Monitoring Patient Reported Outcomes and Neurophysiologic Signals by Integrating Clinical Records with High-Dimensional Digital Data.

Modern healthcare generates heterogeneous data streams that span electronic health records (EHR), patient-reported outcomes (PROs), and high-frequency neurophysiologic signals. This thesis develops and applies an integrated statistical framework that combines Frequentist and Bayesian approaches to monitor patient outcomes across these modalities while preserving interpretability and rigorous uncertainty quantification. Chapter 2 establishes the methodological foundation in an infectious-disease surveillance setting. We analyze linked clinical data using generalized linear models with penalization and Bayesian counterparts for shrinkage and partial pooling; we extend to correlated data via generalized linear mixed models, generalized estimating equations, and Bayesian hierarchical models; we diagnose misspecification and overdispersion for binary and count endpoints; address missingness with multiple imputation and fully Bayesian joint modeling; and we evaluate policy effects using interrupted time series with classical and Bayesian ARIMA, as well as structural time series models. Together, these components illustrate how complementary paradigms support transparent estimation, probabilistic prediction, and principled sensitivity analysis in real-world public-health applications. Chapter 3 turns to high-dimensional intracranial EEG. We perform physiologically guided feature extraction using FOOOF to parameterize each power spectrum into aperiodic and oscillatory components, then rank regional stability via within-subject coefficients of variation using Friedman tests and Kendall’s . Hierarchical Bayesian models (fitted in R/Stan) quantify region- and time-of-day effects while accounting for subject and channel heterogeneity. Posterior predictive checks and PSIS–LOO confirm model adequacy. Results reveal robust stratification of spectral-stability across regions and a positive residual correlation between aperiodic offset and exponent, supporting joint modeling. Chapter 4 addresses perioperative neurophysiology in pediatric anesthesia using the Patient State Index (PSI). We detect abrupt state changes with a two-tier pipeline—global segmentation via PELT and patient-wise Bayesian structural time series—then summarize instability with a phase-normalized, probability-based Variability Ratio Index (VARI). Population-averaged inference uses a logistic GEE with a data-driven working correlation and robust standard errors. Instability concentrates around surgical transitions, whereas other static characteristics (e.g., gender or ethnicity) have little effect. In contrast, children of older ages and larger body sizes (each considered independently) showed significantly more stable PSI dynamics. Across domains, this thesis demonstrates a reproducible, R-first workflow that unifies interpretable feature engineering with hierarchical modeling and probability-focused monitoring. This framework is potentially generalizable to other multi-modal clinical contexts where rigorous uncertainty quantification, biological plausibility, and operational interpretability are essential for data-driven decision-making.