Search

SummaryThis article is part of a For-Discussion-Section of Methods of Information in Medicine about the papers “The Evolution of Boosting Algorithms – From Machine Learning to Statistical Modelling” [1] and “Extending Statistical Boosting – An Overview of Recent Methodological Developments” [2], written by Andreas Mayr and co-authors. It is introduced by an editorial. This article contains the combined commentaries invited to independently comment on the Mayr et al. papers. In sub-sequent issues the discussion can continue through letters to the editor.

SummaryObjectives: Hospital discharge databases store hundreds of thousands of patients. These datasets are usually used by health insurance companies to process claims from hospitals, but they also represent a rich source of information about the patterns of medical care. The proposed subgroup discovery method aims to improve the efficiency of detecting interpretable subgroups in data.Methods: Supervised descriptive rule discovery techniques can prove inefficient in cases when target class samples represent only an extremely small amount of all available samples. Our approach aims to balance the number of samples in target and control groups prior to subgroup discovery process. Additionally, we introduce some improvements to an existing subgroup discovery algorithm enhancing the user experience and making the descriptive data mining process and visualization of rules more user friendly.Results: Instance-based subspace subgroup discovery introduced in this paper is demonstrated on hospital discharge data with focus on medical errors. In general, the number of patients with a recorded diagnosis related to a medical error is relatively small in comparison to patients where medical errors did not occur. The ability to produce comprehensible and simple models with high degree of confidence, support, and predictive power using the proposed method is demonstrated.Conclusions: This paper introduces a subspace subgroup discovery process that can be applied in all settings where a large number of samples with relatively small number of target class samples are present. The proposed method is implemented in Weka machine learning environment and is available at http://ri.fzv.uni-mb.si/ssd.

Summary Objectives: This paper focuses on how we could analyze and interpret filtered back-projection reconstructed signals from low-dose computed tomographic (CT) imaging systems. There exists a growing imbalance between dosage reduction and effective signal interpretation. At the same time, low-dose applications are undergoing alarming growth. Methods: This paper interprets filtered back-projection images in low-dose CT systems and details the possible properties of the artifacts. The interpretation leads to design of a new multi-image filtered back-projection approach that allows artifacts to be effectively identified across multiple images. We use this approach as a building block to propose a new reconstruction method that enables effective artifacts reduction and efficient implementation. Results: Experiments with both clinical and simulated low-dose images demonstrate the validity and effectiveness of the proposed approach. Conclusions: This study discusses a new FBP-based reconstruction approach based on signal interpretation from low-dosage acquisition. This method uses multiple filtered back-projection images from projection subsets to provide clues for distinguishing underlying clinical structure from artifacts. A framework is derived for effective signal interpretation and artifacts reduction. It requires no hardware change and a minimum amount of extra software support compared with current CT systems. Clinical and simulated low-dose CT scans demonstrated effectiveness of the proposed method.

Summary Objectives: There have been many developments and applications of mathematical methods in the context of record linkage as one area of interdisciplinary research efforts. However, comparative evaluations of record linkage methods are still underrepresented. In this paper improvements of the Fellegi-Sunter model are compared with other elaborated classification methods in order to direct further research endeavors to the most promising methodologies. Methods: The task of linking records can be viewed as a special form of object identification. We consider several non-stochastic methods and procedures for the record linkage task in addition to the Fellegi-Sunter model and perform an empirical evaluation on artificial and real data in the context of iterative insertions. This evaluation provides a deeper insight into empirical similarities and differences between different modelling frames of the record linkage problem. In addition, the effects of using string comparators on the performance of different matching algorithms are evaluated. Results: Our central results show that stochastic record linkage based on the principle of the EM algorithm exhibits best classification results when calibrating data are structurally different to validation data. Bagging, boosting together with support vector machines are best classification methods when calibrating and validation data have no major structural differences. Conclusions: The most promising methodologies for record linkage in environments similar to the one considered in this paper seem to be stochastic ones.

