Search

Background Myocardial infarction (MI), stroke and diabetes share underlying risk factors and commonalities in clinical management. We examined if their combined impact on mortality is proportional, amplified or less than the expected risk separately of each disease and whether the excess risk is explained by their associated comorbidities. Methods Using large-scale electronic health records, we identified 2,007,731 eligible patients (51% women) and registered with general practices in the UK and extracted clinical information including diagnosis of myocardial infarction (MI), stroke, diabetes and 53 other long-term conditions before 2005 (study baseline). We used Cox regression to determine the risk of all-cause mortality with age as the underlying time variable and tested for excess risk due to interaction between cardiometabolic conditions. Results At baseline, the mean age was 51 years, and 7% (N = 145,910) have had a cardiometabolic condition. After a 7-year mean follow-up, 146,994 died. The sex-adjusted hazard ratios (HR) (95% confidence interval [CI]) of all-cause mortality by baseline disease status, compared to those without cardiometabolic disease, were MI = 1.51 (1.49–1.52), diabetes = 1.52 (1.51–1.53), stroke = 1.84 (1.82–1.86), MI and diabetes = 2.14 (2.11–2.17), MI and stroke = 2.35 (2.30–2.39), diabetes and stroke = 2.53 (2.50–2.57) and all three = 3.22 (3.15–3.30). Adjusting for other concurrent comorbidities attenuated these estimates, including the risk associated with having all three conditions (HR = 1.81 [95% CI 1.74–1.89]). Excess risks due to interaction between cardiometabolic conditions, particularly when all three conditions were present, were not significantly greater than expected from the individual disease effects. Conclusion Myocardial infarction, stroke and diabetes were associated with excess mortality, without evidence of any amplification of risk in people with all three diseases. The presence of other comorbidities substantially contributed to the excess mortality risks associated with cardiometabolic disease multimorbidity.

Today, despite decades of developments in medicine and the growing interest in precision healthcare, vast majority of diagnoses happen once patients begin to show noticeable signs of illness. Early indication and detection of diseases, however, can provide patients and carers with the chance of early intervention, better disease management, and efficient allocation of healthcare resources. The latest developments in machine learning (including deep learning) provides a great opportunity to address this unmet need. In this study, we introduce BEHRT: A deep neural sequence transduction model for electronic health records (EHR), capable of simultaneously predicting the likelihood of 301 conditions in one’s future visits. When trained and evaluated on the data from nearly 1.6 million individuals, BEHRT shows a striking improvement of 8.0–13.2% (in terms of average precision scores for different tasks), over the existing state-of-the-art deep EHR models. In addition to its scalability and superior accuracy, BEHRT enables personalised interpretation of its predictions; its flexible architecture enables it to incorporate multiple heterogeneous concepts (e.g., diagnosis, medication, measurements, and more) to further improve the accuracy of its predictions; its (pre-)training results in disease and patient representations can be useful for future studies (i.e., transfer learning).

Multicollinearity refers to the presence of collinearity between multiple variables and renders the results of statistical inference erroneous (Type II error). This is particularly important in environmental health research where multicollinearity can hinder inference. To address this, correlated variables are often excluded from the analysis, limiting the discovery of new associations. An alternative approach to address this problem is the use of principal component analysis. This method, combines and projects a group of correlated variables onto a new orthogonal space. While this resolves the multicollinearity problem, it poses another challenge in relation to interpretability of results. Standard hypothesis testing methods can be used to evaluate the association of projected predictors, called principal components, with the outcomes of interest, however, there is no established way to trace the significance of principal components back to individual variables. To address this problem, we investigated the use of sparse principal component analysis which enforces a parsimonious projection. We hypothesise that this parsimony could facilitate the interpretability of findings. To this end, we investigated the association of 20 environmental predictors with all-cause mortality adjusting for demographic, socioeconomic, physiological, and behavioural factors. The study was conducted in a cohort of 379,690 individuals in the UK. During an average follow-up of 8.05 years (3,055,166 total person-years), 14,996 deaths were observed. We used Cox regression models to estimate the hazard ratio (HR) and 95% confidence intervals (CI). The Cox models were fitted to the standardised environmental predictors a) without any transformation b) transformed with PCA, and c) transformed with SPCA. The comparison of findings underlined the potential of SPCA for conducting inference in scenarios where multicollinearity can increase the risk of Type II error. Our analysis unravelled a significant association between average noise pollution and increased risk of all-cause mortality. Specifically, those in the upper deciles of noise exposure have between 5-10% increased risk of all-cause mortality compared to the lowest decile.

Observational causal inference is useful for decision making in medicine when randomized clinical trials (RCT) are infeasible or non generalizable. However, traditional approaches fail to deliver unconfounded causal conclusions in practice. The rise of "doubly robust" non-parametric tools coupled with the growth of deep learning for capturing rich representations of multimodal data, offers a unique opportunity to develop and test such models for causal inference on comprehensive electronic health records (EHR). In this paper, we investigate causal modelling of an RCT-established null causal association: the effect of antihypertensive use on incident cancer risk. We develop a dataset for our observational study and a Transformer-based model, Targeted BEHRT coupled with doubly robust estimation, we estimate average risk ratio (RR). We compare our model to benchmark statistical and deep learning models for causal inference in multiple experiments on semi-synthetic derivations of our dataset with various types and intensities of confounding. In order to further test the reliability of our approach, we test our model on situations of limited data. We find that our model provides more accurate estimates of RR (least sum absolute error from ground truth) compared to benchmarks for risk ratio estimation on high-dimensional EHR across experiments. Finally, we apply our model to investigate the original case study: antihypertensives' effect on cancer and demonstrate that our model generally captures the validated null association., Comment: This work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible

