Artificial Intelligence in Medicine

The biomedical community has developed many ontologies in the last years, which may follow a set of community accepted principles for ontology development such as the ones proposed by the OBO Foundry. One of such principles is the orthogonality of biomedical ontologies, which should be based on the reuse of existing content. Previous works have studied how ontology matching techniques help to increase the number of terms reused. In this paper we investigate to what extent the reuse of terms also mean reuse of logical axioms. For this purpose, our method identifies two different ways of reusing terms, reuse of URIs (implicit reuse) and reuse of concepts (explicit reuse). The method is also able of detecting hidden axioms, that is, axioms associated with a reused term but that are not actually reused. We have developed and applied our method to a corpus of 144 OBO Foundry ontologies. The results show that 75 ontologies implicitly reuse terms, 50% of which also explicitly does it. The characterisation based on reuse enables the visualisation of the corpus as a dependency graph that can be clustered for grouping ontologies by their reuse profile. Finally, the application of a locality-based module extractor reveals that roughly 2 000 terms and 20 000 hidden axioms, on average, could be automatically reused.

The emergence of electronic health records has highlighted the need for semantic standards for representation of observations in laboratory medicine. Two such standards are LOINC, with a focus on detailed encoding of lab tests, and SNOMED CT, which is more general, including the representation of qualitative and ordinal test results. In this paper we will discuss how lab observation entries can be represented using SNOMED CT. We use resources provided by the Regenstrief Institute and SNOMED International collaboration, which formalize LOINC terms as SNOMED CT post-coordinated expressions. We demonstrate the benefits brought by SNOMED CT to classify lab tests. We then propose a SNOMED CT based model for lab observation entries aligned with the BioTopLite2 (BTL2) upper level ontology. We provide examples showing how a model designed with no ontological foundation can produce misleading interpretations of inferred observation results. Our solution based on a BTL2 conformant formal interpretation of SNOMED CT concepts allows representing lab test without creating unintended models. We argue in favour of an ontologically explicit bridge between compositional clinical terminologies, in order to safely use their formal representations in intelligent systems.

As the world population is growing older, more and more peoples are facing health issues. For elderly, leaving alone can be tough and risky, typically, a fall can have serious consequences for them. Consequently, smart homes are becoming more and more popular. Such sensors enriched environment can be exploited for health-care applications, in particular Anomaly Detection (AD). Currently, most AD solutions only focus on detecting anomalies in the user daily activities while omitting the ones from the environment itself. For instance the user may have forgotten the pan on the stove while he/she is phoning. In this paper, we present a novel approach for detecting anomaly occurring in the home environment during user activities: CAREDAS. We propose a combination between ontologies and Markov Logic Network to classify the situations to anomaly classes. Our system is implemented, tested and evaluated using real data obtained from the Hadaptic platform. Experimental results prove our approach to be efficient in terms of recognition rate.

Computer-interpretable implementations of clinical guidelines (CIGs) add knowledge that is outside the scope of the original guideline. This knowledge can customize CIGs to patients’ psycho-social context or address comorbidities that are common in the local population, potentially increasing standardization of care and patient compliance. We developed a two-layered contextual decision-model based on the PROforma CIG formalism that separates the primary knowledge of the original guideline from secondary arguments for or against specific recommendations. In this paper we show how constraint logic programming can be used to verify the layered model for two essential properties: (1) secondary arguments do not rule in recommendations that are ruled out in the original guideline, and (2) the CIG is complete in providing recommendation(s) for any combination of patient data items considered. We demonstrate our approach when applied to the asthma domain.

In this paper we show how we used multiple large knowledge sources to construct a much smaller knowledge graph that is focussed on single disease (in our case major depression disorder). Such a disease-centric knowledge-graph makes it more convenient for doctors (in our case psychiatric doctors) to explore the relationship among various knowledge resources and to answer realistic clinical queries.

