Christopher C. Yang

Clinical prediction models trained on heterogeneous populations may exhibit disparities in predictive performance across patient subgroups, potentially reflecting differences in data availability, feature completeness, and representation in training data. Existing fairness-aware approaches can introduce performance-fairness trade-offs and commonly optimize group-level disparities without explicitly accounting for their subgroup-specific sources, limiting both their effectiveness and their clinical interpretability. This dissertation makes two methodological contributions. The first is a collaborative learning framework that treats demographic subgroups as clients and aggregates their model parameters using strategies that encode different assumptions about group contribution. Rather than optimizing an explicit group-fairness constraint, the framework preserves and integrates subgroup-specific information to improve equity while largely maintaining subgroup predictive performance. The second contribution is the localized conformal classifier, a neighborhood-adaptive conformal prediction framework for instance-level uncertainty quantification. By calibrating prediction sets according to the density and composition of a patient's local neighborhood in the feature space, the method communicates both the model's prediction and the uncertainty associated with that individual prediction. This information is particularly consequential when predictions inform clinical decision-making. Both contributions are evaluated across three clinical prediction tasks: 30-day mortality in ICU patients with sepsis, treatment non-completion in patients with substance use disorder, and hospital readmission after surgical resection in patients with colorectal cancer. Compared with reweighting and constrained optimization, collaborative learning achieves more favorable fairness-performance trade-offs with less degradation in predictive performance, and its variants consistently appear on the Pareto frontier across all three prediction tasks. A novel difficulty decomposition framework, derived from the localized conformal classifier's calibration scores, distinguishes sources of uncertainty arising from the local neighborhood from those specific to an individual prediction. SHAP attribution reveals how collaborative learning and fairness interventions alter the feature-attribution patterns in the underlying prediction models. Surrogate regression trees trained on the localized conformal classifier's difficulty scores relate neighborhood and instance difficulty to clinical features observable at the point of care. Clinical features remain strongly associated with instance-level difficulty under collaborative learning, whereas these associations weaken substantially after reweighting and are absent in some analyses. Across all three clinical contexts, local outcome heterogeneity is the most consistent driver of neighborhood-level predictive difficulty, while proximity to the decision boundary is the principal driver of instance-level difficulty. Together, these contributions provide a framework for developing and evaluating clinical prediction models that are not only accurate, but also more equitable across patient groups and more informative about the reliability of individual predictions.

Adverse Events (AEs) in clinical trials and healthcare settings are undesired reactions to treatment. These data are high-dimensional, sparse, and interrelated, and their characteristics and interactions may evolve over time. These properties make their analysis very challenging and hence impact effective safety monitoring. Conventional methods often overlook the temporal dependencies, and interrelatedness of AEs, and this limits their ability to fully detect emerging safety signals or identify critical intervention points. We introduce a novel framework, GraphCHASUR, which rooted in temporal network frame. It integrates AE incidence rates which are survival-informed, change point detection methodology, and graph structure to model evolving AE dynamics while capturing AEs relations. In GraphCHASUR, the entire course of treatment is segmented into data-informed snapshots by applying change point detection techniques on weighted AE incidence rates. Once the snapshots are detected, a graph develops within each segment by having AEs as nodes in co-occurrence networks. Two AEs form an edge if they occurred together in at least one patient, these edges will next be weighted by severity/toxicity and timing of the co-occurrence onsets, aggregated across all observed AE pairs. This process produces evolving networks that capture shifts in toxicity patterns while accounting for AEs inter-connections. Once networks are developed, relevant metrics such as centrality measures, node and edge strengths, in addition to community structure are extracted and will be used for downstream predictive modeling of clinical outcomes. We assessed the performance of the network derived features using simulated data. The results demonstrated that the GraphCHASUR-derived variables improved mortality prediction compared to the models with only patients features at baseline, such age, gender, ECOG, and biomarker status. The framework also outperformed models incorporating crude count of AE. In comparison with deep learning approaches, such as LSTM, the performance was quite comparable while the GraphCHASUR features remained highly interpretable. The network divided feature modeling achieving the highest F1 score (0.84) and recall (0.86) in predating patient mortality. Application to a phase 3 colorectal cancer trial (N = 1183) revealed clinically meaningful shifts in toxicity patterns, transitioning from early gastrointestinal symptoms to later hematologic and immune-related toxicities. The identified community structures and safety signals aligned to what we observed in the literature for this particular investigational treatment. By addressing the limitations of static AE summaries with dynamic network-based modeling, GraphCHASUR provides an interpretable, time-aware framework for safety analysis. The framework supports timely detection of safety signals and enables a clearer understanding of toxicity progression in high-dimensional AE data.

