Quantifying data quality for clinical trials using electronic data capture.

BACKGROUND: Historically, only partial assessments of data quality have been performed in clinical trials, for which the most common method of measuring database error rates has been to compare the case report form (CRF) to database entries and count discrepancies. Importantly, errors arising from medical record abstraction and transcription are rarely evaluated as part of such quality assessments. Electronic Data Capture (EDC) technology has had a further impact, as paper CRFs typically leveraged for quality measurement are not used in EDC processes. METHODS AND PRINCIPAL FINDINGS: The National Institute on Drug Abuse Treatment Clinical Trials Network has developed, implemented, and evaluated methodology for holistically assessing data quality on EDC trials. We characterize the average source-to-database error rate (14.3 errors per 10,000 fields) for the first year of use of the new evaluation method. This error rate was significantly lower than the average of published error rates for source-to-database audits, and was similar to CRF-to-database error rates reported in the published literature. We attribute this largely to an absence of medical record abstraction on the trials we examined, and to an outpatient setting characterized by less acute patient conditions. CONCLUSIONS: Historically, medical record abstraction is the most significant source of error by an order of magnitude, and should be measured and managed during the course of clinical trials. Source-to-database error rates are highly dependent on the amount of structured data collection in the clinical setting and on the complexity of the medical record, dependencies that should be considered when developing data quality benchmarks.

Computational Journalism: from Answering Question to Questioning Answers and Raising Good Questions

Our media is saturated with claims of ``facts'' made from data. Database research has in the past focused on how to answer queries, but has not devoted much attention to discerning more subtle qualities of the resulting claims, e.g., is a claim ``cherry-picking''? This paper proposes a Query Response Surface (QRS) based framework that models claims based on structured data as parameterized queries. A key insight is that we can learn a lot about a claim by perturbing its parameters and seeing how its conclusion changes. This framework lets us formulate and tackle practical fact-checking tasks --- reverse-engineering vague claims, and countering questionable claims --- as computational problems. Within the QRS based framework, we take one step further, and propose a problem along with efficient algorithms for finding high-quality claims of a given form from data, i.e. raising good questions, in the first place. This is achieved to using a limited number of high-valued claims to represent high-valued regions of the QRS. Besides the general purpose high-quality claim finding problem, lead-finding can be tailored towards specific claim quality measures, also defined within the QRS framework. An example of uniqueness-based lead-finding is presented for ``one-of-the-few'' claims, landing in interpretable high-quality claims, and an adjustable mechanism for ranking objects, e.g. NBA players, based on what claims can be made for them. Finally, we study the use of visualization as a powerful way of conveying results of a large number of claims. An efficient two stage sampling algorithm is proposed for generating input of 2d scatter plot with heatmap, evalutaing a limited amount of data, while preserving the two essential visual features, namely outliers and clusters. For all the problems, we present real-world examples and experiments that demonstrate the power of our model, efficiency of our algorithms, and usefulness of their results.

Persistent Homology Analysis of Brain Artery Trees.

New representations of tree-structured data objects, using ideas from topological data analysis, enable improved statistical analyses of a population of brain artery trees. A number of representations of each data tree arise from persistence diagrams that quantify branching and looping of vessels at multiple scales. Novel approaches to the statistical analysis, through various summaries of the persistence diagrams, lead to heightened correlations with covariates such as age and sex, relative to earlier analyses of this data set. The correlation with age continues to be significant even after controlling for correlations from earlier significant summaries.

Bayesian Learning with Dependency Structures via Latent Factors, Mixtures, and Copulas

Bayesian methods offer a flexible and convenient probabilistic learning framework to extract interpretable knowledge from complex and structured data. Such methods can characterize dependencies among multiple levels of hidden variables and share statistical strength across heterogeneous sources. In the first part of this dissertation, we develop two dependent variational inference methods for full posterior approximation in non-conjugate Bayesian models through hierarchical mixture- and copula-based variational proposals, respectively. The proposed methods move beyond the widely used factorized approximation to the posterior and provide generic applicability to a broad class of probabilistic models with minimal model-specific derivations. In the second part of this dissertation, we design probabilistic graphical models to accommodate multimodal data, describe dynamical behaviors and account for task heterogeneity. In particular, the sparse latent factor model is able to reveal common low-dimensional structures from high-dimensional data. We demonstrate the effectiveness of the proposed statistical learning methods on both synthetic and real-world data.