Background Updatable estimates of COVID-19 onset, progression, and trajectories underpin pandemic mitigation efforts. To identify and characterise disease trajectories, we aimed to define and validate ten COVID-19 phenotypes from nationwide linked electronic health records (EHR) using an extensible framework. Methods In this cohort study, we used eight linked National Health Service (NHS) datasets for people in England alive on Jan 23, 2020. Data on COVID-19 testing, vaccination, primary and secondary care records, and death registrations were collected until Nov 30, 2021. We defined ten COVID-19 phenotypes reflecting clinically relevant stages of disease severity and encompassing five categories: positive SARS-CoV-2 test, primary care diagnosis, hospital admission, ventilation modality (four phenotypes), and death (three phenotypes). We constructed patient trajectories illustrating transition frequency and duration between phenotypes. Analyses were stratified by pandemic waves and vaccination status. Findings Among 57 032 174 individuals included in the cohort, 13 990 423 COVID-19 events were identified in 7 244 925 individuals, equating to an infection rate of 12·7% during the study period. Of 7 244 925 individuals, 460 737 (6·4%) were admitted to hospital and 158 020 (2·2%) died. Of 460 737 individuals who were admitted to hospital, 48 847 (10·6%) were admitted to the intensive care unit (ICU), 69 090 (15·0%) received non-invasive ventilation, and 25 928 (5·6%) received invasive ventilation. Among 384 135 patients who were admitted to hospital but did not require ventilation, mortality was higher in wave 1 (23 485 [30·4%] of 77 202 patients) than wave 2 (44 220 [23·1%] of 191 528 patients), but remained unchanged for patients admitted to the ICU. Mortality was highest among patients who received ventilatory support outside of the ICU in wave 1 (2569 [50·7%] of 5063 patients). 15 486 (9·8%) of 158 020 COVID-19-related deaths occurred within 28 days of the first COVID-19 event without a COVID-19 diagnoses on the death certificate. 10 884 (6·9%) of 158 020 deaths were identified exclusively from mortality data with no previous COVID-19 phenotype recorded. We observed longer patient trajectories in wave 2 than wave 1. Interpretation Our analyses illustrate the wide spectrum of disease trajectories as shown by differences in incidence, survival, and clinical pathways. We have provided a modular analytical framework that can be used to monitor the impact of the pandemic and generate evidence of clinical and policy relevance using multiple EHR sources. Funding British Heart Foundation Data Science Centre, led by Health Data Research UK.

One major impediment to the wider use of deep learning for clinical decision making is the difficulty of assigning a level of confidence to model predictions. Currently, deep Bayesian neural networks and sparse Gaussian processes are the main two scalable uncertainty estimation methods. However, deep Bayesian neural networks suffer from lack of expressiveness, and more expressive models such as deep kernel learning, which is an extension of sparse Gaussian process, captures only the uncertainty from the higher-level latent space. Therefore, the deep learning model under it lacks interpretability and ignores uncertainty from the raw data. In this paper, we merge features of the deep Bayesian learning framework with deep kernel learning to leverage the strengths of both methods for a more comprehensive uncertainty estimation. Through a series of experiments on predicting the first incidence of heart failure, diabetes and depression applied to large-scale electronic medical records, we demonstrate that our method is better at capturing uncertainty than both Gaussian processes and deep Bayesian neural networks in terms of indicating data insufficiency and identifying misclassifications, with a comparable generalization performance. Furthermore, by assessing the accuracy and area under the receiver operating characteristic curve over the predictive probability, we show that our method is less susceptible to making overconfident predictions, especially for the minority class in imbalanced datasets. Finally, we demonstrate how uncertainty information derived by the model can inform risk factor analysis towards model interpretability.

One major impediment to the wider use of deep learning for clinical decision making is the difficulty of assigning a level of confidence to model predictions. Currently, deep Bayesian neural networks and sparse Gaussian processes are the main two scalable uncertainty estimation methods. However, deep Bayesian neural network suffers from lack of expressiveness, and more expressive models such as deep kernel learning, which is an extension of sparse Gaussian process, captures only the uncertainty from the higher level latent space. Therefore, the deep learning model under it lacks interpretability and ignores uncertainty from the raw data. In this paper, we merge features of the deep Bayesian learning framework with deep kernel learning to leverage the strengths of both methods for more comprehensive uncertainty estimation. Through a series of experiments on predicting the first incidence of heart failure, diabetes and depression applied to large-scale electronic medical records, we demonstrate that our method is better at capturing uncertainty than both Gaussian processes and deep Bayesian neural networks in terms of indicating data insufficiency and distinguishing true positive and false positive predictions, with a comparable generalisation performance. Furthermore, by assessing the accuracy and area under the receiver operating characteristic curve over the predictive probability, we show that our method is less susceptible to making overconfident predictions, especially for the minority class in imbalanced datasets. Finally, we demonstrate how uncertainty information derived by the model can inform risk factor analysis towards model interpretability., 21 pages

Over a billion people live in slums, with poor sanitation, education, property rights and working conditions having direct impact on current residents and future generations. A key problem in relation to slums is slum mapping. Without delineations of where all slum settlements are, informed decisions cannot be made by policymakers in order to benefit the most in need. Satellite images have been used in combination with machine learning models to try and fill the gap in data availability of slum locations. Deep learning has been used on RGB images with some success but since labeled satellite images of slums are relatively low quality and the physical/visual manifestation of slums significantly varies within and across countries, it is important to quantify the uncertainty of predictions for reliable application in downstream tasks. Our solution is to train Monte Carlo dropout U-Net models on multispectral 13-band Sentinel-2 images from which we can calculate pixelwise epistemic (model) and aleatoric (data) uncertainty in our predictions. We trained our model on labelled images of Mumbai and verified our epistemic and aleatoric uncertainty quantification approach using altered models trained on modified datasets. We also used SHAP values to investigate how the different features contribute towards the model’s predictions and this showed that certain short-wave infrared and red-edge image bands are powerful features for determining the locations of slums within images. Having created our model with uncertainty quantification, in the future it can be applied to downstream tasks and decision-makers will know where predictions have been made with low uncertainty, giving them greater confidence in its deployment.

Electronic health records (EHR) represent a holistic overview of patients’ trajectories. Their increasing availability has fueled new hopes to leverage them and develop accurate risk prediction models for a wide range of diseases. Given the complex interrelationships of medical records and patient outcomes, deep learning models have shown clear merits in achieving this goal. However, a key limitation of current study remains their capacity in processing long sequences, and long sequence modelling and its application in the context of healthcare and EHR remains unexplored. Capturing the whole history of medical encounters is expected to lead to more accurate predictions, but the inclusion of records collected for decades and from multiple resources can inevitably exceed the receptive field of the most existing deep learning architectures. This can result in missing crucial, long-term dependencies. To address this gap, we present Hi-BEHRT, a hierarchical Transformer-based model that can significantly expand the receptive field of Transformers and extract associations from much longer sequences. Using a multimodal large-scale linked longitudinal EHR, the Hi-BEHRT exceeds the state-of-the-art deep learning models 1% to 5% for area under the receiver operating characteristic (AUROC) curve and 1% to 8% for area under the precision recall (AUPRC) curve on average, and 2% to 8% (AUROC) and 2% to 11% (AUPRC) for patients with long medical history for 5-year heart failure, diabetes, chronic kidney disease, and stroke risk prediction. Additionally, because pretraining for hierarchical Transformer is not well-established, we provide an effective end-to-end contrastive pre-training strategy for Hi-BEHRT using EHR, improving its transferability on predicting clinical events with relatively small training dataset.

