Better models by discarding data?

BackgroundHigh-throughput gene expression profiling technologies generating a wealth of data, are increasingly used for characterization of tumor biopsies for clinical trials. By applying machine learning algorithms to such clinically documented data sets, one hopes to improve tumor diagnosis, prognosis, as well as prediction of treatment response. However, the limited number of patients enrolled in a single trial study limits the power of machine learning approaches due to over-fitting. One could partially overcome this limitation by merging data from different studies. Nevertheless, such data sets differ from each other with regard to technical biases, patient selection criteria and follow-up treatment. It is therefore not clear at all whether the advantage of increased sample size outweighs the disadvantage of higher heterogeneity of merged data sets. Here, we present a systematic study to answer this question specifically for breast cancer data sets. We use survival prediction based on Cox regression as an assay to measure the added value of merged data sets.ResultsUsing time-dependent Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) and hazard ratio as performance measures, we see in overall no significant improvement or deterioration of survival prediction with merged data sets as compared to individual data sets. This apparently was due to the fact that a few genes with strong prognostic power were not available on all microarray platforms and thus were not retained in the merged data sets. Surprisingly, we found that the overall best performance was achieved with a single-gene predictor consisting of CYB5D1.ConclusionsMerging did not deteriorate performance on average despite (a) The diversity of microarray platforms used. (b) The heterogeneity of patients cohorts. (c) The heterogeneity of breast cancer disease. (d) Substantial variation of time to death or relapse. (e) The reduced number of genes in the merged data sets. Predictors derived from the merged data sets were more robust, consistent and reproducible across microarray platforms. Moreover, merging data sets from different studies helps to better understand the biases of individual studies and can lead to the identification of strong survival factors like CYB5D1 expression.

Aircraft structural damage detection is becoming of increased importance. Technologies such as acousto-ultrasonic have been suggested for this application; however, an optimization strategy for sensor network design is required to ensure a high detection probability while minimizing sensor network mass. A methodology for optimizing acousto-ultrasonic transducer placement for adhesive disbond detection on metallic aerospace structures is presented. Experimental data sets were acquired using three-dimensional scanning laser vibrometry enabling in-plane and out-of-plane Lamb wave components to be considered. This approach employs a novel multi-sensor site strategy which is difficult to achieve with physical transducers. Different excitation frequencies and source–damage–sensor paths were considered. A fitness assessment criterion which compared baseline and damaged data sets using cross-correlation coefficients was developed empirically. Efficient sensor network optimization was achieved using a bespoke genetic algorithm for different network sizes with the effectiveness assessed and discussed. A comparable numerical data set was also produced using the local interaction simulation approach and optimized using the same methodology. Comparable results with those of the experimental data set indicated a good agreement. As such, the numerical approach demonstrates that acousto-ultrasonic sensor networks can be optimized using simulation (with some further refinement) during an aircraft design phase, being a useful tool to sensor network designers.

The use of no-observed-effect levels (NOLs) and lowest-observed-effect levels (LOELs) assummary statistics for low toxic effects has been criticized recently. In this paper, we explore the regression-based approach for estimating low toxic effects. Key questions with this approach include: (1) is the approach practical with typical toxicity data sets, (2) can low toxic effects be estimated with confidence, (3) what models are appropriate for typical data sets and endpoints, and (4) how do point estimates of low toxic effects (e.g., EC10) compare with their corresponding LOELs and NOELs? Analyses of 198 toxicity data sets (all but 13 unpublished) revealed that: (1) even with a range of models to choose from (three models in the logistic family, probit model, Weibull model), >80% of the data sets did not produce a single adequate model fit; (2) estimates of low toxic effects were often model dependent when an extrapolation beyond the toxicity data was required; (3) confidence intervals can be quite large at 5% effect and lower; (4) of the five models, the three-parameter logistic equation with a steep slope parameter had the best model fit in the majority of the data sets; and (5) 76.9% of the NOELs and 100% of the LOELs were higher than their corresponding 10% effect point estimates from the best-fit model equations, suggesting that NOELs and LOELs are poor indicators of low toxic effects. We believe that the regression-based approach is a better tool than hypothesis testing for estimating low toxic effects. The approach, however, does not produce adequate model fits unless there is an obvious concentration–response relationship and several treatments with partial effects.

