Infant and youth mortality has seen a steady decline over the years. However, many issues related to sociodemographic factors still persist. In Portugal, while mortality forecasts are regularly disclosed to the general public by specialised public entities, very few studies have focused on its determinants, and none have taken advantage of the modelling capabilities of Machine Learning (ML) techniques. This work makes use of real-world data in order to identify the main determinants of infant and youth mortality in Portugal using some of these techniques. The data used for this study comprised 178 databases from various authorities, encompassing economic, demographic, environmental, health, education, and mortality variables at the municipal level. No data at the individual level was available. Two different approaches were proposed. For the first one, the problem was framed as a regression problem, with the number of deaths as the target variable and the potential determinants as the predictors. Simple regression models were used, mainly due to their interpretability. A neural network was also employed to enable a comparison between linear and nonlinear models. Feature elimination and feature selection methods were devised in order to ascertain which variables were the most relevant. These include a feature selection method specifically custom for the problem at hand which proved particularly effective, as it led to performance improvements for every model used in this work. The second approach made use of the K-means clustering technique to determine which of the previously selected variables led to better clusters with both the number of deaths and the mortality rate. To this end, the silhouette method was chosen as the evaluation metric. The best regression model achieved an R2of 0.846. The foreign population with legal status of residence in the parents’ place of residence and the average monthly earnings of employees were shown to be the features with the greatest impact on mortality.

Dynamic scheduling has received widespread attention from academia and industry due to the increasing complexity in manufacturing systems. Highly dynamic and adaptable behaviours are necessary for an improved production efficiency in unstable and constantly changing environments. This paper proposes an agent-based hybrid tabu-search heuristic (AB-TSH) to solve dynamic flexible job-shop scheduling problems. The solution is fully implemented and tested in an industrial environment for seven distinct dynamic scenarios derived from a static scenario using the benchmark of AIP-PRIMECA Flexible Manufacturing System. The scheduling plan is obtained by exploitation using a greedy heuristic on tabu search solution points. The hybrid tabu-search heuristic is supported by a multi-agent system that react and re-optimize the scheduling plan in case of disturbances and unpredicted events. The proposed solution demonstrated superior performance in terms of makespan in the majority of dynamic scenarios tested when compared to previous studies in the literature. This improved performance is attributed to the solution’s ability to combine the scheduling plan both statically and dynamically.

The First-Order Autonomous Learning Multi-Model (ALMMo) system was initially introduced as a regressor which could be easily adapted to a binary classifier. In this paper, an extension of the ALMMo algorithm is proposed, enabling it to tackle multi-class classification tasks, without escalating the computational demand substantially. Thus, this paper highlights the flexibility of the method by increasing its range of capabilities. The proposed extension is tested in 3 benchmark datasets, and the obtained results are presented as a proof of the concept. Furthermore, these results are compared to 2 benchmark methods, those being shallow neural-networks and support vector machines; as well as to the ALMMo-0 classifier.

The need to ensure that all funding projects made available by the European Union are fully exploited by each country leads to a robust investigation of the main difficulties of applying to these projects. In this context, Portugal has verified that these funds are not being fully used and detected that one of the main setbacks in the European Union funding application system is the identification of the best call for application. Therefore, this work aims at developing a funds assistant framework to suggest the most suitable set of available calls for application based on the written description of the beneficiaries' projects, via the use of Natural Language Processing and text classification algorithms. Moreover, this problem is addressed as a hierarchical multi-class text classification, taking advantage of the hierarchical structure inside the European funds and aiming to improve the overall performance of the proposed framework. In this context, four models are chosen to compare the classification performance: Naive Bayes, Support Vector Machines, Random Forest and k-Nearest Neighbors. Simulation results highlight that the Support Vector Machine outperforms the rest of the algorithms for almost all classification scenarios. In order to enhance the results, the models are also calibrated to return the second prediction when the likelihood of the first prediction is small, achieving even higher performances for all levels of the hierarchy.

Since most branches of the distribution grid may already be close to their maximum capacity, smart management when charging electric vehicles (EVs) is becoming more and more crucial. In fact, office buildings might not be able to handle several transactions at the same time, especially considering the next generation of fast chargers which are very power expensive. Thus, an efficient charging policy needs to be found. This paper proposes the scheduling of real-time EVs charging through deep reinforcement learning (DRL) techniques. DRL has been chosen because it can adaptively learn from interacting with the surrounding environment. The focus of the optimization is to ensure the completion of the charging transactions in a timely manner, while shifting the load from the times of peak demand. The novelty of the proposed approach lies in its innovative framework: pools of electric vehicles with different characteristics are categorized using a clustering algorithm, a tree-based classifier has been developed to sort new instances of EVs, and a multilayer perceptron artificial deep neural network has been trained to predict the expected duration of each charging session. These features are used as inputs to the DRL agent, and are mapped into actions that adjust the maximum power associated to each charging station. The model has been compared to a traditional charging algorithm and increasingly challenging scenarios have been considered. Results have shown that the developed algorithm fails less than the baseline, with a reduction of the load due to EVs charging of 80% during peak times.

