HMS Theses and Dissertations

The glymphatic system, responsible for metabolic waste clearance in the central nervous system, has emerged as a critical pathway in the study of neurodegeneration. Enlarged perivascular spaces (ePVS), visible on standard MRI, represent one of the few accessible imaging biomarkers of glymphatic activity in humans. However, existing approaches to longitudinal glymphatic assessment have relied on coarse-grained summary metrics or highly localized diffusion-based indices, leaving the fine-grained spatiotemporal dynamics of ePVS distribution largely unexplored. In this work, we analyze a 7T MRI dataset of 30 healthy subjects scanned at three time points: before sleep deprivation (Baseline), immediately following 24 hours of total sleep deprivation, and 72 hours into recovery. Longitudinal ePVS changes are investigated at four biological scales: whole-brain, regional, sub-regional, and single-voxel. In our first specific aim, we show that coarse-grained ePVS burden metrics do not reveal significant cohort-level responses to sleep deprivation. However, projecting subject-level trajectories into a novel transition space uncovers a robust elastic behavior of the glymphatic system: perturbations in ePVS burden induced by sleep deprivation tend to resolve following recovery, across both segmentation modalities and most brain regions. In our second specific aim, we introduce two novel computational frameworks. A mathematically principled SE(3) mid-space registration pipeline is developed and validated, ensuring unbiased alignment of longitudinal scan pairs. Building on this, a Deformation-Based Morphometry pipeline and an Optimal Transport framework produce single-voxel maps of volumetric changes in the brain parenchyma and ePVS mass redistribution respectively, summarized into fingerprint representations. Cross-transition correlation analysis reveals patterns consistent with the elastic glymphatic response and suggests a directional association between parenchymal adaptations during sleep deprivation and ePVS dynamics during recovery. These results demonstrate the potential of spatially resolved longitudinal ePVS analysis for advancing our understanding of glymphatic dynamics.

Assessment of anesthesiology crisis management remains challenging due to the multidimensional and context-dependent nature of clinical performance. Competency-based medical education (CBME) requires robust, interpretable measures of competence that extend beyond knowledge or isolated skill domains. We developed a multidimensional Proficiency Index (PI) integrating four domains: workflow, crisis resource management (CRM), clinical skills, and knowledge to provide a comprehensive assessment of crisis management performance. Methods: The PI was developed using a modified Delphi process to establish content validity and define domain structure and weighting. It was subsequently evaluated in a prospective cross-sectional study of anesthesiology residents (CA1-CA3) completing two standardized high-fidelity simulation scenarios (intraoperative anaphylaxis and unexpected difficult airway) and a 45-item knowledge assessment. Each simulation performance was independently rated by two faculty assessors. Descriptive statistics were calculated at the participant level. Inter-rater reliability was assessed using intraclass correlation coefficients (ICC[2,1]) and Bland–Altman analysis. Measurement precision was evaluated using the standard error of measurement (SEM) and minimal detectable change (MDC95). Differences across training levels were analyzed using ANOVA and mixed-effects models with Tukey-adjusted pairwise comparisons. Effect sizes were calculated using Hedges’ g. Results: Forty-eight residents participated (22 CA1, 14 CA2, 12 CA3). Mean PI scores increased progressively across training levels (CA1: 39.7±5.0; CA2: 59.4±3.4; CA3: 71.8±5.2; p.001). Performance domains (workflow, CRM, clinical skills) demonstrated consistent separation across training levels, whereas knowledge demonstrated minimal difference between CA2 and CA3. Inter-rater reliability was excellent for the PI (ICC 0.96, 95% CI 0.92–0.98). Measurement precision was acceptable (SEM 2.9; MDC95 8.0). Mixed-effects models confirmed training level as a significant predictor of PI and performance domains, but not knowledge at higher levels. Effect sizes were large to very large for the PI and performance domains. Conclusions: The Proficiency Index provides a reliable, multidimensional measure of anesthesiology crisis management performance that discriminates across training levels. Divergent trajectories across domains highlight the importance of integrated assessment frameworks in CBME and support the use of domain-specific metrics to guide individualized training and evaluation and inform competency-based progression decisions.

ABSTRACT 1: Accuracy of Tachypnea to Predict Mortality in Young Infants: A Systematic Review and Meta-analysis

Context: Tachypnea in young infants is commonly used in clinical assessment to identify high-risk infants, however the optimal respiratory rate threshold is unclear. Objective: To systematically review the evidence on the accuracy of different thresholds of tachypnea to predict mortality in infants 0-59 days. Data Sources: MEDLINE, Embase, CINAHL, Global Index Medicus, and CENTRAL. Study Selection: Studies reporting the accuracy of tachypnea (≥60, ≥70, or ≥80 breaths per minute (bpm)) to predict mortality in infants 0-59 days. Data Extraction: We followed Cochrane methods for study screening, data extraction, and quality assessment using the Newcastle-Ottawa and QUAPAS scales. Results: Of 7641 studies identified, 8 were included. Tachypnea with threshold of ≥60 bpm had an overall sensitivity of 57% (95% CI: 23%-86%) and specificity of 54% (95% CI: 28%-79%) for predicting future mortality in infants by 59 days of age (6 studies, N = 3819 infants). The pooled odds ratio for the association between tachypnea ≥60 bpm and mortality was 2.00 (95% CI: 1.39-2.87, 7 studies, N = 7122 infants). Tachypnea ≥70 bpm had a 4.6-fold higher odds of mortality (95% CI: 1.60-13.00, 1 study, N = 6924 infants). There was limited data on other thresholds and timing of mortality (early versus late neonatal periods). Limitations: Heterogeneity and the low number of studies limited the evidence. Conclusions: Tachypnea was associated with significantly higher odds of mortality in young infants. Overall sensitivity and specificity were low. Further research is needed to determine the optimal threshold and the accuracy for different postnatal ages.

