Knowledge about the biology, etiology, staging, and treatment of a disease can be found in a growing and disparate set of sources, including observational clinical data, scientific literature, and clinical guidelines. However, effectively utilizing these sources to support clinical decision making remains a challenge. One of the challenges stems from the need to integrate knowledge from multiple sources that have heterogeneous representation, while locating and appraising evidence relevant to an individual patient can also be an issue. The objective of this dissertation is the formulation of an intermediate representation that logically consolidates and standardizes knowledge fragments across these sources, along with the definition of operators on this representation that generate, in a principled manner, the needed elements to facilitate answering clinical queries to support evidence-based medicine (EBM). The contributions of this work are: (1) a standardized representation, called Phenomenon-Centric Data Model Plus (PCDM+), which adopts the probabilistic entity-relationship model and captures and structures information about a disease drawn from scientific literature and patient records, emphasizing population-level observations and evidence; and (2) a set of operators that retrieve and infer information about individual patients from the PCDM+ to inform clinical queries. This work is demonstrated and evaluated in the domain of intracranial aneurysm.

Advances in next-generation sequencing coupled with comparative analyses have hadtremendous implication for oncology. Differentially expressed genomic features have revealed molecular pathways, oncogenic drivers, and resistance patterns targeted by drug development leading to significant decreases in cancer mortality. The introduction of high throughput multiomics, including transcriptomics, epigenomics, and proteomics, promised to scale translational innovation. Increasing feature complexity, however, cannot be comparatively resolved, and many efforts in this space have failed due to inconclusive or conflicting results. Computational systems biology and functional genomics have proposed dynamic integration of multiple molecular pathway models for data harmonization, yet the requirement for complete information has biased discovery. Similarly, the incorporation of probabilistic approaches that constrain features may obscure incremental biologic effects. This problem is exemplified by the several dozen cancer genomic biomarkers and models from peer-reviewed high impact publications that do not meet statistical significance when applied beyond their training and validation datasets.

This dissertation seeks to employ a scientifically rigorous process of evaluation using iterativemodeling to benchmark multiomic comparisons. First, this research develops a framework for characterizing multimodal data based on structural and functional information. Second, this research benchmarks similarity metrics using varied data structures, offering techniques to reveal key biologic differences using probabilistic modeling (hierarchical clustering). This framework is then applied to neoepitope prediction incorporating human leukocyte antigen-B supertypes and used to resolve previously inconclusive and conflicting results, including the difference in survival based on B44 supertype in patients with non-small cell lung cancer (NSCLC) and melanoma treated with immune checkpoint blockade (ICB).

Ultimately, this dissertation advances our understanding of the interaction between featureselection and power in multiomic analyses and offers recommendations to enhance the reliability of these investigations.

A tenet of precision medicine is the ability to predict a clinical response or outcome for a given individual. The creation of predictive disease models as a part of this evidence-based process has increased in recent years as electronic medical data and machine learning techniques have grown in popularity. A number of methods are available for testing internal validity when developing models. However, the performance of developed disease models must also be tested in external populations. This practice of external validation, or transportability, can help determine the extent to which information in a predictive model can be applied generally across population samples. Model performance can suffer when applied in external settings because populations have inherent differences or data was collected with different practices. A number of impediments, particularly in reporting, continue to hinder the widespread application of transportability methods. This dissertation addresses these concerns using a methodology for improving the classification of models and evaluating a proposed technique for partially adjusting problematic models. A set of simulations are proposed, taking advantage of a minimal set of published values to extend reported findings through bootstrap analysis. Interpretations derived from this method are able to guide the assignment and selection of models by transport levels. Causal transportability analysis, a previously proposed transportability theory in graphical models, is also evaluated for use in partial adjustment scenarios. The resulting process allows for evaluation of model transportability with minimal information and for assessing model adjustments. These investigations serve as useful tools for future transportability analysis. In addition, the results of this work introduce new items that can be added to reporting guidelines and support current trends to establish improved validation frameworks.