The segmentation of old documents is a crucial phase for reading and understanding the content of a document automatically. Also, the extraction of words and phrases in a document needs segmentation of each line and word. But, the variations of text lines directions throughout the same document and overlapping characters between two or more text lines, especially in Arabic manuscripts, are the problems that usually found in such documents. For that, this chapter proposes an approach for text segmentation as well as line and word for historical Arabic manuscripts. First, text segmentation is realized using an encoder-decoder deep model to segment the main text and side text in the image. The model has been trained on two Arabic manuscripts dataset including Bukhari and RASM2018 datasets. Then, the segmentation of lines using a smoothing approach followed by thresholding determined automatically according to the size of handwriting. Then, segmentation of words is provided using smoothed Chamfer distance which takes into consideration the handwriting characteristics. The evaluation of the proposed approach is reported on the QUWI Arabic database and very promising results are achieved.

Background Forecasting incident heart failure is a critical demand for prevention. Recent research suggested the superior performance of deep learning models on the prediction tasks using electronic health records. However, even with a relatively accurate predictive performance, the major impediments to the wider use of deep learning models for clinical decision making are the difficulties of assigning a level of confidence to model predictions and the interpretability of predictions. Purpose We aimed to develop a deep learning framework for more accurate incident heart failure prediction, with provision of measures of uncertainty and interpretability. Methods We used a longitudinal linked electronic health records dataset, Clinical Practice Research Datalink, involving 788,880 patients, 8.3% of whom had an incident heart failure diagnosis. To embed the uncertainty estimation mechanism into the deep learning models, we developed a probabilistic framework based on a novel transformer deep learning model: deep Bayesian Gaussian processes (DBGP). We investigated the performance of incident heart failure prediction and uncertainty estimation for the model and validated it using an external held-out dataset. Diagnoses, medications, and age for each encounter were included as predictors. By comparing the uncertainty, we investigated the possibility of identifying the correct predictions from wrong ones to avoid potential misclassification. Using model distillation meant to mimic a well-trained complex model with simple models, we investigated the importance of associations between diagnoses, medications and heart failure with an interpretable linear regression component learned from DBGP. Results The DBGP achieved high precision with 0.941 as AUROC for external validation. More importantly, it showed the uncertainty information could distinguish the correct predictions from wrong ones, with significant difference (p-value with 500 samples) between distribution of uncertainties for negative predictions (3.21e-69 between true negative and false negative), and positive predictions (3.39e-22 between true positive and false positive). Utilising the distilled model, we can specify the contribution of each diagnosis and medication to heart failure prediction. For instance, Losartan/Fosinopril, Bisoprolol and Left bundle-branch block showed strong association to heart failure incidence with coefficient 0.11 (95% CI: 0.10, 0.12), 0.09 (0.08, 0.11) and 0.09 (0.07, 0.11) respectively; Peritoneal adhesions, Trochanteric bursitis and Galactorrhea showed strong disassociations with coefficient −0.07 (−0.09, −0.05), −0.07 (−0.09, −0.04) and −0.06 (−0.08, −0.04) individually. Conclusions Our novel probabilistic deep learning framework adds a measure of uncertainty the prediction and helps to mitigate misclassification. Model distillation provides an opportunity to interpret deep learning models and offers a data-driven perspective for risk factor analysis. Funding Acknowledgement Type of funding source: Public Institution(s). Main funding source(s): Oxford Martin School,University of Oxford; NIHR Oxford Biomedical Research Centre, University of Oxford

Background/Introduction Predicting incident heart failure has been challenging. Deep learning models when applied to rich electronic health records (EHR) offer some theoretical advantages. However, empirical evidence for their superior performance is limited and they remain commonly uninterpretable, hampering their wider use in medical practice. Purpose We developed a deep learning framework for more accurate and yet interpretable prediction of incident heart failure. Methods We used longitudinally linked EHR from practices across England, involving 100,071 patients, 13% of whom had been diagnosed with incident heart failure during follow-up. We investigated the predictive performance of a novel transformer deep learning model, “Transformer for Heart Failure” (BEHRT-HF), and validated it using both an external held-out dataset and an internal five-fold cross-validation mechanism using area under receiver operating characteristic (AUROC) and area under the precision recall curve (AUPRC). Predictor groups included all outpatient and inpatient diagnoses within their temporal context, medications, age, and calendar year for each encounter. By treating diagnoses as anchors, we alternatively removed different modalities (ablation study) to understand the importance of individual modalities to the performance of incident heart failure prediction. Using perturbation-based techniques, we investigated the importance of associations between selected predictors and heart failure to improve model interpretability. Results BEHRT-HF achieved high accuracy with AUROC 0.932 and AUPRC 0.695 for external validation, and AUROC 0.933 (95% CI: 0.928, 0.938) and AUPRC 0.700 (95% CI: 0.682, 0.718) for internal validation. Compared to the state-of-the-art recurrent deep learning model, RETAIN-EX, BEHRT-HF outperformed it by 0.079 and 0.030 in terms of AUPRC and AUROC. Ablation study showed that medications were strong predictors, and calendar year was more important than age. Utilising perturbation, we identified and ranked the intensity of associations between diagnoses and heart failure. For instance, the method showed that established risk factors including myocardial infarction, atrial fibrillation and flutter, and hypertension all strongly associated with the heart failure prediction. Additionally, when population was stratified into different age groups, incident occurrence of a given disease had generally a higher contribution to heart failure prediction in younger ages than when diagnosed later in life. Conclusions Our state-of-the-art deep learning framework outperforms the predictive performance of existing models whilst enabling a data-driven way of exploring the relative contribution of a range of risk factors in the context of other temporal information. Funding Acknowledgement Type of funding source: Private grant(s) and/or Sponsorship. Main funding source(s): National Institute for Health Research, Oxford Martin School, Oxford Biomedical Research Centre

Highlights • Machine learning is making significant contributions towards our understanding of the complex relationships between diseases. • Advanced models take a range of modalities from big datasets with little pre-processing or information loss at study design. • Developments in matrix factorisation and deep learning allow a better understanding of evolving patterns of multimorbidity., The prevalence of multimorbidity has been increasing in recent years, posing a major burden for health care delivery and service. Understanding its determinants and impact is proving to be a challenge yet it offers new opportunities for research to go beyond the study of diseases in isolation. In this paper, we review how the field of machine learning provides many tools for addressing research challenges in multimorbidity. We highlight recent advances in promising methods such as matrix factorisation, deep learning, and topological data analysis and how these can take multimorbidity research beyond cross-sectional, expert-driven or confirmatory approaches to gain a better understanding of evolving patterns of multimorbidity. We discuss the challenges and opportunities of machine learning to identify likely causal links between previously poorly understood disease associations while giving an estimate of the uncertainty on such associations. We finally summarise some of the challenges for wider clinical adoption of machine learning research tools and propose some solutions.

