Gregory Farage

External biological knowledge provides valuable information about relationships among metabolites, yet this information is usually not incorporated directly into statistical estimation procedures. Most existing approaches estimate metabolite effects independently, ignoring known biochemical structure such as shared subclasses and pathway membership. We propose a Bayesian hierarchical framework that improves metabolite effect estimates by incorporating external biological information describing relationships among metabolites. The proposed method improves metabolite-specific estimates by allowing related metabolites to borrow information from one another while preserving metabolite-level inference. We evaluate the methodology using simulation studies across a range of sample sizes and heterogeneity regimes together with three metabolomics applications involving distinct biological annotation structures. Across both simulated and real datasets, incorporating external biological information consistently improves metabolite effect estimation. Gains are most pronounced when sample sizes are small and metabolite classes are informative, i.e. more homogenous within classes.

Background: Social isolation and social support are clinically important but inconsistently represented in structured electronic health record data. Their identification from clinical narratives is complicated by nuanced, context-dependent language and by variation in how social connectedness is documented across clinical settings, patient populations, and geographic locations. Models intended for health care use must therefore distinguish social isolation from documented support, negation, ambiguous social circumstances, and non-social clinical uses of similar terminology while remaining robust to differences in documentation practices. Objective: This study aimed to develop and evaluate locally deployable, fine-tuned language models for context-aware classification of social isolation, social support, and contextually similar nonrelevant clinical text across multiple health systems. Materials and Methods: We conducted a retrospective study using electronic health record data from 326,847 adults aged 50 years or older who received care at 3 Tennessee health systems from 2020 to 2023. Candidate clinical text was identified using social-environment ICD-10-CM Z codes and an established social isolation lexicon. Human annotators classified 9,748 spans from 9,578 clinical notes contributed by 6,881 patients as social isolation, social support, or no social isolation. The annotated corpus included social isolation and support concepts including negations, infection-control and microbiology terminology, institutional living arrangements, and other contextually ambiguous examples. Uncertainty-based active learning was used to prioritize informative spans for annotation. FLAN-T5-Large, BERT, RoBERTa, and Gemma-2-2B were fine-tuned for 3-class span classification. Full fine-tuning and parameter-efficient fine-tuning using low-rank adaptation approaches were evaluated. Performance was assessed using site-based internal-external cross-validation, with macro-F1 as the primary metric. Results: The annotated corpus included 2,318 social isolation spans, 638 social support spans, and 6,792 spans classified as no social isolation. Fully fine-tuned FLAN-T5-Large achieved the highest and most consistent performance, with a mean macro-F1 of 0.92 (SD 0.04). Class-specific F1 scores were 0.91 (SD 0.03) for social isolation, 0.90 (SD 0.04) for social support, and 0.94 (SD 0.05) for no social isolation. Gemma-2-2B achieved a macro-F1 of 0.89 (SD 0.10), compared with 0.80 (SD 0.21) for RoBERTa and 0.77 (SD 0.17) for BERT. Full fine-tuning also outperformed parameter-efficient fine-tuning in the corresponding held-out evaluation. Remaining errors primarily involved patients who lived alone but received assistance, the presence of family members without meaningful support, grief-related language, and non-social clinical uses of isolation terminology. Conclusions: Fine-tuned language models can distinguish social isolation, social support, and contextually similar but unrelated language across heterogeneous clinical settings. FLAN-T5-Large provided the strongest and most stable performance, including for the less frequent social support class. The 3-class annotation framework, inclusion of difficult negative examples, and multisite evaluation offer a practical approach for converting nuanced social-connectedness documentation into structured data. Further work should evaluate end-to-end identification in unselected clinical notes and prospective integration into clinical and population health workflows.