This population-based study aimed to quantify fracture risk after ARPIs in PCa patients by pre-existing health conditions. Patients were identified from the SEER-Medicare files who received abiraterone with prednisone (AAP) or enzalutamide (ENZA) between 1/1/2013 and 12/31/2020. Health and fracture history were based on claims one year before ARPI with follow-up through 12/31/2020. The main outcome of the study was the cumulative fracture risk after first date of ARPI. Fine and Gray’s sub-distribution hazard model was used to obtain adjusted relative risks with confounding factors. This study included 10,463 patients (6,037- AAP; 4,426-ENZA). The 3-year fracture risk after ARPI was high, exceeding 25% among those without a prior fracture. Among 1,445 men with a fracture the year before ARPI, 3-year fracture risk exceeded 50%, and remained high (above 44%) despite using bone health agents (BHA). A recent history of fracture was associated with a 2.84-fold fracture risk (aHR: 2.84, 95% CI 2.58-3.12). Pre-existing osteoporosis and a comorbidity score ≥ 2 were associated with 15% (aHR : 1.15, 95% CI 1.03-1.29) and 11% (aHR : 1.11, 95% CI 1.00 to 1.24) higher fracture risks. Bone health agent (BHA) use was associated with a 23% lower fracture risk (aHR: 0.77, 95% CI 0.70-0.83). Fracture risk after ARPI was high, exceeding 44% within 3 years in those with prior fractures despite BHA, suggesting limited benefit in patients with poor bone quality. Early identification and intervention for patients at high risk of fractures is critical.

Cancer staging is critical for patient prognosis and treatment planning, yet extracting pathologic TNM staging from unstructured pathology reports poses a persistent challenge. Existing natural language processing (NLP) and machine learning (ML) strategies often depend on large annotated datasets, limiting their scalability and adaptability. In this study, we introduce two Knowledge Elicitation methods designed to overcome these limitations by enabling large language models (LLMs) to induce and apply domain-specific rules for cancer staging. The first, Knowledge Elicitation with Long-Term Memory (KEwLTM), uses an iterative prompting strategy to derive staging rules directly from unannotated pathology reports, without requiring ground-truth labels. The second, Knowledge Elicitation with Retrieval-Augmented Generation (KEwRAG), employs a variation of RAG where rules are pre-extracted from relevant guidelines in a single step and then applied, enhancing interpretability and avoiding repeated retrieval overhead. We leverage the ability of LLMs to apply broad knowledge learned during pre-training to new tasks. Using breast cancer pathology reports from the TCGA dataset, we evaluate their performance in identifying T and N stages, comparing them against various baseline approaches on two open-source LLMs. Our results indicate that KEwLTM outperforms KEwRAG when Zero-Shot Chain-of-Thought (ZSCOT) inference is effective, whereas KEwRAG achieves better performance when ZSCOT inference is less effective. Both methods offer transparent, interpretable interfaces by making the induced rules explicit. These findings highlight the promise of our Knowledge Elicitation methods as scalable, high-performing solutions for automated cancer staging with enhanced interpretability, particularly in clinical settings with limited annotated data.

Accurate interpretation of clinical narratives is critical for patient care, but the complexity of these notes makes automation challenging. While Large Language Models (LLMs) show promise, single-model approaches can lack the robustness required for high-stakes clinical tasks. We introduce a collaborative multi-agent system (MAS) that models a clinical consultation team to address this gap. The system is tasked with identifying clinical problems by analyzing only the Subjective (S) and Objective (O) sections of SOAP notes, simulating the diagnostic reasoning process of synthesizing raw data into an assessment. A Manager agent orchestrates a dynamically assigned team of specialist agents who engage in a hierarchical, iterative debate to reach a consensus. We evaluated our MAS against a single-agent baseline on a curated dataset of 420 MIMIC-III notes. The dynamic multi-agent configuration demonstrated consistently improved performance in identifying congestive heart failure, acute kidney injury, and sepsis. Qualitative analysis of the agent debates reveals that this structure effectively surfaces and weighs conflicting evidence, though it can occasionally be susceptible to groupthink. By modeling a clinical team's reasoning process, our system offers a promising path toward more accurate, robust, and interpretable clinical decision support tools.

Healthcare systems generate diverse multimodal data, including Electronic Health Records (EHR), clinical notes, and medical images. Effectively leveraging this data for clinical prediction is challenging, particularly as real-world samples often present with varied or incomplete modalities. Existing approaches typically require complete modality data or rely on manual selection strategies, limiting their applicability in real-world clinical settings where data availability varies across patients and institutions. To address these limitations, we propose MoE-Health, a novel Mixture of Experts framework designed for robust multimodal fusion in healthcare prediction. MoE-Health architecture is specifically developed to handle samples with differing modalities and improve performance on critical clinical tasks. By leveraging specialized expert networks and a dynamic gating mechanism, our approach dynamically selects and combines relevant experts based on available data modalities, enabling flexible adaptation to varying data availability scenarios. We evaluate MoE-Health on the MIMIC-IV dataset across three critical clinical prediction tasks: in-hospital mortality prediction, long length of stay, and hospital readmission prediction. Experimental results demonstrate that MoE-Health achieves superior performance compared to existing multimodal fusion methods while maintaining robustness across different modality availability patterns. The framework effectively integrates multimodal information, offering improved predictive performance and robustness in handling heterogeneous and incomplete healthcare data, making it particularly suitable for deployment in diverse healthcare environments with heterogeneous data availability.

