Using data mining for prediction of hospital length of stay: an application of the CRISP-DM Methodology

Hospitals are nowadays collecting vast amounts of data related with patient records. All this data hold valuable knowledge that can be used to improve hospital decision making. Data mining techniques aim precisely at the extraction of useful knowledge from raw data. This work describes an implementation of a medical data mining project approach based on the CRISP-DM methodology. Recent real-world data, from 2000 to 2013, were collected from a Portuguese hospital and related with inpatient hospitalization. The goal was to predict generic hospital Length Of Stay based on indicators that are commonly available at the hospitalization process (e.g., gender, age, episode type, medical specialty). At the data preparation stage, the data were cleaned and variables were selected and transformed, leading to 14 inputs. Next, at the modeling stage, a regression approach was adopted, where six learning methods were compared: Average Prediction, Multiple Regression, Decision Tree, Artificial Neural Network ensemble, Support Vector Machine and Random Forest. The best learning model was obtained by the Random Forest method, which presents a high quality coefficient of determination value (0.81). This model was then opened by using a sensitivity analysis procedure that revealed three influential input attributes: the hospital episode type, the physical service where the patient is hospitalized and the associated medical specialty. Such extracted knowledge confirmed that the obtained predictive model is credible and with potential value for supporting decisions of hospital managers.

Using customer lifetime value to improve the prediction of bank deposit subscription in telemarketing campaigns

Customer lifetime value (LTV) enables using client characteristics, such as recency, frequency and monetary (RFM) value, to describe the value of a client through time in terms of profitability. We present the concept of LTV applied to telemarketing for improving the return-on-investment, using a recent (from 2008 to 2013) and real case study of bank campaigns to sell long- term deposits. The goal was to benefit from past contacts history to extract additional knowledge. A total of twelve LTV input variables were tested, un- der a forward selection method and using a realistic rolling windows scheme, highlighting the validity of five new LTV features. The results achieved by our LTV data-driven approach using neural networks allowed an improvement up to 4 pp in the Lift cumulative curve for targeting the deposit subscribers when compared with a baseline model (with no history data). Explanatory knowledge was also extracted from the proposed model, revealing two highly relevant LTV features, the last result of the previous campaign to sell the same product and the frequency of past client successes. The obtained results are particularly valuable for contact center companies, which can improve pre- dictive performance without even having to ask for more information to the companies they serve.

An Equation for the Prediction of Oxygen Consumption in a Brazilian Population

Background: The equations predicting maximal oxygen uptake (VO2max or peak) presently in use in cardiopulmonary exercise testing (CPET) softwares in Brazil have not been adequately validated. These equations are very important for the diagnostic capacity of this method. Objective: Build and validate a Brazilian Equation (BE) for prediction of VO2peak in comparison to the equation cited by Jones (JE) and the Wasserman algorithm (WA). Methods: Treadmill evaluation was performed on 3119 individuals with CPET (breath by breath). The construction group (CG) of the equation consisted of 2495 healthy participants. The other 624 individuals were allocated to the external validation group (EVG). At the BE (derived from a multivariate regression model), age, gender, body mass index (BMI) and physical activity level were considered. The same equation was also tested in the EVG. Dispersion graphs and Bland-Altman analyses were built. Results: In the CG, the mean age was 42.6 years, 51.5% were male, the average BMI was 27.2, and the physical activity distribution level was: 51.3% sedentary, 44.4% active and 4.3% athletes. An optimal correlation between the BE and the CPET measured VO2peak was observed (0.807). On the other hand, difference came up between the average VO2peak expected by the JE and WA and the CPET measured VO2peak, as well as the one gotten from the BE (p = 0.001). Conclusion: BE presents VO2peak values close to those directly measured by CPET, while Jones and Wasserman differ significantly from the real VO2peak.

