Deep learning has been shown repeatedly to be a successful method of obtaining accurate classifiers. This also applies to orchid identification from digital photographs. However, deep neural networks possess the major weakness of lack of explainability, missing the ability to explain the reasons behind a decision. Nevertheless, most current research regarding automated orchid identification applies this blackbox approach. By contrast, in this paper we propose a new method for trustworthy automated orchid identification combining two complementary methods: deep neural networks and feature-based Bayesian networks, where the Bayesian network is also utilized for providing an explanation of the generated solutions. We use other deep neural networks to extract flower characteristics, the features, from the images which are subsequently fed into the Bayesian network as uncertain evidence. When combining the deep neural network and the Bayesian network as an ensemble classifier, both reaching the same conclusion, an accuracy of 89.4% is achieved, the most trustworthy outcome. With a human-in-the-loop ensemble classifier, validation results are even better, yielding an accuracy of 98.1%. Our approach also exploits the taxonomic knowledge represented in the Bayesian network to provide an explanation of the solutions for every case, reinforcing further trust in the method. The result is an explainable user-in-the-loop ensemble classifier. Providing explainability can help build user trust in a system and may play a major role when it is used as a learning aid for new orchid enthusiasts. Finally, the proposed method may be also of value in many fields other than plant determination.

OBJECTIVES: Establishing causal dependencies is crucial in applied domains, such as medicine and healthcare, where decision-making must be explainable. In these settings, small sample sizes and missing data call for federated approaches to maximise the amount of information we can use. METHODS: We propose a novel federated causal discovery algorithm capable of pooling information from multiple sources with heterogeneous missing data to learn a graph representing cause-effect relationships. In particular, we learn a causal graph on a centralised server while taking into account both prior knowledge and missingness mechanism specific to each client. RESULTS: We applied the proposed algorithm to synthetic data and real-world data from a multicentric study on endometrial cancer, validating the obtained causal graph through quantitative analyses and a clinical literature review. CONCLUSION: Our approach learns an accurate model despite data missing not-at-random.

Abstract Automated image-based plant identification systems are black-boxes, failing to provide an explanation of a classification. Such explanations are seen as being essential by taxonomists and are part of the traditional procedure of plant identification. In this paper, we propose a different method by extracting explicit features from flower images that can be employed to generate explanations. We take the benefit of feature extraction derived from the taxonomic characteristics of plants, with the orchids as an example domain. Feature classifiers were developed using deep neural networks. Two different methods were studied: (1) a separate deep neural network was trained for every individual feature, and (2) a single, multi-label, deep neural network was trained, combining all features. The feature classifiers were tested in predicting 63 orchid species using naive Bayes (NB) and tree-augmented Bayesian networks (TAN). The results show that the accuracy of the feature classifiers is in the range 83-93%. By combining these features using NB and TAN the species can be predicted with an accuracy of 88.9%, which is better than a standard pre-trained deep neural-network architecture, but inferior to a deep learning architecture after fine-tuning of multiple layers. The proposed novel feature extraction method still performs well for identification and is explainable, as opposed to black-box solutions that only aim for the best performance. Graphical abstract

The development and validation of Clinical Decision Support Models (CDSM) based on Bayesian networks (BN) is commonly done in a collaborative work between medical researchers providing the domain expertise and computer scientists developing the decision support model. Although modern tools provide facilities for data-driven model generation, domain experts are required to validate the accuracy of the learned model and to provide expert knowledge for fine-tuning it while computer scientists are needed to integrate this knowledge in the learned model (hybrid modeling approach). This generally time-expensive procedure hampers CDSM generation and updating. To address this problem, we developed a novel interactive visual approach allowing medical researchers with less knowledge in CDSM to develop and validate BNs based on domain specific data mainly independently and thus, diminishing the need for an additional computer scientist. In this context, we abstracted and simplified the common workflow in BN development as well as adjusted the workflow to medical experts' needs. We demonstrate our visual approach with data of endometrial cancer patients and evaluated it with six medical researchers who are domain experts in the gynecological field.

Bayesian networks (BN) are a popular representation for reasoning under uncertainty. The analysis of many real-world use cases, that in principle can be modeled by BNs, suffers however from the computational complexity of inference. Inference methods based on Weighted Model Counting (WMC) reduce the cost of inference by exploiting patterns exhibited by the probabilities associated with BN nodes. However, these methods require a computationally intensive compilation step in search of these patterns, which effectively prohibits the handling of larger BNs. In this paper, we propose a solution to this problem by extending WMC methods with a framework called Compositional Weighted Model Counting (CWMC). CWMC reduces compilation cost by partitioning a BN into a set of subproblems, thereby scaling the application of state-of-the-art innovations in WMC to scenarios where inference cost could previously not be amortized over compilation cost. The framework supports various target representations that are less or equally succinct as decision-DNNF. At the same time, its inference time complexity O(nexp⁡(w)), where n is the number of variables and w is the tree-width, is comparable to mainstream algorithms based on variable elimination, clustering and conditioning.