SummaryThis editorial is part of a For-Discussion- Section of Methods of Information in Medicine about the papers “The Evolution of Boosting Algorithms – From Machine Learning to Statistical Modelling” [1] and “Ex-tending Statistical Boosting – An Overview of Recent Methodological Developments” [2], written by Andreas Mayr and co authors. It preludes two discussed reviews on developments and applications of boosting in biomedical research. The two review papers, written by Andreas Mayr, Harald Binder, Olaf Gefeller, and Matthias Schmid, give an overview on recently published methods that utilise gradient or likelihood-based boosting for fitting models in the life sciences. The reviews are followed by invited comments [3] by experts in both boosting theory and applications.

Background An accurate and reproducible method to delineate tumor margins is of great importance in clinical diagnosis and treatment. In nasopharyngeal carcinoma (NPC), due to limitations such as high variability, low contrast, and discontinuous boundaries in presenting soft tissues, tumor margin can be extremely difficult to identify in magnetic resonance imaging (MRI), increasing the challenge of NPC segmentation task. Objectives The purpose of this work is to develop a semiautomatic algorithm for NPC image segmentation with minimal human intervention, while it is also capable of delineating tumor margins with high accuracy and reproducibility. Methods In this paper, we propose a novel feature selection algorithm for the identification of the margin of NPC image, named as modified random forest recursive feature selection (MRF-RFS). Specifically, to obtain a more discriminative feature subset for segmentation, a modified recursive feature selection method is applied to the original handcrafted feature set. Moreover, we combine the proposed feature selection method with the classical random forest (RF) in the training stage to take full advantage of its intrinsic property (i.e., feature importance measure). Results To evaluate the segmentation performance, we verify our method on the T1-weighted MRI images of 18 NPC patients. The experimental results demonstrate that the proposed MRF-RFS method outperforms the baseline methods and deep learning methods on the task of segmenting NPC images. Conclusion The proposed method could be effective in NPC diagnosis and useful for guiding radiation therapy.

SummaryObjectives: This paper addresses the problem of decision-making in relation to the administration of noninvasive mechanical ventila tion (NIMV) in intensive care units.Methods: Data mining methods were employed to find out the factors influencing the success/failure of NIMV and to predict its results in future patients. These artificial intelligence-based methods have not been applied in this field in spite of the good results obtained in other medical areas.Results: Feature selection methods provided the most influential variables in the success/ failure of NIMV, such as NIMV hours, PaCO2 at the start, PaO2 / FiO2 ratio at the start, hematocrit at the start or PaO2 / FiO2 ratio after two hours. These methods were also used in the preprocessing step with the aim of improving the results of the classifiers. The algorithms provided the best results when the dataset used as input was the one containing the attributes selected with the CFS method. Conclusions: Data mining methods can be successfully applied to determine the most influential factors in the success/failure of NIMV and also to predict NIMV results in future patients. The results provided by classifiers can be improved by preprocessing the data with feature selection techniques.

SummaryIntroduction: This article is part of the Focus Theme of Methods of Information in Medicine on “Biosignal Interpretation: Advanced Methods for Neural Signals and Images”.Background: Dynamic Causal Modelling (DCM) is a generic formalism to study effective brain connectivity based on neuroimaging data, particularly functional Magnetic Resonance Imaging (fMRI). Recently, there have been attempts at modifying this model to allow for stochastic disturbances in the states of the model.Objectives: This paper proposes the Multiple- Model Kalman Filtering (MMKF) technique as a stochastic identification model discriminating among different hypothetical connectivity structures in the DCM framework; moreover, the performance compared to a similar de terministic identification model is assessed.Methods: The integration of the stochastic DCM equations is first presented, and a MMKF algorithm is then developed to perform model selection based on these equations. Monte Carlo simulations are performed in order to investigate the ability of MMKF to distinguish between different connectivity structures and to estimate hidden states under both deterministic and stochastic DCM.Results: The simulations show that the proposed MMKF algorithm was able to successfully select the correct connectivity model structure from a set of pre-specified plausible alternatives. Moreover, the stochastic approach by MMKF was more effective compared to its deterministic counterpart, both in the selection of the correct connectivity structure and in the estimation of the hidden states.Conclusions: These results demonstrate the applicability of a MMKF approach to the study of effective brain connectivity using DCM, particularly when a stochastic formulation is desirable.