In pharmacovigilance, reported cases are considered suspected adverse drug reactions (ADR). Health authorities have thus adopted structured causality assessment methods, allowing the evaluation of the likelihood that a medicine was the causal agent of an adverse reaction. The aim of this work was to develop and validate a new causality assessment support system used in a regional pharmacovigilance centre. A Bayesian network was developed, for which the structure was defined by an expert, aiming at implementing the current guidelines for causality assessment, while the parameters were learnt from 593 completely-filled ADR reports evaluated by the Portuguese Northern Pharmacovigilance Centre expert between 2000 and 2012. Precision, recall and time to causality assessment (TTA) was evaluated, according to the WHO causality assessment guidelines, in a retrospective cohort of 466 reports (April to September 2014) and a prospective cohort of 1041 reports (January to December 2015). Results show that the network was able to easily identify the higher levels of causality (recall above 80%), although strugling to assess reports with a lower level of causality. Nonetheless, the median (Q1:Q3) TTA was 4 (2:8) days using the network and 8 (5:14) days using global introspection, meaning the network allowed a faster time to assessment, which has a procedural deadline of 30 days, improving daily activities in the centre.

Event-related potentials (ERPs) have been shown to reflect neurodegenerative processes in Alzheimer’s disease (AD) and might qualify as non-invasive and cost-effective markers to facilitate the objectivization of AD assessment in daily clinical practice. Lately, the combination of multivariate pattern analysis (MVPA) and Gaussian process classification (GPC) has gained interest in the neuroscientific community. Here, we demonstrate how a MVPA-GPC approach can be applied to electrophysiological data. Furthermore, in order to account for the temporal information of ERPs, we develop a novel method that integrates interregional synchrony of ERP time signatures. By using real-life ERP recordings of a prospective AD cohort study (PRODEM), we empirically investigate the usefulness of the proposed framework to build neurophysiological markers for single subject classification tasks. GPC outperforms the probabilistic reference method in both tasks, with the highest AUC overall (0.802) being achieved using the new spatiotemporal method in the prediction of rapid cognitive decline.

The complexity of gene expression regulation relies on the synergic nature underlying the molecular interplay among its principal actors, transcription factors (TFs). Exerting a spatiotemporal control on their target genes, they define transcriptional programs across the genome, which are strongly perturbed in a disease context. In order to gain a more comprehensive picture of these complex dynamics, a data fusion approach, aimed at performing the integration of heterogeneous -omics data is fundamental.Bayesian Networks provide a natural framework for integrating different sources of data and knowledge through the priors’ use. In this work, we developed an hybrid structure-learning algorithm with the aim of exploiting TF ChIP-seq and gene expression (GE) data to investigate disease-specific transcriptional regulations in a genome-wide perspective. TF ChIP seq profiles were firstly used for structure learning and then integrated in the model as a prior probability. GE panels were employed to learn the model parameters, trying to find the best heuristic transcriptional network. We applied our approach to a specific pathological case, the chronic myeloid leukemia (CML), a myeloproliferative disorder, whose transcriptional mechanisms have not yet been deeply elucidated.The proposed data-driven method allows to investigate transcriptional signatures, highlighting in the obtained probabilistic network a three-layered hierarchy, as a different TFs influence on gene expression cellular programs.

Bayesian networks are attractive for developing prognostic models in medicine, due to the possibility for modelling the multivariate relationships between variables that come into play in the care process. In practice, the development of these models is hindered due to the fact that medical data is often censored, in particular the survival time. In this paper, we propose to directly integrate Cox proportional hazards models as part of a Bayesian network. Furthermore, we show how such Bayesian network models can be learned from data, after which these models can be used for probabilistic reasoning about survival. Finally, this method is applied to develop a prognostic model for Glioblastoma Multiforme, a common malignant brain tumour.

