Zongqi Xia

Background and Aims: Distinct autoimmune diseases can share symptoms and autoantibodies, but the molecular pathways underlying this shared etiology remain poorly understood. We assessed genetic support for IL12 signaling across 27 diseases. Approach and Results: We identified a rare variant upstream of IL12A with large risk-increasing effects on autoimmune hepatitis [AIH; per-allele odds ratio (OR)=2.98, p =3×10 ⁻⁹ ], primary biliary cholangitis (PBC; OR=2.80, p =2×10 ⁻⁸ ), systemic sclerosis (SSc; OR=2.64, p =2×10 ⁻¹¹ ), Sjogren's disease (SjD; OR=1.94, p =10 ⁻¹¹ ), systemic lupus erythematosus (SLE; OR=1.83, p =5×10 ⁻¹⁰ ) and multiple sclerosis (OR=1.79, p =10 ⁻¹¹ ). Lifetime risk for developing at least one of these conditions increased 6-fold in homozygote carriers (OR=5.73, p =0.003). We then evaluated genetic support for all 4 IL12 signaling genes and found that: (i) a non-coding variant downstream of IL12A (linked to lower gene expression) protected against 5 of the 6 diseases; (ii) an intronic variant in IL12RB2 (linked to higher gene expression) increased risk of all 6 diseases; and (iii) heterozygote carriers of rare predicted loss-of-function or missense variants in IL12RB2 had 18% lower odds of SLE (OR=0.82, p =2×10 ⁻⁵ ). IL23 signaling, which shares ligand and receptor subunits with IL12, was not associated with these conditions. In epidemiologic analyses of co-prevalence, 5 of the 6 diseases strongly co-occurred with each other: AIH, PBC, SSc, SjD and SLE. Conclusions: AIH, PBC, SSc, SjD and SLE represent a distinct subset of autoimmune diseases that share genetic associations with the IL12 pathway, as well as profound co-prevalence, suggesting they may have a shared etiology and may respond to the same targeted therapeutic intervention(s).

The multiple sclerosis (MS) therapeutic landscape has evolved over time. We conducted a knowledge graph-guided analysis of MS-specific disease-modifying therapy (DMT) prescription trends using longitudinal real-world clinical data. We analyzed registry-linked electronic health record (EHR) data from two large independent healthcare systems between 2004 and 2022, including both academic and community practices. We developed an unsupervised phenotyping algorithm that leverages a publicly available knowledge graph and informative EHR features to identify patients with MS. After identifying MS cases, we combined the two cohorts and constructed annual time-varying knowledge graphs that capture co-occurrence patterns between DMTs and MS diagnosis. For each year, we analyzed co-occurrence patterns between DMTs and MS diagnosis using Shifted Positive Pointwise Mutual Information transformation and singular value decomposition to generate embeddings. We computed patient-wise cosine similarities and confidence intervals to quantify MS - specific DMT usage patterns. The phenotyping algorithm achieved robust performance in predicting MS diagnosis (AUROC: MGB = 0.994, UPMC = 0.922), identifying 29,169 MS patients in the combined dataset. Among commonly used standard-effectiveness DMTs, MS-specific prescriptions declined after 2011 for interferon-beta (DMT-MS cosine similarity slope = −0.019 ± 0.011, p = 0.002) and glatiramer acetate (slope = −0.013 ± 0.012, p = 0.026), from 2013–2020 for fumarates (slope = −0.028 ± 0.015, p = 0.004), and after 2014 for S1P receptor modulators (slope = −0.026 ± 0.016, p = 0.005). Among commonly used higher-effectiveness DMTs, B-cell depletion therapies (slope = 0.051 ± 0.027, p = 0.001), particularly ocrelizumab (slope = 0.020 ± 0.016, p =0.032), showed a marked increase since 2018. Natalizumab usage peaked in 2011 (slope _pre-2011 = 0.063 ± 0.013, p _pre-2011 < 0.001; slope _post-2011 = −0.027 ± 0.008, p _post-2011 < 0.001). Other DMT classes such as cell proliferation inhibitors and chemotherapy agents, showed low usage during follow-up. These findings provide real-world evidence from two large EHR-based MS cohorts, highlighting distinct temporal shifts in the complex MS therapeutic landscape toward higher-effectiveness DMTs, particularly B-cell depletion therapy.

