Abstract In this work, we demonstrate the unprecedented value of NIH’s “All of Us Research Program” (AoURP) dataset in studying maternal morbidity and building predictive machine learning (ML) models across heterogeneous populations in the United States. We developed robust and data-driven preprocessing pipelines to curate a longitudinal, multi-site, multimodal, and demographically diverse pregnancy dataset (20,253 subjects; 27,525 pregnancy episodes) from AoURP data, using electronic health records (EHR) (Conditions, Labs, Measurements) and survey responses (Social Determinant of Health (SDoH)), focusing on 7 crucial maternal health adverse outcomes. After characterizing data quality, missingness, and heterogeneity, we performed statistical correlation analysis to identify risk factors. We subsequently developed XGBoost and sequential LSTM models to predict the adverse outcomes, reaching state-of-the-art performance for multiple outcomes. We conducted model interpretability post-hoc analysis to understand success points and fairness analysis to evaluate implications for socio-economic disparities. Four practicing physicians reviewed the set of statistically significant and ML model identified features to assess their clinical validity and novelty. Most features identified through either statistical correlations or ML feature importance analysis aligned with known clinical risk factors. Several features were identified that the ML models used but that are not currently used in clinical practice and may merit further clinical investigation. Fairness analysis revealed certain associations with SDoH and age highlight areas that warrant continued monitoring. Overall, we demonstrate that meaningful populational level patterns can be extracted, and high-performing machine learning models can be trained on this longitudinal, diverse, multi-site dataset. Important risk features, particularly novel ones identified, if validated, could inform new strategies for maternal care or enable development and validation of outcome-specific, clinically deployable ML models.

Artificial intelligence (AI) has the potential to transform healthcare for all patients. Yet, there are disproportionately fewer paediatric AI studies and US Food and Drug Administration approvals relative to adults, indicating less effort focused on AI for children. Here, given that innovations in medical AI are accelerated by community-driven research on public datasets, we hypothesized that the disparity in AI for paediatrics is tied to the lack of public paediatric medical imaging data to support their development and evaluation. To that end, we systematically identified and reviewed 203 datasets, revealing that 33% of datasets lacked metadata on patient ages, and when available, children represented less than 2% of patients. To illustrate how a lack of paediatric data can lead to harmful algorithmic bias, we trained models to classify adult cardiomegaly and evaluated them on healthy children. We found a consistent pattern of age-related bias, reproducible across four large-scale public chest X-ray datasets. These findings suggest that the lack of public paediatric data hinders the development of safe AI for children, producing a landscape of adult-first and adult-only AI models with unknown patterns of bias in children. In an analysis of 203 public medical imaging datasets, children are found to represent less than 2% of the overall images, leading also to disparities in algorithm performance in a case study on cardiomegaly classification.

In high-stakes settings where machine learning models are used to automate decision-making about individuals, the presence of algorithmic bias can exacerbate systemic harm to certain subgroups of people. These biases often stem from the underlying training data. In practice, interventions to "fix the data" depend on the actual additional data sources available – where many are less than ideal. In these cases, the effects of data scaling on subgroup performance become volatile, as the improvements from increased sample size are counteracted by the introduction of distribution shifts in the training set. In this paper, we investigate the limitations of combining data sources to improve subgroup performance within the context of healthcare. Clinical models are commonly trained on datasets comprised of patient electronic health record (EHR) data from different hospitals or admission departments. Across two such datasets, the eICU Collaborative Research Database and the MIMIC-IV dataset, we find that data addition can both help and hurt model fairness and performance, and many intuitive strategies for data selection are unreliable. We compare model-based post-hoc calibration and data-centric addition strategies to find that the combination of both is important to improve subgroup performance. Our work questions the traditional dogma of "better data" for overcoming fairness challenges by comparing and combining data- and model-based approaches.

Despite well-documented healthcare access disparities, their impact on electronic health record reliability and resulting clinical prediction models remains poorly understood. Here, analysing 205,186 participants from the All of Us Research Program, we found that participants with cost-constrained or delayed care had worse electronic health record reliability for 73

Post-training quantization reduces the memory needed to run large language models but alters their social biases in ways that aggregate metrics fail to capture. We present the first large-scale study of 50 quantized models evaluated on PostTrainingBiasBench, a unified benchmark of 13 closed- and open-ended bias datasets. We identify a phenomenon we term quantization-induced bias flipping, in which quantization causes models to change responses from biased to unbiased and vice versa, up to 21

