Adam Cross

Background: Traumatic brain injury (TBI) in children disrupts cognitive, physical, and emotional functioning, posing a significant public health challenge in the United States. Recovery after pediatric TBI is nonlinear, influenced not only by fluctuations in progress but also by social determinants of health (SDoH), including sociodemographic features and health care access. Objective: This study investigates how SDoH influence TBI phenotypes and associated recovery trajectories. Our retrospective cohort analysis evaluates health disparities among pediatric patients and examines how these disparities influence health outcomes following TBI. Methods: We extracted 926 pediatric patients from the Children’s Health After Injury (CHAI) dataset to ensure consistency across 5 key variables: Abbreviated Injury Scale (AIS) scores, vital signs and assessments, comorbidities, complications, and Glasgow Coma Scale (GCS) scores. Dimensionality reduction and density-based hierarchical clustering identified distinct phenotypes and their SDoH characteristics. Further exploratory analysis revealed distinct recovery trajectories across the phenotypes. To confirm the effect of SDoH on the phenotypes, generalized linear modeling and the random forest modeling were used. Results: Four distinct pediatric TBI phenotypes were identified, each with a unique injury-severity profile, recovery trajectory, and pattern of SDoH association: (1) adolescents with high injury severity and stable GCS scores (n=278) showed a significant association between phenotype membership and age. This group experienced a gradual increase in psychological and behavioral impairment over the recovery period, with total impairment scores increasing by 25% from baseline to the 12-month follow-up. (2) Younger children with moderate injuries and stable GCS scores (n=141) showed no significant SDoH associations. Total impairment scores remained stable across the follow-up period, though neurological and physical conditions showed long-term worsening, warranting ongoing monitoring. (3) Critically injured children with severe neurological impairment (n=98) showed the strongest SDoH associations, with race/ethnicity and its interaction with age significantly linked to phenotype membership; Asian children had 1.8 (95% CI 1.02‐3.34) times the odds of belonging to this critically injured phenotype compared with White children. Despite carrying the lowest overall impairment of any cluster at baseline, this group showed escalating psychological burden and fluctuating neurological conditions through the 12-month follow-up. (4) Moderately injured children with stable GCS scores followed by a sharp progression (n=409) showed no significant SDoH associations, yet carried the highest overall impairment burden of any cluster across all time points. Total scores increased 31% from baseline to the 12-month follow-up, driven by a marked escalation in psychological/neurological issues. Conclusions: Age and race/ethnicity were the most influential predictors of cluster membership, with SDoH risk concentrated in the 2 most severely injured groups. Identified SDoH-driven phenotypic disparities can guide targeted interventions for pediatric patients at high risk for TBI sequelae. Future studies should extend follow-up beyond 12 months and validate these phenotypes in diverse cohorts to strengthen generalizability and inform equitable care.

Concussion evaluation technologies have increasingly emphasized objectivity, leveraging AI and machine learning to formalize and standardize diagnostic processes. However, this orientation toward objectivity has not been evaluated via stakeholder engagement, to see if this is what they want . In this paper, we present findings from interviews with 11 concussion evaluators, examining how they make decisions, use healthcare technology, and envision the role of AI in their workflows. Our study surfaces the tensions between evaluators’ values, such as flexibility and coordination, and the assumptions built into existing decision support systems. We show how evaluators seek tools that preserve clinical agency, support collaboration across fragmented care settings, and complement rather than override their interpretive expertise. Drawing on these insights, we propose design considerations for developing AI-based systems that align with concussion evaluation’s subjective and distributed nature. This work contributes to CSCW by challenging objectivity as a default design goal and advocating for stakeholder-centered AI in healthcare.