In this paper, a new approach to the usage of 0-order Autonomous Learning Multi-Model (ALMMo-0) classifiers is proposed. ALMMo-0 classifiers are fully automatic and do not rely on any hyper-parameters. The creation of clouds relies on normalizing data points by their norm, which may remove an important degree of freedom from the data itself. The proposed approach consists of adding the initial radius of the clouds as an hyper-parameter, which makes it possible to skip the normalization step. This approach requires the search for the ideal value of the hyper-parameter. This way, upon training a set of models with different values for the initial radius, the user is expected to be able to choose from several models which range from more accurate to less complex. This approach was tested on three benchmark problems and compared to the results obtained using the original approach. Furthermore, this approach was also tested on a real dataset (Acute Kidney Injury). The obtained results enhance the versatility provided by the proposed method, successfully allowing the user to choose the model that fits better the design demands regarding accuracy, training time, and complexity.

This paper proposes a flexible manufacturing system based on intelligent computational agents. A Multi-Agent System composed of 4 types of reactive agents was designed to control the operation of a real implementation in the Intelligent Automation Lab at Instituto Superior Técnico. This implementation was based and constructed analogously to a known benchmark, AIP-PRIMECA. The agents were modelled using Petri nets and agent communications were defined through the combination of FIPA Interaction Protocols. The system was tested under the conditions of static and dynamic scenarios, having its performance validated whenever possible by comparison with results from a Potential Fields Approach in the same benchmark. Overall, the performance exhibited by the proposed MAS was slightly better and it is worth highlighting the simple behaviour of each agent and ability to respond in real-time to all the dynamic scenarios tested.

Insulin therapy is crucial to control blood sugar levels for critical-care patients. However, there is no consensus on which is the most beneficial glucose control, either intensive or conventional, for these patients to reduce mortality. Through the use of electronic health records and computational intelligent systems, this work intends to increase the understanding of mechanisms that influence survival of adult patients (≥ 16 years old) within the first 24 hours of being admitted ≥to an intensive care unit under insulin therapy to control blood glucose. The cohort for this study is composed by 9098 patients and a total of 187 features. Fuzzy modeling, gradient boosting and logistic regression were selected due to their interpretability and simplicity. Results show that gradient boosting achieved the highest performance (AUC = 0.94), followed by fuzzy modeling (AUC = 0.87) and logistic regression (AUC = 0.86). Logistic regression resulted to be more difficult to interpret than the other models. The most meaningful variables among all models were Glasgow coma scale score, total urine output, both respiratory and infectious diseases, ventilation time and anion gap.

This paper proposes a decision support system based on fuzzy clustering, fuzzy modeling and fuzzy fingerprints, to provide personalized therapy for critically ill patients. It is hypothesized that the ‘collective experience’ from large clinical databases, where clinical decisions are linked with patient outcomes, can be used to identify specific patient sub-groups and build personalized therapy models towards a new era of personalized medicine, allowing the improvement of patient outcomes in the Intensive Care Unit (ICU). The validity of the proposed systems will be tested using the case study of patients admitted to the ICU who then develop acute kidney injury (AKI); Two-thirds of patients with AKI require renal support therapy. Generalized severity scoring systems have consistently performed poorly for patients with AKI.

The fourth industrial revolution, Industrie 4.0, motivates research to adapt concepts from Internet of Things and Cyber-Physical Systems to automation. Focus is placed on integrating all data sources, from sensors and actuators to management software, to achieve interoperability of heterogeneous systems. To assure data integration, a data collection architecture that guarantees data access from all systems, including legacy devices, is needed. This allows companies to gradually evolve to Industrie 4.0 deployments while still keeping their mission-critical systems online. A middleware-based architecture is a valid answer to the presented challenges. Numerous communication technologies can act as a middleware, all with different advantages and disadvantages. A technology-independent implementation ensures that a lock-in to a specific technology is prevented and makes the architecture more robust and adequate for various use-cases. This work contributes with a concept of a middleware-based data collection architecture, which is independent of specific communication protocols. Therefore, a protocol-agnostic communication interface is developed. A technology comparison summarizes relevant protocols in the field of industrial automation. The concept is subsequently implemented for a case-study and compared to a peer-to-peer legacy approach. Results show that the concept is feasible, as well as that it reduces complexity and effort for implementation and migration of such communication systems. Future work should focus on further developing the concepts, in order to support additional communication protocols and verify the suitability of the architecture in scenarios with a multitude of technologies.

