Supervised Machine Learning Methods Research Articles

1614-P: Targeting People with Unknown Diabetes in Health Insurance Databases through Artificial Intelligence

Background: The studies on unknown diabetes are usually based on small samples of population. Information from the entire population can be accessible through health insurance databases like the French National Health Data System (SNDS) comprising data on reimbursement of out-of-hospital dispensed health care and from public and private hospitals from the 66 million people living in France. The objective of our study was to develop an algorithm to identify unknown diabetes cases in the SNDS using artificial intelligence (AI). Methods: Data from the Constances cohort were used to develop the algorithm since individual information recorded in self-administered questionnaire, medical examination and biological tests is linked with data from the SNDS. We applied a supervised machine learning method, involving eight steps. First, the reference database was selected (n=44,185 participants), after excluding known diabetes cases. Then, the unknown diabetes cases - fasting blood glucose ≥7 mmol/L- were identified as target positive (n = 655). The following steps were: codification of the SNDS variables, split the referenced database into training and testing databases, selection of variables and training, validation and selection of algorithms. Results: Among the 3471 variables coded, 12 variables were selected based on their ability to discriminate the target: unknown diabetes cases versus no diabetes. The final algorithm is a logistic regression model based on the 5 most discriminating variables: age, sex and number of out-of-hospital reimbursements in the previous year for blood lipid profile tests, general practitioner consultations and blood glucose tests. The specificity, sensitivity and precision of the algorithm were 70%, 71% and 69%, respectively. Conclusion: AI opens many perspectives in terms of diabetes prevention. Thus, unknown diabetes cases could be ascertained for development and evaluation of prevention policies at a national, regional or local level. Disclosure S. Fuentes: None. R. Hrzic: None. S. Kab: Employee; Spouse/Partner; Sanofi. R. Haneef: None. S. Fosse-Edorh: None. E. Cosson: None.

Machine learning prediction of survival in diffuse large B-cell lymphoma based on gene-expression profiling.

8047 Background: The current clinical risk stratification of Diffuse Large B-cell Lymphoma (DLBCL) relies on the International Prognostic Index (IPI) comprising a limited number of clinical variables but is imperfect in the identification of high-risk disease. Our study aimed to: (1) develop a risk prediction model based on the genetic and clinical features; and (2) evaluate the model’s biological implications in association with the estimated profiles of immune infiltration. Methods: Gene-expression profiling was performed on 718 patients with DLBCL for which RNA sequencing data and clinical covariates were available by Reddy et al (2017). Unsupervised and supervised machine learning methods were used to discover and identify the best set of survival-associated gene signatures for prediction. A multivariate model of survival from these signatures was constructed in the training set and validated in an independent test set. The compositions of the tumor-infiltrating immune cells were enumerated using CIBERSORT for deconvolution analysis. Results: A four gene-signature-based score was developed that separated patients into high- and low-risk groups with a significant difference in survival in the training, validation and complete cohorts (p < 0.001), independently of the IPI. The combination of the gene-expression-based score with the IPI improved the discrimination on the validation and complete sets. The area-under-the-curve at 2 and 5 years increased from 0.71 and 0.69 to 0.75 and 0.74 in the validation set, respectively. Conclusions: By analyzing the gene-expression data with a systematic approach, we developed and validated a risk prediction model that outperforms existing risk assessment methods. Our study, which integrated the profiles of immune infiltration with prognostic prediction, unraveled important associations that have the potential to identify patients who could benefit from the various therapeutic interventions, as well as highlighting possible targets for new drugs.

Applying Machine Learning to Kinematic and Eye Movement Features of a Movement Imitation Task to Predict Autism Diagnosis

Autism is a developmental condition currently identified by experts using observation, interview, and questionnaire techniques and primarily assessing social and communication deficits. Motor function and movement imitation are also altered in autism and can be measured more objectively. In this study, motion and eye tracking data from a movement imitation task were combined with supervised machine learning methods to classify 22 autistic and 22 non-autistic adults. The focus was on a reliable machine learning application. We have used nested validation to develop models and further tested the models with an independent data sample. Feature selection was aimed at selection stability to assure result interpretability. Our models predicted diagnosis with 73% accuracy from kinematic features, 70% accuracy from eye movement features and 78% accuracy from combined features. We further explored features which were most important for predictions to better understand movement imitation differences in autism. Consistent with the behavioural results, most discriminative features were from the experimental condition in which non-autistic individuals tended to successfully imitate unusual movement kinematics while autistic individuals tended to fail. Machine learning results show promise that future work could aid in the diagnosis process by providing quantitative tests to supplement current qualitative ones.