Rare diseases affect 1 in 10 Americans yet remain systematically underdocumented in clinical records. ICD-based systems cannot capture their breadth, over 50% of Orphanet codes lack a direct ICD mapping and only 2.2% of HPO codes have matching ICD codes, leaving patient populations invisible and delaying diagnosis. Mining unstructured clinical notes offers a direct path forward, but real notes are long, noisy, and abbreviation-dense, and limited annotations make fine-tuning infeasible, demanding approaches that generalize without task-specific training. We present Rare Disease Mining Agents (RDMA), an agentic framework equipping smaller quantized LLMs with tools for abbreviation resolution, implicit phenotype reasoning, and ontology grounding against Orphanet and HPO. RDMA substantially outperforms fine-tuned and RAG-based baselines across benchmarks with different data characteristics, without any task-specific training. A small quantized model achieves maximal performance, reducing inference costs by up to 10x and local hardware costs by up to 17x, enabling private deployment on standard hardware without cloud-based PHI exposure. RDMA’s uncertainty-flagging mechanism further reduces expert annotation burden while preserving agreement quality, supporting scalable rare disease documentation in clinical practice. Available at https://github.com/jhnwu3/RDMA .

Diagnosing concussions is challenging due to the absence of accessible, reliable, and interpretable biomarkers of neurocognitive impairment (NCI) that is subject to variability of symptoms throughout recovery. FlightPath is the mobile mixed-reality application in which users continuously follow a virtual agent by engaging visuomotor and cognitive functions. As a sideline screening tool, FlightPath was deployed with university athletes before and after sports-related brain injury. Using FlightPath performance data from 187 participants, this manuscript aimed to characterize behavioral biomarkers that distinguish concussed from non-concussed individuals. A simple three-dimensional object-following error (TE₃ᴰ) proved insufficient to capture behavioral dynamics under NCI. We therefore introduce TRACE (Temporal Reaction and Correction Evaluation), a demand–response framework that quantifies the effectiveness of corrective actions to task demands over time. Diagnostic testing shows that TRACE, particularly during adaptive agent sessions, provides greater discriminative power than TE₃ᴰ alone, supporting reliable concussion identification with interpretable connection to NCI symptoms.

BACKGROUND Traumatic Brain Injury (TBI) in children disrupts cognitive, physical, and emotional functioning, posing a significant public health challenge in the United States. Recovery after pediatric TBI is non-linear, influenced by fluctuations in cognitive and physical progress, which can be further shaped by social determinants of health (SDoH) factors, including sociodemographic features and healthcare access. OBJECTIVE This manuscript aims to investigate how SDoH influences phenotyping and their recovery trajectories. Particularly, the retrospective cohort analysis evaluates health disparities among pediatric patients and examines how these disparities influence health outcomes following TBI. METHODS The cohort of 926 pediatric patients was extracted from the Children’s Health After Injury (CHAI) dataset, to ensure consistency across five key variables: Abbreviated Injury Scale (AIS), Vital Signs and Assessments, Comorbidities, Complications, and Glasgow Coma Scale (GCS) scores. Dimensionality reduction, followed by density-based hierarchical clustering, were applied to identify distinct phenotypes and their SDoH characteristics. Further, exploratory analysis revealed distinct recovery trajectories over the phenotypes. To confirm the effect of SDoH factors on the phenotypes, generalized linear modeling and Random Forest method were used. RESULTS The four distinct TBI phenotypes exhibited unique severity scale correlations and recovery trajectories, which has implications for the SDoH features. Adolescents with high injury severity and stable GCS: This cohort comprised 278 patients and exhibited a significant positive correlation with age group (coefficient = 0.0807, p-value = 0.026). This group demonstrated a gradual progression of psychological and behavioral impairments, suggesting a persistent emotional burden. Younger children with moderate injuries and stable GCS: This cohort consisted of 141 patients, and no significant SDoH associations were observed (all p-values > 0.05). It exhibited stable psychological outcomes but tended to follow long-term worsening of neurological and physical conditions, which necessitates ongoing monitoring. Critically injured children with severe neurological impairment: This critical cohort comprised 98 patients and exhibited significant associations with race/ethnicity (coefficient = 0.4048, p-value = 0.001) and a negative interaction between age group and race/ethnicity (coefficient = -0.1437, p-value = 0.003). It consistently exhibited escalating psychological issues and fluctuating neurological conditions, underscoring the need for continuous support and pain management interventions. Moderately injured children with stable GCS, followed by sharp progression: This cohort consisted of 409 patients and did not exhibit statistically significant SDoH associations (all p-values > 0.05). Patients in this cluster often experienced a rapid escalation of psychological issues and a sharp deterioration of neurological conditions, emphasizing the urgent need for support. CONCLUSIONS Disparities in sociodemographic characteristics evident among the phenotypes will help identify pediatric patients at high risk for persistent sequelae from TBI and develop targeted interventions to treat these individuals. CLINICALTRIAL None