We study how reliably sparse autoencoders (SAEs) support claims about reasoning-related internal features in large language models. We first give a stylized analysis showing that sparsity-regularized decoding can preferentially retain stable low-dimensional correlates while suppressing high-dimensional within-behavior variation, motivating the possibility that contrastively selected "reasoning" features may concentrate on cue-like structure when such cues are coupled with reasoning traces. Building on this perspective, we propose a falsification-based evaluation framework that combines causal token injection with LLM-guided counterexample construction. Across 22 configurations spanning multiple model families, layers, and reasoning datasets, we find that many contrastively selected candidates are highly sensitive to token-level interventions, with 45

Abstract The heterogeneity onset of menopause varies across racial and ethnic groups, yet this heterogeneity may partially reflect methodological differences rather than true biological differences. Using the All of Us Research Program Controlled Tier dataset (v8), we analyzed age at first menopause diagnosis across three progressively refined cohorts: a full ICD-based cohort (Cohort 1, n = 11,306), a survey-linked subcohort adjusted for neighborhood deprivation, smoking, and alcohol use (Cohort 2, n = 10,639), and a confirmatory sensitivity cohort applying SNOMED-based surgical exclusions to the same adjusted framework (Cohort 3, n = 10,222). Asian & Pacific Islander and Indigenous/Other individuals experienced significantly earlier menopause onset than White individuals across both adjusted cohorts. The Black-White heterogeneity was attenuated after covariate adjustment and did not re-emerge after surgical exclusion. Current smoking was the sole significant behavioral predictor across adjusted models. These findings demonstrate that menopause definition and cohort selection critically shape estimates across race and ethnicity, and that EHR-based ascertainment combined with structured surgical exclusion yields broadly consistent adjusted results.

Abstract Menopause affects over one billion women worldwide, yet remains poorly characterized at scale. We apply an ICD-10-based phenotyping algorithm to electronic health records (EHR) from an academic medical center ( n =33,444 women aged 35–64) and a safety-net hospital system ( n =7,041), yielding one of the most racially and socioeconomically diverse menopause cohorts in the literature. Structured EHR fields underrepresent symptom burden: only 38.8% of patients had any documented symptom via natural language processing, despite an estimated prevalence of 90%. Adverse pregnancy outcomes were associated with earlier menopause onset after adjustment ( β = − 1.21 years, p =8.7 × 10 − 45 ). Menopausal women showed elevated risk for osteoporosis (hazard ratio of 12.40), rheumatoid arthritis (HR of 2.43), and mental and behavioral disorders (HR 2.38) relative to age-matched men, with divergence at menopause onset. We show that large-scale EHR can characterize menopause at a scale and diversity that prospective enrollment has not achieved.

Early prediction of severe clinical deterioration and remaining length of stay can enable timely intervention and better resource allocation in high-acuity settings such as the ICU. This has driven the development of machine learning models that leverage continuous streams of vital signs and other physiological signals for real-time risk prediction. Despite their promise, existing methods have important limitations. Contrastive pretraining treats all patients as equally strong negatives, failing to capture clinically meaningful similarity between patients with related diagnoses. Meanwhile, downstream fine-tuning typically ignores complementary modalities such as clinical notes, which provide rich contextual information unavailable in physiological signals alone. To address these challenges, we propose OC-Distill, a two-stage framework that leverages multimodal supervision during training while requiring only vital signs at inference. In the first stage, we introduce an ontology-aware contrastive objective that exploits the ICD hierarchy to quantify patient similarity and learn clinically grounded representations. In the second stage, we fine-tune the pretrained encoder via cross-modal knowledge distillation, transferring complementary information from clinical notes into the model. Across multiple ICU prediction tasks on MIMIC, OC-Distill demonstrates improved label efficiency and achieves state-of-the-art performance among methods that use only vital signs at inference.

The high cost of data labeling presents a major barrier to deploying machine learning systems at scale. Semi-supervised learning (SSL) mitigates this challenge by utilizing unlabeled data alongside limited labeled examples, while the emergence of foundation models (FMs) offers powerful zero-shot capabilities that can further reduce labeling cost. However, directly fine-tuning large FMs is often impractical in resource-constrained settings, and naïvely using their pseudo-labels for unlabeled data can degrade performance due to its unreliablity or domain mismatch with target task. In this work, we introduce ZeroMatch, a novel SSL framework that integrates knowledge distillation with consistency-based learning to jointly leverage labeled data, unlabeled data, and pseudo-labels from FMs. ZeroMatch enables training compact student models using only FM inference, making it suitable for low-resource environments such as personal devices with limited compute. Experiments on six vision and language classification benchmarks show that ZeroMatch consistently outperforms standard SSL and zero-shot augmented methods, demonstrating its effectiveness and robustness across a range of foundation model qualities.