The expected post-operatory weight loss is not always achieved after bariatric surgery. Efforts have been done to describe the causes. Recently, total weight loss (%TWL) has been pointed out to better assess weight loss in bariatric patients. However, there is no cut off point that delimits the patients who successfully achieve their weight goals after a bariatric surgery. In this work, a method based on fuzzy modeling is implemented to help clinicians setting up the best cut-off point in %TWL for a specific population. The best boundary to delimit success and failure will be selected based on the predictive performance of the assessed cut-off points: 25, 30, 35 and 40%TWL after one and two years of surgery. Area under the receiver operating characteristic curve (AUC) values of 0.70 and 0.75 were achieved for the first and second post-surgery periods, respectively. Further, features not previously described as predictors of weight loss were identified as good predictors of the outcome.

Up to two thirds of critically ill patients in Intensive Care Units (ICU) around the world develop Acute Kidney Injury (AKI). This complication is characterized by a reduction of kidney function and is intrinsically linked to an increased risk of mortality. AKI is also connected to higher costs in healthcare due to longer hospital stay times and ICU readmissions for affected patients. The present study developed a model capable of predicting which patients are at a larger risk of death and therefore allow for a more efficient use of hospital resources. Information regarding several physiological parameters of patients with AKI as well as data regarding their survival was used for the creation of a fuzzy model. The model achieved AUC values of up to 0.92 when using mean values for time series and analyzing a subset of patients with AKI at its most advanced stage. When analysing a general population of ICU patients, the algorithm can achieve AUC values of 0.86. This value can be compared with values obtained in similar studies of 0.85 and surpasses values obtained in traditional risk scores such as APACHE II and SOFA. A model such as this, shows the potential of using this type of predictive applications for scheduling human resources and managing monetary resources on critical patients according to their relative need for attention and treatment.

The oscillating-water-column (OWC) is widely regarded as the simplest and most reliable type of wave energy converter. An OWC device comprises a partly submerged hollow structure composed of an aperture below the water surface and an air chamber above the free surface. The oscillating motion of the internal free water surface produced by the incident waves compresses and expands the air in the chamber. The pressure between the air chamber and the atmosphere can be used to drive a turbine connected directly to an electrical generator. Usually the turbines used in OWCs are of the self-rectifying type, i.e., they rotate always in the same sense independently of the flow direction. The aim of this paper is to assess the potential use of experimental data collected in the Mutriku breakwater wave power plant, located in the north of Spain, to perform short-term prediction of the pressure inside the air chamber by a Nonlinear Autoregressive with eXogenous inputs (NARX) Artificial Neural Network (ANN). The NARX model is explored and tuned for a one-step and multi-step-ahead prediction. The results showed that the model is able to capture the OWC system dynamics for a few-steps ahead with a good performance.

In this paper a novel Handwritten Character Identification methodology that performs the recognition of the students identification numbers handwritten in classroom maps has been proposed. A dataset of 60.0000 handwritten training images of the MNIST dataset and a proprietary dataset composed of 3.415 images extracted from 12 classroom maps handwritten by 11 different persons is used in this work. These images were obtained through the segmentation process described in this work, and had suffer some image processing before feature extraction and classification in order to be as similar as possible to the MNIST dataset samples. The algorithm is composed by four main steps: pre-processing, segmentation, feature extraction and classification. It was implemented a segmentation strategy designed for the classroom maps, based on morphological operations and connected components. This strategy can be easily adapted to others handwritten character recognition problems. In classification, three different approaches are used: Support Vector Machine, Convolutional Neural Networks and an hybrid approach that use the CNN extracted features as inputs for a Support Vector Machine. The hybrid approach has the advantage of avoiding the feature extraction step. The dataset is divided into train, test and validation sets and the performance of the three different classifiers is compared and evaluated. The results presented were obtained using real classroom maps. The hybrid CNN+SVM classifier achieved a high performance with overall accuracy of 96.5%. In conclusion, this designed system could successfully reduce the time of creating and accessing the digital storage of students examinations, as well as allow the automatic indexing of student grades towards a full digital evaluation system.

Electricity fraud results in significant losses to utilities. This paper proposes the use of a semi-supervised learning framework to derive an electricity fraud detector from data lacking information on the presence of fraud for the majority of samples. Utilities are only able to make a limited number of inspections, resulting in a lack of data representing cases of fraud. Using a co-training by committee semi-supervised learning framework, the detection performance is improved in comparison to the use of supervised models only trained with labeled data. The framework starts by training a random forest classifier on the labeled data. Next, the unlabeled data that the model can classify with the most confidence is added iteratively to the set of labeled samples, augmenting the data available for model training. The electricity fraud detector achieves a classification performance of 84% true positive rate, 11% false positive rate and 0.89 area under the receiver operating characteristic curve under a positive class balance of 5% and 90% unlabeled samples in the training data.