In recent years, the landscape of digital biomedicine and healthcare has been reshaped due to the disruptive breakthroughs in AI-facilitated by tremendous data and high-performance computers, large language models (LLMs) have transformed information technology from accessing data to performing analytical tasks. While demonstrating unprecedented capabilities, LLMs have been found unreliable in tasks requiring factual knowledge and rigorous reasoning. Biomedicine and healthcare, as an important vertical domain rapidly benefitting from progress in AI, necessitates strict requirements on the accuracy, controllability, and interpretability of analytical models, posing critical challenges for LLMs. Despite recent studies addressing the hallucination problem of LLMs, research on empowering LLMs with the ability to plan, reason, and ground with explicit knowledge has also started to prosper, especially in the biomedicine and healthcare domain. On the other hand, biomedical data are enormous and notoriously complex, coming from various sources (e.g., biomedical knowledge bases, online literature, and hospitals) and bearing various modalities (e.g., tables, texts, images and time-series). Healthcare professionals have spent decades collecting, cleaning, and curating various types of data. The processes are extremely costly, producing various datasets with different data schemas, coding systems, and quality standards, many privately owned by the creators, making their integrative analysis and utilization through unified AI techniques still rather challenging. The generalizability of LLMs across different types of data endow them strong promises in automating the processing of large-scale complex healthcare data such as into unified knowledge graphs (KGs). Our goal in this survey is to systematically investigate and summarize recent studies on the unification of LLMs and KGs, towards fully utilizing the value of complex data, unleashing the power of generative AI, and expediting next-generation AI for biomedicine and healthcare applications.

Social Determinants of Health (SDoH) are economic, social and personal circumstances that affect or influence an individual's health status. SDoHs have shown to be correlated to wellness outcomes, and therefore, are useful to physicians in diagnosing diseases and in decision-making. In this work, we automatically extract SDoHs from clinical text using traditional deep learning and Large Language Models (LLMs) to find the advantages and disadvantages of each on an existing publicly available dataset. Our models outperform a previous reference point on a multilabel SDoH classification by 10 points, and we present a method and model to drastically speed up classification (12X execution time) by eliminating expensive LLM processing. The method we present combines a more nimble and efficient solution that leverages the power of the LLM for precision and traditional deep learning methods for efficiency. We also show highly performant results on a dataset supplemented with synthetic data and several traditional deep learning models that outperform LLMs. Our models and methods offer the next iteration of automatic prediction of SDoHs that impact at-risk patients.