Automatic detection method of cracks from concrete surface imagery using two‐step light gradient boosting machine

AbstractAutomated crack detection based on image processing is widely used when inspecting concrete structures. The existing methods for crack detection are not yet accurate enough due to the difficulty and complexity of the problem; thus, more accurate and practical methods should be developed. This paper proposes an automated crack detection method based on image processing using the light gradient boosting machine (LightGBM), one of the supervised machine learning methods. In supervised machine learning, appropriate features should be identified to obtain accurate results. In crack detection, the pixel values of the target pixels and geometric features of the cracks that occur when they are connected linearly should be considered. This paper proposes a methodology for generating features based on pixel values and geometric shapes in two stages. The accuracy of the proposed methodology is investigated using photos of concrete structures with adverse conditions, such as shadows and dirt. The proposed methodology achieves an accuracy of 99.7%, sensitivity of 75.71%, specificity of 99.9%, precision of 68.2%, and an F‐measure of 0.6952. The experimental results demonstrate that the proposed method can detect cracks with higher performance than the pix2pix‐based approach. Furthermore, the training time is 7.7 times shorter than that of the XGBoost and 2.3 times shorter than that of the pix2pix. The experimental results demonstrate that the proposed method can detect cracks with high accuracy.

Supervised machine learning for coronary artery lumen segmentation in intravascular ultrasound images.

Intravascular ultrasound (IVUS) has been widely used to capture cross sectional lumen frames of inner wall of coronary arteries. This kind of medical imaging modalities is capable of providing detailed and significant information of lumen contour shape, which is very important for clinical diagnosis and analysis of cardiovascular diseases. Numerous learning based techniques have recently become very popular for coronary artery segmentation due to their impressive results. In this work, a supervised machine learning method for coronary artery lumen segmentation with high accuracy and minimal user interaction is designed. The fully discriminative lumen segmentation method jointly learning a classifier the weak learners rely on and the features of the classifier is developed. Additionally, the theoretical supports of the Gradient Boosting framework used in this work and its quadratic approximation are presented. The proposed algorithm is tested on the public datasets of boundary detection of lumen in IVUS challenge held in MICCAI 2011 and achieves a higher average Jaccard similarity of 96.8% and a lower mean error distance of 0.55 (in Cartesian coordinates), which shows higher accuracy compared to the existing learning based methods. Moreover, three real patient IVUS datasets are used to evaluate the performance of the proposed coronary artery lumen segmentation algorithm, which is shown to achieve lower percent error of lumen area of 1.861% ± 0.965%, 1.968% ± 0.864%, and 1.671% ± 0.584%, respectively, compared to the manually measured lumen area (ground truth). The proposed lumen segmentation method is found to be superior to the latest learning based segmentation techniques. Given the efficiency and robustness, our method has great potential in IVUS images processing and coronary artery segmentation and quantification. NOVELTY STATEMENT: The main contributions are summarized in the following aspects: A detailed review of related work about learning based coronary artery lumen segmentation in intravascular ultrasound images is presented. A fully discriminative lumen segmentation method jointly learning a classifier our weak learners rely on and the features of the classifier is developed. The theoretical supports of the Gradient Boosting framework and its quadratic approximation used in this work are presented.