Intensive care treatment presents unique challenges in the medical world. When treating patients, their wide variety leave care providers with few past examples to draw on. Instead of operating in a pure knowledge discovery capacity, decision support systems can be developed to help predict short-term and long-term patient outcome, based upon available data. One area in which generalized severity scoring systems have consistently performed poorly is among patients admitted intensive care units (ICU) who then develop acute kidney injury. Urine output is used to guide fluid resuscitation and is one of the criteria for the diagnosis of acute kidney injury. This paper provides an example application for predicting short-term critical kidney function in an intensive care unit. Feature construction is performed to extract important aspects of the clinical evolution of the patient. Feature selection is performed on several patient features. Classifiers based on support vector machines and Takagi-Sugeno fuzzy models are developed to predict short-term drops in patient urine output rate. Both types of models showed comparable results, with an AUC of 78%. This shows potential in using similar classifiers to build an ICU decision support system with the goal of predicting short-term complication in the patient and augment current guidelines by anticipating treatment.

Logistic regression and Takagi-Sugeno fuzzy models are sequentially trained with categorical and numerical data in an ensemble-based multistage scheme. In the first stage, a logistic regression model is used to transform the binary feature space into a numerical feature that is used to train a second stage of models consisting of an ensemble of two Takagi-Sugeno fuzzy models. In the ensemble, one model is trained in the space of numerical features and first stage prediction values. The other model is trained only with samples that were classified with a low degree of confidence by the first stage model, in the space of numerical variables. The final output is given by the average of the ensemble predictions at second stage. This scheme was devised under the hypothesis that separating binary from numerical features in the modeling process would increase the performance of a single model using both types of features together. The proposed multistage approach is used to solve a clinical classification problem in a Portuguese hospital. The problem consists of predicting comanagement signalling based on patient clinical data, including diagnosis, procedures, comorbidities and numerical scores, collected before surgery. The multistage performed better in the comanagement dataset, and in 2 out of 5 benchmark datasets.

The roll out of electricity grid assets with advanced communications capabilities enables new ways to steal energy, such as false data attacks and remote meter disconnection. On the other hand, data communicated by these devices has the potential to improve utilities ability to combat fraud through computational intelligence techniques. We propose a clustering-based novelty detection scheme to uncover electricity theft. The scheme starts by extracting easily interpreted consumption indicators from data collected by smart meters. Fuzzy clustering is then used to capture the structure of the data that consists of indicators from benign consumers. The extracted clusters provide the basis for a distance-based novelty detection model to uncover abnormal data sent by consumers. The results for the developed use case show that the proposed scheme using Gustafson-Kessel fuzzy clustering best captures the behavior of consumers, achieving good performance with a low number of clusters, in comparison to euclidean distance-based hard and fuzzy C-means.

Data mining in medical databases often involves the comparison of time series, which represent the evolution of a physiological variable. Temporal misalignment of physiological variables can conceal the discovery of patterns and trends shared between different patients. To address this problem, this paper proposes the mixed fuzzy clustering (MFC) algorithm with the dynamic time-warping (DTW) distance. We developed the MFC algorithm by 1) incorporating the DTW distance into the standard fuzzy c-means to handle misaligned time series; 2) introducing a new dimension into the spatiotemporal clustering algorithm to handle P time-variant features; and 3) incorporating unsupervised learning of cluster-dependent attribute weights. The algorithm is designed to simultaneously cluster time-variant and time-invariant data. We demonstrate the advantages of the proposed algorithm in four synthetic datasets and in two real-world applications in intensive care units. The first application is the classification of patients who will need the administration of vasopressors, and the second is the classification of patients with a high risk of mortality. Time-variant features consist of physiological variables collected with different sampling rates at different points in time. Time-invariant features consist of patients' demographics and score records. The performance is evaluated using cluster validity measures, showing that the proposed algorithm outperforms fuzzy c-means.

This paper proposes the use of mixed fuzzy clustering (MFC) algorithm to derive Takagi-Sugeno (T-S) fuzzy models (FMs). Mixed fuzzy clustering handles both time invariant and multivariate time variant features, allowing the user to control the weight of each component in the clustering process. Two model designs based on MFC are investigated. In the first, the antecedent fuzzy sets of the T-S model are obtained from the clusters obtained by the MFC algorithm. In the second, FMs based on fuzzy c-means (FCM) are constructed over the input space of the partition matrix generated by MFC. The proposed fuzzy modeling approaches are used in health care classification problems, where time series of unequal lengths are very common. MFC-based T-S FMs outperform FCM-based T-S FMs in four out of five datasets and k-nearest neighbors classifiers in five out of five datasets. Dynamic time warping performs better than the Euclidean distance in one dataset and similarly in the remaining. Given the different nature of time variant and invariant data, the choice of a clustering algorithm that treats data differently should be considered for model construction.