SummaryObjectives: To develop an adaptive approach to mine frequent semantic tags (FSTs) from heterogeneous clinical research texts.Methods: We develop a “plug-n-play” framework that integrates replaceable un-supervised kernel algorithms with formatting, functional, and utility wrappers for FST mining. Temporal information identification and semantic equivalence detection were two example functional wrappers. We first compared this approach’s recall and efficiency for mining FSTs from ClinicalTrials.gov to that of a recently published tag-mining algorithm. Then we assessed this approach’s adaptability to two other types of clinical research texts: clinical data requests and clinical trial protocols, by comparing the prevalence trends of FSTs across three texts.Results: Our approach increased the average recall and speed by 12.8% and 47.02% respectively upon the baseline when mining FSTs from ClinicalTrials.gov, and maintained an overlap in relevant FSTs with the baseline ranging between 76.9% and 100% for varying FST frequency thresholds. The FSTs saturated when the data size reached 200 documents. Consistent trends in the prevalence of FST were observed across the three texts as the data size or frequency threshold changed.Conclusions: This paper contributes an adaptive tag-mining framework that is scalable and adaptable without sacrificing its recall. This component-based architectural design can be potentially generalizable to improve the adaptability of other clinical text mining methods.

SummaryIntroduction: This article is part of the Focus Theme of Methods of Information in Medicine on “Big Data and Analytics in Healthcare”.Objectives: This paper describes the scale-up efforts at the VA Salt Lake City Health Care System to address processing large corpora of clinical notes through a natural language processing (NLP) pipeline. The use case described is a current project focused on detecting the presence of an indwelling uri-nary catheter in hospitalized patients and subsequent catheter-associated urinary tract infections.Methods: An NLP algorithm using v3NLP was developed to detect the presence of an indwelling urinary catheter in hospitalized patients. The algorithm was tested on a small corpus of notes on patients for whom the presence or absence of a catheter was already known (reference standard). In planning for a scale-up, we estimated that the original algorithm would have taken 2.4 days to run on a larger corpus of notes for this project (550,000 notes), and 27 days for a corpus of 6 million records representative of a national sample of notes. We approached scaling-up NLP pipelines through three techniques: pipeline replication via multi-threading, intra-annotator threading for tasks that can be further decomposed, and remote annotator services which enable annotator scale-out.Results: The scale-up resulted in reducing the average time to process a record from 206 milliseconds to 17 milliseconds or a 12-fold increase in performance when applied to a corpus of 550,000 notes.Conclusions: Purposely simplistic in nature, these scale-up efforts are the straight forward evolution from small scale NLP processing to larger scale extraction without incurring associated complexities that are inherited by the use of the underlying UIMA framework. These efforts represent generalizable and widely applicable techniques that will aid other computationally complex NLP pipelines that are of need to be scaled out for processing and analyzing big data.