Hypertension is the leading cause of cardiovascular-related mortality (CVRM), affecting approximately 1 billion people worldwide. To enable patients at significant risk of CVRM to be treated appropriately, it is essential to correctly diagnose hypertensive patients at an early stage. Our work achieves highly accurate risk scores and classification using 24-h Ambulatory Blood Pressure Monitoring (ABPM) to improve predictions. It involves two stages: (1) time series feature extraction using sliding window clustering techniques and transformations on raw ABPM signals, and (2) incorporation of these features and patient attributes into a probabilistic classifier to predict whether patients will die from cardiovascular-related illness within a median period of 8 years. When applied to a cohort of 5644 hypertensive patients, with 20% held out for testing, a K2 Bayesian network classifier (BNC) achieves 89.67% test accuracy on the final evaluation. We evaluate various BNC approaches with and without ABPM features, concluding that best performance arises from combining APBM features and clinical features in a BNC that represents multiple interactions, learned with some human knowledge in the form of arc constraints.

In this paper, we presented the results of data analysis of 1-year measurements from diabetes patients within the Slovenian healthcare project eCare. We focused on looking for groups/clusters of patients with the similar time profile of the glucose values and describe those patients with their clinical status. We treated in a similar way the WONCA scores (i.e., patients’ functional status). Considering the complexity of the data at hand (time series with a different number of measurements and different time intervals), we used predictive clustering trees with dynamic time warping as the distance between time series. The obtained PCTs identified several groups of patients that exhibit similar behavior. More specifically, we described groups of patients that are able to keep under control their disease, and groups that are less successful in that. Furthermore, we identified and described groups of patients that have similar functional status.

Polysomnograhy is the standard method for objectively measuring sleep, both in patient diagnostics in the sleep laboratory and in clinical research. However, the correspondence between this objective measurement and a person’s subjective assessment of the sleep quality is surprisingly small, if existent. Considering standard sleep characteristics based on the Rechtschaffen and Kales sleep models and the Self-rating Sleep and Awakening Quality scale (SSA), the observed correlations are at most 0.35. An alternative way of sleep modelling - the probabilistic sleep model (PSM) characterises sleep with probability values of standard sleep stages Wake, S1, S2, slow wave sleep (SWS) and REM operating on three second long time segments. We designed sleep features based on the PSM which correspond to the standard sleep characteristics or reflect the dynamical behaviour of probabilistic sleep curves. The main goal of this work is to show whether the continuous sleep representation includes more information about the subjectively experienced quality of sleep than the traditional hypnogram. Using a linear combination of sleep features an improvement in correlation with the subjective sleep quality scores was observed in comparison to the case when a single sleep feature was considered.

Quality of life of patients with Parkinson’s disease degrades significantly with disease progression. This paper presents a step towards personalized medicine management of Parkinson’s disease patients, based on discovering groups of similar patients. Similarity is based on patients’ medical conditions and changes in the prescribed therapy when the medical conditions change. The presented methodology combines multitask learning using predictive clustering trees and short time series analysis to better understand when a change in medications is required. The experiments on PPMI (Parkinson Progression Markers Initiative) data demonstrate that using the proposed methodology we can identify some clinically confirmed patients’ symptoms suggesting medications change.

A clinical decision support system (CDSS) and its components can malfunction due to various reasons. Monitoring the system and detecting its malfunctions can help one to avoid any potential mistakes and associated costs. In this paper, we investigate the problem of detecting changes in the CDSS operation, in particular its monitoring and alerting subsystem, by monitoring its rule firing counts. The detection should be performed online, that is whenever a new datum arrives, we want to have a score indicating how likely there is a change in the system. We develop a new method based on Seasonal-Trend decomposition and likelihood ratio statistics to detect the changes. Experiments on real and simulated data show that our method has a lower delay in detection compared with existing change-point detection methods.

The analysis of surgical motion has received a growing interest with the development of devices allowing their automatic capture. In this context, the use of advanced surgical training systems make an automated assessment of surgical trainee possible. Automatic and quantitative evaluation of surgical skills is a very important step in improving surgical patient care. In this paper, we present a novel approach for the discovery and ranking of discriminative and interpretable patterns of surgical practice from recordings of surgical motions. A pattern is defined as a series of actions or events in the kinematic data that together are distinctive of a specific gesture or skill level. Our approach is based on the discretization of the continuous kinematic data into strings which are then processed to form bags of words. This step allows us to apply discriminative pattern mining technique based on the word occurrence frequency. We show that the patterns identified by the proposed technique can be used to accurately classify individual gestures and skill levels. We also present how the patterns provide a detailed feedback on the trainee skill assessment. Experimental evaluation performed on the publicly available JIGSAWS dataset shows that the proposed approach successfully classifies gestures and skill levels.