Neural networks for online learning of non-stationary data streams: a review and application for smart grids flexibility improvement

Learning efficient predictive models in dynamic environments requires taking into account the continuous changing nature of phenomena generating the data streams, known in machine learning as “concept drift”. Such changes may affect models’ effectiveness over time, requiring permanent updates of parameters and structure to maintain performance. Several supervised machine learning methods have been developed to be adapted to learn in dynamic and non-stationary environments. One of the most well-known and efficient learning methods is neural networks. This paper focuses on the different neural networks developed to build learning models able to adapt to concept drifts on streaming data. Their performance will be studied and compared using meaningful criteria. Their limits to address the challenges related to the problem of the improvement of electrical grid flexibility in presence of distributed Wind–PV renewable energy resources within the context of energy transition will be highlighted. Finally, the study provides a self-adaptive scheme based on the use of neural networks to overcome these limitations and tackle these challenges.

Development and validation of phenotype classifiers across multiple sites in the observational health data sciences and informatics network.

ObjectiveAccurate electronic phenotyping is essential to support collaborative observational research. Supervised machine learning methods can be used to train phenotype classifiers in a high-throughput manner using imperfectly labeled data. We developed 10 phenotype classifiers using this approach and evaluated performance across multiple sites within the Observational Health Data Sciences and Informatics (OHDSI) network.Materials and MethodsWe constructed classifiers using the Automated PHenotype Routine for Observational Definition, Identification, Training and Evaluation (APHRODITE) R-package, an open-source framework for learning phenotype classifiers using datasets in the Observational Medical Outcomes Partnership Common Data Model. We labeled training data based on the presence of multiple mentions of disease-specific codes. Performance was evaluated on cohorts derived using rule-based definitions and real-world disease prevalence. Classifiers were developed and evaluated across 3 medical centers, including 1 international site.ResultsCompared to the multiple mentions labeling heuristic, classifiers showed a mean recall boost of 0.43 with a mean precision loss of 0.17. Performance decreased slightly when classifiers were shared across medical centers, with mean recall and precision decreasing by 0.08 and 0.01, respectively, at a site within the USA, and by 0.18 and 0.10, respectively, at an international site.Discussion and ConclusionWe demonstrate a high-throughput pipeline for constructing and sharing phenotype classifiers across sites within the OHDSI network using APHRODITE. Classifiers exhibit good portability between sites within the USA, however limited portability internationally, indicating that classifier generalizability may have geographic limitations, and, consequently, sharing the classifier-building recipe, rather than the pretrained classifiers, may be more useful for facilitating collaborative observational research.

Towards Piston Fine Tuning of Segmented Mirrors through Reinforcement Learning

Unlike supervised machine learning methods, reinforcement learning allows an entity to learn how to deploy a task from experience rather than labeled data. This approach has been used in this paper to correct piston misalignment between segments in a segmented mirror telescope. It was proven in simulations that the algorithm converges to a point where it learns how to move the piston actuators in order to maximize the Strehl ratio of the wavefront at the intersection.

Prediction of Signed Protein Kinase Regulatory Circuits.

Complex networks of regulatory relationships between protein kinases comprise a major component of intracellular signaling. Although many kinase-kinase regulatory relationships have been described in detail, these tend to be limited to well-studied kinases whereas the majority of possible relationships remains unexplored. Here, we implement a data-driven, supervised machine learning method to predict human kinase-kinase regulatory relationships and whether they have activating or inhibiting effects. We incorporate high-throughput data, kinase specificity profiles, and structural information to produce our predictions. The results successfully recapitulate previously annotated regulatory relationships and can reconstruct known signaling pathways from the ground up. The full network of predictions is relatively sparse, with the vast majority of relationships assigned low probabilities. However, it nevertheless suggests denser modes of inter-kinase regulation than normally considered in intracellular signaling research. A record of this paper's transparent peer review process is included in the Supplemental Information.

A perspective on machine learning in turbulent flows

The physical complexity and the large number of degrees of freedom that can be resolved today by direct numerical simulations of turbulent flows, and by the most sophisticated experimental techniques, require new strategies to reduce and analyse the data so generated, and to model the turbulent behaviour. We discuss a few concrete examples for which the turbulence data have been analysed by machine learning tools. We also comment on work in neighbouring fields of physics, particularly astrophysical (and astronomical) work, where Big Data has been the paradigm for some time. We discuss unsupervised, semi-supervised and supervised machine learning methods to direct numerical simulations data of homogeneous isotropic turbulence, Rayleigh-Bénard convection, and the minimal flow unit of a turbulent channel flow; for the last case, we discuss in some detail the application of echo state networks, this being one implementation of reservoir computing. The paper also provides a brief perspective on machine learning applications more broadly.