Background Aortic valve stenosis is commonly considered a degenerative disorder with no recommended preventive intervention, with only valve replacement surgery or catheter intervention as treatment options. Purpose We sought to assess the causal association between exposure to lipid levels and risk of aortic stenosis. Methods Causality of association was assessed using two-sample Mendelian Randomization (MR) framework through different statistical methods. MR approach uses instrumental variable analysis to mimic the randomization process that underpins causal inference in clinical trials. It takes advantage of the naturally-occurring random allocation of alleles inherited by offspring from their parents during the formation of the zygote. We retrieved summary estimations of 157 genetic variants that have been shown to be associated with plasma lipid levels in the Global Lipids Genetics Consortium that included 188,577 participants, mostly European ancestry, and genetic association with aortic stenosis as the main outcome from a total of 432,173 participants in the UK Biobank. Secondary negative control outcomes included aortic regurgitation and mitral regurgitation. Results The odds ratio (OR) for developing aortic stenosis per unit increase in lipid parameter was 1.52 (95% confidence interval [CI], 1.22 to 1.90; per 0.98 mmol/L) for low-density lipoprotein (LDL) cholesterol, 1.03 (95% CI, 0.80 to 1.31; per 0.41 mmol/L) for high-density lipoprotein (HDL) cholesterol, and 1.38 (95% CI 0.92 to 2.07; per 1 mmol/L) for triglycerides. There was no evidence of a causal association between any of the lipid parameters and aortic or mitral regurgitation. Conclusion Lifelong exposure to high LDL-cholesterol increases the risk of symptomatic aortic stenosis, suggesting that LDL-lowering treatment may be effective in its prevention. Funding Acknowledgement Type of funding source: Foundation. Main funding source(s): British Heart Foundation

This paper focuses on detecting inconsistencies within text corpora. It is a very interesting area with many applications. Most existing methods deal with this problem using complicated textual analysis, which is known for not being accurate enough. We propose a new methodology that consists of two steps, the first one being a machine learning step that performs multilevel text categorization. The second one applies conceptual reasoning on the predicted categories in order to detect inconsistencies. This paper has been validated on a set of Islamic advisory opinions (also known as fatwas). This domain is gaining a large interest with users continuously checking the authenticity and relevance of such content. The results show that our method is very accurate and can complement existing methods using the linguistic analysis.

Background How measures of long‐term exposure to elevated blood pressure might add to the performance of “current” blood pressure in predicting future cardiovascular disease is unclear. We compared incident cardiovascular disease risk prediction using past, current, and usual systolic blood pressure alone or in combination. Methods and Results Using data from UK primary care linked electronic health records, we applied a landmark cohort study design and identified 80 964 people, aged 50 years (derivation cohort=64 772; validation cohort=16 192), who, at study entry, had recorded blood pressure, no prior cardiovascular disease, and no previous antihypertensive or lipid‐lowering prescriptions. We used systolic blood pressure recorded up to 10 years before baseline to estimate past systolic blood pressure (mean, time‐weighted mean, and variability) and usual systolic blood pressure (correcting current values for past time‐dependent blood pressure fluctuations) and examined their prospective relation with incident cardiovascular disease (first hospitalization for or death from coronary heart disease or stroke/transient ischemic attack). We used Cox regression to estimate hazard ratios and applied Bayesian analysis within a machine learning framework in model development and validation. Predictive performance of models was assessed using discrimination (area under the receiver operating characteristic curve) and calibration metrics. We found that elevated past, current, and usual systolic blood pressure values were separately and independently associated with increased incident cardiovascular disease risk. When used alone, the hazard ratio (95% credible interval) per 20–mm Hg increase in current systolic blood pressure was 1.22 (1.18–1.30), but associations were stronger for past systolic blood pressure (mean and time‐weighted mean) and usual systolic blood pressure (hazard ratio ranging from 1.39–1.45). The area under the receiver operating characteristic curve for a model that included current systolic blood pressure, sex, smoking, deprivation, diabetes mellitus, and lipid profile was 0.747 (95% credible interval, 0.722–0.811). The addition of past systolic blood pressure mean, time‐weighted mean, or variability to this model increased the area under the receiver operating characteristic curve (95% credible interval) to 0.750 (0.727–0.811), 0.750 (0.726–0.811), and 0.748 (0.723–0.811), respectively, with all models showing good calibration. Similar small improvements in area under the receiver operating characteristic curve were observed when testing models on the validation cohort, in sex‐stratified analyses, or by using different landmark ages (40 or 60 years). Conclusions Using multiple blood pressure recordings from patients’ electronic health records showed stronger associations with incident cardiovascular disease than a single blood pressure measurement, but their addition to multivariate risk prediction models had negligible effects on model performance., See Editorial Ahmad and Oparil

Multimorbidity, or the presence of several medical conditions in the same individual, has been increasing in the population — both in absolute and relative terms. Nevertheless, multimorbidity remains poorly understood, and the evidence from existing research to describe its burden, determinants and consequences has been limited. Previous studies attempting to understand multimorbidity patterns are often cross-sectional and do not explicitly account for multimorbidity patterns’ evolution over time; some of them are based on small datasets and/or use arbitrary and narrow age ranges; and those that employed advanced models, usually lack appropriate benchmarking and validations. In this study, we (1) introduce a novel approach for using Non-negative Matrix Factorisation (NMF) for temporal phenotyping (i.e., simultaneously mining disease clusters and their trajectories); (2) provide quantitative metrics for the evaluation of these clusters and trajectories; and (3) demonstrate how the temporal characteristics of the disease clusters that result from our model can help mine multimorbidity networks and generate new hypotheses for the emergence of various multimorbidity patterns over time. We trained and evaluated our models on one of the world’s largest electronic health records (EHR) datasets, containing more than 7 million patients, from which over 2 million where relevant to, and hence included in this study.