Importance: Multiple myeloma (MM) is the second most common hematologic malignancy in the U.S., with a higher incidence among Black patients than White patients. Chimeric antigen receptor T-cell (CAR-T) therapies show clinical promise, but their limited availability raises concerns about access. Objective: To examine associations between disease characteristics, treatment location, and patient demographics with receipt of CAR-T therapy among patients with MM. Design: Retrospective cohort study using electronic health record data from the University of California Health Data Warehouse (UCHDW) between January 2021 and January 2025. Setting: Six academic health centers and twelve affiliated hospitals within the UCHDW. Participants: A population-based cohort of 12,360 adult patients diagnosed with MM and treated at a University of California facility offering CAR-T administration. Analyses were conducted from February 2025 to March 2025. Exposures: Receipt of multiple cancer therapies following MM diagnosis. Main Outcomes and Measures: Logistic regression was used to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for associations between disease characteristics, treatment locations, and patient demographics with receipt of CAR-T therapy. A zero-shot GPT-4 inference model was applied to UCSF clinical notes to assess whether CAR-T therapy was discussed, determine documented eligibility, and classify rationale for eligibility determinations. Results: Among 12,360 patients with MM (mean age, 68.5 years; 51.6% male), 320 (2.6%) received CAR-T therapy. Disease characteristics at diagnosis, measured by the International Staging System (ISS), was distributed as follows: Stage I (65.3%), Stage II (24.4%), Stage III (2.8%), and Unknown (7.5%). Patients treated at UC-1 (49.3%), and UC-2 (50.0%) were more frequently diagnosed with ISS Stage II, whereas patients treated at UC-3 (55.5%) were more frequently diagnosed with ISS Stage I. Our model showed that patients identifying as Black or African American had lower odds of receiving CAR-T therapy compared with White patients (OR, 0.33; [95% CI, 0.17-0.62]). Patients treated at UC-3 also had lower odds of receiving CAR-T therapy compared with UC-1 (OR, 0.42; [95% CI, 0.30-0.59]). Among 270 UCSF patients assessed for CAR-T eligibility using clinical notes, the proportion of patients deemed eligible without documented CAR-T discussions was highest among those identifying as Other Pacific Islander (50%), followed by Black or African American (4.2%), Asian (3.2%), and White patients (0.6%). Conclusions and Relevance: Within a large academic health system, receipt of CAR-T therapy varied by treatment location and patient-reported race. A subset of patients with documented eligibility lacked recorded discussions of CAR-T therapy, suggesting potential differences in referral, documentation, or care pathways influencing observed treatment patterns.

BACKGROUND:The Model for End-Stage Liver Disease (MELD) score guides liver transplant (LT) allocation, but transplant programs also evaluate psychosocial risks that may affect outcomes such as graft survival, adherence, and follow-up. How medical urgency influences consideration of these risks remains unclear. METHODS:We conducted a retrospective analysis of 2959 adults listed for LT at a large academic center (2012-2023). Using a validated large language model-based techniques, we extracted 23 predefined psychosocial factors from clinical documentation. We compared the prevalence of these factors across MELD categories and assessed associations with successful transplantation. RESULTS:Among listed patients, 10 psychosocial factors increased significantly with higher MELD scores. Low transplant knowledge rose from 3.1% (MELD <15) to 13.3% (MELD >35; adjusted odds ratio [aOR], 1.35; 95% confidence interval [CI], 1.22-1.51), history of nonadherence from 7.8% to 22.2% (aOR, 1.25; 95% CI, 1.16-1.35), and past trauma from 5.4% to 12.2% (aOR, 1.19; 95% CI, 1.09-1.30). Among LT recipients, 11 factors showed similar trends, including low transplant knowledge (1.8%-11.1%; aOR, 1.43; 95% CI, 1.23-1.68), nonadherence (5.2%-22.2%; aOR, 1.37; 95% CI, 1.23-1.53), and past trauma (5.2%-13.9%; aOR, 1.22; 95% CI, 1.08-1.37). Adjusting for demographics minimally changed results. CONCLUSIONS:Transplant teams appear to tolerate greater psychosocial risk as medical urgency increases. This suggests psychosocial criteria are applied more flexibly at higher MELD scores. Standardized frameworks that explicitly incorporate medical urgency may improve equity and transparency in transplant decision-making.

While large language models (LLMs) hold promise for transforming clinical healthcare, current comparisons and benchmark evaluations of large language models in medicine often fail to capture real-world efficacy. Specifically, we highlight how key discrepancies arising from choices of data, tasks, and metrics can limit meaningful assessment of translational impact and cause misleading conclusions. Therefore, we advocate for rigorous, context-aware evaluations and experimental transparency across both research and deployment.