Acute kidney injury (AKI) affects 5 to 7% of all hospitalized patients, with a much higher incidence in the critically ill. Although AKI patients have increased risks of death in the intensive care units (ICU) and in the short term, few studies examined the association between the diagnosis of AKI and long-term outcomes. The aim of this study is to address the prediction of short and long-term mortality in patients that presented AKI diagnosis at their hospital admission. Demographic, clinical, physiologic, and date of death data of ICU patients were extracted from the MIMIC-III database to develop predictive models. Fuzzy models are developed using fuzzy c-means and Gustafson-Kessel algorithms to predict mortality in the ICU within 24 hours, when the patients had an ICD-9 admission diagnosis of AKI. The proposed models achieved an AUC of 0.77, which is slightly better than the results obtained by Celi et al. [1] that obtained an AUC of 0.74. Thus, AKI revealed to be a significant risk factor for in-hospital and long-term mortality for ICU patients.

Patients are often discharged prematurely from Intensive Care Units (ICU) due to clinical resource limitations, economic pressure or poor discharge planning. The readmission of such patients is associated with an increased risk of death and is currently viewed as a marker for poor quality care. Several studies have focused on predicting which patients are likely to be readmitted, using techniques such as logistic regression or machine learning algorithms, and based on physiological data measured during the patients' stay at the ICU. So far, no published algorithms have been able to predict readmissions to a satisfactory degree. In this work we hypothesize that physicians' and nurses' notes could give a better explanation of both ICU discharges and readmissions, and propose using the text notes in an ICU database in order to build classification models for the prediction of readmissions. We tested the use of Fuzzy Fingerprints and other traditional text classifiers and compared them to a previously proposed model based on numerical data, obtaining very relevant improvements in the classification results, namely an AUC=0.8.

This paper proposes a methodology to define the load profiles of residential natural gas consumers using smart metering data. A detailed clustering analysis is performed using the fuzzy c-means clustering algorithm and multiple clustering validity indices. The analysis is based on a sample of more than one thousand households over one year. Compact and well defined annual clusters of natural gas consumers are obtained. The results provide evidence that the consumers' representative profiles are mainly characterized by a morning and an evening peak consumption, the time at which the consumption starts to rise and to decline and the off-peak consumption. The knowledge obtained with this methodology can assist decision makers in the energy utility industry in order to develop demand side management programs, consumer engagement strategies, marketing, demand forecasting tools as well as in designing more innovative and sophisticated tariff systems.

Two new methods for the optimization of probabilistic fuzzy classifiers are proposed. Probabilistic fuzzy systems are specially attractive due to their explicit and simultaneous modelling of two kinds of uncertainty, namely vagueness in linguistic terms (fuzziness) and probabilistic uncertainty. The current method uses the maximization of the likelihood with the stochastic gradient descent, which not only converges to local minima but also does not guarantee the minimization of the misclassification error. The proposed methods address this specific problem by incorporating global search techniques. The first algorithm proposed is a genetic algorithm with simple crossover and mutation operations. The other is a first generation memetic algorithm which combines the genetic algorithm with the stochastic gradient descent. A total of five benchmarks were used to compare the three algorithms. The results show that the proposed methods have an average relative improvement of 2% and 6% for the accuracy with the genetic and memetic algorithms, respectively.

This papers proposes two novel approaches for the identification of Takagi-Sugeno fuzzy models with time variant and invariant features. The proposed Mixed Fuzzy Clustering algorithm is proposed for determining the parameters of Takagi-Sugeno fuzzy models in two different ways: (1) the antecedent fuzzy sets are determined based on the partition matrix generated by the Mixed Fuzzy Clustering algorithm; (2) the input features are transformed using the same algorithm and the antecedent fuzzy sets are derived using Fuzzy C-Means clustering. The proposed approaches are tested on four different health care applications: readmissions in intensive care units, administration of vasopressors and mortality. The results show that the proposed clustering algorithm resulted in an increase of the performance of the fuzzy models in three out of four applications in comparison to the use of Fuzzy C-Means.

The rationale behind ensemble machine learning systems is the creation of many classifiers and the combination of their output such that the combination improves the performance of each single classifier. There are two key issues in the creation of ensemble classifiers: one is how two select and group the data samples to train the individual models and the other is how to select or combine the multiple outputs.

The current demands on the energy market, such as efficiency, sustainability and affordability increase the need for customer understanding and data analysis. This paper presents an analysis of the segmentation of electricity consumers based on the fuzzy clustering of time variant electricity consumption data and invariant features like the demographic customer information. The algorithm used is mixed fuzzy clustering (MFC), which allows integrating both variant and invariant features into one clustering. The clustering is evaluated both in its stability over the two years of data, using a entropy measurement and in its general quality given by the three clustering validity indices, Calinski-Harabasz, Davies-Bouldin and Silhouette index.