SummaryBackground: An aging population and people’s higher survival to diseases and traumas that leave physical consequences are challenging aspects in the context of an efficient health management. This is why telerehabilitation systems are being developed, to allow monitoring and support of physiotherapy sessions at home, which could reduce healthcare costs while also improving the quality of life of the users.Objectives: Our goal is the development of a Kinect-based algorithm that provides a very accurate real-time monitoring of physical rehabilitation exercises and that also provides a friendly interface oriented both to users and physiotherapists.Methods: The two main constituents of our algorithm are the posture classification method and the exercises recognition method. The exercises consist of series of movements. Each movement is composed of an initial posture, a final posture and the angular trajectories of the limbs involved in the movement. The algorithm was designed and tested with datasets of real movements performed by volunteers. We also explain in the paper how we obtained the optimal values for the trade-off values for posture and trajectory recognition.Results: Two relevant aspects of the algorithm were evaluated in our tests, classification accuracy and real-time data processing. We achieved 91.9% accuracy in posture classification and 93.75% accuracy in trajectory recognition. We also checked whether the algorithm was able to process the data in real-time. We found that our algorithm could process more than 20,000 postures per second and all the required trajectory data-series in real-time, which in practice guarantees no perceptible delays. Later on, we carried out two clinical trials with real patients that suffered shoulder disorders. We obtained an exercise monitoring accuracy of 95.16%.Conclusions: We present an exercise recognition algorithm that handles the data provided by Kinect efficiently. The algorithm has been validated in a real scenario where we have verified its suitability. Moreover, we have received a positive feedback from both users and the physiotherapists who took part in the tests.

SummaryBackground: Drugs can treat human diseases through chemical interactions between the ingredients and intended targets in the human body. However, the ingredients could unexpectedly interact with off-targets, which may cause adverse drug side effects. Notifying patients and physicians of potential drug effects is an important step in improving healthcare quality and delivery.Objective: With the increasing popularity of Web 2.0 applications, more and more patients start discussing drug side effects in many online sources. These online discussions form a valuable source for mining interesting knowledge about side effects. The main goal of this paper is to investigate the feasibility of exploiting these discussions to discover unrecognized drug side effects.Methods: We propose methods that can 1) build a knowledge base for drug side effects by automatically integrating the in -formation related to drug side effects from different sources; and 2) monitor online discussions about drugs and discover potential unrecognized drug side effects.Results: Experiment results show that the online discussions indeed provide useful information discovering unrecognized drug side effects. We find that the integrated knowledge base contains more information than individual online sources. Moreover, both proposed detection methods can identi -fy the side effects related to the four recently recalled drugs, and the information from online discussions makes it possible to make the detection much earlier than official announcements. Finally, the proposed genera -tive modeling method is shown to be more effective than the discriminative method.Conclusions: We find that it is possible to monitor online discussions to detect un -recognized drug side effects. The developed system is expected to serve as a com -plementary tool for drug companies and FDA to receive feedbacks from the patients, and it has the potentials to expedite the discovery process of unrecognized drug side effects and to improve the quality of healthcare.

SummaryBackground: Multi-class molecular cancer classification has great potential clinical implications. Such applications require statistical methods to accurately classify cancer types with a small subset of genes from thousands of genes in the data.Objectives: This paper presents a new functional gradient descent boosting algorithm that directly extends the HingeBoost algorithm from the binary case to the multi-class case without reducing the original problem to multiple binary problems.Methods: Minimizing a multi-class hinge loss with boosting technique, the proposed Hinge-Boost has good theoretical properties by implementing the Bayes decision rule and providing a unifying framework with either equal or unequal misclassification costs. Furthermore, we propose Twin HingeBoost which has better feature selection behavior than Hinge-Boost by reducing the number of ineffective covariates. Simulated data, benchmark data and two cancer gene expression data sets are utilized to evaluate the performance of the proposed approach.Results: Simulations and the benchmark data showed that the multi-class HingeBoost generated accurate predictions when compared with the alternative methods, especially with high-dimensional covariates. The multi-class Hinge-Boost also produced more accurate prediction or comparable prediction in two cancer classification problems using gene expression data.Conclusions: This work has shown that the HingeBoost provides a powerful tool for multi-classification problems. In many applications, the classification accuracy and feature selection behavior can be further improved when using Twin HingeBoost.