Public health surveillance is a basic function of an effective modern state. Yet this surveillance is a deeply political phenomenon; the specifics of which data are collected, about whom, and the speed and format of how they are aggregated and disseminated are shaped by social and political priorities. Surveillance, then, can serve as a powerful tool for tracking social inequalities, or alternatively, a mechanism by which they remain invisible. Here, using a series of case studies, I develop and implement a model of public health surveillance that is rapid, critically-applied, mixed-methods in nature, and especially focused on the measurement of emerging or shifting health inequalities. A key aim is the advancement of new techniques to fill data gaps, especially by drawing on mixed methods research. While public health surveillance has traditionally been nearly synonymous with quantitative methods, here I seek to combine data science approaches with qualitative social science techniques, especially ethnography. A nuanced data science approach especially drawing on novel data sources, can offer robust prediction in a real-time fashion, which can further yield detailed insights about emerging health disparities. Similarly, ethnography represents a manner of accessing common-sense, on-the-ground logics around social dynamics and health disparities that can often be surprisingly difficult to characterize with traditional survey methodologies and other quantitative approaches. This model of public health surveillance is also critically-applied, meaning grounded in social theory (but not fully subservient to it) and ultimately oriented towards affecting real change. In this fashion, critically-applied surveillance work can be thought of as a first step in assessing, interrogating, and challenging the structural violence that pervades modern societies. In sum, in the 14 case studies contained within this body of work, I develop and implement a critically-applied and inequalities-oriented, heavily mixed-methods approach to public health surveillance, which can be employed widely to improve inequality monitoring across many areas of public health and clinical practice. This model is applied here especially to the study of shifting trends in the North American overdose crisis, studying clandestine, illicit, and complex dynamics that are difficult to assess with traditional approaches.

Glioblastoma multiforme (GBM) is the most common type of primary brain tumor, characterized by a short survival period after diagnosis. As with most other cancers, treatment and follow-up decisions are made largely based on observed changes in tumor size and appearance during imaging studies.

The quantification of tumor measurements is problematic due to the systematic variability introduced while attempting to quantify tumor characteristics in uncertain regions. This issue is primarily observed around the tumor boundary, where it is often hard to differentiate whether a given region is part of the tumor (e.g., active, necrotic, edema, etc.) or part of normal brain tissue (e.g., grey matter, white matter). This problem has significant implications because this uncertainty can affect ensuing quantitative/computational analyses. Current approaches for the segmentation of glioblastoma multiforme still face multiple challenges, often failing to consistently identify the tumor region so as to be clinically useful and reliable; moreover, these different techniques tend to produce results that differ significantly from each other (i.e., measurement variability).

To address these problems, this dissertation describes a framework to help characterize factors that influence variability in brain tumor boundaries and to optimize their performance through methods that calculate an estimate of expected variability arising from different automated segmentation approaches, setting the bases for the development of better knowledge-based methods. Additionally, a novel automated method was developed to generate more robust brain tumor segmentations by taking into consideration the inherent variability of brain tumors and statistical priors that provide context-relevant information about the different brain and tumor tissues.

Altogether, this dissertation project provides further understanding of the sources of variability that arise in GBM across different image analysis methodologies and the integration of these insights into the development of tumor variability maps that can provide a better characterization of tumors.

Identifying high-risk lung cancer individuals at an early disease stage is the most effective way of improving survival. The landmark National Lung Screening Trial (NLST) demonstrated the utility of low-dose computed tomography (LDCT) imaging to reduce mortality (relative to x-ray screening). As a result of the NLST and other studies, imaging-based lung cancer screening programs are now being implemented. However, LDCT interpretation results in a high number of false positives. A set of dynamic Bayesian networks (DBN) were designed and evaluated to provide insight into how longitudinal data can be used to help inform lung cancer screening decisions.

The LDCT arm of the NLST dataset was used to build and explore five DBNs for high-risk individuals. Three of these DBNs were built using a backward construction process, and two using structure learning methods. All models employ demographic, smoking status, cancer history, family lung cancer history, exposure risk factors, comorbidities related to lung cancer, and LDCT screening outcome information. Given the uncertainty arising from lung cancer screening, a cancer state-space model based on lung cancer staging was utilized to characterize the cancer status of an individual over time. The models were evaluated on balanced training and test sets of cancer and non-cancer cases to deal with data imbalance and overfitting.

Results were comparable to expert decisions. The average area under the curve (AUC) of the receiver operating characteristic (ROC) for the three intervention points of the NLST trial was higher than 0.75 for all models. Evaluation of the models on the complete LDCT arm of the NLST dataset (N = 25, 486) demonstrated satisfactory generalization. Consensus of predictions over similar cases is reported in concordance statistics between the models’ and the physicians’ predictions. The models’ predictive ability with respect to missing data was also evaluated with the sample of cases that missed the second screening exam of the trial (N = 417). The DBNs outperformed comparison models such as logistic regression and naïve Bayes.

The lung cancer screening DBNs demonstrated high discrimination and predictive power with the majority of cancer and non-cancer cases.