Radiological reporting generates a large amount of free-text clinical narrative, a potentially valuable source of information for improving clinical care and supporting research. The use of automatic techniques to analyze such reports is necessary to make their content effectively available to radiologists in an aggregated form. In this paper we focus on the classification of chest computed tomography reports according to a classification schema proposed by radiologists of the Italian hospital ASST Spedali Civili di Brescia. At the time of writing, 346 reports have been annotated by a radiologist. Each report is classified according to the schema developed by radiologists and textual evidences are marked in the report. The annotations are then used to train different machine learning based classifiers. We present in this paper a method based on a cascade of classifiers which make use of a set of syntactic and semantic features. By testing the classifiers in cross-validation on manually annotated reports, we obtained a range of accuracy of 81–96%.

Many medical tasks such as self-diagnosis, health-care assessment, and clinical trial patient recruitment involve the usage of information access tools. A key underlying step to achieve such tasks is the document-to-document matching which mostly fails to bridge the gap identified between raw level representations of information in documents and high-level human interpretation. In this paper, we study how to optimize the document representation by leveraging neural-based approaches to capture latent representations built upon both validated medical concepts specified in an external resource as well as the used words. We experimentally show the effectiveness of our proposed model used as a support of two different medical search tasks, namely health search and clinical search for cohorts.

Substance abuse poses many negative health risks. Tobacco use increases the rates of many diseases such as coronary heart disease and lung cancer. Clinical notes contain rich information detailing the history of substance abuse from caregivers perspective. In this work, we present our work on automatic identification of substance abuse from clinical text. We created a publicly available dataset that has been annotated for three types of substance abuse including tobacco, alcohol, and drug, with 7 entity types per event, including status, type, method, amount, frequency, exposure-history and quit-history. Using a combination of machine learning and natural language processing approaches, our results on an unseen test set range from 0.51–0.58 F1 on stringent, full event, identification, and from 0.80–0.91 F1 for identification of the substance abuse event and status. These results indicate the feasibility of extracting detailed substance abuse information from clinical records.

The increasing population with chronic diseases and highly engaged in online communication has triggered an urge in healthcare to understand this phenomenon. We propose an automatic approach to analyze the perceived intentions behind public tweets. Our long-term goal is to create high-level, behavioral models of the health information consumers and disseminators, relevant to studies in narrative medicine and health information dissemination. The contributions of this paper are: (1) a validated intention taxonomy, derived from pragmatics and empirically adjusted to Twitter public communication; (2) a tagged health-related corpus of 1100 tweets; (3) an effective approach to automatically discover intentions from text, using supervised machine learning with discourse features only, independent of domain vocabulary. Reasoning on the results, we claim the transferability of our solution to other healthcare corpora, enabling thus more extensive studies in the concerned domains.

In clinical practice, physicians often need to take decisions based both on previous experience and medical evidence. Such evidence is usually available in the form of clinical practice guidelines, which elaborate and summarize the knowledge contained in multiple documents. During clinical practice the synthetic format of medical guidelines is an advantage. However, when guidelines are used for educational purposes or when a clinician wants to gain deeper insight into a recommendation, it could be useful to examine all the guideline references relevant to a specific question. In this work we explored IBM Watson services available on the Bluemix cloud to automatically retrieve information from the wide corpus of documents referenced in a recent Italian compendium on emergency neurology. We integrated this functionality in a web application that combines multiple Watson services to index and query the referenced corpus. To evaluate the proposed approach we use the original guideline to check whether the retrieved text matches the actions mentioned in the recommendations.