ABSTRACT 2: Accuracy of Infant Clinical Signs to Predict Neonatal Mortality in Rural Bangladesh Background: Clinical signs provide early warning of illness in neonates at high risk of dying in community settings in low- and middle- income countries (LMICs). There is limited validation of current clinical signs and/or algorithms to predict neonatal mortality in LMICs. Objectives: 1) To examine the diagnostic accuracy of existing clinical sign algorithms to predict neonatal mortality, and 2) To determine the association of individual clinical signs with neonatal mortality and their diagnostic accuracy in predicting neonatal death. Methods: We conducted a secondary analysis of a birth cohort in Sylhet, Bangladesh (NCT01572532). Of 7788 live births, 6251 newborns had a clinical examination during a community health worker home visit within 3 days of birth, and 5289 were followed up until 28 days of age with available vital status. We validated existing newborn clinical sign algorithms identified from our prior systematic review (Shafiq 2024), including four Integrated Management of Childhood Illness (IMCI)-like checklist algorithms and the Score of Essential Neonatal Symptoms and Signs (SENSS). We also estimated the odds ratios (ORs) for neonatal mortality associated with individual clinical signs exploring optimal thresholds using univariable logistic regression; and calculated sensitivity and specificity of individual clinical signs for identifying neonatal death. Results: The WHO Young Infant Study 7-sign modification Z checklist algorithm had the sensitivity of 66.7% (95% CI: 56.6% to 75.7%) and specificity of 68.6% (95% CI: 67.3% to 69.9%) for predicting future mortality. The SENSS algorithm had an Area Under the Curve (AUC) of 75.9% (95% CI: 69.9% to 82.0%), calibration intercept of -1.21 (95% CI: -1.69 to -0.72), and calibration slope of 0.94 (95% CI: 0.78 to 1.11). Among individual clinical signs at birth, fast breathing ≥80 breaths per minute (bpm), fever ≥38.5˚C, and hypothermia .5˚C showed the strongest association with neonatal death, with ORs were 44.4 (95% CI: 13.8 to 142.6) for fast breathing ≥80 bpm, 39.3 (95% CI: 3.5 to 441.3) for fever ≥38.5˚C, and 30.3 (95% CI: 17.1 to 53.9) for hypothermia .5˚C. Conclusions: Four IMCI-like checklist algorithms had low sensitivity and high specificity to predict neonatal mortality. The Score of Essential Neonatal Symptoms and Signs (SENSS) algorithm reported fair discrimination to identifying neonates at high risk of dying. Further research is needed to optimize clinical sign algorithms to identify high-risk infants in different age groups and settings.

Radiology reports are the primary medium through which radiologists communicate, conveying critical clinical information in natural language. AI holds the potential to both generate and analyze these reports, yet two key challenges persist. First, as AI systems increasingly generate radiology reports, evaluating their accuracy remains an open problem. Second, a fundamental disconnect exists between the findings described in radiology reports and their corresponding locations in the imaging studies, limiting referring physicians, patients, and trainees who must interpret findings without explicit visual guidance. This thesis addresses these challenges through three interconnected contributions spanning report evaluation, report visualization, and clinical application.

First, we introduce CRIMSON, a clinically grounded evaluation metric for radiology report generation, along with two new benchmarks: RadJudge, a 30-case clinical judgment test suite, and RadPref, a 100-case radiologist preference benchmark. CRIMSON incorporates patient context and weights errors by clinical significance when comparing generated against reference reports. Validated against radiologist error counts from the ReXVal dataset, RadJudge, and RadPref, CRIMSON achieves stronger alignment with expert judgment than prior metrics.

Second, we introduce ReXGroundingCT, the first publicly available dataset linking free-text radiology findings to manually annotated 3D segmentation masks in chest CT scans. Designed through a multi-stage annotation pipeline, the dataset comprises 3,142 scans and 8,028 segmented findings. We benchmark state-of-the-art text-prompted segmentation models, demonstrating that current approaches fall substantially short of clinical utility, even after fine-tuning. Subsequently, we release the dataset, along with a public leaderboard to drive continued progress on this task.

Third, we present RadGame, an AI-powered platform for radiology education that brings together both report evaluation and grounding as a concrete downstream application. The platform teaches two core skills: localizing findings through interactive bounding-box annotation, and writing radiology reports with automated structured feedback. In a prospective multi-institutional study with 18 medical students, participants using RadGame achieved a 68% improvement in localization accuracy and a 31% improvement in report-writing scores, outperforming traditional passive learning methods.

Together, these contributions address the radiology AI pipeline from multiple angles: evaluating generated reports against clinical standards, spatial grounding of reports in imaging, and applying both to advance radiology education.

Introduction: Assessment of resident competency in obstetric neuraxial procedures is limited by the absence of clearly defined, observable criteria within existing competency−based frameworks. Although ACGME milestones and entrustable professional activities (EPAs) emphasize performance, they provide limited guidance on what competent procedural execution actually looks like in clinical practice. This lack of specificity contributes to variability in teaching, reliance on a hidden curriculum, and inconsistent assessment.

Methods: A three−round Delphi study with one modification was conducted to develop expert−informed assessment criteria for obstetric neuraxial procedures. Twenty−seven U.S. experts in obstetric anesthesiology were recruited through purposive sampling. Panelists rated statements derived from ACGME Milestones 2.0, ACGME Milestones 2.0, Supplemental Guide, Tanaka’s neuraxial EPA, and three expert interviews using a 7−point Likert scale. Consensus was defined a priori as ≥85% agreement. Investigator−led iterative revisions were guided by expert feedback to improve clarity and alignment with competency at graduation.

Results: Consensus increased across rounds from 29.17% (Round 1) to 82.35% (Round 2), to 100.0% in Round 3. Ratings were consistently clustered in the “essential” range, with minimal disagreement regarding clinical importance. Variability in early rounds was primarily related to wording and expectations for graduation−level performance rather than disagreement about underlying core clinical importance. The final set of criteria reflects convergence on key domains including patient selection, communication, procedural technique, troubleshooting, and clinical decision−making.

Conclusion: This study provides a set of expert−informed, observable assessment criteria that offer a more practical and clinically grounded definition of competency for obstetric neuraxial anesthesia. Beyond standardizing assessment, these criteria may support more consistent teaching, targeted feedback, and progression from competence toward higher levels of performance. This framework may serve as a model for defining procedural competency across anesthesiology and other medical specialties.

Project 1 Abstract

Title: Association of Diabetes with Changes in Blood Pressure during Hemodialysis:A Secondary Analysis of the Frequent Hemodialysis Network Daily Trial.

Introduction Diabetes mellitus is a common cause of kidney failure and is often complicated by autonomic neuropathy, which may have implications for blood pressure (BP) homeostasis during hemodialysis (HD).

