies and clinical trials of knee OA for the assessment of OPs in all three knee compartments and the femoral notch using CT as a reference standard. The Strontium Ranelate Efficacy in Knee Osteoarthritis Trial (SEKOIA) clinical trial explored the effect of 3 years of treatment with strontium ranelate in patients with primary knee OA. All patients included in the current study participated in the CT sub-study of SEKOIA. In the current investigation, baseline data from three different centers were used. OPs were scored using a modified MOAKS scoring system in the patellofemoral (PFJ), the medial tibio-femoral (TFJ) and the lateral TFJ. Size was assessed from 0 to 3 in 18 locations. Descriptive statistics were used to describe differences in ordinal grading between CT and MRI. In addition, weighted-kappa statistics were employed to assess agreement between scoring using the two methods. Finally, sensitivity, specificity, positive predictive value and negative predictive value as well as AUC measures of diagnostic performance were employed using CT as the reference standard. Mean age of the randomized patients was 62.9 ±7.5 years with a majority of women (70%). Body mass index was 29.9±5.0 kg/m 2 . The target knee JSW was 3.50±0.84 mm at baseline, patients were Kellgren and Lawrence grade 2 (62%) or 3 (38%). Included were 74 patients with available MRI and CT data. Altogether 1332 locations were evaluated. For the PFJ MRI detected 141 (72%) of 197 CT-defined OPs with a w-kappa of 0.58 (95% CI [0.52-0.65]). In the medial TFJ, MRI detected 178 (81%) of 219 CT-OPs with a w-kappa of 0.58 (95% CI [0.51-0.64]). For the lateral compartment these numbers were 84 (70%) of 120 CT-OPs with a w-kappa of 0.58 (95% CI [0.50-0.66]). Regarding disagreement, for the PFJ MRI underestimated OP size in 94 cases (21%) and over-estimated in 27 cases (6%). For the medial TFJ these numbers were 90 (20%) and 29 (7%), and for the lateral TFJ 67 (15%) and 15 (3%). Details of the diagnostic performance analysis are presented in Table 1 . MRI underestimates presence and size of osteophytes in all 3 knee compartments. However, the sensitivity of >0.9 for OP detection in the medial compartment, which is commonly the compartment of relevance in clinical DMOAD trials, reflects that MRI is adequately able to detect osteophytes that commonly define structural disease. CT may be helpful particularly reading assessment of small osteophytes particularly in early disease. No funding was received for this sub-analysis of the SEKOIA Study AG has received consultancies, speaking fees, and/or honoraria from Pfizer, Novartis, AstraZeneca, Merck Serono, and TissueGene and is President and shareholder of Boston Imaging Core Lab (BICL), LLC a company providing image assessment services. FWR is Chief Medical Officer and shareholder of BICL, LLC. and has received consultancies, speaking fees, and/or honoraria from Calibr –California Institute of Biomedical Research and Grünenthal, GmbH CORRESPONDENCE ADDRESS: frank.roemer@uk-erlangen.de .

Sarcopenia Diagnosis: Reliability of the Ultrasound Assessment of the Tibialis Anterior Muscle as an Alternative Evaluation Tool.

Sarcopenia is a skeletal muscle disorder characterized by reduced muscle mass, strength, and performance. Muscle ultrasound can be helpful in assessing muscle mass, quality, and architecture, and thus possibly useful for diagnosing or screening sarcopenia. The objective of this study was to evaluate the reliability of ultrasound assessment of tibialis anterior muscle in sarcopenia diagnosis. We included subjects undergoing total or partial hip replacement, comparing measures with a healthy control group. We measured the following parameters: tibialis anterior muscle thickness, echogenicity, architecture, stiffness, skeletal muscle index (SMI), hand grip strength, and sarcopenia related quality of life evaluated through the SarQoL questionnaire. We included 33 participants with a mean age of 54.97 ± 23.91 years. In the study group we found reduced tibialis anterior muscle thickness compared to the healthy control group (19.49 ± 4.92 vs. 28.94 ± 3.63 mm, p < 0.05) with significant correlation with SarQoL values (r = 0.80, p < 0.05), dynamometer hand strength (r = 0.72, p < 0.05) and SMI (r = 0.76, p < 0.05). Moreover, we found reduced stiffness (32.21 ± 12.31 vs. 27.07 ± 8.04 Kpa, p < 0.05). AUC measures of ROC curves were 0.89 predicting reduced muscle strength, and 0.97 predicting reduced SMI for tibialis anterior muscle thickness, while they were 0.73 and 0.85, respectively, for muscle stiffness. Our findings showed that ultrasound assessment of tibialis anterior muscle might be considered a reliable measurement tool to evaluate sarcopenia.