SummaryBackground: Automated analysis of imaged histopathology specimens could potentially provide support for improved reliability in detection and classification in a range of investigative and clinical cancer applications. Automated segmentation of cells in the digitized tissue microarray (TMA) is often the prerequisite for quantitative analysis. However overlapping cells usually bring significant challenges for traditional segmentation algorithms.Objectives: In this paper, we propose a novel, automatic algorithm to separate overlapping cells in stained histology specimens acquired using bright-field RGB imaging.Methods: It starts by systematically identifying salient regions of interest throughout the image based upon their underlying visual content. The segmentation algorithm subsequently performs a quick, voting based seed detection. Finally, the contour of each cell is obtained using a repulsive level set deformable model using the seeds generated in the previous step. We compared the experimental results with the most current literature, and the pixel wise accuracy between human experts’ annotation and those generated using the automatic segmentation algorithm.Results: The method is tested with 100 image patches which contain more than 1000 overlapping cells. The overall precision and recall of the developed algorithm is 90% and 78%, respectively. We also implement the algorithm on GPU. The parallel implementation is 22 times faster than its C/C++ sequential implementation.Conclusion: The proposed segmentation algorithm can accurately detect and effectively separate each of the overlapping cells. GPU is proven to be an efficient parallel platform for overlapping cell segmentation.

SummaryBackground: In the concept of cloud-computing-based systems, various authorized users have secure access to patient records from a number of care delivery organizations from any location. This creates a growing need for remote visualization, advanced image processing, state-of-the-art image analysis, and computer aided diagnosis.Objectives: This paper proposes a system of algorithms for automatic detection of anatomical landmarks in 3D volumes in the cloud computing environment. The system addresses the inherent problem of limited bandwidth between a (thin) client, data center, and data analysis server.Methods: The problem of limited bandwidth is solved by a hierarchical sequential detection algorithm that obtains data by progressively transmitting only image regions required for processing. The client sends a request to detect a set of landmarks for region visualization or further analysis. The algorithm running on the data analysis server obtains a coarse level image from the data center and generates landmark location candidates. The candidates are then used to obtain image neighborhood regions at a finer resolution level for further detection. This way, the landmark locations are hierarchically and sequentially detected and refined.Results: Only image regions surrounding landmark location candidates need to be trans- mitted during detection. Furthermore, the image regions are lossy compressed with JPEG 2000. Together, these properties amount to at least 30 times bandwidth reduction while achieving similar accuracy when compared to an algorithm using the original data.Conclusions: The hierarchical sequential algorithm with progressive data transmission considerably reduces bandwidth requirements in cloud-based detection systems.

SummaryBackground: Diffusion-MRI provides a unique window on brain anatomy and insights into aspects of tissue structure in living humans that could not be studied previously. There is a major effort in this rapidly evolving field of research to develop the algorithmic tools necessary to cope with the complexity of the datasets.Objectives: This work illustrates our strategy that encompasses the development of a modularized and open software tool for data processing, visualization and interactive exploration in diffusion imaging research and aims at reinforcing sustainable evaluation and progress in the field.Methods: In this paper, the usability and capabilities of a new application and toolkit component of the Medical Imaging and Interaction Toolkit (MITK, www.mitk.org), MITKDI, are demonstrated using in-vivo datasets.Results: MITK-DI provides a comprehensive software framework for high-performance data processing, analysis and interactive data exploration, which is designed in a modular, extensible fashion (using CTK) and in adherence to widely accepted coding standards (e.g. ITK, VTK). MITK-DI is available both as an open source software development toolkit and as a ready-to-use in stallable application.Conclusions: The open source release of the modular MITK-DI tools will increase verifiability and comparability within the research community and will also be an important step towards bringing many of the current techniques towards clinical application.