Text categorization is an important research field that finds many applications nowadays. It is usually performed in two steps: feature extraction and classification. In the feature extraction step, discriminating keywords are extracted in order to distinguish between different categories of documents. In the classification step, the extracted keywords are fed to a classifier in order to detect the category of each document. In this paper, we use the hyper rectangle method which represents the corpus of documents using a binary relation in which the documents correspond to objects and words to attributes. The hyper rectangle method extracts a tree of keywords such that most discriminative keywords are at the top levels and less discriminative keywords are in the deep levels. We are particularly interested to study different proposed weighting metrics that yield different orderings of keywords. We study how these weighting metrics impact the categorization performance. For the classification step we used both a logistic regression and random forests classifiers. We tested our method on both the 20 newsgroups dataset as well as the Reuters R8 dataset. Our method achieves high performance on both datasets which compete very well with state-of-the-art methods. ACKNOWLEDGMENT This contribution was made possible by NPRP grant #06-1220-1-233 from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the authors. Scopus

Most modern search engines feature keyword based search interfaces. These interfaces are usually found on websites belonging to enterprises or governments or sites related to news articles, blogs and social media that contain a large corpus of documents. These collections of documents are not easily indexed by web search engines, and are considered as hidden web databases. These databases provide opportunities for data analysis for many third-parties through their keyword search interfaces. A significant amount of research has already been carried out on analyzing and extracting aggregate information about these hidden document corpora. But most of these research focus on the high level big-picture information of the database. Not enough focus has been done on extracting analytical information which is specific to individual queries. This paper focuses on that analysis gap and takes ideas from other existing research to formulate a query cardinality estimation technique i.e. The count of documents matching a query in the document corpus of a search engine. We experimentally assess the effectiveness of our method by building a search engine on the Reuters-21578 document corpus. For a given keyword the corresponding documents' count is estimated only by sending search queries using the interface. 1 2017 IEEE. ACKNOWLEDGMENT This contribution was made possible by NPRP grant #07-794-1-145 from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the authors. Scopus

Ranking raters based on their agreement or disagreement is a very active research field with many applications. It has been applied for instance for ranking automatic essay grading algorithms based on their agreement with human graders. Many metrics have been introduced for measuring agreement between graders with different performances depending on the type of data. In this paper, we apply the agreement measures on a new field which is the hadith authentication problem. Indeed, reported hadiths (actions or sayings of Prophet Muhamad PBUH) are not always authentic and scholars who study their authenticity often disagree on some hadiths. In this study, many metrics have been proposed and studied on a simulated data in order to determine which metric best reflects the measure of agreement especially in the presence of missing data. Best performing metric has been applied to real data in order to rank the hadith scholars. This study is a breakthrough in the science of hadith as it makes it possible to assign a confidence score to each authenticity judgment that reflects how often the hadith scholar in question agrees or disagrees with other scholars. Qatar National Research Fund, QNRF& Qatar Foundation, QF Scopus

Prediction of date and Ideology (aka Islamic school or Madhab) is a very important research field. It allows scholars and researchers to find out more about what is specific to a particular period of time and what the main topics of interest were during each period. It also makes it possible to see how each Islamic school differs from others. In this paper, we describe a new system for date and Islamic school prediction for one single page of text. Our method extracts the keywords using the hyper concept algorithm which outputs the keywords in a structured tree based on their importance. The keywords are then fed to a classification algorithm which predicts both the date and the Islamic school associated to each document. Experiments have been conducted on a large database crawled from Shamela online library. The results show that date prediction can be done with a root mean square error of about 200 years. Moreover, we study how the error varies with respect to the time period. This study is the first in its kind that gives a global overview of the Islamic jurisprudence heritage. Qatar National Research Fund, QNRF& Qatar Foundation, QF Scopus

Offline signature verification is undeniably a prominent aspect of the biometric research. It has many applications, including banking and forensics. Signature verification task involves a comparison of questioned signature with a set of one or more reference signatures. The questioned signature may be genuine (written by the authentic writer), a forgery (written by a different person) or a disguise (written by the authentic writer with some modifications with the intent of a later denial). Signature verification generally encompasses the two main steps: feature extraction and classification. The system performance is primarily dependent on the feature extraction step because the characterising features distinguish between genuine, disguised and forged signatures. In this study, we propose several geometric and colour features to characterise the signatures. The features are combined using random forests, logistic regression and generalised linear models. The results are reported on the datasets of several competitions, including ICDAR 2009, ICFHR 2010, ICDAR 2011 and ICFHR 2012 signature verification competitions. The proposed method generally outperforms the other participating methods.

Objective:To compare the performance of predicting incident cardiovascular disease (CVD) using past, current, and usual systolic (SBP), separately and in combination, using large-scale routine electronic health records (EHR).Design and method:Using data from UK primary care linked EHR, we applied a

Authentication of Hadiths (sayings of Prophet Muhammad) is very important field for religious scholars as well as historians. Authenticity verification is traditionally conducted by studying how trustworthy is each person in the narration chain. In this study, we propose a novel approach completely based on the content of each Hadith. For each category of hadiths (authentic and non-authentic), we create a binary relation in which the hadiths correspond to the objects of the relation and the words correspond to its attributes. Keywords for each category are then obtained in a hierarchical ordering of importance using the hyper rectangular decomposition. Classification is done by feeding the extracted keywords to a logistic regression classifier. The method has been validated on a database of about 1600 hadiths. Results show that classification accuracy increases with the number of annotators who agreed on the authenticity of each hadith. These findings suggest that our method successfully extracts relevant keywords and can be combined with other traditional methods. 2016 IEEE. Scopus

Named entity extraction is an important step in natural language processing. It aims at finding the entities which are present in text such as organizations, places or persons. Named entities extraction is of a paramount importance when it comes to automatic translation as different named entities are translated differently. Named entities are also very useful for advanced search engines which aim at searching for a detailed information regarding a specific entity. Named entity extraction is a difficult problem as it usually requires a disambiguation step as the same word might belong to different named entities depending on the context.This work has been conducted on the ANERCorp named entities database. This Arabic database contains four different named entities: person, organization, location and miscellaneous. The database contains 6099 sentences, out of which 60% are used for training 20% for validation and 20% for testing.Our method for named entity extraction contains two main steps: the first step predicts the list of named entities which are present at the sentence level. The second step predicts the named entity of each word of the sentence.The prediction of the list of named entities at the sentence level is done through separating the document into sentences using punctuation marks. Subsequently, a binary relation between the set of sentences (x) and the set of words (y) is created from the obtained list of sentences. A relation exists between the sentence (x) and the word (y) if, and only if, (x) contains (y). A binary relation is created for each category of named entities (person, organization, location and miscellaneous). If a sentence contains several named entities, it is duplicated in the relation corresponding to each one of them. Our method then extracts keywords from the obtained binary relations using the hyper concept method [1]. This method decomposes the original relation into non-overlapping rectangles and highlights for each rectangle the most representative keyword. The output is a list of keywords sorted in a hierarchical ordering of importance. The obtained keyword list associated with each category of named entities are fed into a random forest classifier of 10000 random trees in order to predict the list of named entities associated with each sentence. The random forest classifier produces for each sentence the list of probabilities corresponding to the existence of each category of named entities within the sentence.Random Forest [sentence(i)] = (P(Person),P(Organization),P(Location),P(miscellaneous)).Subsequently, the sentence is associated with the named entities for which the corresponding probability is larger than a threshold set empirically on the validation set.In the second step, we create a lookup table associating to each word in the database, the list of named entities to which it corresponds in the training set.For unseen sentences of the test set, the list of named entities predicted at the sentence level is produced, and for each word, the list of predicted named entities is also produced using the lookup table previously built. Ultimately, for each word, the intersection between the two predicted lists of named entities (at the sentence and the word level) will give the final predicted named entity. In the case where more than one named entity is produced at this stage, the one with the maximum probability is kept.We obtained an accuracy of 76.58% when only considering lookup tables of named entities produced at the word level. When performing the intersection with the list produced at the sentence level the accuracy reaches 77.96%.In conclusion, the hierarchical named entity extraction leads to improved results over direct extraction. Future work includes the use of other linguist features and larger lookup table in order to improve the results. Validation on other state of the art databases is also considered.AcknowledgementsThis contribution was made possible by NPRP grant #06-1220-1-233 from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the authors.Reference[1] A. Hassaine, S. Mecheter, and A. Jaoua. “Text Categorization Using Hyper Rectangular Keyword Extraction: Application to News Articles Classification”. Relational and Algebraic Methods in Computer Science. Springer International Publishing, 2015. 312–325.