BACKGROUND Electronic Health Records (EHRs), including datasets like MIMIC-IV, often lack explicit links between medications and diagnoses, complicating clinical decision-making and research efforts. Even when such links are present, diagnosis lists can be incomplete or inaccurate, particularly during early patient visits when diagnostic uncertainty is high. Discharged summaries, documented at the end of patient care, may offer more detailed explanations of patient visits, potentially aiding in inferring the most likely accurate diagnoses for prescribed medications, especially if we can exploit Large Language Models (LLMs). LLMs have shown promise in processing unstructured medical text, but systematic evaluations are necessary to determine their effectiveness in extracting meaningful medication-diagnosis relationships. OBJECTIVE This study explores the use of LLMs to predict implicitly mentioned diagnoses from clinical notes and link them to corresponding medications. We evaluate their effectiveness and investigate strategies to improve prediction performance. Specifically, we examine two research questions: (1) Does majority voting across diverse LLM configurations enhance diagnostic prediction accuracy compared to the best single-model configuration? (2) How sensitive is the diagnostic prediction accuracy of majority voting to the LLM's hyperparameters, including temperature, top-p, and clinical note summary length? METHODS A new dataset of 240 expert-annotated medication-diagnosis pairs from 20 MIMIC-IV clinical notes was created to evaluate predictive accuracy, as no such dataset previously existed. We hypothesized that combining deterministic, balanced, and exploratory configurations could enhance prediction performance. Key hyperparameters—temperature, top-p, and summary length—were systematically varied. Two levels of summarization, short and long, were tested to assess context length impact. Using GPT-3.5 Turbo, 18 configurations were generated, and random subsets of five were selected, resulting in 8,568 test cases. Majority voting was applied to select the most frequent diagnosis. Performance was evaluated using accuracy scores, comparing majority voting with the best single-model configuration and analyzing hyperparameters contributing to the highest accuracy. RESULTS The majority voting achieved 75% accuracy, outperforming the best single configuration (66%). No single parameter setting consistently excelled; instead, combining diverse configurations aligned with deterministic, balanced, and exploratory strategies yielded better performance. Shorter summaries (2000 tokens) generally improved accuracy. Longer summaries (4000 tokens) were effective only with deterministic settings. CONCLUSIONS Majority voting across LLM configurations enhances diagnostic prediction accuracy in EHRs, demonstrating the potential of ensemble methods for improving medication-diagnosis associations. By leveraging diverse configurations, this approach mitigates model biases and improves robustness in predictive analytics. Future work should explore scalability with larger datasets, additional LLM architectures, and broader clinical applications to refine its effectiveness in real-world settings.

Electronic Health Records (EHRs) often lack explicit links between medications and diagnoses, making clinical decision-making and research more difficult. Even when links exist, diagnosis lists may be incomplete, especially during early patient visits. Discharge summaries tend to provide more complete information, which can help infer accurate diagnoses, especially with the help of large language models (LLMs). This study investigates whether LLMs can predict implicitly mentioned diagnoses from clinical notes and link them to corresponding medications. We address two research questions: (1) Does majority voting across diverse LLM configurations outperform the best single configuration in diagnosis prediction? (2) How sensitive is majority voting accuracy to LLM hyperparameters such as temperature, top-p, and summary length? To evaluate, we created a new dataset of 240 expert-annotated medication-diagnosis pairs from 20 MIMIC-IV notes. Using GPT-3.5 Turbo, we ran 18 prompting configurations across short and long summary lengths, generating 8568 test cases. Results show that majority voting achieved 75 percent accuracy, outperforming the best single configuration at 66 percent. No single hyperparameter setting dominated, but combining deterministic, balanced, and exploratory strategies improved performance. Shorter summaries generally led to higher accuracy.In conclusion, ensemble-style majority voting with diverse LLM configurations improves diagnosis prediction in EHRs and offers a promising method to link medications and diagnoses in clinical texts.

This paper presents the Pulmonary Acoustic Sensor Telemetry Array (PASTA), a novel hardware prototype developed for remote pulmonary monitoring, diagnosis, and prognostication of pulmonary sounds in clinical settings. PASTA employs an array of acoustic sensors strategically positioned at standard auscultation sites on the patient’s chest and back, enabling real-time acquisition and analysis of pulmonary sounds with an edge computing device. This task is impossible using digital stethoscopes available in the market. The proposed system addresses critical challenges in pulmonary care by providing continuous, high-fidelity, multi-site acoustic data collection, a capability not offered by conventional stethoscopes. This functionality is pivotal for enhancing diagnostic accuracy and enabling early detection of respiratory abnormalities, potentially reducing treatment delays in acute respiratory conditions. PASTA integrates a compact design, multi-channel monitoring, and real-time telemetry, distinguishing itself from existing commercial medical devices. Performance evaluations against established digital stethoscopes, conducted using a high-fidelity medical simulation training manikin, demonstrate sensitivity and signal quality. The potential applications of PASTA extend beyond inpatient care, making it suitable for remote monitoring in outpatient programs, such as hospital-at-home and outpatient hospice care. By bridging a significant gap in current medical device offerings, PASTA sets a new standard in pulmonary auscultation and serves as a foundational platform for future advancements in Internet-of-Things (IoT)-enabled healthcare solutions.