Efficient knowledge retrieval to calibrate input variables in forest fire prediction

Forest fires are a serious threat to humans and nature from an ecological, social and economic point of view. Predicting their behaviour by simulation still delivers unreliable results and remains a challenging task. Latest approaches try to calibrate input variables, often tainted with imprecision, using optimisation techniques like Genetic Algorithms. To converge faster towards fitter solutions, the GA is guided with knowledge obtained from historical or synthetical fires. We developed a robust and efficient knowledge storage and retrieval method. Nearest neighbour search is applied to find the fire configuration from knowledge base most similar to the current configuration. Therefore, a distance measure was elaborated and implemented in several ways. Experiments show the performance of the different implementations regarding occupied storage and retrieval time with overly satisfactory results.

Psychoses a l'adolescence, les neurosciences ameliorent-elles la prediction? [Adolescent psychosis, can neuroscience improve prediction?]

Developments in the field of neuroscience have created a high level of interest in the subject of adolescent psychosis, particularly in relation to prediction and prevention. As the medical practice of adolescent psychosis and its treatment is characterised by a heterogeneity which is both symptomatic and evolutive, the somewhat poor prognosis of chronic development justifies the research performed: apparent indicators of schizophrenic disorders on the one hand and specific endophenotypes on the other are becoming increasingly important. The significant progresses made on the human genome show that the genetic predetermination in current psychiatric pathologies is complex and subject to moderating effects and there is therefore significant potential for nature-nurture interactions (between the environment and the genes). The road to be followed in researching the phenotypic expression of a psychosis gene is long and winding and is susceptible to many external influences at various levels with different effects. Neurobiological, neurophysiological, neuropsychological and neuroanatomical studies help to identify endophenotypes, which allow researchers to create identifying "markers" along this winding road. The endophenotypes could make it possible to redefine the nosological categories and enhance understanding of the physiopathology of schizophrenia. In a predictive approach, large-scale retrospective and prospective studies make it possible to identify risk factors, which are compatible with the neurodevelopmental hypothesis of schizophrenia. However, the predictive value of such markers or risk indicators is not yet sufficiently developed to offer a reliable early-detection method or possible schizophrenia prevention measures. Nonetheless, new developments show promise against the background of a possible future nosographic revolution, based on a paradigm shift. It is perhaps on the basis of homogeneous endophenotypes in particular that we will be able to understand what protects against, or indeed can trigger, psychosis irrespective of the clinical expression or attempts to isolate the common genetic and biological bases according to homogeneous clinical characteristics, which have to date, proved unsuccessful

Prediction and discrimination of osteoporotic hip fracture in postmenopausal women.

Osteoporotic hip fractures increase dramatically with age and are responsible for considerable morbidity and mortality. Several treatments to prevent the occurrence of hip fracture have been validated in large randomized trials and the current challenge is to improve the identification of individuals at high risk of fracture who would benefit from therapeutic or preventive intervention. We have performed an exhaustive literature review on hip fracture predictors, focusing primarily on clinical risk factors, dual X-ray absorptiometry (DXA), quantitative ultrasound, and bone markers. This review is based on original articles and meta-analyses. We have selected studies that aim both to predict the risk of hip fracture and to discriminate individuals with or without fracture. We have included only postmenopausal women in our review. For studies involving both men and women, only results concerning women have been considered. Regarding clinical factors, only prospective studies have been taken into account. Predictive factors have been used as stand-alone tools to predict hip fracture or sequentially through successive selection processes or by combination into risk scores. There is still much debate as to whether or not the combination of these various parameters, as risk scores or as sequential or concurrent combinations, could help to better predict hip fracture. There are conflicting results on whether or not such combinations provide improvement over each method alone. Sequential combination of bone mineral density and ultrasound parameters might be cost-effective compared with DXA alone, because of fewer bone mineral density measurements. However, use of multiple techniques may increase costs. One problem that precludes comparison of most published studies is that they use either relative risk, or absolute risk, or sensitivity and specificity. The absolute risk of individuals given their risk factors and bone assessment results would be a more appropriate model for decision-making than relative risk. Currently, a group appointed by the World Health Organization and lead by Professor John Kanis is working on such a model. It will therefore be possible to further assess the best choice of threshold to optimize the number of women needed to screen for each country and each treatment.