In consistency-based diagnosis (CBD), abnormal behavior is sorted out based on de- viation from a normal behavior specification. Probabilities have been added to CBD for quantifying uncertainty on, e.g., the behavior of faulty components. While resulting in more complete models, the requirement of such uncertainty parameters goes in opposition to the original CBD motivation. The conflict measure stands closer to CBD by comput- ing solutions without the need of priors on candidates, however, its results might not be suitable when only partial observations are available. In this paper, we propose a method called the diagnostic coefficient, which better solves the partial observability case, while needing the same parameters as the conflict measure. The diagnostic coefficient is based on the idea that observations are conflicting if the observed outputs are discrepant with respect to alternative outputs that could have been observed. We report experiments with logical circuits where the diagnostic coefficient shows promising results compared to the conflict measure under various settings with missing observations.

Learning Bayesian network structures from data is known to be hard, mainly because the number of candidate graphs is super-exponential in the number of variables. Furthermore, using observational data alone, the true causal graph is not discernible from other graphs that model the same set of conditional independencies. In this paper, it is investigated whether Bayesian network structure learning can be improved by exploiting the opinions of multiple domain experts regarding cause-effect relationships. In practice, experts have different individual probabilities of correctly labeling the inclusion or exclusion of edges in the structure. The accuracy of each expert is modeled by three parameters. Two new scoring functions are introduced that score each candidate graph based on the data and experts' opinions, taking into account their accuracy parameters. In the first scoring function, the experts' accuracies are estimated using an expectation-maximization-based algorithm and the estimated accuracies are explicitly used in the scoring process. The second function marginalizes out the accuracy parameters to obtain more robust scores when it is not possible to obtain a good estimate of experts' accuracies. The experimental results on simulated and real world datasets show that exploiting experts' knowledge can improve the structure learning if we take the experts' accuracies into account.

The theory of causal independence is frequently used to facilitate the assessment of the probabilistic parameters of discrete probability distributions of complex Bayesian networks. Although it is possible to include continuous parameters in Bayesian networks as well, such parameters could not, so far, be modeled by means of causal-independence theory, as a theory of continuous causal independence was not available. In this paper, such a theory is developed and generalized such that it allows merging continuous with discrete parameters based on the characteristics of the problem at hand. This new theory is based on the discovered relationship between the theory of causal independence and convolution in probability theory, discussed in detail for the first time in this paper. Furthermore, the new theory is used as a basis to develop a relational theory of probabilistic interactions. It is also illustrated how this new theory can be used in connection with special probability distributions.

Approaches for extending logic to deal with uncertainty immanent to many real-world problems are often on the one side purely qualitative, such as modal logics, or on the other side quantitative, such as probabilistic logics. Research on combinations of qualitative and quantitative extensions to logic which put qualitative constraints on probability distributions, has mainly remained theoretical until now. In this paper, we propose a practically useful logic, which supports qualitative as well as quantitative uncertainty and can be extended with modalities with varying level of quantitative precision. This language has a solid semantic foundation based on imprecise probability theory. While in general imprecise probabilistic inference is much harder than the precise case, this is the first expressive imprecise probabilistic formalism for which probabilistic inference is shown to be as hard as corresponding precise probabilistic problems. A second contribution of this paper is an inference algorithm for this language based on the translation to a weighted model counting (WMC) problem, an approach also taken by state-of-the-art probabilistic inference methods for precise problems.

MoSHCA is a mHealth project designed to improve patient-doctor interaction and to promote the self-management of chronic diseases by the patients themselves. The number of people with a chronic disease is dramatically increasing worldwide. This is becoming a major obstacle for economic stability and growth and the sustainability of national health care systems. The introduction of self-management by patients with a chronic disease seems inevitable as a countermeasure against these developments. MoSHCA provides intelligent, user-friendly and secure, medical and well-being decision support through embedded software in mobile devices by utilizing specific sensors and data from customized information systems.

Performing safety and security tasks requires the continuous gathering and interpretation of information about objects to detect and predict events of interest. Especially, reasoning about objects' identities and intentions is crucial, but requires making use of heterogeneous information inherent with uncertainty. This makes such tasks very challenging for human operators and decision support systems are clearly needed. In this paper, we present a generic probabilistic logic model for the systematic representation and uncertainty reasoning about the identity and intentions of monitored objects. We apply the model to the area of maritime safety and security after incorporating domain specific knowledge. Experiments with simulated and real-world vessel data demonstrate the model's capabilities to draw conclusions about objects using uncertain information. The first-order probabilistic logic we use shows to be a very powerful tool to deal with dynamic amounts of information and objects. To our knowledge it is one of the few real-world applications of probabilistic logic.