Artificial intelligence (AI) is rapidly transforming healthcare, but its benefits are not reaching all patients equally. Children remain overlooked with only 17% of FDA-approved medical AI devices labeled for pediatric use. In this work, we demonstrate that this exclusion may stem from a fundamental data gap. Our systematic review of 181 public medical imaging datasets reveals that children represent just under 1% of available data, while the majority of machine learning imaging conference papers we surveyed utilized publicly available data for methods development. Much like systematic biases of other kinds in model development, past studies have demonstrated the manner in which pediatric representation in data used for models intended for the pediatric population is essential for model performance in that population. We add to these findings, showing that adult-trained chest radiograph models exhibit significant age bias when applied to pediatric populations, with higher false positive rates in younger children. This work underscores the urgent need for increased pediatric representation in publicly accessible medical datasets. We provide actionable recommendations for researchers, policymakers, and data curators to address this age equity gap and ensure AI benefits patients of all ages.

Psychosocial risk factors and social determinants of health (SDOH) contribute to persistent disparities in liver transplantation access. We developed a large language model framework to extract and analyze how these factors influence care trajectories. Prevalence of key modifiable barriers varied by demographics: social support gaps (35.4%, disproportionately affecting females), recent substance use (14.2–22.7%), and mental health challenges (17.6%, with Hispanic/Latino treatment gaps). Each factor was associated with 5–14 percentage point reductions in listing probability, comparable to clinical metrics. Psychosocial risk and SDOH factors explained 42.6% of racial disparities in listing decisions for Asian patients, exceeding liver health metrics (36.8%) and contributing to 94.6% collective explanation of differences. Priority interventions should target caregiver support, substance use, mental health, and patient education. This framework for systematically analyzing patient circumstances could enhance understanding of care decisions and health disparities.

Objective:Multiple myeloma (MM) is the second most common hematologic malignancy in the U.S., with Black patients being diagnosed at twice the rate of White patients. MM treatment options are limited and ineffective, but CAR-T therapies show promise. However, their limited availability results in disparities in access. This study aimed to explore disparities in Multiple Myeloma disease risk and CAR-T therapy access. Methods:Our study included a population-based cohort of 12,360 patients diagnosed with Multiple Myeloma who received more than one cancer therapy extracted from the University of California Health Data Warehouse (UCHDW) between January 2021 and November 2024. Regression models were used to compute odds ratio (OR) and 95% confidence intervals (CI) associating disease severity, UC-Location, and baseline demographics with CAR-T therapy access. The GPT-4 inference model was prompted with a zero-shot learning approach to analyze UCSF clinical notes with the following objectives: (1) Was CAR-T discussed? [yes/no], (2) Is the patient eligible for CAR-T? [yes/no/unclear], and (3) Provide the rationale for the eligibility determination. Results:Our study included 12,360 patients (mean age 68.5 years, SD 12.8 years) treated for multiple myeloma across the University of California Health System, 320 of which received CAR-T (Table-1). Overall, 51.6% of MM patients identified as Male, and 48.4% as Female. Disease Severity was measured by the International Staging System (ISS) and was distributed by ISS Stage: I (65.3%), II (24.4%), III (2.8%), and None (7.5%). Patients treated at UC-1 (49.3%), and UC-2 (50.0%) were primarily diagnosed with Stage II, while patients at UC-3 (55.5%) were primarily diagnosed with Stage I. Our model indicated that patients who identified as Black or African American (OR= 0.33, [95% CI, 0.17-0.62) were less likely to receive CAR-T therapy when compared to White patients. Patients treated at UC-3 with predominantly Black or African American patients (OR = 0.42, [95% CI, 0.30-0.59]) were less likely to receive CAR-T therapy when compared to UC-1. We identified CAR-T eligibility for 270 UCSF patients and found those who identified as other Pacific Islander had the highest rate of eligibility without discussions at 50%, followed by Black or African American (4.2%), Asian (3.2%), and White (0.6%). Conclusion and Relevance:This study emphasizes the influence of race and UC-Location on disparities in CAR-T therapy access.

Understanding reasons for treatment switching is of significant medical interest, but these factors are often only found in unstructured clinical notes and can be difficult to extract. We evaluated the zero-shot abilities of GPT-4 and eight other open-source large language models (LLMs) to extract contraceptive switching information from 1964 clinical notes derived from the UCSF Information Commons dataset. GPT-4 extracted the contraceptives started and stopped at each switch with microF1 scores of 0.85 and 0.88, respectively, compared to 0.81 and 0.88 for the best open-source model. When evaluated by clinical experts, GPT-4 extracted reasons for switching with an accuracy of 91.4% (2.2% hallucination rate). Transformer-based topic modeling identified patient preference, adverse events, and insurance coverage as key reasons. These findings demonstrate the value of LLMs in identifying complex treatment factors and provide insights into reasons for contraceptive switching in real-world settings.

The MI-CLAIM checklist has been revised to take account of new capabilities of large language models and other generative artificial intelligence tools.