Methods In this post hoc analysis of the Frequent Hemodialysis Network (FHN) Daily Trial, we used random effects Poisson and linear regression models to estimate the association of diabetes (versus not) with intra-dialytic hypotension (IDH) and peri-dialytic BP parameters, respectively. We tested for differential associations according to the randomized treatment (6/week vs 3/week HD) and pre-HD systolic BP.

Results Of the 244 patients with intra-dialytic BP data, 100 (41%) had diabetes at baseline. The mean age was 51 ± 14 years; 39% were female. In adjusted models, diabetes (vs. not) was associated with a 93% higher risk of developing IDH (IRR 1.93; 95% CI 1.26, 2.95). There was no evidence that the randomized treatment assignment modified the association between diabetes and IDH (P-interaction=0.32), but more potent associations were noted among those with higher pre-HD systolic BP (P-interaction .001). Diabetes (vs. not) was associated with a lower adjusted nadir intra-HD BP (-4.2; 95%CI -8.3, -0.2 mmHg), but not with the pre- or post-HD systolic BP.

Conclusions Among participants of the FHN Daily trial, patients with diabetes had a higher risk of intra-dialytic hypotension and lower nadir intra-HD systolic BP than patients without diabetes, even when undergoing HD up to 6 times per week.

Project 2 Abstract

Title: Association of Diabetes with Heart Rate Variability during Hemodialysis: Insights from the Frequent Hemodialysis Network Daily Trial.

Introduction Autonomic dysfunction is common among patients with diabetes receiving hemodialysis (HD). We wished to explore the association of diabetes with heart rate variability (HRV; a surrogate of autonomic dysfunction) and whether HRV mediates the association of diabetes with intra-dialytic hypotension (IDH).

Methods In this secondary analysis of the Frequent Hemodialysis Network Daily Trial, we performed: 1) random effects linear regression to estimate the association of diabetes with log-transformed low-frequency power [LF, proxy of sympathetic activity], high-frequency power [HF, proxy of parasympathetic activity], ratio of LF/HF (proxy for sympathovagal balance), and standard deviation of the normal-to-normal R-R interval [SDNN] measured at baseline and 12-months); 2) linear regression to explore the association of diabetes with changes in HRV parameters over 12 months. Models were adjusted for age, sex, designated race, height, access type, HD vintage, history of heart failure, pre-HD systolic BP, heart rate, ultrafiltration rate, hemoglobin, serum albumin, beta-blocker use, calcium channel blocker use, diuretic use, left ventricular mass, and randomized treatment assignment.

Results Of the 198 patients without baseline atrial fibrillation, 82 (41%) had self-reported diabetes. In adjusted random effects models, diabetes (vs. no diabetes) was associated with lower SDNN –18% (95%CI –27, –9) on a per session basis. The presence of diabetes was not associated with differences in LF 7% (95%CI –20, 43), HF 10% (95%CI –10, 33), or LF/HF –4% (95%CI –19, 14). Diabetes (vs. no diabetes) was not associated with a change from baseline to 12 months in any HRV parameter. SDNN did not attenuate the observed association of diabetes with IDH.

Conclusions Among participants in the FHN Daily Trial, diabetes (vs. no diabetes) was associated with 18% lower SDNN. The association of diabetes with IDH did not appear to be mediated by SDNN. The reasons for higher rates of IDH in patients with diabetes remain elusive.

Background: Pediatric Hospital Medicine (PHM) is one of the newest pediatric subspecialties, and the number of PHM fellowship programs continues to grow rapidly. Many PHM attendings have limited experience supervising fellows. Currently, the PHM fellowship program at Boston Children’s Hospital lacks a structured curriculum to prepare attendings for clinical supervision. This study aims to conduct an organizational needs assessment among PHM attendings and fellows to identify faculty development needs related to fellow supervision. This assessment represents the second step in Kern’s six-step approach to curriculum development, as incorporating faculty and fellow input is essential for designing a curriculum that meets attendings’ needs. Methods: This cross-sectional study utilized three surveys targeting: (1) attendings at the main academic site (BCH), (2) attendings at the community site (SSH), and (3) current and graduated PHM fellows. The attending surveys assessed overall preparedness for fellow supervision, key topics to include in a curriculum, and overall interest in a curriculum. Fellow surveys explored their experiences working with attendings. Surveys were distributed anonymously via Qualtrics. Data analysis included descriptive statistics and a modified qualitative analysis using the framework method for free-response questions. Results: Results from BCH and SSH surveys varied given the inherent differences in training environments. The academic site (BCH) had a survey completion rate of 71.1% (32/45). Attendings expressed a need for more training on supervising both residents and fellows while tailoring teaching to the fellow’s level. Additionally, the majority of attendings at both sites shared interest in participating in a curriculum on fellow supervision. At the community site, the completion rate was 53.8% (7/13), with attendings identifying teaching to the fellow’s level and providing feedback as key training needs. The fellow survey had a completion rate of 78.6% (11/14). Fellows identified promoting fellow autonomy as a critical topic at both training sites. Qualitative responses highlighted the need to educate community site faculty on the overall goals of fellowship training. Across all groups, the preferred curriculum delivery method was a shared document or resource folder. Conclusion: Our organizational needs assessment highlights the need for enhanced attending training on PHM fellow supervision from the perspectives of both attendings and fellows. While training priorities may differ between the academic and community sites due to the inherent differences in training environment, ongoing faculty development is essential for effectively preparing the next generation of PHM fellows.