SummaryBackground: Falls are a predominant problem in our aging society, often leading to severe somatic and psychological consequences, and having an incidence of about 30% in the group of persons aged 65 years or above. In order to identify persons at risk, many assessment tools and tests have been developed, but most of these have to be conducted in a supervised setting and are dependent on an expert rater.Objectives: The overall aim of our research work is to develop an objective and unobtrusive method to determine individual fall risk based on the use of motion sensor data. The aims of our work for this paper are to derive a fall risk model based on sensor data that may potentially be measured during typical activities of daily life (aim #1), and to evaluate the resulting model with data from a one-year follow-up study (aim #2).Methods: A sample of n = 119 geriatric inpatients wore an accelerometer on the waist during a Timed ‘Up & Go’ test and a 20 m walk. Fifty patients were included in a one-year follow-up study, assessing fall events and scoring average physical activity at home in telephone interviews. The sensor data were processed to extract gait and dynamic balance parameters, from which four fall risk models – two classification trees and two logistic regression models – were computed: models CT#1 and SL#1 using accelerometer data only, models CT#2 and SL#2 including the physical activity score. The risk models were evaluated in a ten-times tenfold cross-validation procedure, calculating sensitivity (SENS), speci ficity (SPEC), positive and negative predictive values (PPV, NPV), classification accuracy, area under the curve (AUC) and the Brier score.Results: Both classification trees show a fair to good performance (models CT#1/ CT#2): SENS 74% / 58%, SPEC 96% / 82%, PPV 92% / 74%, NPV 77%/82%, accuracy 80% / 78%, AUC 0.83 / 0.87 and Brier scores 0.14 / 0.14. The logistic regression models (SL#1/ SL#2) perform worse: SENS 42% / 58%, SPEC 82% / 78%, PPV 62% / 65%, NPV 67% / 72%, accuracy 65% /70%, AUC 0.65 / 0.72 and Brier scores 0.23 / 0.21.Conclusions: Our results suggest that accelerometer data may be used to predict falls in an unsupervised setting. Furthermore, the parameters used for prediction are measurable with an unobtrusive sensor device during normal activities of daily living. These promising results have to be validated in a larger, long-term prospective trial.

SummaryBackground: In this paper, we describe a new method for the study of clinical information system (CIS) logfiles joined with information in the clinical data warehouse. This method uses heatmap representations and clustering techniques to examine clinicians’ viewing patterns of laboratory test results. The context of our application of these techniques is to inform the creation of a widget-based interface to the CIS.Objectives: We address the rationale, feasibility, and usefulness of our method through examination of three hypotheses:1) The frequency distribution of laboratory test viewing will follow a ’long tail’ pattern, indicating that patterns are highly variable and supporting the rationale for a widget-based configurable system.2) Patterns of laboratory testing viewing (by clinician, specialty, clinician/patient/day, and ICD-9-CM codes) can be distinguished by our methods.3) The identified clusters will include more than 80% of the laboratory test elements found in 30 randomly selected patient records for one day.Methods: The data were plotted as heatmaps and clustered using hierarchical clustering software. Various parameters were tested to give the optimal clusters.Results: All the hypotheses were supported. For Hypothesis 3, 91.4% of information elements in the records were covered by the generated clusters.Conclusions: Study findings support the rationale, feasibility, and usefulness of our methods to examine patterns of information access among clinicians and to inform the creation of widget-based interfaces. The results also contribute to our general understanding of clinicians’ CIS use.

Summary Objectives: To many people, home is a sanctuary. With the maturing of smart home technologies, many people with cognitive and physical disabilities can lead independent lives in their own homes for extended periods of time. In this paper, we investigate the design of machine learning algorithms that support this goal. We hypothesize that machine learning algorithms can be designed to automatically learn models of resident behavior in a smart home, and that the results can be used to perform automated health monitoring and to detect anomalies. Methods: Specifically, our algorithms draw upon the temporal nature of sensor data collected in a smart home to build a model of expected activities and to detect unexpected, and possibly health-critical, events in the home. Results: We validate our algorithms using synthetic data and real activity data collected from volunteers in an automated smart environment. Conclusions: The results from our experiments support our hypothesis that a model can be learned from observed smart home data and used to report anomalies, as they occur, in a smart home.