Medical reports include many occurrences of relevant events in the form of free-text. To make data easily accessible and improve medical decisions, clinical information extraction is crucial. Traditional extraction methods usually rely on the availability of external resources, or require complex annotated corpora and elaborate designed features. Especially for languages other than English, progress has been limited by scarce availability of tools and resources. In this work, we explore recurrent neural network (RNN) architectures for clinical event extraction from Italian medical reports. The proposed model includes an embedding layer and an RNN layer. To find the best configuration for event extraction, we explored different RNN architectures, including Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU). We also tried feeding morpho-syntactic information into the network. The best result was obtained by using the GRU network with additional morpho-syntactic inputs.

Clinical trials are fundamental for evaluating therapies and diagnosis techniques. Yet, recruitment of patients remains a real challenge. Eligibility criteria are related to terms but also to patient laboratory results usually expressed with numerical values. Both types of information are important for patient selection. We propose to address the processing of numerical values. A set of sentences extracted from clinical trials are manually annotated by four annotators. Four categories are distinguished: C (concept), V (numerical value), U (unit), O (out position). According to the pairs of annotators, the inter-annotator agreement on the whole annotation sequence CVU goes up to 0.78 and 0.83. Then, an automatic method using CFRs is exploited for creating a supervised model for the recognition of these categories. The obtained F-measure is 0.60 for C, 0.82 for V, and 0.76 for U.

Depression has been consistently linked with alterations in speech motor control characterised by changes in formant dynamics. However, potential differences in the manifestation of depression between male and female speech have not been fully realised or explored. This paper considers speech-based depression classification using gender dependant features and classifiers. Presented key observations reveal gender differences in the effect of depression on vowel-level formant features. Considering this observation, we also show that a small set of hand-crafted gender dependent formant features can outperform acoustic-only based features (on two state-of-the-art acoustic features sets) when performing two-class (depressed and non-depressed) classification.

The generation of medical terminologies is an important activity. A flexible and structured terminology both helps professionals in everyday manual classification of clinical texts and is crucial to build knowledge bases for encoding tools implementing software to support medical tasks. For these reasons, it would be nice to “enforce” medical dictionaries such as MedDRA with sets of locutions semantically related to official terms. Unfortunately, the manual generation of medical terminologies is time consuming. Even if the human validation is an irreplaceable step, a significative set of “high-quality” candidate terminologies can be automatically generated from clinical documents by statistical methods for linguistic. In this paper we adapt and use a co-occurrence based technique to generate new MedDRA locutions, starting from some large sets of narrative documents about adverse drug reactions. We describe here the methodology we designed and results of some first experiments.

Duration constraints are among the most subtle and important aspects of clinical practice. When designing healthcare processes, it is important to incorporate such constraints into process diagrams and to provide suitable mechanisms for managing their violations during process run-time. Nonetheless, the Business Process Model and Notation (BPMN 2.0) fails in providing proper support for duration constraint modeling. In this paper, we propose a set of BPMN patterns to foster the modeling and management of shifted duration, that constitutes an unexplored kind of duration constraints, recurrent in medical practice. Specifically, this constraint is suitable for dealing with the real-time adjustments required by pharmacological therapies and monitoring tasks.

Pharmaco-epidemiology (PE) is the study of uses and effects of drugs in well defined populations. As medico-administrative databases cover a large part of the population, they have become very interesting to carry PE studies. Such databases provide longitudinal care pathways in real condition containing timestamped care events, especially drug deliveries. Temporal pattern mining becomes a strategic choice to gain valuable insights about drug uses. In this paper we propose DCM, a new discriminant temporal pattern mining algorithm. It extracts chronicle patterns that occur more in a studied population than in a control population. We present results on the identification of possible associations between hospitalizations for seizure and anti-epileptic drug switches in care pathway of epileptic patients.

Evidence is lacking for patient-reported effectiveness of treatments for most medical conditions and specifically for lower back pain. In this paper, we examined a consumer-based social network that collects patients’ treatment ratings as a potential source of evidence. Acknowledging the potential biases of this data set, we used propensity score matching and generalized linear regression to account for confounding variables. To evaluate validity, we compared results obtained by analyzing the patient reported data to results of evidence-based studies. Overall, there was agreement on the relationship between back pain and being obese. In addition, there was agreement about which treatments were effective or had no benefit. The patients’ ratings also point to new evidence that postural modification treatment is effective and that surgery is harmful to a large proportion of patients.