Background. Early diagnosis and intervention are crucial in multiple sclerosis (MS), yet diagnostic delays are common. Large language models (LLMs) such as generative pre-trained transformers (GPTs) may help streamline diagnostic workflows by extracting MS diagnostic signals from clinical notes. Objective. To derive MS diagnosis status from the first neurology note using a computable algorithm based on the 2017 McDonald criteria and applying GPT-4 for node-level reasoning within a structured decision framework. Methods. We analyzed first neurology notes from 125 randomly selected patients (including those with MS, related disorders, and controls) enrolled in a clinic cohort between 2017 and 2023. We included the clinical history and diagnostic testing sections but redacted the assessment and plan. We converted the 2017 McDonald criteria into a decision tree and provided expert-curated clinical knowledge to guide GPT-4 reasoning at each decision node. GPT-4 generated binary decisions at each node to traverse the tree and classified MS diagnoses at terminal nodes. We evaluated performance against neurologist-assessed diagnoses and characterized hallucinations (non-factual, incongruent, irrelevant, over-reliant, and logical reasoning errors). Results. In this study cohort (mean age 40±13 years; 81% women) representative of the clinic population, GPT-4 performed well in predicting MS diagnosis (84% accuracy, 79% precision, 74% recall, 91% specificity) using first neurology notes. Hallucinations occurred in 32 cases (26%), most commonly incoherence (75%) and overreliance (47%). Conclusion. A structured, LLM-guided decision framework can flag MS diagnoses from early clinical documentation. Large-scale studies are needed to mitigate hallucinations, validate this approach, and test implementation in clinical settings.

Background: Cognitive assessments are sparsely documented in electronic health records (EHRs), limiting scalable detection of cognitive worsening in real-world clinical settings. Methods: We applied a deep neural network optimized for identifying clinical event timing from sparsely labeled gold-standard data ( label-efficient incident phenotyping from longitudinal EHR , LATTE) to predict time-to-first sustained cognitive worsening in AD patients from a large healthcare system (2011–2022) with linkage to an AD Research Center registry in a subset. Sustained cognitive worsening was defined as cognitive decline persisting over ≥2 consecutive visits within 3 years. Separate LATTE models were trained with worsening labels from Clinical Dementia Rating (CDR), Mini-Mental Status Examination (MMSE), and Montreal Cognitive Assessment (MoCA) scores; semi-supervised learning scaled predictions to larger imputation cohorts lacking sufficient longitudinal scores. We evaluated model performance using average time-specific area under the receiver operating characteristic curve (AUC), area between curves (ABC), and Brier scores. To demonstrate clinical utility, we examined whether predicted time-to-worsening differentiated clinically meaningful patient subgroups using competing-risk Cox proportional hazards models accounting for death. Findings: The cohort comprised 27,614 AD patients (65% women, 91% non-Hispanic White, mean [SD] age at start of follow-up 78.76 [9.53] years). In gold-standard cohorts (n: CDR=632, MMSE=710, MoCA=752; remaining patients formed imputation cohorts), LATTE demonstrated robust predictive performance (average time-AUC: CDR 0.816, MMSE 0.694, MoCA 0.710; ABC: CDR 0.067, MMSE 0.293, MoCA 0.078; Brier score: CDR 0.252, MMSE 0.437, MoCA 0.295). APOE -ε4 carriers had shorter predicted time-to-worsening compared to non-carriers across all assessments in the imputation cohorts (HRs 1.241–1.376, all p p <.001). Interpretation: LATTE enables scalable prediction of sustained cognitive worsening timing, differentiating clinically meaningful patient subgroups. This approach could improve AD clinical monitoring and decision-making in routine care and support targeted clinical trial enrichment. RESEARCH IN CONTEXT: Evidence before this study: The growing burden of Alzheimer’s disease (AD) creates an urgent unmet need for pragmatic tools to monitor cognitive decline at the point of care and identify target patient populations for clinical trial recruitment. However, cognitive assessments are sparsely documented in electronic health records (EHRs), and fluctuating scores can obscure true worsening, whereas specialized fluid and neuroimaging biomarkers are rarely available outside research settings, limiting scalable real-world utility. Added value of this study: We applied a deep neural network algorithm optimized for identifying clinical event timing from sparsely labeled longitudinal EHR data (LATTE) to predict time-to-first sustained cognitive worsening across three complementary cognitive assessments in AD patients from a large healthcare system: the Clinical Dementia Rating (CDR), Mini-Mental State Examination (MMSE), and Montreal Cognitive Assessment (MoCA). We defined sustained cognitive worsening as clinically meaningful decline (without improvement) persisting over ≥2 consecutive visits within 3 years. We trained LATTE on gold-standard cohorts with sparse outcome labels derived from longitudinal cognitive assessments, then leveraged a semi-supervised framework to scale predictions to larger imputation cohorts lacking gold-standard cognition outcome labels. The algorithm robustly predicted the timing of first sustained cognitive worsening, with CDR outperforming MMSE and MoCA. For orthogonal validation, predicted time-to-worsening differentiated clinically meaningful subgroups defined by APOE-ε4 carrier status and knowledge graph-guided patient clustering. Implications of all the available evidence: Scalable prediction of cognitive worsening from sparsely labeled EHR data could identify patients at higher risk of sustained cognitive worsening. In routine clinical practice, this could inform timely disease-modifying therapy initiation and care planning. Targeted enrichment of clinical trial populations with higher-risk patients could increase statistical power and reduce sample size requirements.