Splitting statistical potentials into meaningful scoring functions: testing the prediction of near-native structures from decoy conformations

Background: Recent advances on high-throughput technologies have produced a vast amount of protein sequences, while the number of high-resolution structures has seen a limited increase. This has impelled the production of many strategies to built protein structures from its sequence, generating a considerable amount of alternative models. The selection of the closest model to the native conformation has thus become crucial for structure prediction. Several methods have been developed to score protein models by energies, knowledge-based potentials and combination of both.Results: Here, we present and demonstrate a theory to split the knowledge-based potentials in scoring terms biologically meaningful and to combine them in new scores to predict near-native structures. Our strategy allows circumventing the problem of defining the reference state. In this approach we give the proof for a simple and linear application that can be further improved by optimizing the combination of Zscores. Using the simplest composite score () we obtained predictions similar to state-of-the-art methods. Besides, our approach has the advantage of identifying the most relevant terms involved in the stability of the protein structure. Finally, we also use the composite Zscores to assess the conformation of models and to detect local errors.Conclusion: We have introduced a method to split knowledge-based potentials and to solve the problem of defining a reference state. The new scores have detected near-native structures as accurately as state-of-art methods and have been successful to identify wrongly modeled regions of many near-native conformations.

Prediction of enzyme function by combining sequence similarity and protein interactions

Background: A number of studies have used protein interaction data alone for protein function prediction. Here, we introduce a computational approach for annotation of enzymes, based on the observation that similar protein sequences are more likely to perform the same function if they share similar interacting partners. Results: The method has been tested against the PSI-BLAST program using a set of 3,890 protein sequences from which interaction data was available. For protein sequences that align with at least 40% sequence identity to a known enzyme, the specificity of our method in predicting the first three EC digits increased from 80% to 90% at 80% coverage when compared to PSI-BLAST. Conclusion: Our method can also be used in proteins for which homologous sequences with known interacting partners can be detected. Thus, our method could increase 10% the specificity of genome-wide enzyme predictions based on sequence matching by PSI-BLAST alone.

Urinary incontinence in the prediction of falls in hospitalized elderly

Objective Analyzing the effect of urinary incontinence as a predictor of the incidence of falls among hospitalized elderly. Method Concurrent cohort study where 221 elderly inpatients were followed from the date of admission until discharge, death or fall. The Kaplan-Meier methods, the incidence density and the Cox regression model were used for the survival analysis and the assessment of the association between the exposure variable and the other variables. Results Urinary incontinence was a strong predictor of falls in the surveyed elderly, and was associated with shorter time until the occurrence of event. Urinary incontinence, concomitant with gait and balance dysfunction and use of antipsychotics was associated with falls. Conclusion Measures to prevent the risk of falls specific to hospitalized elderly patients who have urinary incontinence are necessary.



Improving generalized regression analysis for the spatial prediction of forest communities

Aim This study used data from temperate forest communities to assess: (1) five different stepwise selection methods with generalized additive models, (2) the effect of weighting absences to ensure a prevalence of 0.5, (3) the effect of limiting absences beyond the environmental envelope defined by presences, (4) four different methods for incorporating spatial autocorrelation, and (5) the effect of integrating an interaction factor defined by a regression tree on the residuals of an initial environmental model. Location State of Vaud, western Switzerland. Methods Generalized additive models (GAMs) were fitted using the grasp package (generalized regression analysis and spatial predictions, http://www.cscf.ch/grasp). Results Model selection based on cross-validation appeared to be the best compromise between model stability and performance (parsimony) among the five methods tested. Weighting absences returned models that perform better than models fitted with the original sample prevalence. This appeared to be mainly due to the impact of very low prevalence values on evaluation statistics. Removing zeroes beyond the range of presences on main environmental gradients changed the set of selected predictors, and potentially their response curve shape. Moreover, removing zeroes slightly improved model performance and stability when compared with the baseline model on the same data set. Incorporating a spatial trend predictor improved model performance and stability significantly. Even better models were obtained when including local spatial autocorrelation. A novel approach to include interactions proved to be an efficient way to account for interactions between all predictors at once. Main conclusions Models and spatial predictions of 18 forest communities were significantly improved by using either: (1) cross-validation as a model selection method, (2) weighted absences, (3) limited absences, (4) predictors accounting for spatial autocorrelation, or (5) a factor variable accounting for interactions between all predictors. The final choice of model strategy should depend on the nature of the available data and the specific study aims. Statistical evaluation is useful in searching for the best modelling practice. However, one should not neglect to consider the shapes and interpretability of response curves, as well as the resulting spatial predictions in the final assessment.