In this paper, we present a novel process mining approach, specifically tailored to medical applications, which allows the user to build an initial process model from the hospital event log, and then supports further model refinements, by directly exploiting her knowledge-based model evaluation. In such a way, it supports the interactive construction of the process model at multiple and user-defined levels of abstraction, ranging from a model which perfectly adheres to the input traces (i.e., all of its paths correspond to at least one trace in the log) to models which increasingly loose precision, but gain generality. Our results in the field of stroke management, reported as a case study in this paper, show that our approach can provide relevant advantages with respect to traditional process mining techniques.

Sequential pattern mining algorithms are widely used to explore care pathways database, but they generate a deluge of patterns, mostly redundant or useless. Clinicians need tools to express complex mining queries in order to generate less but more significant patterns. These algorithms are not versatile enough to answer complex clinician queries. This article proposes to apply a declarative pattern mining approach based on Answer Set Programming paradigm. It is exemplified by a pharmaco-epidemiological study investigating the possible association between hospitalization for seizure and antiepileptic drug switch from a french medico-administrative database.

Many hospital information systems nowadays record data about the executed medical process instances in the form of traces in an event log. In this paper we present a framework able to convert actions found in the traces into higher level concepts, on the basis of domain knowledge. Abstracted traces are then provided as an input to semantic process mining. The approach has been tested in stroke care, where we show how the abstraction mechanism allows the user to mine process models that are easier to interpret, since unnecessary details are hidden, but key behaviors are clearly visible.

The problem of incompatible pairs in living-donor kidney transplant can be solved using paired kidney exchange, i.e., two incompatible patient-donor pairs interchange donors, creating a cycle, which can be extended to three or even more pairs. Finding a set of cycles that maximizes the number of successful transplants is a complex task.The Organización Nacional de Trasplantes (ONT) is responsible for donation and transplantation processes in Spain. In this paper we compare the current ONT heuristic finding-cycles procedure with an integer programming approach by means of a true-data-based empirical simulation. The obtained results show that, although the two methods provide quite different solutions, they both exhibit weak and strong points.

Comparing care trajectories helps improve health services. Medico-administrative databases are useful for automatically reconstructing the patients’ history of care. Care trajectories can be compared by determining their overlapping parts. This comparison relies on both semantically-rich representation formalism for care trajectories and an adequate similarity measure. The longest common subsequence (LCS) approach could have been appropriate if representing complex care trajectories as simple sequences was expressive enough. Furthermore, by failing to take into account similarities between different but semantically close medical events, the LCS overestimates differences. We propose a generalization of the LCS to a more expressive representation of care trajectories as sequences of sets. A set represents a medical episode composed by one or several medical events, such as diagnosis, drug prescription or medical procedures. Moreover, we propose to take events’ semantic similarity into account for comparing medical episodes. To assess our approach, we applied the method on a care trajectories’ sample from patients who underwent a surgical act among three kinds of acts. The formalism reduced calculation time, and introducing semantic similarity made the three groups more homogeneous.

Dealing with missing data is a crucial step in the preprocessing stage of most data mining projects. Especially in healthcare contexts, addressing this issue is fundamental, since it may result in keeping or loosing critical patient information that can help physicians in their daily clinical practice. Over the years, many researchers have addressed this problem, basing their approach on the implementation of a set of imputation techniques and evaluating their performance in classification tasks. These classic approaches, however, do not consider some intrinsic data information that could be related to the performance of those algorithms, such as features’ distribution. Establishing a correspondence between data distribution and the most proper imputation method avoids the need of repeatedly testing a large set of methods, since it provides a heuristic on the best choice for each feature in the study. The goal of this work is to understand the relationship between data distribution and the performance of well-known imputation techniques, such as Mean, Decision Trees, k-Nearest Neighbours, Self-Organizing Maps and Support Vector Machines imputation. Several publicly available datasets, all complete, were selected attending to several characteristics such as number of distributions, features and instances. Missing values were artificially generated at different percentages and the imputation methods were evaluated in terms of Predictive and Distributional Accuracy. Our findings show that there is a relationship between features’ distribution and algorithms’ performance, although some factors must be taken into account, such as the number of features per distribution and the missing rate at state.