Influence of the characteristics of the experimental data set used to identify anisotropy parameters

In this paper, a 2D least square FIR differentiation filter method is proposed in the context of tensorial elastography. Displacements are estimated from ultrasonic images obtained during freehand compression. Two 2D filters are used for providing all partial derivative maps according to the estimated motion obtained with a Bilinear Deformable Block Matching method. From these results, several strain tensor images are built, revealing the tissue's elasticity properties. Among these tensors, the infinitesimal strains, the rigid rotation and principal strains tensors are retained, thus improving the contrast between the tumor and the background tissue. The 2D filters were first applied to displacement maps estimated from simulated data and two experimental RF data sets and then compared to the 1D differentiation filter LSQSE used in elastography. The first experimental data set was collected from a homogeneous phantom with a cylindrical hard inclusion and the second data set was collected from a thyroid gland with a malignant tumor. For all studied RF data sets, the contrast between the tumor and the background tissue calculated on strain tensor maps was increased up to a factor of 3 with our method compared to the LSQSE method. We also observed that the rigid rotation tensor and principal strains tensor were well adapted to thyroid elastography.

Mining frequent patterns has been a topic of active research because it is computationally the most expensive step in association rule discovery. In this paper, we discuss the use of compact data structure design for improving the efficiency of frequent pattern mining. It is based on our work in developing efficient algorithms that outperform the best available frequent pattern algorithms on a number of typical data sets. We discuss improvements to the data structure design that has resulted in faster frequent pattern discovery. The performance of our algorithms is studied by comparing their running times on typical test data sets against the fastest Apriori, Eclat, FP-Growth and OpportuneProject algorithms. We discuss the performance results as well as the strengths and limitations of our algorithms.

In health research discourse, there is the assumption that a controlled, planned clinical research naturally produces cleaner datasets, while real-world data suffers from quality deficiencies due to its spontaneous collection without anticipation of secondary research use. However, this binary view simplifies a complex reality. Both clinical research and real-world data exhibit quality issues, challenging the convention that data origin determines data integrity. What proves to be more vital is the way data quality (DQ) is being assessed—particularly when examining the exhibited relationships between data items that reflect the true complexity of healthcare information. Health data is inherently interconnected e.g., height relates to weight and age; medication dosages connect to patient characteristics and medical conditions. Yet most data quality assessment (DQA) studies limit their logical evaluations to only a handful of interdependent items, missing the broader web of relationships that could reveal critical quality issues. Contradictions within datasets serve as an important DQ indicator (measurable attribute of data) that extend far beyond simple data entry errors. When values contradict each other, they reveal misunderstandings of domain knowledge, failures in data capture infrastructure, and problems in data transformation processes. Given the complexity of the contradiction patterns, a comprehensive assessment becomes essential. Comprehensive DQA in health research requires evaluating data on multiple dimensions, such as completeness, conformance, and plausibility, to determine the suitability of data for intended use. These dimensions are used to group related measurable attributes of data referred to as indicators. While some DQ indicators are implemented as generic rules (e.g., completeness), contradiction, which denotes pair(s) of measurements or facts that cannot logically coexist, are multi-dimensional. Some contradictions require common knowledge for their evaluation—e.g., the disparity in age information at baseline and during follow-up visits. However, contradictions in the pre-analytic states of blood samples require knowledge of experts. Although various tools have been developed to assess contradictions in health data, the evolving nature of contradiction rules presents the following challenges for existing tools: 1) limited reusability across diverse health datasets due to incompatible predefined rules, 2) insufficient contextual information required for comprehensive contradiction assessment, and 3) lack of consistent representation of contradiction patterns to support transparent rule implementations and computational efficiency. To address these gaps in a systematic manner, the first step of this thesis was to adapt an existing tool developed based on an existing framework to assess broader contradiction rules introduced in the German Corona Consensus (GECCO) dataset. Building on the experience from the modification of Schmidt et al.'s tool, this thesis developed a new framework that generalizes the DQA tool to accept custom contradiction rules compatible with diverse health datasets. This is intended to empower domain experts in defining effective contradiction rules on an ad-hoc basis while maintaining alignment with established DQ indicators. Although the primary dependencies between health data items form the basis for contradiction rule definition, the required context for a comprehensive analysis is equally relevant. For example, diabetes mellitus (DM) and insulin medication (INS) are two interdependent items, however, it is the timestamps that indicate if the INS was administered before the DM diagnosis, which is a prelude for a conclusive contradiction finding. Through a qualitative grading scheme for analyzing metadata defined in a study database, this work created a mechanism to measure the availability of the contextual information required to support a context-aware contradiction analysis. However, having access to rich contextual information is only part of the solution—the way contradiction rules are structured and implemented is equally critical for effective DQA. While researchers have worked extensively on harmonizing DQ indicators, the challenge extends beyond categorization to implementation efficiency. Though the taxonomies for contradictions, such as logical and empirical contradictions, reflect the domain description, the rules implemented in the DQA systems do not rely on semantics. Extracting the structure of the varied contradiction patterns ensures transparency and consistency in rule implementation. This work evaluated the performance of different rule implementations and offered an optimization method that integrates the fastest and traceable units of the rules in one unified implementation. This is relevant because broadly defined DQA rules can cause performance degradation in large database infrastructures. With these several perspectives on contradictions as a DQ indicator, the contributions in this thesis enhance contradiction assessment by improving reusability across diverse health datasets, enabling context-aware assessments, and optimizing rule implementation for transparency and computational efficiency. This systematic approach supports high-quality health data assessment while maintaining traceability to support data cleansing efforts.