Applying artificial intelligence techniques to management of chronic diseases - smart monitoring - has great potential to improve chronic disease care. Probabilistic models offer powerful methods for automatic data interpretation, and thus play a potentially large role in mobile, personalised care. In particular in the context of disease monitoring one needs clinical time-series data that include data of multiple patient parameters, to allow building such models. However, in practice clinical time-series data of patients with chronic disease are only limited available, and when they are available usually only of a few patients. In this paper, we explore different ways to build predictive models for the detection of COPD exacerbations and related hospitalisation, focusing on the temporal aspect of monitoring data while taking into account data sparsity. Preliminary results indicate that even with the limited data available some predictions can be made about hospitalisation.

In this paper we present innovative research for the automatic interpretation of biochemical test strip color by a smartphone using image processing techniques. Urinalysis is the current application for these techniques. Our mobile application captures images of the color pads on strips using the camera phone, then analyzes automatically the images within the device itself and compares these against reference color pads to obtain a final classification. As a test scenario we focus on the detection of proteinuria, i.e., leakage of protein into the urine, for the detection of which strips for protein and creatinine are used. We performed an initial laboratory evaluation using specially prepared concentrations to check the accuracy and precision of detection. The results obtained are encouraging and show that the proposed technique has a good potential for the development of cheap, mobile and smart home-based readers for the early detection of health problems.

Control engineering is a field of major industrial importance as it offers principles for engineering controllable physical devices, such as cell phones, television sets, and printing systems. Control engineering techniques assume that a physical system's dynamic behaviour can be completely described by means of a set of equations. However, as modern systems are often of high complexity, drafting such equations has become more and more difficult. Moreover, to dynamically adapt the system's behaviour to a changing environment, observations obtained from sensors at runtime need to be taken into account. However, such observations give an incomplete picture of the system's behaviour; when combined with the incompletely understood complexity of the device, control engineering solutions increasingly fall short. Probabilistic reasoning would allow one to deal with these sources of incompleteness, yet in the area of control engineering such AI solutions are rare. When using a Bayesian network in this context the required model can be learnt, and tuned, from data, uncertainty can be handled, and the model can be subsequently used for stochastic control of the system's behaviour. In this paper we discuss industrial research in which Bayesian networks were successfully used to control complex printing systems.

Computer-aided detection (CAD) is used in medical science as a means of supporting a doctor's observations and interpretations. While X-ray imaging techniques, such as mammography, yield a great deal of information, it is not always easy to evaluate detected mammographic regions as being suspicious for cancer, which results in a number of cancers to be misinterpreted or missed in an image. In this sense, CAD systems have as aim the increase of detection rates when analysing mammograms, by identifying features that are characteristic for breast cancer. In this research we aim at using the features extracted from mammographic images in order to analyse the development of suspicious lesions. Differently from other breast cancer models, the data modelling exploits object orientation. This allows not only for a natural description of domain entities and their intrinsic relations, but also the application of relational learning techniques, which handles our heterogeneous data instances both in terms of learning and inference.

Conflict-based diagnosis is a recently proposed probabilistic method for model-based diagnosis, inspired by consistencybased diagnosis, that uses a measure of data conflict, called the diagnostic conflict measure, to rank diagnoses. In this paper, this method is refined using an abductive method that reuses part of the computation of the diagnostic conflict measure.

The goal of breast cancer screening programs is to detect cancers at an early (preclinical) stage, by using periodic mammographic examinations in asymptomatic women. In evaluating cases, mammographers insist on reading multiple images (at least two) of each breast as a cancerous lesion tends to be observed in different breast projections (views). Most computer-aided detection (CAD) systems, on the other hand, only analyze single views independently, and thus fail to account for the interaction between the views. In this paper, we propose a Bayesian framework for exploiting multi-view dependencies between the suspected regions detected by a single-view CAD system. The results from experiments with real-life data show that our approach outperforms the singleview CAD system in distinguishing between normal and abnormal cases. Such a system can support screening radiologists to improve the evaluation of breast cancer cases.

Abstract Requirements about the quality of clinical guidelines can be represented by schemata borrowed from the theory of abductive diagnosis, using temporal logic to model the time-oriented aspects expressed in a guideline. Previously, we have shown that these requirements can be verified using interactive theorem proving techniques. In this paper, we investigate how this approach can be mapped to the facilities of a resolution-based theorem prover,otterand a complementary program that searches for finite models of first-order statements,mace-2. It is shown that the reasoning required for checking the quality of a guideline can be mapped to such a fully automated theorem-proving facilities. The medical quality of an actual guideline concerning diabetes mellitus 2 is investigated in this way.

The application of a medical guideline to the treatment of a patient's disease can be seen as the execution of tasks, sequentially or in parallel, in the face of patient data. It has been shown that many of such guidelines can be represented as a "network of tasks," that is, as a sequence of steps that have a specific function or goal. In this paper, a novel methodology for verifying the quality of such guidelines is introduced. To investigate the quality of such guidelines, we propose to include medical background knowledge to task networks and to formalize criteria for good medical practice that a guideline should comply with. This framework was successfully applied to a guideline dealing with the management of diabetes mellitus type 2 by using KIV.