Prediction of subcutaneous fatty acid composition from ham of CLA fed boars and gilts using a NIRS fiber optic probe

The aim of this work was the use of NIR technology by direct application of a fiber optic probe on back fat to analyze the fatty acid composition of CLA fed boars and gilts. 265 animals were fed 3 different diets and the fatty acid profile of back fat from Gluteus medius was analyzed using gas chromatography and FT-NIR. Spectra were acquired using a Bruker Optics Matrix-F duplex spectrometer equipped with a fiber optic probe (IN-268-2). Oleic and stearic fatty acids were predicted accurately; myristic, vaccenic and linoleic fatty acids were predicted with lower accuracy, while palmitic and α-linolenic fatty acids were poorly predicted. The relative percentage of fatty acids and NIR spectra showed differences in fatty acid composition of back fat from pigs fed CLA which increased the relative percentage of SFA and PUFA while MUFA decreased. Results suggest that a NIR fiber optic probe can be used to predict total saturated and unsaturated fatty acid composition, as well as the percentage of stearic and oleic. NIR showed potential as a rapid and easily implemented method to discriminate carcasses from animals fed different diets.

Runoff measurement and prediction for a watershed under natural vegetation in central Brazil

This work aimed to measure and analyze total rainfall (P), rainfall intensity and five-day antecedent rainfall effects on runoff (R); to compare measured and simulated R values using the Soil Conservation Service Curve Number method (CN) for each rainfall event; and to establish average R/P ratios for observed R values. A one-year (07/01/96 to 06/30/97) rainfall-runoff data study was carried out in the Capetinga watershed (962.4 ha), located at the Federal District of Brazil, 47° 52' longitude West and 15° 52' latitude South. Soils of the watershed were predominantly covered by natural vegetation. Total rainfall and runoff for the period were 1,744 and 52.5 mm, respectively, providing R/P of 3% and suggesting that watershed physical characteristics favored water infiltration into the soil. A multivariate regression analysis for 31 main rainfall-runoff events totaling 781.9 and 51.0 mm, respectively, indicated that the amount of runoff was only dependent upon rainfall volume. Simulated values of total runoff were underestimated about 15% when using CN method and an area-weighted average of the CN based on published values. On the other hand, when average values of CN were calculated for the watershed, total runoff was overestimated about 39%, suggesting that CN method shoud be used with care in areas under natural vegetation.

Modelling of weather radar echoes from anomalous propagation using a hybrid parabolic equation method and NWP model data

Contamination of weather radar echoes by anomalous propagation (anaprop) mechanisms remains a serious issue in quality control of radar precipitation estimates. Although significant progress has been made identifying clutter due to anaprop there is no unique method that solves the question of data reliability without removing genuine data. The work described here relates to the development of a software application that uses a numerical weather prediction (NWP) model to obtain the temperature, humidity and pressure fields to calculate the three dimensional structure of the atmospheric refractive index structure, from which a physically based prediction of the incidence of clutter can be made. This technique can be used in conjunction with existing methods for clutter removal by modifying parameters of detectors or filters according to the physical evidence for anomalous propagation conditions. The parabolic equation method (PEM) is a well established technique for solving the equations for beam propagation in a non-uniformly stratified atmosphere, but although intrinsically very efficient, is not sufficiently fast to be practicable for near real-time modelling of clutter over the entire area observed by a typical weather radar. We demonstrate a fast hybrid PEM technique that is capable of providing acceptable results in conjunction with a high-resolution terrain elevation model, using a standard desktop personal computer. We discuss the performance of the method and approaches for the improvement of the model profiles in the lowest levels of the troposphere.