Background: Resuscitation skills are fundamental to the practice of Emergency Medicine, yet there is a lack of structured, formalized hands-on learning sessions for emergency medicine trainees. To address this need, we developed and implemented a novel, multimodal advanced resuscitation curriculum for PGY-2 Emergency Medicine residents. This study aims to assess the impact of this dedicated advanced resuscitation curriculum on residents’ medical knowledge, clinical skills, and attitudes. Methods: This mixed-methods study used explanatory sequential design. Quantitative data were first collected through resident pre- and post- curriculum self-assessment surveys, knowledge assessments, and faculty evaluations of residents’ clinical performance. Following quantitative data analysis, qualitative semi-structured interviews were conducted with participants to explore their experiences with the curriculum and its role in shaping their learning, clinical skills, and attitudes. Interview transcripts were analyzed thematically using a deductive-inductive hybrid coding approach. Results: Resident surveys indicated highly positive reactions to the curriculum (Kirkpatrick 1) and statistically significant improvements in confidence across key domains, including resuscitation leadership, airway management, and procedural skills (Kirkpatrick 2). Additionally, residents’ post-course assessment scores improved significantly compared to pre-course assessment scores, and follow-up assessment scores outperformed the national average (Kirkpatrick 2). Faculty surveys reflected observable on-shift behavior change, with residents demonstrating improved resuscitation leadership, procedural skills, and application of evidence-based medicine (Kirkpatrick 3). Qualitative analysis provided deeper insights as to how the curriculum was impactful. Residents emphasized the effectiveness of the curriculum’s individualized structure and alignment with adult learning principles. They also noted that the curriculum improved their confidence in resuscitation leadership, clinical decision-making, and ability to apply evidence-based medicine to practice. Participants also expressed a stronger appreciation for reflective practice and reported more receptiveness to feedback after taking this course. Conclusion: Implementation of a structured resuscitation curriculum significantly improved EM residents’ confidence, knowledge, and clinical decision-making skills. The mixed-methods approach provided a comprehensive evaluation of curriculum impact, suggesting that structured resuscitation training grounded in adult learning principles is effective strategy to prepare EM residents for high-stakes clinical care.

Manuscript 1: Longitudinal Dynamics of Plasma Neuro lament Light Chain in Hereditary Spastic Paraplegia Type 11 (HSP-SPG11) and Type 15 (HSP-ZFYVE26) ABSTRACT Background: HSP-SPG11 and HSP-ZFYVE26 are autosomal-recessive forms of hereditary spastic paraplegias (HSPs). As therapeutic trials emerge, validated biomarkers are critically needed. Objectives: To evaluate plasma neurofilament light chain (pNfL) as a biomarker for neurodegeneration and disease progression. Methods: We analyzed pNfL levels in 57 patients (36 HSP-SPG11, 21 HSP-ZFYVE26) and matched controls using single-molecule array technology. Longitudinal clinical and biomarker data were collected over 5 years. Results: Baseline pNfL levels were significantly elevated in patients: 33.85 pg/mL (IQR 25.15-47.38) in HSP-SPG11, 46.70 pg/mL (IQR 29.95-54.84) in HSP-ZFYVE26, and 4.90 pg/mL (IQR 3.48-6.90) in controls (P 0.001). No significant difference was observed between HSP-SPG11 and HSP-ZFYVE26. In matched pair analysis, pNfL showed inverse correlation with age (ρ = -0.463, P 0.001). Baseline pNfL did not predict future clinical progression. Conclusions: Elevated pNfL reflects early neuroaxonal injury in HSP-SPG11 and HSP-ZFYVE26; however, it could not be used as a surrogate for disease progression.

KEYWORDS: axonal degeneration; biomarker; hereditary spastic paralysis; neurofilament light chain; spasticity.

Manuscript 2: Plasma Proteomic Signatures in Prototypical Forms of Hereditary Spastic Paraplegia ABSTRACT Background: Hereditary spastic paraplegias are a group of neurogenetic disorders caused by axonopathy of corticospinal tract. To date, no validated fluid biomarkers exist for HSP, limiting clinical trial readiness. This study aims to identify plasma proteomic signatures associated with HSP subtypes and disease severity using a proximity extension assay approach. Methods: We performed a case-control study of 105 genetically confirmed HSP patients (SPG4, n = 42; AP-4-HSP, n = 41; SPG11/15, n = 22) and 45 age- and sex-matched healthy controls. Plasma proteins were measured using the Olink Target 48 Neurodegeneration panel. Differentially abundant proteins were identified using limma linear models adjusted for age, with Benjamini-Hochberg false discovery rate correction. Subtype-specific analyses with phenotype correlation were explored. Linear regression and logistic regression models associating protein levels with clinical scores were performed. Results: We discovered that there were no shared differentially abundant proteins in prototypical forms of HSP vs matched controls. Subtype-level analyses revealed partially overlapping but distinct profiles: AP-4-HSP was characterized by co-downregulation of BACE1 and Aβ40 with elevation of GLRX, HLA-DRA, and FOXO3; SPG4 by nine DEPs (NEFL, GLRX, TP53, EIF2AK2, FOXO3, Aβ40, Aβ42, BMP7, pTau217); and SPG11/15 by marked upregulation of NEFL and HLA-DRA. Some proteins are correlated with phenotypes in each subtypes. Within-subtype severity regressions identified subtype-dependent protein–severity associations. Cross-platform validation confirmed strong concordance between Olink NEFL and Simoa pNfL. Conclusion: Plasma proteomic profiling reveals distinct and diagnostically informative signatures across three prototypical forms of HSP. These findings provide a framework for integrating circulating protein biomarkers into disease monitoring and clinical trial design in HSP. Replication in larger longitudinal cohorts is needed.

KEYWORDS: hereditary spastic paraplegia; neurodegeneration; plasma proteomics; proximity extension assay; biomarker

Panoramic radiographs (PANs) are widely used in dental practice for their broad coverage of the maxillofacial region, low cost, and clinical convenience. However, compared with periapical radiographs (PAs), PANs offer lower spatial resolution and reduced visibility of fine anatomical structures, including the lamina dura, periodontal ligament space, root morphology, and subtle periapical changes, which can limit their diagnostic utility for precision-dependent evaluations. Developing a super-resolution (SR) and enhancement method to enhance PAN images toward PA-like anatomical clarity could ultimately aid in early detection and prevention of potential diseases and reduce the need for additional intraoral imaging. A major challenge in developing an SR solution to align the level of detail in PAN to the PA standard is the lack of paired PAN and PA images that depict the same anatomy at their own levels of detail. A two-step method was designed to solve this problem. First, the image style translation model CycleGAN was used to transform HR PA images into synthetic PAN-style counterparts that simulate panoramic-like degradation in contrast, noise, and detail level. These generated images were paired with their original PA to create supervised training data for SR training. Next, two SR architectures, SwinIR and NAFNet, were then trained and evaluated using quantitative image quality metrics, including peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), feature similarity index (FSIM), and mean squared error (MSE). Both models demonstrated strong reconstruction performance and successfully recovered finer anatomical structures from degraded PAN-styled inputs. NAFNet achieved better overall performance and produced outputs that have both a quantitatively higher score and a closer visual appearance to the ground truth PA images. These results suggest that the proposed framework is a viable approach for enhancing dental radiographic detail and improving anatomical visibility in PANs. Although further validation on real PAN images with clinically verified reference labels is still needed, this work represents an important first step toward enhancing the anatomical detail visible in panoramic radiographs. With future clinical validation and expert assessment, this framework may help improve PAN-based screening and expand the diagnostic value of routinely acquired dental imaging.