Failure in achieving a function-based resection related to the insufficient patient's participation is a drawback of awake surgery. To assess preoperative parameters predicting the risk of patient insufficient intraoperative cooperation leading to the arrest of the awake resection. Observational, retrospective, multicentric cohort analysis enrolling 384 (experimental dataset) and 100 (external validation dataset) awake surgeries. In the experimental data set, an insufficient intraoperative cooperation occurred in 20/384 patients (5.2%), leading to awake surgery failure in 3/384 patients (ie, no resection, 0.8%), and precluded the achievement of the function-based resection in 17/384 patients (ie, resection limitation, 4.4%). The insufficient intraoperative cooperation significantly reduced the resection rates (55.0% vs 94.0%, P < .001) and precluded a supratotal resection (0% vs 11.3%, P = .017). Seventy years or older, uncontrolled epileptic seizures, previous oncological treatment, hyperperfusion on MRI, and mass effect on midline were independent predictors of insufficient cooperation during awake surgery ( P < .05). An Awake Surgery Insufficient Cooperation score was then assessed: 96.9% of patients (n = 343/354) with a score ≤2 presented a good intraoperative cooperation, while only 70.0% of patients (n = 21/30) with a score >2 presented a good intraoperative cooperation. In the experimental data set, similar date were found: 98.9% of patients (n = 98/99) with a score ≤2 presented a good cooperation, while 0% of patients (n = 0/1) with a score >2 presented a good cooperation. Function-based resection under awake conditions can be safely performed with a low rate of insufficient patient intraoperative cooperation. The risk can be assessed preoperatively by a careful patient selection.

We describe a method for assessing data set complexity based on the estimation of the underlining probability distribution and Hellinger distance. Contrary to some popular measures it is not focused on the shape of decision boundary in a classification task but on the amount of available data with respect to attribute structure. Complexity is expressed in terms of graphical plot, which we call complexity curve. We use it to propose a new variant of learning curve plot called generalisation curve. Generalisation curve is a standard learning curve with x-axis rescaled according to the data set complexity curve. It is a classifier performance measure, which shows how well the information present in the data is utilised. We perform theoretical and experimental examination of properties of the introduced complexity measure and show its relation to the variance component of classification error. We compare it with popular data complexity measures on 81 diverse data sets and show that it can contribute to explaining the performance of specific classifiers on these sets. Then we apply our methodology to a panel of benchmarks of standard machine learning algorithms on typical data sets, demonstrating how it can be used in practice to gain insights into data characteristics and classifier behaviour. Moreover, we show that complexity curve is an effective tool for reducing the size of the training set (data pruning), allowing to significantly speed up the learning process without reducing classification accuracy. Associated code is available to download at: https://github.com/zubekj/complexity_curve (open source Python implementation).