Summary Objectives: The purpose of this paper is to introduce a new method for spatial-time-frequency analysis of multichannel biomedical data. We exemplify the method for data recorded with a 31-channel Philips biomagnetometer. Methods: The method creates approximations and decompositions of spatiotemporal signal distributions using elements (atoms) chosen from a very large and redundant set (dictionary). The method is based on the Matching Pursuit algorithm, but it uses atoms that are distributed both in time and space (instead of only time-distributed atoms in standard Matching Pursuit). The time-frequency distribution of signal components is described by Gabor atoms and their spatial distribution is modeled by spatial modes. The spatial modes are created with the help of Bessel functions. Two versions of the method, differing in the definition of spatial properties of the atoms, are presented. Results: The technique was validated on simulated data and real magnetic field data. It was used for removal of powerline noise from multichannel magnetoencephalography data, extraction of high-frequency somatosensory evoked fields and for separation of partially overlapping T- and U-waves in magnetocardiography. Conclusions: The method allows for parameterization of multichannel data in the time-frequency as well as in the spatial domains. It thus can be used for signal preserving filtering simultaneously in time, frequency, and space. Applications are e.g. the elimination of artifact components, extraction of components with biological meaning, and data exploration.

Summary Objectives : The intrauterine growth restriction (IUGR) is a pathological state: the fetus is at risk of hypoxia and this condition is associated with increased perinatal morbidity and mortality. However, evidence-based guidelines for clinical surveillance are poor and lack reliable indexes. This study introduces new procedures to extract parameters from the fetal heart rate signal in order to identify severe intrauterine growth restricted (IUGR) fetuses Methods : Standard parameters (time domain and frequency domain indexes) are compared to a new parameter, the Lempel Ziv complexity, and to two regularity estimators (approximate entropy and sample entropy). The paper analyzes the robustness of the indexes coming from the parameter extraction procedure. Results and Conclusions : The results show that the LZ complexity is a stable parameter and it is able to significantly discriminate the severe IUGR (preterm delivered) from moderate IUGR (at term delivered) and from healthy fetuses.

Summary Objectives: Modern tomographic imaging devices enable the acquisition of spatial and temporal image sequences. But, the spatial and temporal resolution of such devices is limited and therefore image interpolation techniques are needed to represent images at a desired level of discretization. This paper presents a method for structure-preserving interpolation between neighboring slices in temporal or spatial image sequences. Methods: In a first step, the spatiotemporal velocity field between image slices is determined using an optical flow-based registration method in order to establish spatial correspondence between adjacent slices. An iterative algorithm is applied using the spatial and temporal image derivatives and a spatiotemporal smoothing step. Afterwards, the calculated velocity field is used to generate an interpolated image at the desired time by averaging intensities between corresponding points. Three quantitative measures are defined to evaluate the performance of the interpolation method. Results: The behaviorand capability of the algorithm is demonstrated by synthetic images. A population of 17 temporal and spatial image sequences are utilized to compare the optical flow-based interpolation method to linear and shape-based interpolation. The quantitative results show that the optical flow-based method outperforms the linear and shape-based interpolation statistically significantly. Conclusions: The interpolation method presented is able to generate image sequences with appropriate spatial or temporal resolution needed for image comparison, analysis or visualization tasks. Quantitative and qualitative measures extracted from synthetic phantoms and medical image data show that the new method definitely has advantages over linear and shape-based interpolation.

Summary Objectives: Methodology based on expert panels has been commonly used to evaluate the appropriateness of interventions. An important issue is the adequate synthesis of the generated information in an applicable way to clinical decision making. This paper shows how statistical procedures help synthesize the results of an expert panel. Methods: Three statistical techniques were applied to an expert panel that developed explicit criteria to assess the appropriateness of total hip joint replacement: classification tree, regression tree and multiple correspondence analysis combined with automatic classification. Results: Results provided by the three models were shown in graphical displays and were compared to the original panel results using crude and weighted probability of misclassification. Results were also applied to real interventions in order to know the implication of the misclassification on real patients. Conclusions: The statistical techniques help summarize data from panels of experts and provide useful decision models for clinical practice, especially when the number of indications is big. However, degree of misclassification and its implication should be taken into account.