Accurate immune profiling of the tumor microenvironment (TME) requires integrating nuclei, tissue, and spatial-level information across sequential reasoning steps. Recent advances in vision language models (VLMs) have enabled natural language interaction with histopathology images, yet existing pathology VLMs are predominantly limited to single-turn question-answering and lack the structured, multi-step reasoning required for clinical immune profiling. We present a dataset construction, fine-tuning, and evaluation framework for multi-turn visual question answering (VQA), applied to computational immune profiling of the TME. Our multi-turn conversational pipeline decomposes immune profiling into sequential, clinically-grounded subtasks: nucleus detection and classification, cell-type quantification, tissue region analysis, and TME summarization. We introduce a structured approach to generate multi-turn VQA training data from annotated histopathology images. Our dataset comprises 238,747 patches spanning 22 organ types from 7 datasets, including hematoxylin and eosin (H&E) and immunohistochemistry (IHC) slides, from which we construct 515,382 multi-turn conversation samples across 13 turn types that span pixel-level and patch-level spatial scales. We further introduce a modular, turn-aware evaluation framework that assesses turn-level accuracy, cross-turn consistency, and conversational coherence. To our knowledge, this is the first such framework for multi-turn histopathology VQA. On an external test set, our fine-tuned model achieves a median nuclei detection F1 of 0.667, outperforming HoverNet on nuclei detection and segmentation. On a held-out test set, it surpasses all four VLM baselines across multi-turn evaluation metrics. This work establishes the feasibility of fine-tuning a general-purpose VLM for structured immune profiling. We envision this framework as a generalizable basis for developing and benchmarking conversational VLMs across immune profiling tasks and broader digital pathology applications.

In April 2006, an international working group revised Bax’s widely used modern definition of Cerebral Palsy (CP), characterizing it as: “A group of permanent disorders of the development of movement and posture, causing activity limitation, that are attributed to non-progressive disturbances that occurred in the developing fetal or infant brain. The motor disorders of CP are often accompanied by disturbances of sensation, perception, cognition, communication and behavior, by epilepsy, and by secondary musculoskeletal problems”. CP remains the most common cause of chronic childhood disability, with a prevalence of 2-2.5 per 1000 live births in developed countries. This thesis will address two intertwined and critical aspects of CP, aligned with its contemporary holistic definition. First, due to the non-progressive but permanent nature of the primary brain lesion in CP, many acquired clinical manifestations such as musculoskeletal pathology progress over time in the setting of growth and spasticity. Hip displacement, second only to Achilles contracture in prevalence, effects approximately 35% of all children with CP, and is positively correlated with the severity of functional impairment as defined by the Gross Motor Function Classification System (GMFCS). This is a five-level ordinal rating scale based on the assessment of self-initiated movement with emphasis on function with regard to sitting and walking. Incidence of hip displacement in non-ambulatory patients (GMFCS IV and V) ranges from 60% up to 90%. The adverse natural history of CP hip displacement necessitates surgery in the majority of children, which aims to obtain a pain-free, mobile, and concentrically reduced hip with symmetric range of motion, prior to the onset of painful arthritis. Hip reconstruction surgery typically involves soft tissue releases, femoral varus derotation osteotomy (VDRO), pelvic osteotomy (PO), or a combination thereof. In the literature, clear surgical indications for performing an additional acetabuloplasty after VDRO are elusive, and in practice the decision remains driven by surgeons’ experience and training. While combining VDRO and PO generally lowers resubluxation rates, some patients can safely undergo VDRO alone, thus avoiding additional surgical risk and morbidity. The first part of this thesis evaluates the effect of intraoperative hip arthrography – successfully used in Developmental Dysplasia of the Hip - as a tool for acetabuloplasty decision-making in neuromuscular hip reconstruction. Second, the revised CP definition emphasizes a holistic approach; shifting from the historical musculoskeletal focus towards incorporating comorbidities that influence patients’ health-related quality of life (HRQoL). Hip reconstruction surgery and spinal fusion aim to alleviate pain, ease caregiving, and improve HRQoL, yet their effect has been debated due to surgical complexity, postoperative risk of morbidity and patient mortality. Today, orthopedic surgeons recognize that HRQoL improvements are more meaningful to patients and caregivers than radiographic outcomes, underscoring the importance of HRQoL outcome measures. The Caregiver Priorities and Child Health Index of Life with Disabilities (CPCHILD) questionnaire, validated for assessing HRQoL and caregiver burden in GMFCS levels IV and V CP patients, is increasingly used to evaluate the impact of hip and spine surgery. However, before becoming a gold standard, effective outcome instruments must also demonstrate responsiveness - the tools’ ability to detect change and remain constant across diverse cohorts. The second part of thesis assesses the responsiveness of the CPCHILD questionnaire in children with severe non-ambulatory CP undergoing hip reconstruction surgery and spinal fusion.

The analysis of stereo-EEG (SEEG) recordings is integral to the surgical treatment of patients with drug-resistant epilepsy. These data, recorded from intracerebral depth electrodes, provide the spatiotemporal resolution necessary to identify the seizure onset zone and characterize the epileptogenic network. While prior machine learning approaches have demonstrated strong performance on seizure detection benchmarks, they lack the resolution and computational scalability required to precisely localize seizure onset and offset for each individual recording channel. We propose a two-stage framework for automated channel-level seizure detection in SEEG recordings with sub-second precision. Our model combines a permutation-based coarse-grain detector that flags potential seizure events within the recording and a fine-grain detector that performs window-level classification within each detected interval to identify seizure onset and offset for each channel. Our dataset contained of 5723 hours of SEEG recordings and 677 annotated seizure events from 26 patients who underwent SEEG at Massachusetts General Hospital. When evaluated on a held-out test cohort of 12 seizure events from three patients, the coarse-grain detector achieved an F1 score of 0.608 and the fine-grain detector achieved an F1 score of 0.16 on the window level classification task. The fine-grain detector's low F1 score was driven by low precision, reflecting a tendency to over-predict seizure windows within intervals flagged by the coarse-detector. We also train and evaluate the fine-grain detector on the Bonn EEG dataset, where it achieves an F1 score of 0.981. Although the Bonn dataset is not directly comparable to our task of sub-second channel-level seizure onset and offset detection, it demonstrates strong performance on standard seizure classification tasks. In summary, this work combines a computationally efficient coarse-grain detector with a transformer-based fine-grain classifier for automated channel-level seizure detection in clinical SEEG recordings. This framework enables the generation channel-level seizure onset and offset predictions at the spatial and temporal resolution required for clinical analyses of the epileptic network while also addressing the computational constraints of analyzing large-scale multi-channel SEEG data.