Multifractal nature of network induced time delay in networked control systems

Macromolecular crystallography has been, for the last few decades, the main source of structural information of biological macromolecular systems and it is one of the most powerful techniques for the analysis of enzyme mechanisms and macromolecular interactions at the atomic level. In addition, it is also an extremely powerful tool for drug design. Recent technological and methodological developments in macromolecular X-ray crystallography have allowed solving structures that until recently were considered difficult or even impossible, such as structures at atomic or subatomic resolution or large macromolecular complexes and assemblies at low resolution. These developments have also helped to solve the 3D-structure of macromolecules from twin crystals. Recently, this technique complemented with cryo-electron microscopy and neutron crystallography has provided the structure of large macromolecular machines with great precision allowing understanding of the mechanisms of their function.

After more than half a century of evolution, machine learning and artificial intelligence, in general, are entering a truly exciting era of broad application in commercial and research sectors. In X-ray crystallography, and its application to structural biology, machine learning is finding a home within expert and automated systems, is forecasting experiment and data analysis outcomes, is predicting whether crystals can be grown and even generating macromolecular structures. This review provides a historical perspective on AI and machine learning, offers an introduction and guide to its application in crystallography and concludes with topical examples of how it is currently influencing macromolecular crystallography.

The histidine kinase (HK) domain belonging to the light-oxygen-voltage histidine kinase (LOV-HK) from Brucella abortus is a member of the HWE family, for which no structural information is available, and has low sequence identity (20%) to the closest HK present in the PDB. The `off-edge' S-SAD method in macromolecular X-ray crystallography was used to solve the structure of the HK domain from LOV-HK at low resolution from crystals in a low-symmetry space group (P21) and with four copies in the asymmetric unit (∼108 kDa). Data were collected both from multiple crystals (diffraction limit varying from 2.90 to 3.25 Å) and from multiple orientations of the same crystal, using the κ-geometry goniostat on SOLEIL beamline PROXIMA 1, to obtain `true redundancy'. Data from three different crystals were combined for structure determination. An optimized HK construct bearing a shorter cloning artifact yielded crystals that diffracted X-rays to 2.51 Å resolution and that were used for final refinement of the model. Moreover, a thorough a posteriori analysis using several different combinations of data sets allowed us to investigate the impact of the data-collection strategy on the success of the structure determination.

With the rapid development of technology, this development led to the emergence of microarray technology. It has the effect of studying the levels of gene expression in a way that makes it easier for researchers to observe the expression levels of millions of genes at the same time in a single experiment. Development also helped in the emergence of powerful tools to identify interactions between target genes and regulatory factors. The main aim of this study is to build models to predicate the relationship (Interaction) between Transcription Factors (TFs) proteins and target genes by selecting the subset of important genes (Relevant genes) from original dataset. The proposed methodology comprises into three major stages: the genes selection, merge datasets and the prediction stage. The process of reducing the computational space of gene data has been accomplished by using proposed mutual information method for genes selection based on the data of gene expression. In the prediction, the proposed prediction regression techniques are utilized to predict with binding rate between single TF-target gene. It has been compared the efficiency of two different proposed regression techniques including: Linear Regression and Random Forest Regression. Two available data sets have been utilized to achieve the objectives of this study: Gene’s expression data of Yeast Cell Cycle dataset and Transcription Factors dataset. The evaluation of predictions performance has been performed depending on two performance prediction measures (Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) with (10) Folds-Cross Validation.

A comparison of turbine mass flow models based on pragmatic identification data sets for turbogenerator model development