B-PO02-079 PERFORMANCE OF ATRIAL FIBRILLATION BURDEN PATTERNS DETECTED VIA INSERTABLE CARDIAC MONITOR FOR STROKE RISK STRATIFICATION

Atrial fibrillation (AF) is associated with increased stroke risk, yet limitations in conventional monitoring restrict our understanding of AF burden risk thresholds. Predictive algorithms incorporating continuous monitoring data may be useful for stroke prediction. To evaluate the relative importance of temporal AF burden patterns as a predictor of stroke from a large retrospective cohort with insertable cardiac monitors (ICMs). Using the Optum© claims de-identified database (01/2007-6/2019) linked with the Medtronic CareLink ICM database, we identified patients indicated for AF management (AFM; N = 1,209), suspected AF (SAF; N = 1,629), and cryptogenic stroke (CS; N = 2,263) with at least 12 months continuous enrollment pre- and post-ICM implantation. AF burden, defined as daily (24-hr) AF/AT via the ICM, was transformed into a set of simple moving averages (SMAs) of different periods for each follow-up. Temporal pattern of AF burden was defined as the comparison of unique SMA pairs. The same definition was applied to the other diagnostic parameters. The occurrence of ischemic stroke was treated as a binary classification problem using baseline characteristics (including anticoagulant use), CHA2DS2-VASc score, and diagnostic parameters & patterns as predictors. Variable importance was calculated as the goodness-of-fit for each predictor in a decision tree model, and its bootstrapped mean was used to quantify significance of AF burden. A total of 5,101 patients (age 69.2±11.7 yrs, 50% male, CHA2DS2-VASc 3.9±1.9) met the inclusion criteria. Overall, 879 had an ischemic stroke over 2,445,925 days follow-up (151±166 days with a stroke outcome and 548±200 days without a stroke outcome). Other than prior stroke, AF burden patterns were the most important predictors of stroke in follow-up. The optimal moving average period and AF burden thresholds differed by device indication (AFM & SAF <8 days, CS 8-21 days). A combination of baseline characteristics and diagnostic patterns performed optimally on measures of AUC (0.69), specificity (0.60), and reliability. In this real-world dataset of patients with an ICM, AF burden patterns may provide incremental prognostic value to conventional risk stratification.

Interactive Decision Tree Learning and Decision Rule Extraction Based on the ImbTreeEntropy and ImbTreeAUC Packages

This paper presents two new R packages ImbTreeEntropy and ImbTreeAUC for building decision trees, including their interactive construction and analysis, which is a highly regarded feature for field experts who want to be involved in the learning process. ImbTreeEntropy functionality includes the application of generalized entropy functions, such as Renyi, Tsallis, Sharma-Mittal, Sharma-Taneja and Kapur, to measure the impurity of a node. ImbTreeAUC provides non-standard measures to choose an optimal split point for an attribute (as well the optimal attribute for splitting) by employing local, semi-global and global AUC measures. The contribution of both packages is that thanks to interactive learning, the user is able to construct a new tree from scratch or, if required, the learning phase enables making a decision regarding the optimal split in ambiguous situations, taking into account each attribute and its cut-off. The main difference with existing solutions is that our packages provide mechanisms that allow for analyzing the trees’ structures (several trees simultaneously) that are built after growing and/or pruning. Both packages support cost-sensitive learning by defining a misclassification cost matrix, as well as weight-sensitive learning. Additionally, the tree structure of the model can be represented as a rule-based model, along with the various quality measures, such as support, confidence, lift, conviction, addedValue, cosine, Jaccard and Laplace.

Software fault prediction based on the dynamic selection of learning technique: findings from the eclipse project study

An effective software fault prediction (SFP) model could help developers in the quick and prompt detection of faults and thus help enhance the overall reliability and quality of the software project. Variations in the prediction performance of learning techniques for different software systems make it difficult to select a suitable learning technique for fault prediction modeling. The evaluation of previously presented SFP approaches has shown that single machine learning-based models failed to provide the best accuracy in any context, highlighting the need to use multiple techniques to build the SFP model. To solve this problem, we present and discuss a software fault prediction approach based on selecting the most appropriate learning techniques from a set of competitive and accurate learning techniques for building a fault prediction model. In work, we apply the discussed SFP approach for the five Eclipse project datasets and nine Object-oriented (OO) project datasets and report the findings of the experimental study. We have used different performance measures, i.e., AUC, accuracy, sensitivity, and specificity, to assess the discussed approach’s performance. Further, we have performed a cost-benefit analysis to evaluate the economic viability of the approach. Results showed that the presented approach predicted the software’s faults effectively for the used accuracy, AUC, sensitivity, and specificity measures with the highest achieved values of 0.816, 0.835, 0.98, and 0.903 for AUC, accuracy, sensitivity, and specificity, respectively. The cost-benefit analysis of the approach showed that it could help reduce the overall software testing cost.

Learning to recognize actionable static code warnings (is intrinsically easy)

Static code warning tools often generate warnings that programmers ignore. Such tools can be made more useful via data mining algorithms that select the “actionable” warnings; i.e. the warnings that are usually not ignored. In this paper, we look for actionable warnings within a sample of 5,675 actionable warnings seen in 31,058 static code warnings from FindBugs. We find that data mining algorithms can find actionable warnings with remarkable ease. Specifically, a range of data mining methods (deep learners, random forests, decision tree learners, and support vector machines) all achieved very good results (recalls and AUC(TRN, TPR) measures usually over 95% and false alarms usually under 5%). Given that all these learners succeeded so easily, it is appropriate to ask if there is something about this task that is inherently easy. We report that while our data sets have up to 58 raw features, those features can be approximated by less than two underlying dimensions. For such intrinsically simple data, many different kinds of learners can generate useful models with similar performance. Based on the above, we conclude that learning to recognize actionable static code warnings is easy, using a wide range of learning algorithms, since the underlying data is intrinsically simple. If we had to pick one particular learner for this task, we would suggest linear SVMs (since, at least in our sample, that learner ran relatively quickly and achieved the best median performance) and we would not recommend deep learning (since this data is intrinsically very simple).