Immunosenescence refers to an age-related decline in immune function, primarily characterized by the gradual degeneration of the thymus, also known as thymic involution. Transforming growth factor-β (TGF-β) signaling is integral to the formation and normal function of the thymus. As a member of the TGF-β superfamily, Growth differentiation factor 11 (GDF11) and its receptors are expressed during thymocyte development. Moreover, GDF11 plays a critical role in the immune system by reducing inflammation through the regulation of specific pathways and immune cell functions. More importantly, GDF11 has been shown to have potential rejuvenating effects on various organs, including the heart and brain. However, its impact on thymic rejuvenation remains unknown. To study the effect of GDF11 on thymic involution, here we utilized the Gdf8^Gdf11MD mice, which have increased levels of circulating GDF11, and analyzed changes in their thymocyte and splenocyte populations. Our results showed that young adult female (4-month-old) and male (3-month-old) Gdf8^Gdf11MD mice exhibited a trend toward accelerated thymic involution compared to controls. Although this difference diminished with age, the reduction in thymic cellularity persisted into late adulthood (8-month-old) in male mutants. Notably, significant changes in the absolute number of multiple thymic subsets were observed in young (3-month-old) and pre-middle-aged (8-month-old) Gdf8^Gdf11MD males, indicating potential alterations in thymocyte development. Using flow cytometry, distinct immune cell subsets within the spleen were analyzed to further explore the effects of accelerated thymic involution on peripheral T cells in mutant mice. We found that, despite a significant increase in spleen weight in 12-month-old female and 8-month-old male mutants, there were no differences in the frequency or absolute number of splenocytes compared to controls. In addition, 4-month-old female Gdf8^Gdf11MD mutants exhibited an increased spleen-to-body weight ratio, accompanied by a decreased proportion but increased number of CD4⁺ central memory T cells (TCM). These findings demonstrate a possible correlation between GDF11 expression levels and thymic involution, highlighting the important role of GDF11 in immunosenescence.

Sarcoma is a rare group of cancers originating from mesenchymal lineages, with high metastatic potential and poor clinical outcomes. Previous studies have linked collagen and extracellular matrix (ECM) remodeling to sarcoma invasion, metastasis, and immune evasion. However, with bulk RNA-seq datasets, it remains unclear whether these signals come from tumor cells themselves or from the tumor microenvironment (TME). In this study, I designed and implemented a single-cell RNA sequencing (scRNA-seq) analysis pipeline to address this question, utilizing publicly available datasets from three sarcoma subtypes: undifferentiated pleomorphic sarcoma (UPS), leiomyosarcoma (LMS), and dedifferentiated liposarcoma (DDLPS). The designed pipeline identifies malignant cells without relying on marker genes by using an iterative inferCNV reference refinement strategy. The pipeline achieved good annotation concordance with original study labels in UPS/LMS samples, with reduced concordance in liposarcoma samples, potentially attributable to the focal copy number architecture characteristic of DDLPS and substantial inter-patient transcriptional heterogeneity. Downstream differential gene expression analysis showed that major fibrillar collagen genes, including COL1A1, COL1A2, and COL3A1, were upregulated in malignant cells compared with non-malignant cells across all analyzed datasets, while other ECM-related genes and collagen subtypes followed subtype- or patient-specific patterns. Collagen biosynthesis and ECM assembly pathways were upregulated in malignant cells, while Hallmark oncogenic pathways showed patient-specific patterns. In conclusion, these results support tumor-intrinsic collagen expression as a consistent feature of sarcoma malignant cell states. The study may have important implications for malignant progression, immune evasion, and the development of matrix-targeted therapeutic strategies.

MANUSCRIPT 1:Artificial Intelligence to Extract Structured Details from Unstructured Medical Records in a Global Heart Failure Trial Background: Global clinical trials collect extensive unstructured medical records that richly describe participants’ clinical presentation, but their narrative format precludes quantitative analysis. Converting these records into structured data could reveal insights into events like heart failure hospitalization (HFH) but is prohibitively labor intensive. Varied documentation styles, translated text, and scanned or handwritten documents pose challenges to data extraction in clinical trials. Large language models (LLMs) may be able to extract structured presenting features from unstructured trial records. Methods: We developed an LLM prompting workflow to extract 51 variables (symptoms, signs, laboratory and imaging results, and treatments) from HFH dossiers in the DELIVER trial. LLM outputs were validated against physician chart review in a random sample of 125 dossiers, a second physician reviewed 25 dossiers for inter-reviewer agreement. Accuracy and positive and negative predictive values were assessed against the physician gold-standard. The model was then applied across the full trial to quantify the frequency of HFH presenting features. Results: The LLM achieved an overall accuracy of 0.96 (95% CI, 0.95–0.97), PPV of 0.94 (95% CI, 0.92–0.95), and NPV of 0.97 (95% CI, 0.97–0.98). Accuracy exceeded 0.90 for 49 of 51 variables. The inter-reviewer concordance was 98%. In the full trial (1,237 HFHs), the most common features were dyspnea (92%), peripheral edema (68%), radiographic congestion (59%), elevated natriuretic peptides (65%), and administration of intravenous diuretics (84%). Conclusion: A prompt-engineered LLM accurately extracted structured data, including signs and symptoms, laboratory and imaging data from adjudication dossiers at scale in a global clinical trial to generate a structured, useable database. LLM-based data extraction could be extended to unlock quantitative insights from a wide range of narrative trial records.