Sentiment analysis is a very important research task that aims at understanding the general sentiment of a specific community or group of people. Sentiment analysis of Arabic content is still in its early development stages. In the scope of Islamic content mining, sentiment analysis helps understanding what topics Muslims around the world are discussing, which topics are trending and also which topics will be trending in the future.This study has been conducted on a dataset of 5000 comments on news articles collected from Al Jazeera Arabic website. All articles were about the recent war against the Islamic State. The database has been annotated using Crowdflower which is website for crowdsourcing annotations of datasets. Users manually selected whether the sentiment associated with the comment was positive or negative or neutral. Each comment has been annotated by four different users and each annotation is associated with a confidence level between 0 and 1. The confidence level corresponds to whether the users who annotated the same comment agreed or not (1 corresponds to full agreement between the four annotators and 0 to full disagreement).Our method represents the corpus by a binary relation between the set of comments (x) and the set of words (y). A relation exists between the comment (x) and the word (y) if, and only if, (x) contains (y). Three binary relations are created for comments associated with positive, negative and neutral sentiments. Our method then extracts keywords from the obtained binary relations using the hyper concept method [1]. This method decomposes the original relation into non-overlapping rectangles and highlights for each rectangle the most representative keyword. The output is a list of keywords sorted in a hierarchical ordering of importance. The obtained keyword list associated with positive, negative and neutral comments are fed into a random forest classifier of 1000 random trees in order to predict the sentiment associated with each comment of the test set.Experiments have been conducted after splitting the database into 70% training and 30% testing subsets. Our method achieves a correct classification rate of 71% when considering annotations with all values of confidence and even 89% when only considering the annotation with a confidence value equal to 1. These results are very promising and testify of the relevance of the extracted keywords.In conclusion, the hyper concept method extracts discriminative keywords which are used in order to successfully distinguish between comments containing positive, negative and neutral sentiments. Future work includes performing further experiments by using a varying threshold level for the confidence value. Moreover, by applying a part of speech tagger, it is planned to perform keyword extraction on words corresponding to specific grammatical roles (adjectives, verbs, nouns… etc.). Finally, it is also planned to test this method on publicly available datasets such as the Rotten Tomatoes Movie Reviews dataset [2].AcknowledgmentThis contribution was made possible by NPRP grant #06-1220-1-233 from the Qatar National Research Fund (a member of Qatar Foundation). The statements made herein are solely the responsibility of the authors.References[1] A. Hassaine, S. Mecheter, and A. Jaoua. “Text Categorization Using Hyper Rectangular Keyword Extraction: Application to News Articles Classification.” Relational and Algebraic Methods in Computer Science. Springer International Publishing, 2015. 312–325.[2] B. Pang and L. Lee. 2005. “Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales”. In ACL, pages 115–124.

In this paper we present results from the Mapping Dark Matter competition that expressed the weak lensing shape measurement task in its simplest form and as a result attracted over 700 submissions in 2 months and a factor of 3 improvement in shape measurement accuracy on high signal to noise galaxies, over previously published results, and a factor 10 improvement over methods tested on constant shear blind simulations. We also review weak lensing shape measurement challenges, including the Shear TEsting Programmes (STEP1 and STEP2) and the GRavitational lEnsing Accuracy Testing competitions (GREAT08 and GREAT10)., Comment: 20 pages, 12 figures. Solicited submission to New Astronomy Reviews

Automatic text categorization is still a very important research topic. Typical applications include assisting end-users in archiving existing documents, or helping them in browsing existing corpus of documents in a hierarchical way. Text categorization is usually composed of two main steps: keyword extraction and classification. In this paper, a corpus of documents is represented by a binary relation linking each document to the words it contains. From this relation, the Hyper Rectangle Algorithm extracts the list of the most representative words in a hierarchical way. A hyper-Rectangle associated to an element of the range of a binary relation is the union of all non-enlargeable rectangles containing it. The extracted keywords are fed into the random forest classifier in order to predict the category of each document. The method has been validated on the popular Reuters 21578 news articles database. Results are very promising and show the effectiveness of the Hyper Rectangular method in extracting relevant keywords.

The classification of Islamic content on the Internet is a very important step towards authenticity verification. Many Muslims complain that the information they get from the Internet is either inaccurate or simply wrong. With the content growing in an exponential way, its manual labeling and verification is simply an impossible task. To the extent of our knowledge, no previous work has been carried out regarding his task. In this study, we propose a new method for automatic classification of Islamic content on the Internet. A dataset of four Islamic groups has been created containing texts from four different Islamic groups, namely: Sunni (Content representing Sunni Islam), Shia (Content representing Shia Islam), Madkhali (Content forbidding politics and warning against all scholars with different views) and Jihadi (Content promoting Jihad). We collected a dataset containing 20 different texts for each of those groups, totalizing 80 texts, out of which 56 are used for training and 24 for testing. In order to classify those contents automatically, we first preprocessed the texts using normalization, stop words removal, stemming and segmentation into words. Then, we used the hyper-concepts method which makes it possible to represent any corpus through a relation and to decompose it into non-overlapping rectangular relations and to highlight the most representative attributes or keywords in a hierarchical way. The hyper concept keywords extracted from the training set are subsequently used as predictors (containing either 1 when the text contains the keyword and 0 otherwise). Those predictors are fed to a random forest classifier of 5000 random trees. The number of extracted keywords varies according to the depth of the hyper concept tree, ranging from 47 keywords (depth 1) to 296 keywords (depth 15). The average classification accuracy starts at 45.79% for depth 1 and remains roughly stable at 68.33% from depth 10. This result is very interesting as there four different classes (a random predictor would therefore score around 25%). This study is a great step towards the automatic classification of Islamic content on the Internet. The results show that the hyper concept method successfully extracts relevant keywords for each group and helps in categorizing them automatically. The method needs to be combined with some semantic method in order to reach even higher classification rates. The results of the method are also to be compared with manual classification in order to foresee the improvement one can expect as some texts might indifferently belong to more than one category. qscience