Background: Comorbidities are common in multiple sclerosis (MS) and may influence disability outcomes, but their dynamic impact on bidirectional disability transitions and long-term disability remains incompletely understood. Better understanding of this longitudinal relationship could inform personalized disability management strategies for people with MS. Methods: We leveraged two large electronic health record (EHR)-linked MS registries and applied multi-state Markov models (MSMs) to examine the extent to which individual comorbidities and overall comorbidity burden were associated with short-term disability transitions, long-term disability transition probabilities, and expected time spent in each disability state. We additionally compared MSM-based predictions of confirmed disability worsening (CDW) with Cox proportional hazards (CoxPH) model-based predictions using the integrated Brier score with bootstrap validation. Results: Among 3,723 patients with MS (74.6% female; 86.2% non-Hispanic White; mean age=41.9 years; mean disease duration=5.4 years) contributing 41,860 disability assessments over a mean follow-up of 7.3 years, higher cardiometabolic and psychiatric comorbidity burden was associated with increased transition intensity toward worse disability states and decreased transition intensity toward improvement, with a stepwise gradient across burden levels. Compared with patients without comorbidities, those with ≥4 comorbidities had a 28% higher risk of worsening (HR=1.28 [1.06, 1.55]) and a 20% lower risk of improvement (HR=0.80 [0.67, 0.95]). Each individual comorbidity was significantly associated with worse disability transitions. Long-term estimates indicated a higher 5-year probability of severe disability and fewer years spent in the no-disability state among patients with greater comorbidity burden. CoxPH models showed directionally consistent associations but lower predictive accuracy for CDW compared with MSMs. Conclusion: Cardiometabolic and psychiatric comorbidities are associated with worse disability trajectories in MS, reducing improvement and accelerating progression. By providing a nuanced framework to quantify short-term disability transitions and long-term disability patterns, MSMs may have real-world clinical utility in disability prediction.