MANUSCRIPT 2:Prognostic Significance of Clinical Features at admission for Heart Failure: Insights from DELIVER and PARAGON-HF Background: Hospitalization for heart failure (HFH) is a sentinel event associated with high mortality. Whether specific symptoms, signs, and diagnostic findings during HFH are associated with subsequent mortality is unknown. Objective: To determine whether individual clinical features at HFH predict subsequent all-cause mortality in HFmrEF/HFpEF. Methods: We analyzed patients with HFH from two international randomized trials (DELIVER and PARAGON-HF). Sixteen clinical features spanning symptoms, signs, and diagnostic findings were extracted from HFH records using a validated artificial intelligence tool. Cox proportional hazards models were used to assess associations with all-cause mortality. Multiple testing was controlled using the false-discovery rate (FDR). Results: Among 1,534 patients, 507 post-HFH deaths occurred over a median follow-up of 1.2 years. At admission dyspnea (93%) and peripheral edema (70%) were commonly documented; in contrast, other signs of congestion including significant weight gain (19%), ascites (14%), and hepatojugular reflux (3%) were less frequently reported. Radiography showed pulmonary congestion in 59% and pleural effusion in 40%, and worsening kidney function relative to pre-hospitalization baseline occurred in 27%. Intravenous diuretics were administered in 83% of patients, and 10% required mechanical/procedural interventions. Five admission clinical features showed nominal associations with mortality, of which two remained FDR-significant: fatigue (HR 1.59, 95% CI 1.24–2.04) and worsening kidney function assessed as percent creatinine change from baseline (HR per 25% increase 1.10, 95% CI 1.04–1.17). Conclusions: Many classic signs of congestion were infrequently documented in adjudication records during HFH admissions. Fatigue and worsening kidney function were associated with greater subsequent mortality.

PAPER 1 ABSTRACT

BACKGROUND: Coronary artery disease (CAD) polygenic risk score (PRS), low-density-lipoprotein cholesterol (LDL-C), lipoprotein(a) (Lp(a)), and high-sensitivity C-reactive protein (hsCRP) are biomarkers that predict CAD. It is unclear whether integrating genomics with lipid and inflammatory biomarkers could complement traditional risk scores in identifying people at risk of CAD.

OBJECTIVES: This study assesses the predictive value of CAD PRS, LDL-C, Lp(a), and hsCRP for incident CAD across different age and sex groups.

METHODS: Participants (n = 215,695) from the UK Biobank aged 40 to 69 years with baseline CAD PRS, LDL-C, Lp(a), and hsCRP values were followed for 12 years to assess the incidence of CAD. We evaluated a multivariable-adjusted Cox model that included all 4 biomarkers, net reclassification index, C-statistics, and population attributable risk across different age and sex groups.

RESULTS: Over a 12-year follow-up, 4,721 men and 2,425 women developed CAD. The HRs for incident CAD associated with each biomarker elevation were 1.79 (95% CI: 1.70-1.89) for CAD PRS, 1.60 (95% CI: 1.48-1.66) for LDL-C, 1.20 (95% CI: 1.12-1.29) for Lp(a), and 1.64 (95% CI: 1.57-1.72) for hsCRP. CAD PRS demonstrated a stronger association in men (HR per SD: 1.49; 95% CI: 1.45-1.54) than women (HR per SD: 1.37; 95% CI: 1.31-1.44; P-interaction # 0.001). All biomarkers conferred greater HRs at younger ages (P 0.0001). Individuals with all biomarkers elevated had a 4.65-fold increased risk of CAD compared with those with no elevated biomarkers. A combined 4-biomarker model had a higher C-statistic of 0.753 compared with the pooled cohort equations (C-statistic of 0.740). The C-statistic of the combined 4-biomarker model was also higher in younger individuals in both sexes and yielded a 32.0% continuous net reclassification index when compared with the pooled cohort equations.

CONCLUSIONS CAD PRS, LDL-C, hsCRP, and Lp(a) show independent age- and sex-specific associations with CAD. Measuring all 4 biomarkers may improve midlife CAD risk prediction for both male and female patients. (JACC. 2026) © 2026 by the American College of Cardiology Foundation.

PAPER 2 ABSTRACT BACKGROUND: Many patients develop coronary artery disease (CAD) despite being free from standard modifiable risk factors (SMuRFlesss), including diabetes, smoking, hyperlipidemia, and hypertension. Prevention beyond traditional risk factors is needed for primary prevention of SMuRFless individuals. Coronary artery disease polygenic risk score (CAD PRS), low-density lipoprotein cholesterol (LDL-C), lipoprotein(a) [Lp(a)], and high-sensitivity C-reactive protein (hs-CRP) are residual blood biomarkers that predict CAD. This research aimed to investigate the utility of inflammatory, genetic, and Lp(a) residual biomarkers and the prediction of CAD in SMuRFless participants of the UK Biobank (UKB).

METHODS: Baseline measurements of biomarkers hs-CRP, CAD PRS, and Lp(a) were obtained from 40,723 healthy participants of the UKB, a prospective cohort study enrolling individuals ages 40–69 years from 2006–2010, who were SMURFless at enrollment and were followed for a median of 13.7 (12.0–14.4) years for incident CAD and major adverse cardiovascular events (MACE). A Cox model, adjusted for age, sex, estimated glomerular filtration fraction, body mass index, systolic blood pressure, Townsend deprivation index and the first 10 ancestral principal components (only in a model with CAD PRS), was calculated across quintiles of biomarkers. Hazards ratios were also calculated according to commonly used clinical thresholds and continuous scales of hs-CRP, CAD PRS and Lp(a). Additionally, a 13.7-year adjusted cumulative incidence curve for each biomarker and joint effects analyses between two biomarkers (CAD PRS and Lp(a), CAD PRS and hs-CRP, Lp(a) and hs-CRP), adjusted for age and sex, were plotted for incident CAD and MACE.