Random forest-based evaluation technique for internal damage in reinforced concrete featuring multiple nondestructive testing results

The evaluation of internal damage in concrete structures is related to not only its rapid repair and reinforcement but also its safe usage, and is therefore essential for ensuring its longevity. This paper proposes a method to evaluate the extent of internal damage due to rebar corrosion using Random Forest, one of the supervised machine learning methods. In supervised machine learning, appropriate inputs should be identified to obtain accurate results. This research uses air permeability coefficient, electrical resistivity, ultrasonic velocity, and compressive strength which is obtained by nondestructive tests as inputs. In order to acquire a large number of data for the training, the rebar corrosion was promoted by electrolytic corrosion, and the data was frequently acquired. Then, the high accuracy of the model was confirmed by cross-validation. In addition, the proposed method was applied to the inspection of actual bridges, and it can detect internal damage that is otherwise invisible on the exterior surface.

Active Learning Approaches for Labeling Text: Review and Assessment of the Performance of Active Learning Approaches

Supervised machine learning methods are increasingly employed in political science. Such models require costly manual labeling of documents. In this paper, we introduce active learning, a framework in which data to be labeled by human coders are not chosen at random but rather targeted in such a way that the required amount of data to train a machine learning model can be minimized. We study the benefits of active learning using text data examples. We perform simulation studies that illustrate conditions where active learning can reduce the cost of labeling text data. We perform these simulations on three corpora that vary in size, document length, and domain. We find that in cases where the document class of interest is not balanced, researchers can label a fraction of the documents one would need using random sampling (or “passive” learning) to achieve equally performing classifiers. We further investigate how varying levels of intercoder reliability affect the active learning procedures and find that even with low reliability, active learning performs more efficiently than does random sampling.

Discriminating stress from rest based on resting-state connectivity of the human brain: A supervised machine learning study.

Acute stress induces large‐scale neural reorganization with relevance to stress‐related psychopathology. Here, we applied a novel supervised machine learning method, combining the strengths of a priori theoretical insights with a data‐driven approach, to identify which connectivity changes are most prominently associated with a state of acute stress and individual differences therein. Resting‐state functional magnetic resonance imaging scans were taken from 334 healthy participants (79 females) before and after a formal stress induction. For each individual scan, mean time‐series were extracted from 46 functional parcels of three major brain networks previously shown to be potentially sensitive to stress effects (default mode network (DMN), salience network (SN), and executive control networks). A data‐driven approach was then used to obtain discriminative spatial linear filters that classified the pre‐ and post‐stress scans. To assess potential relevance for understanding individual differences, probability of classification using the most discriminative filters was linked to individual cortisol stress responses. Our model correctly classified pre‐ versus post‐stress states with highly significant accuracy (above 75%; leave‐one‐out validation relative to chance performance). Discrimination between pre‐ and post‐stress states was mainly based on connectivity changes in regions from the SN and DMN, including the dorsal anterior cingulate cortex, amygdala, posterior cingulate cortex, and precuneus. Interestingly, the probability of classification using these connectivity changes were associated with individual cortisol increases. Our results confirm the involvement of DMN and SN using a data‐driven approach, and specifically single out key regions that might receive additional attention in future studies for their relevance also for individual differences.

Welding penetration monitoring for pulsed GTAW using visual sensor based on AAM and random forests

Abstract On-line welding penetration state monitoring based on passive visual sensing technology in pulsed gas tungsten arc welding (GTAW) has been a hot research topic in industry and academia. Passive vision sensing technology has some disadvantages. It is easy to be affected by some common disturbance during pulsed GTAW, like spattering, the change of arc intensity, the slightly change of focal length and viewing angle of visual sensor, etc. This paper proposed an intelligent penetration monitoring methodology for GTAW by combining pattern-based visual feature extraction method and supervised machine learning method. The proposed methodology mainly includes two parts: (a) visual features extraction for weld pool images based on active appearance model (AAM); (b) the implementation of penetration monitoring based on supervised machine learning method, random forests (RF). The extracted weld pool visual features include width, length and area size of weld pool. Then the extracted features are utilized to train RF model. Further, the trained RF model can implement the evaluation of penetration states and backside weld seam width. The experiments verify that the proposed method is effective with high accuracy and robustness, which also lays the foundation for online welding penetration control.