Monitoring mental fatigue is of increasing importance for improving cognitive performance and health outcomes. Previous models using eye-tracking data allow inference of fatigue in cognitive tasks, such as driving, but they require us to engage in a specific cognitive task. A model capable of estimating fatigue from eye-tracking data in natural-viewing situations when an individual is not performing cognitive tasks has many potential applications. Here, we collected eye-tracking data from 18 adults as they watched video clips (simulating the situation of watching TV programs) before and after performing cognitive tasks. Using this data, we built a fatigue-detection model including novel feature sets and an automated feature selection method. With eye-tracking data of individuals watching only 30-seconds worth of video, our model could determine whether that person was fatigued with 91.0% accuracy in 10-fold cross-validation (chance 50%). Through a comparison with a model incorporating the feature sets used in previous studies, we showed that our model improved the detection accuracy by up to 13.9% (from 77.1 to 91.0%).

Optical Coherence Tomography (OCT) is, nowadays, one of the most referred ophthalmological imaging techniques. OCT imaging offers a window to the eye fundus in a non-invasive way, permitting the inspection of the retinal layers in a cross sectional visualization. For that reason, OCT images are frequently used in the analysis of relevant diseases such as hypertension or diabetes. Among other pathological structures, a correct identification of cystoid regions is a crucial task to achieve an adequate clinical analysis and characterization, as in the case of the analysis of the exudative macular disease.This paper proposes a new methodology for the automatic identification of intraretinal cystoid fluid regions in OCT images. Firstly, the method identifies the Inner Limitant Membrane (ILM) and Retinal Pigment Epithelium (RPE) layers that delimit the region of interest where the intraretinal cystoid regions are placed. Inside these limits, the method analyzes windows of a given size and determine the hypothetical presence of cysts. For that purpose, a large and heterogeneous set of features were defined to characterize the analyzed regions including intensity and texture-based features. These features serve as input for representative classifiers that were included in the analysis.The proposed methodology was tested using a set of 50 OCT images. 502 and 539 samples of regions with and without the presence of cysts were selected from the images, respectively. The best results were provided by the LDC classifier that, using a window size of $$61 \times 61$$ and 40 features, achieved satisfactory results with an accuracy of 0.9461.

When diagnosing patients suffering from dementia based on imaging data like PET scans, the identification of suitable predictive regions of interest (ROIs) is of great importance. We present a case study of 3-D Convolutional Neural Networks (CNNs) for the detection of ROIs in this context, just using voxel data, without any knowledge given a priori. Our results on data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) suggest that the predictive performance of the method is on par with that of state-of-the-art methods, with the additional benefit of potential insights into affected brain regions.

Dermoscopic images are useful tools towards the diagnosis and classification of skin lesions. One of the first steps to automatically study them is the reduction of noise, which includes bubbles caused by the immersion fluid and skin hair. In this work we provide an effective hair removal algorithm for dermoscopic imagery employing soft color morphology operators able to cope with color images. Our hair removal filter is essentially composed of a morphological curvilinear object detector and a morphological-based inpainting algorithm. Our work is aimed at fulfilling two goals. First, to provide a successful yet efficient hair removal algorithm using the soft color morphology operators. Second, to compare it with other state-of-the-art algorithms and exhibit the good results of our approach, which maintains lesion’s features.

We propose a model for the incorporation of risk in association rule application. We validate this model using data gathered in a randomized controlled trial from a recommender system for medication reviews in primary care. The model’s outcomes are found to have predictive value when tested against decisions made by physicians on 261 patients’ health records.

