No description available.

In this thesis, we describe an approach that won a psychiatric symptom severity prediction challenge. The challenge was to correctly predict the severity of psychiatric symptoms on a 4-point scale. Our winning submission uses a novel stacked machine learning architecture in which (i) a base data ingestion/cleaning step was followed by the (ii) derivation of a base set of features defined using text analytics, after which (iii) association rule learning was used in a novel way to generate new features, followed by a (iv) feature selection step to eliminate irrelevant features, followed by a (v) classifier training algorithm in which a total of 22 classifiers including new classifier variants of AdaBoost and RandomForest were trained on seven different data views, and (vi) finally an ensemble learning step, in which ensembles of best learners were used to improve on the accuracy of individual learners. All of this was tested via standard 10-fold cross-validation on training data provided by the N-GRID challenge organizers, of which the three best ensembles were selected for submission to N-GRID's blind testing. The best of our submitted solutions garnered an overall final score of 0.863 according to the organizer's measure. All 3 of our submissions placed within the top 10 out of the 65 total submissions. The challenge constituted Track 2 of the 2016 Centers of Excellence in Genomic Science (CEGS) Neuropsychiatric Genome-Scale and RDOC Individualized Domains (N-GRID) Shared Task in Clinical Natural Language Processing.

In many practical scenarios, prediction for high-dimensional observations can be accurately performed using only a fraction of the existing features. However, the set of relevant predictive features, known as the sparsity pattern, varies across data. For instance, features that are informative for a subset of observations might be useless for the rest. In fact, in such cases, the dataset can be seen as an aggregation of samples belonging to several low-dimensional sub-models, potentially due to different generative processes. My thesis introduces several techniques for identifying sparse predictive structures and the areas of the feature space where these structures are effective. This information allows the training of models which perform better than those obtained through traditional feature selection. We formalize Informative Projection Recovery, the problem of extracting a set of low-dimensional projections of data which jointly form an accurate solution to a given learning task. Our solution to this problem is a regression-based algorithm that identifies informative projections by optimizing over a matrix of point-wise loss estimators. It generalizes to a number of machine learning problems, offering solutions to classification, clustering and regression tasks. Experiments show that our method can discover and leverage low-dimensional structure, yielding accurate and compact models. Our method is particularly useful in applications involving multivariate numeric data in which expert assessment of the results is of the essence. Additionally, we developed an active learning framework which works with the obtained compact models in finding unlabeled data deemed to be worth expert evaluation. For this purpose, we enhance standard active selection criteria using the information encapsulated by the trained model. The advantage of our approach is that the labeling effort is expended mainly on samples which benefit models from the hypothesis class we are considering. Additionally, the domain experts benefit from the availability of informative axis aligned projections at the time of labeling. Experiments show that this results in an improved learning rate over standard selection criteria, both for synthetic data and real-world data from the clinical domain, while the comprehensible view of the data supports the labeling process and helps preempt labeling errors.

In the past three years, evidence based medicine emerged as a powerful force in an effort to improve quality and health outcomes, and to reduce cost of care. Computerized physician order entry (CPOE) applications brought safety and efficiency features to clinical settings, including ease of ordering medications via pre-defined sets. Order sets offer promise of standardized care beyond convenience features through evidence-based practices built upon a growing and powerful knowledge of clinical professionals to achieve potentially more consistent health outcomes with patients and to reduce frequency of medical errors, adverse drug effects, and unintended side effects during treatment. While order sets existed in paper form prior to the introduction of CPOE, their true potential was only unleashed with support of clinical informatics, at those healthcare facilities that installed CPOE systems and reap rewards of standardized care. Despite ongoing utilization of order sets at facilities that implemented CPOE, there is a lack of quantitative evidence behind their benefits. Comprehensive research into their impact requires a history of electronic medical records necessary to produce large population samples to achieve statistically significant results. The study, conducted at a large Midwest healthcare system consisting of several community and academic hospitals, was aimed at quantitatively analyzing benefits of the order sets applied to prevent venous thromboembolism (VTE) and treat pneumonia, congestive heart failure (CHF), and acute myocardial infarction (AMI) - testing hospital mortality, readmission, complications, and length of stay (LOS) as health outcomes. Results indicated reduction of acute VTE rates among non-surgical patients in the experimental group, while LOS and complications benefits were inconclusive. Pneumonia patients in the experimental group had lower mortality, readmissions, LOS, and complications rates. CHF patients benefited from order sets in terms of mortality and LOS, while there was no sufficient data to display results for readmissions and complications. Utilization of AMI order sets was insufficient to produce statistically significant results. Results will (1) empower health providers with evidence to justify implementation of order sets due to their effectiveness in driving improvements in health outcomes and efficiency of care and (2) provide researchers with new ideas to conduct health outcomes research.

In this dissertation, we present a collection of manuscripts describing the development of prognostic models designed to assist clinical decision making. This work is motivated by limitations of commonly used techniques to produce accessible prognostic models with easily interpretable and clinically credible results. Such limitations hinder prognostic model widespread utilization in medical practice. Our methodology is based on Rough Set Theory (RST) as a mathematical tool for clinical data anal- ysis. We focus on developing rule-based prognostic models for end-of life care decision making in an effort to improve the hospice referral process. The development of the prognostic models is demonstrated using a retrospective data set of 9,103 terminally ill patients containing physiological characteristics, diagnostic information and neurological function values. We develop four RST-based prognostic models and compare them with commonly used classification techniques including logistic regression, support vector machines, random forest and decision trees in terms of characteristics related to clinical credibility such as accessibility and accuracy. RST based models show comparable accuracy with other methodologies while providing accessible models with a structure that facilitates clinical interpretation. They offer both more insight into the model process and more opportunity for the model to incorporate personal information of those making and being affected by the decision.

In the BDAS project one of the goals is to create a new solution for monitoring bone healing to complement current techniques. Data have been collected in clinical trials from infants treated for Craniosynostosis by a craniotomic surgery. The data are collected with a biomedical sensor based in microwave technology. This sensor could be able to sense changes in the composition of the different tissues in the upper hemisphere of the head, by noticing a difference in the propagation of the microwaves, as the bone injury from the craniectomy heals over time. In this thesis I analyze the validity of a proposed analytical model for the biosensor and extend the clinical data analysis in BDAS project. The validity of the model is analyzed by comparing its outcomes to available measurements from phantoms mimicking living tissues and to numerical simulations. In the data analysis two hypotheses are formulated and tested regarding the location of the measurement points with respect to a positioning grid and the healing over time too. By deriving a set of parameters for each collected dataset in the clinical trials, a distinct pattern was found which shows visible changes over the course of the healing process with this technique.