Tracing theory diffusion: a text mining and citation-based analysis of TAM

PurposeTheory is a kind of condensed human knowledge. This paper is to examine the mechanism of interdisciplinary diffusion of theoretical knowledge by tracing the diffusion of a representative theory, the Technology Acceptance Model (TAM).Design/methodology/approachBased on the full-scale dataset of Web of Science (WoS), the citations of Davis's original work about TAM were analysed and the interdisciplinary diffusion paths of TAM were delineated, a supervised machine learning method was used to extract theory incidents, and a content analysis was used to categorize the patterns of theory evolution.FindingsIt is found that the diffusion of a theory is intertwined with its evolution. In the process, the role that a participating discipline play is related to its knowledge distance from the original disciplines of TAM. With the distance increases, the capacity to support theory development and innovation weakens, while that to assume analytical tools for practical problems increases. During the diffusion, a theory evolves into new extensions in four theoretical construction patterns, elaboration, proliferation, competition and integration.Research limitations/implicationsThe study does not only deepen the understanding of the trajectory of a theory but also enriches the research of knowledge diffusion and innovation.Originality/valueThe study elaborates the relationship between theory diffusion and theory development, reveals the roles of the participating disciplines played in theory diffusion and vice versa, interprets four patterns of theory evolution and uses text mining technique to extract theory incidents, which makes up for the shortcomings of citation analysis and content analysis used in previous studies.

P17 Prediction of response of methotrexate in patients with rheumatoid arthritis using serum lipidomics

Abstract Background For patients with rheumatoid arthritis (RA), introduction of early, effective therapy has consistently been shown to improve long-term outcomes. Low-dose methotrexate (MTX) is commonly prescribed as first-line treatment for RA. However, MTX is not effective for a large minority of patients and there is currently no way to determine ahead of therapy which patients are most likely to benefit. Metabolomics and lipidomics are emerging approaches for studying patient stratification in RA and have the potential to identify disease processes that underpin treatment outcomes. Here we apply state-of-the-art machine learning algorithms to predict MTX treatment response, by testing serum lipid levels measured at two time-points (pre-treatment and following 4 weeks on drug) to predict MTX response by 6 months. Methods This study included patients from the Rheumatoid Arthritis Medication Study (RAMS), a UK multi-centre one-year prospective observational study investigating predictors of response to MTX in patients with RA. Since 2008, patients who are about to start MTX for the first time are asked to provide demographic and clinical data, as well as blood samples to permit DNA, RNA and serum-based biomarker studies. Patients about to commence MTX treatment were followed longitudinally and those categorised as good or non-responders following 6 months on-drug using EULAR response criteria were analysed. Serum lipid levels were measured at pre-treatment and following 4 weeks on drug using ultra-performance liquid chromatography tailored for complex lipid analysis, coupled to mass spectrometry. State-of-the-art supervised machine learning methods were then applied to predict EULAR response at 6 months. Models including lipid levels were compared to models including clinical covariates (including: MTX start dose, steroid use at inclusion, BMI, number of swollen joints, number of tender joints, CRP levels, patients’ assessment of their overall wellbeing, gender, age-at-inclusion, age-at-onset, disease duration, HAQ score and pre-treatment smoking habits). Results Following quality control, 3,366 features (1,060 in negatively-charged mode and 2,306 in positive mode) were available for analysis at pre-treatment and 4 weeks from 100 RA patients categorised as good (GR, n = 50) or poor (NR, n = 50) responders to MTX following 6 months on drug. The best model performance for the classifier including clinical covariates was observed using L1/L2-regularised logistic regression (ROC AUC 0.68 ± 0.02). However, the clinical covariate model outperformed the classifier including lipid levels when either pre- or on-treatment time-points were investigated (ROC AUC 0.61 ± 0.02). Conclusion These data do not support the utility of early treatment lipidomic monitoring in routine clinical practice in patients started on MTX for their RA. Disclosures M. Maciejewski: Shareholder/stock ownership; owns stock or stock options in Pfizer. C. Sands None. N. Nair None. S. Ling None. S. Verstappen None. K. Hyrich None. A. Barton None. D. Ziemek Shareholder/stock ownership; owns stock or stock options in Pfizer. M. Lewis None. D. Plant None.