Islamic web content offers a very convenient way for people to learn more about Islam religion and the correct practices. For instance, via these web sites they could ask for fatwas (Islamic advisory opinion) with more facilities and serenity. Regarding the sensitivity of the subject, large communities of researchers are working on the evaluation of these web sites according to several criteria. In particular there is a huge effort to check the consistency of the content with respect to the Islamic shariaa (or Islamic law). In this work we are proposing a semiautomatic approach for evaluating the web sites Islamic content, in terms of inconsistency detection, composed of the following steps: (i) Domain selection and definition: It consists of identifying the most relevant named entities related to the selected domain as well as their corresponding values or keywords (NEV). At that stage, we have started building the Fatwas ontology by analyzed around 100 fatwas extracted from the online system. (ii) Formal representation of the Islamic content: It consists of representing the content as formal context relating fatwas to NEV. Here, each named entity is split into different attributes in the database where each attribute is associated to a possible instantiation of the named entity. (iii) Rules extraction: by applying the ConImp tools, we extract a set of implications (or rules) reflecting cause-effect relations between NEV. As an extended option aiming to provide more precise analysis, we have proposed the inclusion of negative attributes. For example for word "licit", we may associate "not licit" or "forbidden", for word "recommended" we associated "not recommended", etc. At that stage by using an extension of Galois Connection we are able to find different logical associations in a minimal way by using the same tool ConImp. (iv) Conceptual reasoning: the objective is to detect a possibly inconsistency between the rules and evaluate their relevance. Each rule is mapped to a binary table in a relational database model. By joining obtained tables we are able to detect inconsistencies. We may also check if a new law is not contradicting existing set of laws by mapping the law into a logical expression. By creating a new table corresponding to its negation we have been able to prove automatically its consistencies as soon as we obtain an empty join of the total set of joins. This preliminary study showed that the logical representation of fatwas gives promising results in detecting inconsistencies within fatwa ontology. Future work includes using automatic named entity extraction and automatic transformation of law into a formatted database; we should be able to build a global system for inconsistencies detection for the domain. ACKNOWLEDGMENT: This publication was made possible by a grant from the Qatar National Research Fund through National Priority Research Program (NPRP) No. 06-1220-1-233. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the Qatar National Research Fund or Qatar University.

In the context of web Islamic data analysis and authentication an important task is to be able to authenticate the holy book if published in the net. For that purpose, in order to detect texts contained in the holy book, it seems obvious to first characterize words which are specific to existing chapters (i.e. "Sourat") and words characterizing each sentence in any chapter (i.e. "Aya"). In this current research, we have first mapped the text of the Quran to a binary context R linking each chapter to all words contained in it, and by calculating the fringe relation F of R, we have been able to discover in a very short time all specific words in each chapter of the holy book. By applying the same approach we have found all specific words of each sentence (i.e. "Aya") in the same chapter whenever it is possible. We have found that almost all sentences in the same chapter have one or many specific words. Only sentences repeated in the same chapter or those sentences included in each other might not have specific words. Observation of words simultaneously specific to a chapter in the holy book and to the sentence in the same chapter gave us the idea for characterizing all specific sentences in each chapter with respect to the whole Quran. We found that for 42 chapters all specific words of a chapter are also specific of some sentence in the same chapter. Such specific words might be used to detect in a shorter time website containing some part of the Quran and therefore should help for checking their authenticity. As a matter of fact by goggling only two or three specific words of a chapter, we observed that search results are directly related to the corresponding chapter in the Quran. Al results have been obtained for Arabic texts with or without vowels. Utilization of adequate data structures and threads enabled us to have efficient software written in Java language. The present tool is directly useful for the recognition of different texts in any domain. In the context of our current project, we project to use the same methods to characterize Islamic books in general. ACKNOWLEDGMENT: This publication was made possible by a grant from the Qatar National Research Fund through National Priority Research Program (NPRP) No. 06-1220-1-233. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the Qatar National Research Fund or Qatar University.

In forensics, the handedness detection or the classification of writers into left or right-handed helps investigators focusing more on a certain category of suspects. However, only a few studies have been carried out in this field. Classification of handwriting into a demographic category is generally performed in two steps: feature extraction and classification. In this study, we propose a system which extract characterizing features from handwritings and use those features to perform the classification of handwritings with regards to handedness. Classification rates are reported on the QUWI dataset, reaching almost 70% for Left and right Handwriting.

The prediction of gender from handwriting is a very interesting research field. However, no standard benchmark is available for researchers in this field. The aim of this competition is to gather researchers and compare recent advances in gender prediction from handwriting. This competition has been hosted on Kaggle, it has attracted 194 teams from both academia and industry. This paper gives details on this competition, including the dataset used, the evaluation procedure and description of participating methods and their performances.

Stroke recovery from offline handwriting is a very interesting research field. However, no standard benchmark is available for researchers in this field. The aim of this competition is to gather researchers and compare recent advances in stroke recovery from offline handwriting. This competition has been hosted on Kaggle, it has attracted 45 teams from both academia and industry. This paper gives details on this competition, including the dataset used, the evaluation procedure and description of participating methods and their performances.

Line and word segmentation is a key-step in any document image analysis system. It can be used for instance in handwriting recognition when separating words before their recognition. Line segmentation can also serve as a prior step before extracting the geometric characteristics of lines which are unique to each writer. Text line and word segmentation is not an easy task because of the following problems: 1) text lines do not all have the same direction in the handwritten text; 2) text lines are not always horizontal which makes their separation more difficult; 3) characters may overlap between successive text lines; 4) it is often confusing to distinguish between inter and intra word distances. In our method, line segmentation is done by using a smoothed version of the handwritten document which makes it possible to detect the main line components using a subsequent thresholding algorithm. The connected components of the resulting image are then assigned to a separate label which represents a line component. Then, each text region which intersects only with one line component is assigned to the same label of that line component. The Voronoi diagram of the image thus obtained is then computed in order to label the remaining text pixels. Word segmentation is performed by computing a generalized Chamfer distance in which the horizontal distance is slightly favored. This distance is subsequently smoothed in order to reflect the distances between word components and neglect the distance to dots and diacritics. Word segmentation is then performed by thresholding the distance thus obtained. The threshold depends on the characteristics of the handwriting. We have therefore computed several features in order to predict it, including: the sum of maximum distances within each line component, the number of connected components within the document and the average width and height of lines. The optimal threshold is then obtained by training a linear regression of those features on a training set of about 100 documents. This method achieved the best performance on the ICFHR Handwriting Segmentation Contest dataset reaching a matching score of 97.4% on line segmentation and 91% on word segmentation. The method has also been tested on the QUWI Arabic dataset reaching 97.1% on line segmentation and 49.6% on word segmentation. The relatively low performance of word segmentation in Arabic script is due to the fact that words are very close to each other with respect to English script. The proposed method tackles most of the problems of line and word segmentation and achieves high segmentation results. It can however be improved by combining it with a handwriting recognizer which will eliminates words which are not recognized.