Background: Poor sleep quality is common in people with multiple sclerosis (pwMS) and reduces quality of life. Objectives: To examine associations between modifiable factors and sleep quality in pwMS. Methods: In a prospective clinic cohort (2017–2023), we evaluated whether baseline measures of disability, depression, fatigue, and pain were associated with poor sleep quality (Pittsburgh Sleep Quality Index, PSQI) cross-sectionally using covariate-adjusted linear regression, structural equation modeling (SEM), and LASSO logistic regression, and longitudinally using mixed-effects models. Results: In this cohort (n=750; mean age 48.9 years; 80.3% women, 88.7% relapsing type), higher body mass index (β [95% CI]: 0.06 [0.01, 0.12], p=.001) and area deprivation index (6.78 [2.17, 11.39], p<.001) were associated with worse baseline PSQI scores. In adjusted analyses (n=730), disability, depression, fatigue, and pain were each associated with worse sleep. In SEM, pain had a moderate direct effect on sleep (β [95% CI]: 0.56 [0.48, 0.64], p<.001). LASSO models that included pain outperformed the benchmark (AUROC 0.741 vs 0.517). Longitudinally (n=382), time and higher baseline pain predicted worse sleep (β [95% CI]: time in months 0.04 [0.02, 0.06], p<.001; pain 0.36 [0.31, 0.41], p<.001). Conclusion: Pain is a key, potentially modifiable driver of poor sleep quality in pwMS. Key Messages: ◼ Sleep disturbance is common in people with MS and often worsens over time. ◼ Higher baseline pain severity was consistently associated with worse sleep quality, both cross-sectionally and longitudinally. ◼ Pain is a key modifiable driver of sleep disturbance and a potentially important therapeutic target for improving overall quality of life.

Background Prognostication for people with Alzheimer’s disease (AD) at the point of care could improve clinical management. Methods In this retrospective cohort study using the electronic health record (EHR) data from a large healthcare system (2011-2022), we applied an unsupervised latent factor clustering approach guided by knowledge graph embeddings to stratify AD patients into two groups at diagnosis (baseline) using clinical features in the two years preceding diagnosis. We prognosticated the risk of AD-related outcomes (nursing home admission and mortality) for the clusters in survival analyses adjusted for baseline confounders (age, gender, race, ethnicity, healthcare utilization, and comorbidities). To reflect real-world evolution in clinical trajectories, we updated patient stratification for patients remaining at risk one year post-diagnosis and repeated prognostication. Results We stratify 16,411 AD patients into two groups at baseline (41% Group 1, 59% Group 2). Baseline Group 2 has a significantly lower risk of nursing home admission (HR [95% CI] = 0.804 [0.765, 0.844], p < .001) but comparable mortality risk to baseline Group 1 (HR [95% CI] = 1.008 [0.963, 1.056], p = 0.733). We re-stratify the 12,606 patients remaining at risk one year post-diagnosis (46% Group 1, 54% Group 2). Consistent with baseline, the updated Group 2 has a lower risk of nursing home admission (HR [95% CI] = 0.815 [0.766, 0.868], p < .001) but comparable mortality risk (HR [95% CI] = 0.977 [0.922, 1.035], p = .430) to Group 1. Conclusions Patient stratification enables outcome prognosis for AD patients. While baseline prognostication can guide early treatment and tailored management, dynamic prognostication may inform more timely interventions to improve long-term outcomes.

The widespread adoption of electronic health records has created new opportunities for translational clinical research, yet this promise remains constrained by fragmented data across privacy-siloed institutions and substantial heterogeneity in local coding practices. While privacy-preserving collaborative learning allows institutions to work together without sharing patient-level data, it does not address inconsistencies in how clinical concepts are represented across sites. We introduce a graph-based framework that addresses this gap by treating data harmonization as a scalable representation learning problem. Rather than relying on fixed standards or manual mappings, the framework integrates institution-specific summary statistics from health records, curated biomedical knowledge graphs, and semantic information derived from large language models to learn a shared semantic space. This joint learning approach aligns diverse, site-specific vocabularies while preserving patient privacy. Evaluated across seven institutions and two languages, the framework provides a robust, data-centric foundation for training and deploying clinical models across heterogeneous healthcare systems.