Parkinson’s, a progressive neural disorder, is difficult to identify due to the hidden nature of the symptoms associated. We present a machine learning approach that uses a definite set of features obtained from the Parkinson’s Progression Markers Initiative (PPMI) study as input and classifies them into one of two classes: PD (Parkinson’s disease) and HC (Healthy Control). As far as we know this is the first work in applying machine learning algorithms for classifying patients with Parkinson’s disease with the involvement of domain expert during the feature selection process. We evaluate our approach on 1194 patients acquired from Parkinson’s Progression Markers Initiative and show that it achieves a state-of-the-art performance with minimal feature engineering.

Myoelectric signals (EMG) provide an intuitive and rapid interface for controlling technical devices, in particular bionic arm prostheses. However, inferring the intended movement from a surface EMG recording is a non-trivial pattern recognition task, especially if the data stems from low-cost sensors. At the same time, overly complex models are prohibited by strict speed, data parsimony and robustness requirements. As a compromise between high accuracy and strict requirements we propose to apply Echo State Networks (ESNs), which extend standard linear regression with (1) a memory and (2) nonlinearity. Results show that both features, memory and nonlinearity, independently as well as in conjunction, improve the prediction accuracy on simultaneous movements in two degrees of freedom (hand opening/closing and pronation/supination) recorded from four able-bodied participants using a low-cost 8-electrode-array. However, it was also shown that the model is still not sufficiently resistant to external disturbances such as electrode shift.

Terminologies or ontologies to describe patient-reported information are lacking. The development and maintenance of ontologies is usually a manual, lengthy, and resource-intensive process. To support the development of medical specialty-specific ontologies, we created a semi-automated ontology development and management system (SEAM). SEAM supports ontology development by automatically extracting terms, concepts, and relations from narrative text, and then offering a streamlined graphical user interface to edit and create content in the ontology and finally export it in OWL format. The graphical user interface implements card sorting for synonym grouping and concept laddering for hierarchy construction. We used SEAM to create ontologies to support medically unexplained syndromes detection and management among veterans in the U.S.

Process Mining is an emerging discipline investigating tasks related with the automated identification of process models, given real-world data (Process Discovery). The analysis of such models can provide useful insights to domain experts. In addition, models of processes can be used to test if a given process complies (Conformance Checking) with specifications. For these capabilities, Process Mining is gaining importance and attention in healthcare.In this paper we introduce pMineR, an R library specifically designed for performing Process Mining in the medical domain, and supporting human experts by presenting processes in a human-readable way.

In spite the important advantages of influence diagrams over decision trees, including the possibility of solving much more complex problems, the medical literature still contains around 10 decision trees for each influence diagram. In this paper we analyse the reasons for the low acceptance of influence diagrams in health decision analysis, in contrast with its success in artificial intelligence. One of the reasons is the difficulty of representing asymmetric problems. Another one was the lack of algorithms for explaining the reasoning and performing cost-effectiveness analysis, as well as the scarcity of user-friendly software tools for sensitivity analysis. In this paper we review the research conducted by our group in the last 25 years, crystallised in the open-source software tool OpenMarkov, explaining how it has tried to address those challenges.

Quality-adjusted life years (QALYs) are a popular measure employed in cost-utility analysis (CUA) for informing decisions about competing healthcare programs applicable to a target population.CUA is often performed using decision trees (DTs), i.e. probabilistic models that allow calculating the outcome related to different decision options (e.g., two different therapeutic strategies) considering all their expected effects. DTs may in fact include a measure of the quality of life, namely a utility coefficient (UC), for every health state patients might experience as a result of the healthcare interventions. Eliciting reliable UCs from patients poses several challenges, and it is not a common procedure in clinical practice.We recently developed UceWeb, a tool that supports users in that elicitation process. In this paper we describe the public repository where UceWeb collects the elicited UCs, and how this repository can be exploited by researchers interested in performing DT-based CUAs on a specific population. To this aim, we also describe the UceWeb integration with a commercial software for DTs management, which allows to automatically run the models quantified with the mean value of the target population UCs.