This work presents heuristic and metaheuristic approaches for addressing the real-world steel cutting problem of a retail steel distributor as a cutting & packing problem. It consists of the cutting of large steel blocks in order to obtain smaller pieces ordered by clients. The problem was formulated as a 3-dimensional residual bin packing problem for minimization of scrap generation, with guillotine cutting constraint and chips scrap generation. A tabu search and bestfit decreasing (BFD) approaches are proposed and their performance compared to an heuristic and ant colony optimization (ACO) algorithms. It's shown that the tabu search and best-fit decreasing algorithm are able to reduce the generated scrap by up to 52% in comparison with the heuristic in [1]. The orders to suppliers were also reduced by up to 35%. The analysis of the results of the different approaches provide insight onto the most important factors in the problem's scrap minimization.

The aim of this work is to identify groups of patients with similar patterns that are related to a higher risk of readmission to an Intensive Care Unit (ICU). Patients readmissions to ICUs are introduced as a problem associated with increased mortality, morbidity and costs, which complicates the performance of a good clinical management and medical diagnosis. To approach the readmissions classification problem, Fuzzy C-Means (FCM) clustering algorithm was implemented to find the different groups of patients. A multimodel approach was developed using these groups and the best clustering division was assessed through different objective functions. Two decision criteria were used for the multimodel approach, an a priori decision and an a posteriori decision. The data used, from MIMIC II database, consisted on the arithmetic means of time series of variables — acquired during the last 24 hours before discharge. The multimodel using the a priori and the aposteriori decisions were able to predict readmissions with an average AUC of 0.74 and 0.75, respectively. Consequently, the multimodel results overcame the results of previous predictive models developed for the classification of readmissions outcome.

This paper proposes a novel feature selection approach formulated based on the Fish School Search (FSS) optimization algorithm, intended to cope with premature convergence. In order to use this population based optimization algorithm in feature selection problems, we propose the use of a binary encoding scheme for the internal mechanisms of the fish school search, emerging the binary fish school search (BFSS). The suggested algorithm was combined with fuzzy modeling in a wrapper approach for Feature Selection (FS) and tested over three benchmark databases. This hybrid proposal was applied to an ICU (Intensive Care Unit) readmission problem. The purpose of this application was to predict the readmission of ICU patients within 24 to 72 hours after being discharged. We assessed the experimental results in terms of performance measures and the number of features selected by each used FS algorithms. We observed that our proposal can correctly select the discriminating input features.

We propose the application of probabilistic fuzzy systems (PFS) to model the prediction of early readmission in intensive care unit patients and compare it with the gold-standard method - logistic regression based on the APACHE II score. PFS are characterized by the combination of the linguistic description of the system with the statistical properties of data. On one hand, results point that PFS models perform comparably to the gold-standard method, with AUC values of 0.66±0.03. On the other hand, results also show that PFS models use a significant lower number of variables which, from the clinical practice point of view, suggests improved gains in terms of simplicity.

Shock is a life-threatening medical condition requiring the administration of powerful drugs - vasopressors. Early identification of these patients is a worthy goal in order to timely prepare them for therapy. A subset composed of the most frequently sampled and readily available variables in an intensive care unit (ICU) was used for clustering patients. Then, a data exploration process was started through the use of fuzzy clustering with the fuzzy cmeans algorithm, where four clusters were obtained and the groups characteristics were analyzed. A relationship between the clusters obtained and the use of vasopressors was found out and these results were visualized with the help of histograms. First, a single model was derived. Then, four models were trained and used for a multi model approach, one for each identified group of patients. In both cases fuzzy models were used as they are universal approximators. For the multi-model approach, two decision criteria were used. First a decision a priori based on the distance from the clusters centers to the patient characteristics was used. Lastly a decision a posteriori approach where each model was used and the final outcome used is based on the uncertainty of the output response to the threshold of each model. The multi model approach with a posteriori decision had a better performance of the two schemes tested, and also performed better than the single general model approach.

This paper proposes the application of a new binary particle swarm optimization (BPSO) method to feature selection problems. Two enhanced versions of binary particle swarm optimization, designed to cope with premature convergence of the BPSO algorithm, are proposed. These methods control the swarm variability using the velocity and the similarity between best swarm solutions. The proposed PSO methods use neural networks, fuzzy models and support vector machines in a wrapper approach, and are tested in a benchmark database. It was shown that the proposed BPSO approaches require an inferior simulation time, less selected features and increase accuracy. The best BPSO is then compared with genetic algorithms (GA) and applied to a real medical application, a sepsis patient database. The objective is to predict the outcome (survived or deceased) of the sepsis patients. It was shown that the proposed BPSO approaches are similar in terms of model accuracy when compared to GA, while requiring an inferior simulation time and less selected features.