Summary Objectives: Spatial-temporal MR image sequences of the heart contain information about shape and motion changes and pathological structures after myocardial infarction. In this paper a Heart Analysis Tool (HeAT) for the quantitative analysis of 4D MR image sequences of infarct patients is presented. Methods: HeAT supports interactive segmentation of anatomical and pathological structures. Registration of Cine- and DE-MR image data is applied to enable their combined evaluation during the analysis process. Partitioning of the myocardium in segments enables the analysis with high local resolution. Corresponding segments are generated and used for inter/intra-patient comparison. Quantitative parameters were extracted and visualized. Results: Parameters like endocard movement in the infarcted area of six infarct patients were computed in HeAT. Parameters in the infarct area show the expected dysfunctional characteristics. Based on theses parameters passive endocardial movement and myocardial areas with decreased contraction could be identified. Conclusion: In contrast to other software tools HeAT supports the combination of contour information of Cine-MR and DE-MR, local analysis with high resolution and inter/intra patient comparison. HeAT enables an observer-independent evaluation of the complex cardiac image data. Using HeAT in further studies can increase the understanding of left ventricle(LV) remodeling.

Summary Objectives: Menu planning is an important part of per-sonalized lifestyle counseling. The paper describes the results of an automated menu generator (MenuGene) of the web-based lifestyle counseling system Cordelia that provides personalized advice to prevent cardiovascular diseases. Methods: The menu generator uses genetic algorithms to prepare weekly menus for web users. The objectives are derived from personal medical data collected via forms in Cordelia, combined with general nutritional guidelines. The weekly menu is modeled as a multilevel structure. Results: Results show that the genetic algorithm-based method succeeds in planning dietary menus that satisfy strict numerical constraints on every nutritional level (meal, daily basis, weekly basis). The rule-based assessment proved capable of manipulating the mean occurrence of the nutritional components thus providing a method for adjusting the variety and harmony of the menu plans. Conclusions: By splitting the problem into well determined sub-problems, weekly menu plans that satisfy nutritional constraints and have well assorted components can be generated with the same method that is for daily and meal plan generation.

Summary Objectives: In this paper, we propose a novel registration technique, which combines the concepts of landmark and automatic, non-rigid intensity-based approaches. A general framework, which might be used for many different registration problems is presented. The novel approach enables the incorporation of different distance measures as well as different smoothers. Methods: The proposed scheme minimizes a regular-ized distance measure subject to some interpolation constraints. The desired deformation is computed iteratively using an Euler-scheme for the first variation of the chosen objective functional. Results: A fast and robust numerical scheme for the computation of the wanted minimizer is developed, implemented, and applied to various registration tasks. This includes the registration of pre- and post-intervention images of the human eye. Conclusions: A novel framework for a parameter-free, non-rigid registration scheme which allows for the additional incorporation of user-defined landmarks is proposed. It enhances the reliability of conventional approaches considerably and thereby their acceptability by practitioners in a clinical environment.

Knowledge-based systems (KBS) have been proposed to solve a large variety of medical problems. A strategic issue for KBS development and maintenance are the efforts required for both knowledge engineers and domain experts. The proposed solution is building efficient knowledge acquisition (KA) tools. This paper presents a set of KA tools we are developing within a European Project called GAMES II. They have been designed after the formulation of an epistemological model of medical reasoning. The main goal is that of developing a computational framework which allows knowledge engineers and domain experts to interact cooperatively in developing a medical KBS. To this aim, a set of reusable software components is highly recommended. Their design was facilitated by the development of a methodology for KBS construction. It views this process as comprising two activities: the tailoring of the epistemological model to the specific medical task to be executed and the subsequent translation of this model into a computational architecture so that the connections between computational structures and their knowledge level counterparts are maintained. The KA tools we developed are illustrated taking examples from the behavior of a KBS we are building for the management of children with acute myeloid leukemia.