This paper presents a new offline dataset called the Qatar University Writer Identification dataset (QUWI). This dataset contains both Arabic and English handwritings and can be used to evaluate the performance of offline writer identification systems. It consists of handwritten documents of 1017 volunteers of different ages, nationalities, genders and education levels. The writers were asked to copy a specific text and to generate a random text, which allows the dataset to be used for both text-dependent and text-independent writer identification tasks. We describe the gathering and processing steps and define several evaluation tasks regarding the use of this dataset.

Arabic writer identification is a very active research field. However, no standard benchmark is available for researchers in this field. The aim of this competition is to gather researchers and compare recent advances in Arabic writer identification. This competition has been hosted on Kaggle, it has attracted forty-three teams from both academia and industry. This paper gives details on this competition, including the dataset used, the evaluation procedure and description of participating methods and their performances.

Online signatures are acquired using a digital tablet which provides all the trajectories of the signature, as well as the variation in pressure with respect to time. Therefore, online signature verification achieves higher recognition rates than offline signature verification. Nowadays, forensic document examiners distinguish between forgeries, in which an impostor tries to imitate a given signature of another person and disguised signatures, in which the authentic author deliberately tries to hide his/her identity with the purpose of denial at a later stage. The disguised signatures play an important role in real forensic cases but are not considered in recent literature. In this study, we used online signatures acquired using a WACOM Intuos4 digitizing tablet with a sampling rate of 200Hz, a resolution of 2000 lines/cm and a precision of 0.25mm. The pressure information is available in 1024 levels. Online signatures contain a set of samples, each sample corresponds to the point coordinates on the digitizing tablet along with the corresponding pressure (Xt; Yt; Pt) where t corresponds to time. From those three basic signals, four other are extracted: distances, angles, speeds and angular speed. In order to compare the questioned signature with the reference signature, the differences between their corresponding features are computed at both the signal level and the histogram level. This study has been evaluated on ICDAR2009 signature verification competition dataset and a new dataset of online signatures collected at Qatar University (QU-dataset). This dataset contains signatures of genuine signatures, forgeries and disguised of 194 persons. To the extent of our knowledge, this dataset is the only one that contains disguised online signatures. The best individual performing feature for the ICDAR2009 dataset is the pressure histogram difference which reaches 8% equal error rate (EER). The pressure signal difference is the best individual performing feature for the QU-dataset (29% EER). The combination of features led to 7% EER on the ICDAR2009 dataset and 22% EER on the QU-dataset. This online signature verification system deals with both forgeries and disguised signatures. Several features have been proposed and combined using different classifiers reaching promising performance for disguised signatures detection.

Face and signature based multimodal biometric systems are often required in various areas, such as banking biometric systems and secured mobile phone operating systems, among others. Our system combines these two biometric traits and provides better recognition performance compared with the systems based on a single biometric trait or modality. In multimodal biometric system, the most common fusion approach is integration at the matching score level because of the ease of combining and accessing the scores generated by different matchers. In this paper, we study the performance of a max-of-scores fusion technique based on the face and signature traits of a user. The experiments that were conducted on a database of 40 users indicate that the max-of-scores fusion-based method yields better authentication performance than single-face, single-signature, simple-sum or min-of-scores fusion-based biometric systems.

Writer identification consists of identifying the writer of a certain handwritten document and is of high importance in forensic document examination. Indeed, numerous cases over the years have dealt with evidence provided by handwritten documents such as wills and ransom notes. Automatic methods for writer identification can be divided into codebook-based and feature-based approaches. In codebook-based approaches, the writer is assumed to act as a stochastic generator of graphemes. The probability distribution of grapheme usage is used to distinguish between writers. Feature-based approaches compare the handwritings according to some geometrical, structural or textural features. Feature-based approaches prove to be efficient and are generally preferred when there is a limited amount of available handwriting. Therefore, we are more interested in this study in this category of approaches. Writer identification is performed by comparing a query document to a set of known documents and assigning it to the closest document in term of similarity of handwriting. This is done by extracting characterizing features from the documents including: directions, curvatures, tortuosities (or smoothness), chain codes distributions and edge based directional features. These features correspond to probability distribution functions (PDF) extracted from the handwriting images to characterize writer individuality. When matching a query document against any other document, the ¬χ2 distance between their corresponding features is computed. The smaller the distance, the more likely the two documents are written by the same writer. Therefore, the identified writer is the one of the document with the smallest distance to the query document. The writer is said to be correctly identified when the identified writer corresponds to the actual writer of the query document. The performance has been evaluated on the IAM handwriting dataset, with chain code based features generally outperforming the other features reaching 71% correct identification rate. The combination of all the features lead to 76% correct identification rate. The proposed system also won the music scores writer identification contest reaching 77% identification rate. The proposed method automatically extracts features used by forensic experts in order to identify writers of handwritten document. The results show that the method is efficient and language-independent.

Background and Objectives: There is high requirement of face and signature based multimodal biometric systems in various areas, such as banking, biometric systems and secured mobile phone operating systems. Few studies have been carried out in this area to enhance the performance of identification and authentication based on the fusion of those modalities. In multimodal biometric systems, the most common fusion approach is integration at the matching score level, but it is necessary to compare this strategy of fusion to the other strategies, like fusion at feature level. Our system combines these two biometric traits and provides better recognition performance compared with single biometric systems. Multimodal Authentication Systems: The first monomodal verification system is based on face verification using Gabor filters for feature extraction. The second system is based on online signature verification using Nalwa's method. The classification is released using the Cosine Mahalano-bis distance. Due to its efficiency, we used max-of-scores strategy to fuse face and online signature scores. The second proposed system is based on fusion at the feature level. Results and Conclusions: The performance of feature-level fusion and max-of-scores fusion techniques using face and online signature modalities were evaluated on ORL face database and the QU signature database. The lowest equal error rate is obtained by using a fusion strategy based on max-of-monomodal systems scores. Additionally, feature-level fusion based methods demonstrate a low equal error rate compared with the monomodal systems and have not been affected by the increase in the features vector dimension in term of time of verification; on the contrary, the fusion at the score level is clearly affected and it takes more in-time verification because it's necessary to get scores from each biometric trait before the fusion step.