In the present work, we propose the application of probabilistic fuzzy systems (PFS) to model the prediction of mortality in septic shock patients. This technique is characterized by the combination of the linguistic description of the system with the statistical properties of data. Preliminary results for this particular clinical problem point that PFS models, besides performing as accurately as first order Takagi-Sugeno fuzzy models, also provide probability measures that provide additional clinical information upon which physicians can act on.

The amount of data generated in the intensive care environment nowadays prohibits the storage of all the information available. The validation process is time consuming, since nurses have to check every certain periods the data acquired from bedside monitors in order to assess their validity and integrity. This work presents an automatic method for data validation in the intensive care environment, based on an artificial intelligence approach, namely artificial neural networks (ANNs). A real world dataset acquired at Beth Israel Deaconess Medical Center (BIDMC) neonatal intensive care unit (NICU) is used to obtain the validation model and assess its performance. The dataset consists of high frequency sampled data of the level of oxygen saturation (SpO2) of neonates. A subset of 100 neonates was considered for modeling purposes. A total of 7,018,662 samples were available, containing 129,075 validated ones. The performance of the validation model, assessed in terms of its AUC, was of up to 0.75. Both the sensitivity and specificity reached acceptable values according to medical review. Future work would involve a prospective study and validation of the methods proposed in this work.

One consequence of the increasing amount of data stored during acquisition processes is that sampled time series are more prone to be collected in a misaligned uneven fashion and/or be partly lost or unavailable (missing data). Due to their severe impact on data mining techniques, this work proposes methods to (a) align misaligned unevenly sampled data, (b) differentiate absent values related to low sampling frequencies, compared to those resulting from missingness mechanisms, and (c) to classify recoverable and non-recoverable segments of missing data by using statistical and fuzzy modeling approaches. These methods were evaluated against randomly simulated test datasets containing different amounts of missing data. Results show that: (1) using the variable most frequently sampled as a template, combined with cubic interpolation, allowed to unshift misaligned uneven data without significant errors; (2) the differentiation of absent values due to low sampling frequencies from those truly missing, can be successfully performed using 95% confidence intervals relative to the mean sampling time; (3) fuzzy modeling returned better classification results for recoverable segments, while the statistical approach performed better in classifying non-recoverable segments. All three methods proposed in this work decreased their performance when the amount of missing data was increased in the test datasets.

Laboratory testing is a frequent activity for patients in intensive care units (ICU). Recent studies demonstrate that frequent laboratory testing does not necessarily relate to better outcomes. We hypothesize that unnecessary laboratory testing can be reduced by predicting which tests are unlikely to influence clinical management. Reducing unnecessary tests could reduce morbidity and hospitalization costs. We analyzed an ICU database containing 26,665 patient records at Beth Israel Deaconess Medical Center, Boston, and selected a subset of patients with gastrointestinal bleeding. Database knowledge discovery was applied involving data preprocessing, feature selection, and classification. Conventional soft computing tools such as fuzzy models and neural networks were utilized in this work, combined with statistical and mathematical tools. The input variables included bedside monitor trends, lab tests, arterial/central catheter information, urine collections, transfusions, indexes and scores calculated for the patients. The outcome variable was a binary classification based on falling levels of hematocrit. Feature selection was performed by a bottom-up strategy, maximizing the area under the ROC curve (AUC), the integrated discrimination improvement (IDI) and a multiobjective function primarily pondering the sensitivity of the models. A leave-one-out cross validation process was used to evaluate the overall models' performance, as well as the additional predictive value of the variables selected. Urine output was selected by all models as the best predictor of useful hematocrit testing. Our results show that it is possible to correctly classify the usefulness of a hematocrit lab test up to 81% of the time by using fuzzy models and neural networks.

Vasopressors belong to a powerful class of drugs used in the management of systemic shock in ill patients. The administration of a vasopressor involves the non-trivial process of inserting a central venous catheter. This procedure carries with it inherent risks which are increased when done under urgency such as in the case of unexpected systemic shock. The ability to predict the transition to vasopressor dependence could be expected to improve overall outcomes associated with the procedure. We use three different approaches combining fuzzy modeling with bottom-up (BU), top-town (TD) and ant feature selection (AFS), to classify requirements for vasopressors in shock. We observe that fuzzy models combined with BU feature selection return higher values of sensitivity; fuzzy models with no feature selection and fuzzy models with TD feature selection return higher values of AUC and specificity; features most commonly selected to classify impending use of vasopressores in pancreatitis patients include levels of Sodium and White Blood Cell counts, while for pneumonia patients include levels of Lactid Acid and White Blood Cell Count; and finally, fuzzy models combined with BU and fuzzy models combined with AFS demonstrate the lowest number of selected variables with no significant loss in accuracy.