BACKGROUND Recent advancements in large language models (LLMs) have improved the identification of social isolation from clinical narratives, which vary widely in linguistic patterns and documentation practices. However, LLMs fine-tuned on a single dataset often show reduced performance when applied to different healthcare settings or clinical note types. Rigorous evaluation of cross-site generalizability and fairness is therefore essential to ensure accurate and equitable detection of social isolation across diverse populations and clinical contexts. OBJECTIVE This study aimed to evaluate a span-level fine-tuned FLAN-T5-Large model for extracting ‘social isolation’ indicators from unstructured clinical text and to assess its generalizability and fairness across diverse populations and healthcare data sources. METHODS A total of 2,967 unique annotated spans from 9,578 clinical notes across three healthcare systems were used to fine-tune an FLAN-T5-Large model using a contextualized span classification framework. A Gemma‑2-2B model was evaluated in a sensitivity analysis to assess architecture‑related performance differences. Performance was assessed using precision, recall, and macro F1. Fairness was evaluated across demographic variables, social vulnerability strata, and note types using statistical parity difference (SPD) and equal opportunity difference (EOD). RESULTS Incorporating contextual windows around annotated spans improved macro-F1 from 0.90 to 0.94 during validation. In full note evaluation across 900 manually reviewed notes, FLAN-T5-Large achieved high recall for social isolation (0.94 – 0.98) and macro F1 values ranging from 0.69 to 0.81 across sites. Fairness analysis showed generally consistent performance across age, gender, race, and social vulnerability groups, with equitable sensitivity (EOD 0.02 – 0.04) and moderate variation in positive prediction rates (SPD). Performance variability was strongly associated with documentation type, with note type driving substantially greater variability in both performance and fairness metrics than patient demographic factors. CONCLUSIONS Fine-tuned FLAN‑T5-Large demonstrated strong capability in detecting ‘social isolation’ from clinical narratives, while maintaining sensitivity parity across subgroups. The observed heterogeneity was largely driven by documentation context rather than by patient characteristics, highlighting the importance of note type-aware evaluation in clinical NLP. These findings support the use of instruction‑tuned LLMs for equitable extraction of social context information from EHR text.

Background/Objectives: High-throughput metabolomics data provide a detailed molecular window into biological processes. We consider the problem of assessing how association of metabolite levels with individual (sample) characteristics, such as sex or treatment, depend on metabolite characteristics such as pathways. Typically, this is done using a two-step process. In the first step, we assess the association of each metabolite with individual characteristics. In the second step, an enrichment analysis is performed by metabolite characteristics. Methods: We combine the two steps using a bilinear model based on the matrix linear model (MLM) framework previously developed for high-throughput genetic screens. Our method can estimate relationships in metabolites sharing known characteristics, whether categorical (such as type of lipid or pathway) or numerical (such as number of double bonds in triglycerides). Results: We demonstrate the flexibility and interoperability of MLMs by applying them to three metabolomic studies. We show that our approach can separate the contribution of the overlapping triglyceride characteristics, such as the number of double bonds and the number of carbon atoms. Conclusion: The matrix linear model offers a flexible, efficient, and interpretable framework for integrating external information and examining complex relationships in metabolomics data. Our method has been implemented in the open-source Julia package, MatrixLM. Data analysis scripts with example data analyses are also available.

Genetic studies often collect data using high-throughput phenotyping. That has led to the need for fast genomewide scans for large number of traits using linear mixed models (LMMs). Computing the scans one by one on each trait is time consuming. We have developed new algorithms for performing genome scans on a large number of quantitative traits using LMMs, BulkLMM, that speeds up the computation by orders of magnitude compared to one trait at a time scans. On a mouse BXD Liver Proteome data with more than 35,000 traits and 7,000 markers, BulkLMM completed in a few seconds. We use vectorized, multi-threaded operations and regularization to improve optimization, and numerical approximations to speed up the computations. Our soft-ware implementation in the Julia programming language also provides permutation testing for LMMs and is available at https://github.com/senresearch/BulkLMM.jl .