Background: Older adults with advanced prostate cancer and type 2 diabetes mellitus are underrepresented in trials of androgen receptor pathway inhibitors. This study examined changes in unplanned hospitalization rates in patients receiving androgen receptor pathway inhibitors by type 2 diabetes mellitus status and assessed if unplanned hospitalization varies according to androgen receptor pathway inhibitors. Methods: This population-based study of advanced prostate cancer patients aged older than 66 years used Surveillance, Epidemiology, and End Results–Medicare data. Prepost androgen receptor pathway inhibitor initiation changes and androgen receptor pathway inhibitor differences in unplanned hospitalization rates were estimated by adjusted incidence rate ratio with considerations for interactions between period, androgen receptor pathway inhibitor, and type 2 diabetes mellitus status. Linear contrasts were used to estimate and test conditional incidence rate ratios. Tests were 2-sided, and a P value less than .05 was considered statistically significant. Results: The study included 12 240 patients: 3160 (25.8%) with type 2 diabetes mellitus, 7191 (58.8%) received abiraterone acetate with prednisone, and 5049 (41.2%) received enzalutamide. Unplanned hospitalization rates increased after androgen receptor pathway inhibitor initiation by 65% among patients with type 2 diabetes mellitus complications (adjusted incidence rate ratio = 1.65, 95% confidence interval [CI] = 1.37 to 1.98) and 109% in nondiabetics (adjusted incidence rate ratio = 2.09, 95% CI = 1.94 to 2.26). Among patients with type 2 diabetes mellitus without complications, the increase in unplanned hospitalization rates depended on the androgen receptor pathway inhibitor initiated: 103% after abiraterone acetate with prednisone (adjusted incidence rate ratio = 2.03, 95% CI = 1.70 to 2.43) and 47% after enzalutamide (adjusted incidence rate ratio = 1.47, 95% CI = 1.21 to 1.80) and a 38% greater increase in unplanned hospitalization rates after abiraterone acetate with prednisone than enzalutamide (ratio of abiraterone acetate with prednisone adjusted incidence rate ratio divided by enzalutamide adjusted incidence rate ratio = 1.38, 95% CI = 1.06 to 1.80). Conclusions: All patients had higher unplanned hospitalization rates after androgen receptor pathway inhibitor. Our findings highlight the importance of using real-world data to better understand the interplay between preexisting health conditions and treatment outcomes, a critical step toward precision medicine.

Clinical reasoning is a core competency for physicians, and deficiencies in this area can seriously affect patient care. Step 3 of the United States Licensing Exam, which all allopathic physicians must pass for licensure, assesses clinical reasoning through its Computer-based Case Simulations (CCS). CCS allows test-takers to demonstrate aspects of clinical reasoning by managing a patient in a simulated environment. Each test-taker's actions are recorded in a transaction log, providing a rich source of data for investigating various aspects of clinical reasoning, particularly data-gathering skills. The current study examines transaction lists for evidence of data-gathering skills to provide feedback to test-takers, helping them improve their clinical reasoning skills. This study employed data mining and process data methods to identify common data-gathering patterns in two CCS cases. Analytic methods included creating N-grams, applying weights, and examining how test takers employed them with k-means clustering. Because of the importance of time in patient care, sequence similarity calculations included common actions, the time elapsed between those actions, and the overall time spent on the case. The sequences were also analyzed to uncover relationships with demographic variables. Results showed that, despite content differences, both cases shared the most common eight N-grams, and the order of actions in the sequences matched the order of physical exam components that are displayed with checkboxes. Combining action and time information showed that test takers have common sequences, but those sequences are relatively unimportant in the context of their total actions. As anticipated, group differences were evident only in comparisons between U.S. and international medical graduates. These results support the proposition that, since the most frequent data-gathering patterns related to how test-takers selected physical examination components from a menu using checkboxes, test-takers may be demonstrating test-taking strategies, rather than data-gathering skills. Recommendations for future research include modifying real and simulated timestamp recording to more accurately reflect real-world constraints and resources, and encouraging thoughtful data-gathering actions. This study also informs future CCS design efforts to collect and record test-taker actions more accurately, enhancing feedback aimed at improving clinical reasoning skills.