Real-world databases often contain missing data and existing correction algorithms deliver varying performance. Also, most modeling techniques are not suitable to deal with them automatically. In this study we examine different approaches to predicting septic shock in the presence of missing data. Some preprocessing techniques for managing missing data include disregarding data, or replacing it with information that by design introduces bias. In this study, we show that predictive performance improves by employing a minimum pre-processing technique, the Zero-Order-Hold (ZOH) method, by applying a Fuzzy C-Means clustering technique based on the partial distance calculation strategy (FCM-PDS) and by computing the final classification regarding the samples from each patient. Performance improvements continue to occur where up to approximately 60% of the data is missing, though for higher percentage the classification performance still is statistically improved. We further validate this approach by making comparisons with previous studies.

Measuring the performance of a given classifier is not a straightforward or easy task. Depending on the application, the overall classification rate may not be sufficient if one, or more, of the classes fail in prediction. This problem is also reflected in the feature selection process, especially when a wrapper method is used. Cohen's kappa coefficient is a statistical measure of inter-rater agreement for qualitative items. It is generally thought to be a more robust measure than simple percent agreement calculation, since it takes into account the agreement occurring by chance. Considering that kappa is a more conservative measure, then its use in wrapper feature selection is suitable to test the performance of the models. This paper proposes the use of the kappa measure as an evaluation measure in a feature selection wrapper approach. In the proposed approach, fuzzy models are used to test the feature subsets and fuzzy criteria are used to formulate the feature selection problem. Results show that using the kappa measure leads to more accurate classifiers, and therefore it leads to feature subset solutions with more relevant features.

One of the most important techniques in data preprocessing for data mining is feature selection. Real-world data analysis, data mining, classification and modeling problems usually involve a large number of candidate inputs or features. Less relevant or highly correlated features decrease, in general, the classification accuracy, and enlarge the complexity of the classifier. The goal is to find a reduced set of features that reveals the best classification accuracy for a fuzzy classifier. This paper proposes an ant colony optimization (ACO) algorithm for feature selection, which minimizes two objectives: the number of features and the error classification. Two pheromone matrices and two different heuristics are used for each objective. The performance of the method is compared to other features selection methods, revealing higher performance.

The process under study has four manipulating variables: feed flow rate, washing water, air and rejected flow rates. Column flotation process is an example of a complex, nonlinear and multivariable system. Fuzzy modeling is a well-known modeling technique, which has been applied to complex and nonlinear processes. Fuzzy multivariable modeling with fuzzy model predictive control is used in this paper to control a laboratory setup of a flotation column. Moreover, the control strategy is tested in real-time. Results show that the applied methods led to a good control performance of all the controlled variables.

This paper proposes the application of fault-tolerant control (FTC) using fuzzy predictive control. The FTC approach is based on two steps, fault detection and isolation (FDI) and fault accommodation. Fault detection is performed by a model-based approach using fuzzy modeling. Fault isolation uses a fuzzy decision making approach. The model of the isolated fault is used in fault accommodation with a model predictive control (MPC) scheme. This paper uses a fuzzy predictive control scheme, where fuzzy goals and fuzzy constraints are described in a fuzzy objective function. The criteria (goals or constraints) have an associated weight factor, which are chosen by the decision-maker. Two faults were simulated, and the respective fuzzy models were identified. The fuzzy FTC scheme proposed in this paper was able to accommodate the simulated faults.

In practice, it is very difficult to achieve an accurate modeling for real-world systems. If the system structure is not precisely known and the number of potential inputs is large, the model has to be based primarily on data or heuristic information. The inherent characteristics of fuzzy logic theory makes it suitable for modeling this type of systems. In this paper a modification to the regularity criterion algorithm is proposed, aiming the reduction of computational time without loss of accuracy. The fuzzy models obtained using the modified regularity criterion algorithm are optimized by a real-coded genetic algorithm. Real data is used for the design and validation of two industrial processes used as examples. The proposed algorithm improves both the accuracy of the models and it reduces the computational time to obtain them

The column flotation process is a very complex, nonlinear and multivariable system. Fuzzy modeling is a well-known modeling technique, which has been applied to complex and nonlinear processes. This paper proposes a fuzzy modeling identification technique, where the structure of the model is determined using a regularity criterion, and the rules are identified using fuzzy clustering optimized by a real-coded genetic algorithm. Real data is used for the design and validation of the column flotation fuzzy model. The results are compared to other well-known fuzzy modeling techniques. The validation results show that is possible to find a better model using the identification procedure proposed in this paper.