The Dynamics of Political Incivility on Twitter

Online incivility and harassment in political communication have become an important topic of concern among politicians, journalists, and academics. This study provides a descriptive account of uncivil interactions between citizens and politicians on Twitter. We develop a conceptual framework for understanding the dynamics of incivility at three distinct levels: macro (temporal), meso (contextual), and micro (individual). Using longitudinal data from the Twitter communication mentioning Members of Congress in the United States across a time span of over a year and relying on supervised machine learning methods and topic models, we offer new insights about the prevalence and dynamics of incivility toward legislators. We find that uncivil tweets represent consistently around 18% of all tweets mentioning legislators, but with spikes that correspond to controversial policy debates and political events. Although we find evidence of coordinated attacks, our analysis reveals that the use of uncivil language is common to a large number of users.

Performance improvement for machine learning‐based cooperative spectrum sensing by feature vector selection

To explore the potential of machine learning-based cooperative spectrum sensing (CSS) in training time, classification speed and classification performance, this study mainly focuses on studying the problem of the feature vectors selecting for machine learning-based CSS. First, a new machine learning-based CSS framework is presented, in which, energy vector forming module, feature vector conversion module, training module, classification module and training sample database are included. Second, a new two-dimensional distance vector is developed, and it is converted by an m-dimensional energy vector according to the distance measurement between vectors. Furthermore, six combination modes are obtained by combining three feature vectors (energy, probability and distance vectors) with two supervised machine learning methods, which are support vector machine (SVM) and weighted K-nearest-neighbour, respectively. From the proposed experimental simulations, the authors can find that the distance vector is obviously superior to the probability vector in computation time. Moreover, the probability vector and distance vector are superior to the energy vector in training time except for the case of poor signal and fewer users, and obviously superior to the energy vector in classification speed. At last, the probability vector and distance vector with SVM classifier show the best classification performance in six combination modes.

Exploring anomalous behaviour detection and classification for insider threat identification

SummaryRecently, malicious insider threats represent one of the most damaging threats to companies and government agencies. Insider threat detection is a highly skewed data analysis problem, where the huge class imbalance makes the adaptation of learning algorithms to the real‐world context very difficult. This study proposes a new system for user‐centred machine learning‐based anomaly behaviour and insider threat detection on multiple data granularity levels. System evaluations and analysis are performed not only on individual data instances but also on normal and malicious users. Our results show that the proposed system, which is a combination of unsupervised anomaly detection and supervised machine learning methods, can learn from unlabelled data and a very small amount of labelled data. Furthermore, it can generalize to bigger datasets for detecting anomalous behaviours and unseen malicious insiders with a high detection and a low false‐positive rate.

Development of a periodontitis risk assessment model for primary care providers in an interdisciplinary setting.

Periodontitis (PD), a form of gum disease, is a major public health concern as it is globally prevalent and harms both individual quality of life and economic productivity. Global cost in lost productivity is estimated at US$54 billion annually. Moreover, current PD assessment applies only after the damage has already occurred. This study proposes and tests a new PD risk assessment model applicable at point-of-care, using supervised machine learning methods. We compare the performance of five algorithms using retrospective clinical data: Naïve Bayes (NB), Logistic Regression (LR), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Decision Tree (DT). DT and ANN demonstrated higher accuracy in classifying the patients with high or low PD risk as compared to NB, LR and SVM. The resultant model with DT showed a sensitivity of 87.08% (95% CI 84.12% to 89.76%) and specificity of 93.5% (95% CI 91% to 95.49%). A predictive model with high sensitivity and specificity to stratify individuals into low and high PD risk tiers was developed. Validation in other populations will inform translational value of this approach and its potential applicability as clinical decision support tool.