RESULTS: Over a follow-up of 13.7 years, 521 of 40723 (1.00%) developed incident CAD. Median hs-CRP and Lp(a) and mean CAD PRS were higher among SMuRFless individuals with incident CAD compared to SMuRFless individuals with no CAD (median hs-CRP 1.22 vs. 0.88 mg/L, P.001, mean CAD PRS 0.40 vs -0.01 standardized units, P.001, median Lp(a) 20.1 vs. 17.0 nmol/L, P.001). Compared to the lowest quintile, the HR in an adjusted Cox model for incident CAD for quintiles 2 to 5 were associated with increasing HRs for hs-CRP (1.21, 1.13, 1.49, and 1.58), CAD PRS (1.56, 1.80, 2.62, 4.00) and for Lp(a) (0.91, 0.86, 0.90, 1.23). The risk of CAD increased by 12% (HR 1.12, 95% CI 1.03–1.23), 40% (1.40, 95% CI 1.31–1.49), and 5% (HR 1.05, 95% CI 1.01–1.12%) for each increasing quintile in hs-CRP, CAD PRS, and Lp(a), respectively. Joint effect analysis showed independent contribution of each biomarker to incident CAD, with the highest risk observed when there was a joint increase in biomarker in the fifth quintile. Results remain consistent for our secondary outcome MACE but not for Lp(a).

CONCLUSION: Among SMuRFless individuals, elevated levels of hs-CRP, CAD PRS, and Lp(a) were associated with incident CAD, suggesting these biomarkers can aid in clinical stratification of SMuRless individuals with otherwise silent elevated risk.

In our first project, we evaluated the comparative effectiveness of first-line (1L) versus second-line (2L) use of CDK4/6i in HR+/HER2-mBC using a systematic review and reconstructed individual patient data meta-analysis of 7,602 patients. We found that deferring CDK4/6i to the second-line setting was associated with worse PFS2 in the overall population; however, no significant difference in OS was observed between treatment strategies. Notably, when analyses were restricted to randomized clinical trials, both PFS2 and OS were similar between 1L and 2L use, challenging the assumption that earlier use of CDK4/6i universally improves outcomes. These findings have important clinical implications. Early use of CDK4/6i is associated with increased treatment duration, higher toxicity, and substantial cost, without clear survival benefit. Together, these results question the “always first-line” paradigm and highlight the need for more tailored treatment strategies. In selected patients, initial endocrine therapy alone may be a reasonable approach, reserving CDK4/6i for later lines of therapy. In our second project, we investigated genomic biomarkers of response and resistance to the AKT inhibitor, ipatasertib, in endocrine-resistant metastatic breast cancer using data from the phase Ib TAKTIC trial. Among 47 heavily pretreated patients, median progression-free survival was 5.72 months. PI3K/AKT pathway alterations were associated with improved outcomes specifically in patients receiving triplet therapy (ipatasertib, endocrine therapy, and CDK4/6i), but not in the overall cohort. In contrast, ESR1 mutations, FGFR1 amplifications, and EGFR alterations were associated with shorter PFS. Paired ctDNA analyses further demonstrated the emergence of genomic alterations at disease progression, suggesting potential mechanisms of acquired resistance. These findings highlight the dynamic nature of tumor evolution and the potential role of ctDNA in the identification of novel therapeutic targets. Overall, this thesis addresses two critical challenges in metastatic breast cancer: optimizing treatment sequencing and advancing biomarker-driven precision oncology. Together, these studies support a more individualized approach to therapy, emphasizing that effective treatment depends not only on selecting the right drug, but also on identifying the right timing and the right patient.

Missense variants are by far the most prevalent type of pathogenic coding variants, but they are predominantly annotated as variants of uncertain significance (VUS) in ClinVar. This categorization prevents any medical intervention despite the variant being causal. Ex- isting computational tools predict pathogenicity well but cannot answer the question that drives genetic diagnosis: harmful for which disease? We present PheMART2, a contrastive retrieval framework that ranks candidate diseases for each query missense variant. To solve these problems, we implement two major architectural designs to improve general- ization. Diseases are encoded using weighted attention-based combinations of HPO terms instead of relying on concept IDs, allowing the model to generalize to disease labels outside the training distribution. Population-scale variant–phenotype relationships from the MVP phenome-wide association study are mapped to the HPO disease space through a curated CUI mapping and SapBERT-based semantic translation pipeline, and then distilled into our model using bidirectional Kullback-Leibler divergence in both the variant→disease and disease→variant directions. Model evaluation is performed by randomly splitting data on the basis of genes instead of per-variant random splits with negative sampling, across all 1,563 diseases in the dataset. In comparison with this more stringent evaluation approach, PheMART2 demonstrates an MRR of 0.272 for unseen genes, an 18.1% improvement over the no-KL, no-auxiliary version (p = 0.001), with the clinical prior distillation being the key contribution. Disease label out-of-distribution inference shows an MRR score of 0.39, suggesting that our compositional encoding successfully generalizes to diseases constructed from novel phenotypic primitives. Amongst 676 independently curated Phenopackets cases with no overlap between variants used in the training data, the top 10 candidate list in- cludes the true disease for 51.5% and the top 1 for 18.5%. Performance concentrates on variants in genes with existing biological annotations (MRR = 0.418) and falls sharply on genes without much connection in the knowledge graph (MRR = 0.089), which reflects the inherent constraint of the transductive graph setting and points to inductive extension as the natural next step.

Melanoma is known as the deadliest skin cancer and mortality rate of recurrent patients is two times that of non-recurrent patients. Although adjuvant immunotherapy can be applied to reduce recurrence risk for early-stage melanoma patients, the existing staging system based on Breslow thickness and ulceration is insufficient to stratify patients who are actually at high risk of recurrence. Histologic features like mitotic figures (MF) and tumor-infiltrating lymphocytes (TIL) are clinically established and computationally validated prognostic factors of melanoma recurrence, but the current clinical workflow of manual extraction of these features is subject to high inter-observer variability. Despite various MF and TIL detection algorithms having been established, the prognostic value of computationally derived histologic engineered features using MF and TIL is yet to be validated in diverse cohorts. This thesis project aims to fill in the gap by leveraging MF- and TIL-derived features for recurrence prediction in the largest early-stage melanoma recurrence prediction cohort reported to date. Six cell morphology features (mitotic ratios, eTIL score, and TIL density variations) are significantly associated with recurrence risk and reflect clinical features of mitotic rate and TIL infiltration type. In addition, a cell morphology risk combining MF and TIL features remains statistically significant in a multi-variate Cox regression model after adjusting for clinical covariates, indicating that a combined risk score of tumor proliferation and immune response encodes independent prognostic significance beyond existing clinical features. This project provides a computational pathology framework for detecting MFs and TILs, engineering cell morphology features, and using these features for early-stage melanoma recurrence risk prediction.