A bstract: High-throughput metabolomics data provide a detailed molecular window into biological processes. We consider the problem of assessing how the association of metabolite levels with individual (sample) characteristics such as sex or treatment may depend on metabolite characteristics such as pathway. Typically this is one in a two-step process: In the first step we assess the association of each metabolite with individual characteristics. In the second step an enrichment analysis is performed by metabolite characteristics among significant associations. We combine the two steps using a bilinear model based on the matrix linear model (MLM) framework we have previously developed for high-throughput genetic screens. Our framework can estimate relationships in metabolites sharing known characteristics, whether categorical (such as type of lipid or pathway) or numerical (such as number of double bonds in triglycerides). We demonstrate how MLM offers flexibility and interpretability by applying our method to three metabolomic studies. We show that our approach can separate the contribution of the overlapping triglycerides characteristics, such as the number of double bonds and the number of carbon atoms. The proposed method have been implemented in the open-source Julia package, MatrixLM . Data analysis scripts with example data analyses are also available.

The BXD family of mouse strains are an important reference population for systems biology and genetics that have been fully sequenced and deeply phenotyped. To facilitate interactive use of genotype–phenotype relations using many massive omics data sets for this and other segregating populations, we have developed new algorithms and code that enable near-real-time whole-genome quantitative trait locus (QTL) scans for up to one million traits. By using easily parallelizable operations including matrix multiplication, vectorized operations, and element-wise operations, our method is more than 700 times faster than a R/qtl linear model genome scan using 16 threads. We used parallelization of different CPU threads as well as GPUs. We found that the speed advantage of GPUs is dependent on problem size and shape (the number of cases, number of genotypes, and number of traits). Our approach is ideal for interactive web services, such as GeneNetwork.org that need to display results in real-time. Our implementation is available as the Julia language package LiteQTL at https://github.com/senresearch/LiteQTL.jl.

Background: Electronic self-monitoring technology has the potential to provide unique insights into important behaviors for inducing weight loss. Objective: The aim of this study is to investigate the effects of electronic self-monitoring behavior (using the commercial Lose It! app) and weight loss interventions (with differing amounts of counselor feedback and support) on 4- and 12-month weight loss. Methods: In this secondary analysis of the Fit Blue study, we compared the results of two interventions of a randomized controlled trial. Counselor-initiated participants received consistent support from the interventionists, and self-paced participants received assistance upon request. The participants (N=191), who were active duty military personnel, were encouraged to self-monitor their diet and exercise with the Lose It! app or website. We examined the associations between intervention assignment and self-monitoring behaviors. We conducted a mediation analysis of the intervention assignment for weight loss through multiple mediators—app use (calculated from the first principal component [PC] of electronically collected variables), number of weigh-ins, and 4-month weight change. We used linear regression to predict weight loss at 4 and 12 months, and the accuracy was measured using cross-validation. Results: On average, the counselor-initiated–treatment participants used the app more frequently than the self-paced–treatment participants. The first PC represented app use frequencies, the second represented calories recorded, and the third represented reported exercise frequency and exercise caloric expenditure. We found that 4-month weight loss was partially mediated through app use (ie, the first PC; 60.3%) and the number of weigh-ins (55.8%). However, the 12-month weight loss was almost fully mediated by 4-month weight loss (94.8%). Linear regression using app data from the first 8 weeks, the number of self–weigh-ins at 8 weeks, and baseline data explained approximately 30% of the variance in 4-month weight loss. App use frequency (first PC; P=.001), self-monitored caloric intake (second PC; P=.001), and the frequency of self-weighing at 8 weeks (P=.008) were important predictors of 4-month weight loss. Predictions for 12-month weight with the same variables produced an R2 value of 5%; only the number of self–weigh-ins was a significant predictor of 12-month weight loss. The R2 value using 4-month weight loss as a predictor was 31%. Self-reported exercise did not contribute to either model (4 months: P=.77; 12 months: P=.15). Conclusions: We found that app use and daily reported caloric intake had a substantial impact on weight loss prediction at 4 months. Our analysis did not find evidence of an association between participant self-monitoring exercise information and weight loss. As 12-month weight loss was completely mediated by 4-month weight loss, intervention targets should focus on promoting early and frequent dietary intake self-monitoring and self-weighing to promote early weight loss, which leads to long-term success. Trial Registration: ClinicalTrials.gov NCT02063178; https://clinicaltrials.gov/ct2/show/NCT02063178