Prediction of difficult intubation with advanced face recognition techniques: a preliminary study

Introduction: Difficult tracheal intubation remains a constant and significant source of morbidity and mortality in anaesthetic practice. Insufficient airway assessment in the preoperative period continues to be a major cause of unanticipated difficult intubation. Although many risk factors have already been identified, preoperative airway evaluation is not always regarded as a standard procedure and the respective weight of each risk factor remains unclear. Moreover the predictive scores available are not sensitive, moderately specific and often operator-dependant. In order to improve the preoperative detection of patients at risk for difficult intubation, we developed a system for automated and objective evaluation of morphologic criteria of the face and neck using video recordings and advanced techniques borrowed from face recognition. Method and results: Frontal video sequences were recorded in 5 healthy volunteers. During the video recording, subjects were requested to perform maximal flexion-extension of the neck and to open wide the mouth with tongue pulled out. A robust and real-time face tracking system was then applied, allowing to automatically identify and map a grid of 55 control points on the face, which were tracked during head motion. These points located important features of the face, such as the eyebrows, the nose, the contours of the eyes and mouth, and the external contours, including the chin. Moreover, based on this face tracking, the orientation of the head could also be estimated at each frame of the video sequence. Thus, we could infer for each frame the pitch angle of the head pose (related to the vertical rotation of the head) and obtain the degree of head extension. Morphological criteria used in the most frequent cited predictive scores were also extracted, such as mouth opening, degree of visibility of the uvula or thyreo-mental distance. Discussion and conclusion: Preliminary results suggest the high feasibility of the technique. The next step will be the application of the same automated and objective evaluation to patients who will undergo tracheal intubation. The difficulties related to intubation will be then correlated to the biometric characteristics of the patients. The objective in mind is to analyze the biometrics data with artificial intelligence algorithms to build a highly sensitive and specific predictive test.

Prediction markets supporting technology assessment

In this thesis, we study the use of prediction markets for technology assessment. We particularly focus on their ability to assess complex issues, the design constraints required for such applications and their efficacy compared to traditional techniques. To achieve this, we followed a design science research paradigm, iteratively developing, instantiating, evaluating and refining the design of our artifacts. This allowed us to make multiple contributions, both practical and theoretical. We first showed that prediction markets are adequate for properly assessing complex issues. We also developed a typology of design factors and design propositions for using these markets in a technology assessment context. Then, we showed that they are able to solve some issues related to the R&D portfolio management process and we proposed a roadmap for their implementation. Finally, by comparing the instantiation and the results of a multi-criteria decision method and a prediction market, we showed that the latter are more efficient, while offering similar results. We also proposed a framework for comparing forecasting methods, to identify the constraints based on contingency factors. In conclusion, our research opens a new field of application of prediction markets and should help hasten their adoption by enterprises. Résumé français: Dans cette thèse, nous étudions l'utilisation de marchés de prédictions pour l'évaluation de nouvelles technologies. Nous nous intéressons plus particulièrement aux capacités des marchés de prédictions à évaluer des problématiques complexes, aux contraintes de conception pour une telle utilisation et à leur efficacité par rapport à des techniques traditionnelles. Pour ce faire, nous avons suivi une approche Design Science, développant itérativement plusieurs prototypes, les instanciant, puis les évaluant avant d'en raffiner la conception. Ceci nous a permis de faire de multiples contributions tant pratiques que théoriques. Nous avons tout d'abord montré que les marchés de prédictions étaient adaptés pour correctement apprécier des problématiques complexes. Nous avons également développé une typologie de facteurs de conception ainsi que des propositions de conception pour l'utilisation de ces marchés dans des contextes d'évaluation technologique. Ensuite, nous avons montré que ces marchés pouvaient résoudre une partie des problèmes liés à la gestion des portes-feuille de projets de recherche et développement et proposons une feuille de route pour leur mise en oeuvre. Finalement, en comparant la mise en oeuvre et les résultats d'une méthode de décision multi-critère et d'un marché de prédiction, nous avons montré que ces derniers étaient plus efficaces, tout en offrant des résultats semblables. Nous proposons également un cadre de comparaison des méthodes d'évaluation technologiques, permettant de cerner au mieux les besoins en fonction de facteurs de contingence. En conclusion, notre recherche ouvre un nouveau champ d'application des marchés de prédiction et devrait permettre d'accélérer leur adoption par les entreprises.